Merge branch 'dev' into otto/secrt-2157-fix-select-empty-string

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

.agents/skills

@@ -0,0 +1 @@
+../.claude/skills

.claude/settings.json

@@ -0,0 +1,10 @@
+{
+  "permissions": {
+    "allowedTools": [
+      "Read", "Grep", "Glob",
+      "Bash(ls:*)", "Bash(cat:*)", "Bash(grep:*)", "Bash(find:*)",
+      "Bash(git status:*)", "Bash(git diff:*)", "Bash(git log:*)", "Bash(git worktree:*)",
+      "Bash(tmux:*)", "Bash(sleep:*)", "Bash(branchlet:*)"
+    ]
+  }
+}

.claude/skills/open-pr/SKILL.md

@@ -0,0 +1,106 @@
+---
+name: open-pr
+description: Open a pull request with proper PR template, test coverage, and review workflow. Guides agents through creating a PR that follows repo conventions, ensures existing behaviors aren't broken, covers new behaviors with tests, and handles review via bot when local testing isn't possible. TRIGGER when user asks to "open a PR", "create a PR", "make a PR", "submit a PR", "open pull request", "push and create PR", or any variation of opening/submitting a pull request.
+user-invocable: true
+args: "[base-branch] — optional target branch (defaults to dev)."
+metadata:
+  author: autogpt-team
+  version: "1.0.0"
+---
+
+# Open a Pull Request
+
+## Step 1: Pre-flight checks
+
+Before opening the PR:
+
+1. Ensure all changes are committed
+2. Ensure the branch is pushed to the remote (`git push -u origin <branch>`)
+3. Run linters/formatters across the whole repo (not just changed files) and commit any fixes
+
+## Step 2: Test coverage
+
+**This is critical.** Before opening the PR, verify:
+
+### Existing behavior is not broken
+- Identify which modules/components your changes touch
+- Run the existing test suites for those areas
+- If tests fail, fix them before opening the PR — do not open a PR with known regressions
+
+### New behavior has test coverage
+- Every new feature, endpoint, or behavior change needs tests
+- If you added a new block, add tests for that block
+- If you changed API behavior, add or update API tests
+- If you changed frontend behavior, verify it doesn't break existing flows
+
+If you cannot run the full test suite locally, note which tests you ran and which you couldn't in the test plan.
+
+## Step 3: Create the PR using the repo template
+
+Read the canonical PR template at `.github/PULL_REQUEST_TEMPLATE.md` and use it **verbatim** as your PR body:
+
+1. Read the template: `cat .github/PULL_REQUEST_TEMPLATE.md`
+2. Preserve the exact section titles and formatting, including:
+   - `### Why / What / How`
+   - `### Changes 🏗️`
+   - `### Checklist 📋`
+3. Replace HTML comment prompts (`<!-- ... -->`) with actual content; do not leave them in
+4. **Do not pre-check boxes** — leave all checkboxes as `- [ ]` until each step is actually completed
+5. Do not alter the template structure, rename sections, or remove any checklist items
+
+**PR title must use conventional commit format** (e.g., `feat(backend): add new block`, `fix(frontend): resolve routing bug`, `dx(skills): update PR workflow`). See CLAUDE.md for the full list of scopes.
+
+Use `gh pr create` with the base branch (defaults to `dev` if no `[base-branch]` was provided). Use `--body-file` to avoid shell interpretation of backticks and special characters:
+
+```bash
+BASE_BRANCH="${BASE_BRANCH:-dev}"
+PR_BODY=$(mktemp)
+cat > "$PR_BODY" << 'PREOF'
+<filled-in template from .github/PULL_REQUEST_TEMPLATE.md>
+PREOF
+gh pr create --base "$BASE_BRANCH" --title "<type>(scope): short description" --body-file "$PR_BODY"
+rm "$PR_BODY"
+```
+
+## Step 4: Review workflow
+
+### If you have a workspace that allows testing (docker, running backend, etc.)
+- Run `/pr-test` to do E2E manual testing of the PR using docker compose, agent-browser, and API calls. This is the most thorough way to validate your changes before review.
+- After testing, run `/pr-review` to self-review the PR for correctness, security, code quality, and testing gaps before requesting human review.
+
+### If you do NOT have a workspace that allows testing
+This is common for agents running in worktrees without a full stack. In this case:
+
+1. Run `/pr-review` locally to catch obvious issues before pushing
+2. **Comment `/review` on the PR** after creating it to trigger the review bot
+3. **Poll for the review** rather than blindly waiting — check for new review comments every 30 seconds using `gh api repos/Significant-Gravitas/AutoGPT/pulls/{N}/reviews --paginate` and the GraphQL inline threads query. The bot typically responds within 30 minutes, but polling lets the agent react as soon as it arrives.
+4. Do NOT proceed or merge until the bot review comes back
+5. Address any issues the bot raises — use `/pr-address` which has a full polling loop with CI + comment tracking
+
+```bash
+# After creating the PR:
+PR_NUMBER=$(gh pr view --json number -q .number)
+gh pr comment "$PR_NUMBER" --body "/review"
+# Then use /pr-address to poll for and address the review when it arrives
+```
+
+## Step 5: Address review feedback
+
+Once the review bot or human reviewers leave comments:
+- Run `/pr-address` to address review comments. It will loop until CI is green and all comments are resolved.
+- Do not merge without human approval.
+
+## Related skills
+
+| Skill | When to use |
+|---|---|
+| `/pr-test` | E2E testing with docker compose, agent-browser, API calls — use when you have a running workspace |
+| `/pr-review` | Review for correctness, security, code quality — use before requesting human review |
+| `/pr-address` | Address reviewer comments and loop until CI green — use after reviews come in |
+
+## Step 6: Post-creation
+
+After the PR is created and review is triggered:
+- Share the PR URL with the user
+- If waiting on the review bot, let the user know the expected wait time (~30 min)
+- Do not merge without human approval

.claude/skills/setup-repo/SKILL.md

@@ -0,0 +1,195 @@
+---
+name: setup-repo
+description: Initialize a worktree-based repo layout for parallel development. Creates a main worktree, a reviews worktree for PR reviews, and N numbered work branches. Handles .env creation, dependency installation, and branchlet config. TRIGGER when user asks to set up the repo from scratch, initialize worktrees, bootstrap their dev environment, "setup repo", "setup worktrees", "initialize dev environment", "set up branches", or when a freshly cloned repo has no sibling worktrees.
+user-invocable: true
+args: "No arguments — interactive setup via prompts."
+metadata:
+  author: autogpt-team
+  version: "1.0.0"
+---
+
+# Repository Setup
+
+This skill sets up a worktree-based development layout from a freshly cloned repo. It creates:
+- A **main** worktree (the primary checkout)
+- A **reviews** worktree (for PR reviews)
+- **N work branches** (branch1..branchN) for parallel development
+
+## Step 1: Identify the repo
+
+Determine the repo root and parent directory:
+
+```bash
+ROOT=$(git rev-parse --show-toplevel)
+REPO_NAME=$(basename "$ROOT")
+PARENT=$(dirname "$ROOT")
+```
+
+Detect if the repo is already inside a worktree layout by counting sibling worktrees (not just checking the directory name, which could be anything):
+
+```bash
+# Count worktrees that are siblings (live under $PARENT but aren't $ROOT itself)
+SIBLING_COUNT=$(git worktree list --porcelain 2>/dev/null | grep "^worktree " | grep -c "$PARENT/" || true)
+if [ "$SIBLING_COUNT" -gt 1 ]; then
+  echo "INFO: Existing worktree layout detected at $PARENT ($SIBLING_COUNT worktrees)"
+  # Use $ROOT as-is; skip renaming/restructuring
+else
+  echo "INFO: Fresh clone detected, proceeding with setup"
+fi
+```
+
+## Step 2: Ask the user questions
+
+Use AskUserQuestion to gather setup preferences:
+
+1. **How many parallel work branches do you need?** (Options: 4, 8, 16, or custom)
+   - These become `branch1` through `branchN`
+2. **Which branch should be the base?** (Options: origin/master, origin/dev, or custom)
+   - All work branches and reviews will start from this
+
+## Step 3: Fetch and set up branches
+
+```bash
+cd "$ROOT"
+git fetch origin
+
+# Create the reviews branch from base (skip if already exists)
+if git show-ref --verify --quiet refs/heads/reviews; then
+  echo "INFO: Branch 'reviews' already exists, skipping"
+else
+  git branch reviews <base-branch>
+fi
+
+# Create numbered work branches from base (skip if already exists)
+for i in $(seq 1 "$COUNT"); do
+  if git show-ref --verify --quiet "refs/heads/branch$i"; then
+    echo "INFO: Branch 'branch$i' already exists, skipping"
+  else
+    git branch "branch$i" <base-branch>
+  fi
+done
+```
+
+## Step 4: Create worktrees
+
+Create worktrees as siblings to the main checkout:
+
+```bash
+if [ -d "$PARENT/reviews" ]; then
+  echo "INFO: Worktree '$PARENT/reviews' already exists, skipping"
+else
+  git worktree add "$PARENT/reviews" reviews
+fi
+
+for i in $(seq 1 "$COUNT"); do
+  if [ -d "$PARENT/branch$i" ]; then
+    echo "INFO: Worktree '$PARENT/branch$i' already exists, skipping"
+  else
+    git worktree add "$PARENT/branch$i" "branch$i"
+  fi
+done
+```
+
+## Step 5: Set up environment files
+
+**Do NOT assume .env files exist.** For each worktree (including main if needed):
+
+1. Check if `.env` exists in the source worktree for each path
+2. If `.env` exists, copy it
+3. If only `.env.default` or `.env.example` exists, copy that as `.env`
+4. If neither exists, warn the user and list which env files are missing
+
+Env file locations to check (same as the `/worktree` skill — keep these in sync):
+- `autogpt_platform/.env`
+- `autogpt_platform/backend/.env`
+- `autogpt_platform/frontend/.env`
+
+> **Note:** This env copying logic intentionally mirrors the `/worktree` skill's approach. If you update the path list or fallback logic here, update `/worktree` as well.
+
+```bash
+SOURCE="$ROOT"
+WORKTREES="reviews"
+for i in $(seq 1 "$COUNT"); do WORKTREES="$WORKTREES branch$i"; done
+
+FOUND_ANY_ENV=0
+for wt in $WORKTREES; do
+  TARGET="$PARENT/$wt"
+  for envpath in autogpt_platform autogpt_platform/backend autogpt_platform/frontend; do
+    if [ -f "$SOURCE/$envpath/.env" ]; then
+      FOUND_ANY_ENV=1
+      cp "$SOURCE/$envpath/.env" "$TARGET/$envpath/.env"
+    elif [ -f "$SOURCE/$envpath/.env.default" ]; then
+      FOUND_ANY_ENV=1
+      cp "$SOURCE/$envpath/.env.default" "$TARGET/$envpath/.env"
+      echo "NOTE: $wt/$envpath/.env was created from .env.default — you may need to edit it"
+    elif [ -f "$SOURCE/$envpath/.env.example" ]; then
+      FOUND_ANY_ENV=1
+      cp "$SOURCE/$envpath/.env.example" "$TARGET/$envpath/.env"
+      echo "NOTE: $wt/$envpath/.env was created from .env.example — you may need to edit it"
+    else
+      echo "WARNING: No .env, .env.default, or .env.example found at $SOURCE/$envpath/"
+    fi
+  done
+done
+
+if [ "$FOUND_ANY_ENV" -eq 0 ]; then
+  echo "WARNING: No environment files or templates were found in the source worktree."
+  # Use AskUserQuestion to confirm: "Continue setup without env files?"
+  # If the user declines, stop here and let them set up .env files first.
+fi
+```
+
+## Step 6: Copy branchlet config
+
+Copy `.branchlet.json` from main to each worktree so branchlet can manage sub-worktrees:
+
+```bash
+if [ -f "$ROOT/.branchlet.json" ]; then
+  for wt in $WORKTREES; do
+    cp "$ROOT/.branchlet.json" "$PARENT/$wt/.branchlet.json"
+  done
+fi
+```
+
+## Step 7: Install dependencies
+
+Install deps in all worktrees. Run these sequentially per worktree:
+
+```bash
+for wt in $WORKTREES; do
+  TARGET="$PARENT/$wt"
+  echo "=== Installing deps for $wt ==="
+  (cd "$TARGET/autogpt_platform/autogpt_libs" && poetry install) &&
+  (cd "$TARGET/autogpt_platform/backend" && poetry install && poetry run prisma generate) &&
+  (cd "$TARGET/autogpt_platform/frontend" && pnpm install) &&
+  echo "=== Done: $wt ===" ||
+  echo "=== FAILED: $wt ==="
+done
+```
+
+This is slow. Run in background if possible and notify when complete.
+
+## Step 8: Verify and report
+
+After setup, verify and report to the user:
+
+```bash
+git worktree list
+```
+
+Summarize:
+- Number of worktrees created
+- Which env files were copied vs created from defaults vs missing
+- Any warnings or errors encountered
+
+## Final directory layout
+
+```
+parent/
+  main/              # Primary checkout (already exists)
+  reviews/           # PR review worktree
+  branch1/           # Work branch 1
+  branch2/           # Work branch 2
+  ...
+  branchN/           # Work branch N
+```

.github/workflows/classic-autogpt-ci.yml

@@ -6,11 +6,19 @@ on:
     paths:
       - '.github/workflows/classic-autogpt-ci.yml'
       - 'classic/original_autogpt/**'
+      - 'classic/direct_benchmark/**'
+      - 'classic/forge/**'
+      - 'classic/pyproject.toml'
+      - 'classic/poetry.lock'
   pull_request:
     branches: [ master, dev, release-* ]
     paths:
       - '.github/workflows/classic-autogpt-ci.yml'
       - 'classic/original_autogpt/**'
+      - 'classic/direct_benchmark/**'
+      - 'classic/forge/**'
+      - 'classic/pyproject.toml'
+      - 'classic/poetry.lock'
 
 concurrency:
   group: ${{ format('classic-autogpt-ci-{0}', github.head_ref && format('{0}-{1}', github.event_name, github.event.pull_request.number) || github.sha) }}
@@ -19,47 +27,22 @@ concurrency:
 defaults:
   run:
     shell: bash
-    working-directory: classic/original_autogpt
+    working-directory: classic
 
 jobs:
   test:
     permissions:
       contents: read
     timeout-minutes: 30
-    strategy:
-      fail-fast: false
-      matrix:
-        python-version: ["3.10"]
-        platform-os: [ubuntu, macos, macos-arm64, windows]
-    runs-on: ${{ matrix.platform-os != 'macos-arm64' && format('{0}-latest', matrix.platform-os) || 'macos-14' }}
+    runs-on: ubuntu-latest
 
     steps:
-      # Quite slow on macOS (2~4 minutes to set up Docker)
-      # - name: Set up Docker (macOS)
-      #   if: runner.os == 'macOS'
-      #   uses: crazy-max/ghaction-setup-docker@v3
-
-      - name: Start MinIO service (Linux)
-        if: runner.os == 'Linux'
+      - name: Start MinIO service
         working-directory: '.'
         run: |
           docker pull minio/minio:edge-cicd
           docker run -d -p 9000:9000 minio/minio:edge-cicd
 
-      - name: Start MinIO service (macOS)
-        if: runner.os == 'macOS'
-        working-directory: ${{ runner.temp }}
-        run: |
-          brew install minio/stable/minio
-          mkdir data
-          minio server ./data &
-
-      # No MinIO on Windows:
-      # - Windows doesn't support running Linux Docker containers
-      # - It doesn't seem possible to start background processes on Windows. They are
-      #   killed after the step returns.
-      #   See: https://github.com/actions/runner/issues/598#issuecomment-2011890429
-
       - name: Checkout repository
         uses: actions/checkout@v4
         with:
@@ -71,41 +54,23 @@ jobs:
           git config --global user.name "Auto-GPT-Bot"
           git config --global user.email "github-bot@agpt.co"
 
-      - name: Set up Python ${{ matrix.python-version }}
+      - name: Set up Python 3.12
         uses: actions/setup-python@v5
         with:
-          python-version: ${{ matrix.python-version }}
+          python-version: "3.12"
 
       - id: get_date
         name: Get date
         run: echo "date=$(date +'%Y-%m-%d')" >> $GITHUB_OUTPUT
 
       - name: Set up Python dependency cache
-        # On Windows, unpacking cached dependencies takes longer than just installing them
-        if: runner.os != 'Windows'
         uses: actions/cache@v4
         with:
-          path: ${{ runner.os == 'macOS' && '~/Library/Caches/pypoetry' || '~/.cache/pypoetry' }}
-          key: poetry-${{ runner.os }}-${{ hashFiles('classic/original_autogpt/poetry.lock') }}
+          path: ~/.cache/pypoetry
+          key: poetry-${{ runner.os }}-${{ hashFiles('classic/poetry.lock') }}
 
-      - name: Install Poetry (Unix)
-        if: runner.os != 'Windows'
-        run: |
-          curl -sSL https://install.python-poetry.org | python3 -
-
-          if [ "${{ runner.os }}" = "macOS" ]; then
-            PATH="$HOME/.local/bin:$PATH"
-            echo "$HOME/.local/bin" >> $GITHUB_PATH
-          fi
-
-      - name: Install Poetry (Windows)
-        if: runner.os == 'Windows'
-        shell: pwsh
-        run: |
-          (Invoke-WebRequest -Uri https://install.python-poetry.org -UseBasicParsing).Content | python -
-
-          $env:PATH += ";$env:APPDATA\Python\Scripts"
-          echo "$env:APPDATA\Python\Scripts" >> $env:GITHUB_PATH
+      - name: Install Poetry
+        run: curl -sSL https://install.python-poetry.org | python3 -
 
       - name: Install Python dependencies
         run: poetry install
@@ -116,12 +81,13 @@ jobs:
             --cov=autogpt --cov-branch --cov-report term-missing --cov-report xml \
             --numprocesses=logical --durations=10 \
             --junitxml=junit.xml -o junit_family=legacy \
-            tests/unit tests/integration
+            original_autogpt/tests/unit original_autogpt/tests/integration
         env:
           CI: true
           PLAIN_OUTPUT: True
           OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
-          S3_ENDPOINT_URL: ${{ runner.os != 'Windows' && 'http://127.0.0.1:9000' || '' }}
+          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
+          S3_ENDPOINT_URL: http://127.0.0.1:9000
           AWS_ACCESS_KEY_ID: minioadmin
           AWS_SECRET_ACCESS_KEY: minioadmin
 
@@ -135,11 +101,11 @@ jobs:
         uses: codecov/codecov-action@v5
         with:
           token: ${{ secrets.CODECOV_TOKEN }}
-          flags: autogpt-agent,${{ runner.os }}
+          flags: autogpt-agent
 
       - name: Upload logs to artifact
         if: always()
         uses: actions/upload-artifact@v4
         with:
           name: test-logs
-          path: classic/original_autogpt/logs/
+          path: classic/logs/

.github/workflows/classic-autogpt-docker-ci.yml

@@ -148,7 +148,7 @@ jobs:
             --entrypoint poetry ${{ env.IMAGE_NAME }} run \
             pytest -v --cov=autogpt --cov-branch --cov-report term-missing \
             --numprocesses=4 --durations=10 \
-            tests/unit tests/integration 2>&1 | tee test_output.txt
+            original_autogpt/tests/unit original_autogpt/tests/integration 2>&1 | tee test_output.txt
 
           test_failure=${PIPESTATUS[0]}
 

.github/workflows/classic-autogpts-ci.yml

@@ -10,10 +10,9 @@ on:
       - '.github/workflows/classic-autogpts-ci.yml'
       - 'classic/original_autogpt/**'
       - 'classic/forge/**'
-      - 'classic/benchmark/**'
-      - 'classic/run'
-      - 'classic/cli.py'
-      - 'classic/setup.py'
+      - 'classic/direct_benchmark/**'
+      - 'classic/pyproject.toml'
+      - 'classic/poetry.lock'
       - '!**/*.md'
   pull_request:
     branches: [ master, dev, release-* ]
@@ -21,10 +20,9 @@ on:
       - '.github/workflows/classic-autogpts-ci.yml'
       - 'classic/original_autogpt/**'
       - 'classic/forge/**'
-      - 'classic/benchmark/**'
-      - 'classic/run'
-      - 'classic/cli.py'
-      - 'classic/setup.py'
+      - 'classic/direct_benchmark/**'
+      - 'classic/pyproject.toml'
+      - 'classic/poetry.lock'
       - '!**/*.md'
 
 defaults:
@@ -35,13 +33,9 @@ defaults:
 jobs:
   serve-agent-protocol:
     runs-on: ubuntu-latest
-    strategy:
-      matrix:
-        agent-name: [ original_autogpt ]
-      fail-fast: false
     timeout-minutes: 20
     env:
-      min-python-version: '3.10'
+      min-python-version: '3.12'
     steps:
       - name: Checkout repository
         uses: actions/checkout@v4
@@ -55,22 +49,22 @@ jobs:
           python-version: ${{ env.min-python-version }}
 
       - name: Install Poetry
-        working-directory: ./classic/${{ matrix.agent-name }}/
         run: |
           curl -sSL https://install.python-poetry.org | python -
 
-      - name: Run regression tests
+      - name: Install dependencies
+        run: poetry install
+
+      - name: Run smoke tests with direct-benchmark
         run: |
-          ./run agent start ${{ matrix.agent-name }}
-          cd ${{ matrix.agent-name }}
-          poetry run agbenchmark --mock --test=BasicRetrieval --test=Battleship --test=WebArenaTask_0
-          poetry run agbenchmark --test=WriteFile
+          poetry run direct-benchmark run \
+            --strategies one_shot \
+            --models claude \
+            --tests ReadFile,WriteFile \
+            --json
         env:
           OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
-          AGENT_NAME: ${{ matrix.agent-name }}
+          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
           REQUESTS_CA_BUNDLE: /etc/ssl/certs/ca-certificates.crt
-          HELICONE_CACHE_ENABLED: false
-          HELICONE_PROPERTY_AGENT: ${{ matrix.agent-name }}
-          REPORTS_FOLDER: ${{ format('../../reports/{0}', matrix.agent-name) }}
-          TELEMETRY_ENVIRONMENT: autogpt-ci
-          TELEMETRY_OPT_IN: ${{ github.ref_name == 'master' }}
+          NONINTERACTIVE_MODE: "true"
+          CI: true

.github/workflows/classic-benchmark-ci.yml

@@ -1,18 +1,24 @@
-name: Classic - AGBenchmark CI
+name: Classic - Direct Benchmark CI
 
 on:
   push:
     branches: [ master, dev, ci-test* ]
     paths:
-      - 'classic/benchmark/**'
-      - '!classic/benchmark/reports/**'
+      - 'classic/direct_benchmark/**'
+      - 'classic/original_autogpt/**'
+      - 'classic/forge/**'
       - .github/workflows/classic-benchmark-ci.yml
+      - 'classic/pyproject.toml'
+      - 'classic/poetry.lock'
   pull_request:
     branches: [ master, dev, release-* ]
     paths:
-      - 'classic/benchmark/**'
-      - '!classic/benchmark/reports/**'
+      - 'classic/direct_benchmark/**'
+      - 'classic/original_autogpt/**'
+      - 'classic/forge/**'
       - .github/workflows/classic-benchmark-ci.yml
+      - 'classic/pyproject.toml'
+      - 'classic/poetry.lock'
 
 concurrency:
   group: ${{ format('benchmark-ci-{0}', github.head_ref && format('{0}-{1}', github.event_name, github.event.pull_request.number) || github.sha) }}
@@ -23,95 +29,16 @@ defaults:
     shell: bash
 
 env:
-  min-python-version: '3.10'
+  min-python-version: '3.12'
 
 jobs:
-  test:
-    permissions:
-      contents: read
+  benchmark-tests:
+    runs-on: ubuntu-latest
     timeout-minutes: 30
-    strategy:
-      fail-fast: false
-      matrix:
-        python-version: ["3.10"]
-        platform-os: [ubuntu, macos, macos-arm64, windows]
-    runs-on: ${{ matrix.platform-os != 'macos-arm64' && format('{0}-latest', matrix.platform-os) || 'macos-14' }}
     defaults:
       run:
         shell: bash
-        working-directory: classic/benchmark
-    steps:
-      - name: Checkout repository
-        uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-          submodules: true
-
-      - name: Set up Python ${{ matrix.python-version }}
-        uses: actions/setup-python@v5
-        with:
-          python-version: ${{ matrix.python-version }}
-
-      - name: Set up Python dependency cache
-        # On Windows, unpacking cached dependencies takes longer than just installing them
-        if: runner.os != 'Windows'
-        uses: actions/cache@v4
-        with:
-          path: ${{ runner.os == 'macOS' && '~/Library/Caches/pypoetry' || '~/.cache/pypoetry' }}
-          key: poetry-${{ runner.os }}-${{ hashFiles('classic/benchmark/poetry.lock') }}
-
-      - name: Install Poetry (Unix)
-        if: runner.os != 'Windows'
-        run: |
-          curl -sSL https://install.python-poetry.org | python3 -
-
-          if [ "${{ runner.os }}" = "macOS" ]; then
-            PATH="$HOME/.local/bin:$PATH"
-            echo "$HOME/.local/bin" >> $GITHUB_PATH
-          fi
-
-      - name: Install Poetry (Windows)
-        if: runner.os == 'Windows'
-        shell: pwsh
-        run: |
-          (Invoke-WebRequest -Uri https://install.python-poetry.org -UseBasicParsing).Content | python -
-
-          $env:PATH += ";$env:APPDATA\Python\Scripts"
-          echo "$env:APPDATA\Python\Scripts" >> $env:GITHUB_PATH
-
-      - name: Install Python dependencies
-        run: poetry install
-
-      - name: Run pytest with coverage
-        run: |
-          poetry run pytest -vv \
-            --cov=agbenchmark --cov-branch --cov-report term-missing --cov-report xml \
-            --durations=10 \
-            --junitxml=junit.xml -o junit_family=legacy \
-            tests
-        env:
-          CI: true
-          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
-
-      - name: Upload test results to Codecov
-        if: ${{ !cancelled() }}  # Run even if tests fail
-        uses: codecov/test-results-action@v1
-        with:
-          token: ${{ secrets.CODECOV_TOKEN }}
-
-      - name: Upload coverage reports to Codecov
-        uses: codecov/codecov-action@v5
-        with:
-          token: ${{ secrets.CODECOV_TOKEN }}
-          flags: agbenchmark,${{ runner.os }}
-
-  self-test-with-agent:
-    runs-on: ubuntu-latest
-    strategy:
-      matrix:
-        agent-name: [forge]
-      fail-fast: false
-    timeout-minutes: 20
+        working-directory: classic
     steps:
       - name: Checkout repository
         uses: actions/checkout@v4
@@ -124,53 +51,120 @@ jobs:
         with:
           python-version: ${{ env.min-python-version }}
 
+      - name: Set up Python dependency cache
+        uses: actions/cache@v4
+        with:
+          path: ~/.cache/pypoetry
+          key: poetry-${{ runner.os }}-${{ hashFiles('classic/poetry.lock') }}
+
       - name: Install Poetry
         run: |
-          curl -sSL https://install.python-poetry.org | python -
+          curl -sSL https://install.python-poetry.org | python3 -
+
+      - name: Install dependencies
+        run: poetry install
+
+      - name: Run basic benchmark tests
+        run: |
+          echo "Testing ReadFile challenge with one_shot strategy..."
+          poetry run direct-benchmark run \
+            --fresh \
+            --strategies one_shot \
+            --models claude \
+            --tests ReadFile \
+            --json
+
+          echo "Testing WriteFile challenge..."
+          poetry run direct-benchmark run \
+            --fresh \
+            --strategies one_shot \
+            --models claude \
+            --tests WriteFile \
+            --json
+        env:
+          CI: true
+          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
+          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+          NONINTERACTIVE_MODE: "true"
+
+      - name: Test category filtering
+        run: |
+          echo "Testing coding category..."
+          poetry run direct-benchmark run \
+            --fresh \
+            --strategies one_shot \
+            --models claude \
+            --categories coding \
+            --tests ReadFile,WriteFile \
+            --json
+        env:
+          CI: true
+          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
+          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+          NONINTERACTIVE_MODE: "true"
+
+      - name: Test multiple strategies
+        run: |
+          echo "Testing multiple strategies..."
+          poetry run direct-benchmark run \
+            --fresh \
+            --strategies one_shot,plan_execute \
+            --models claude \
+            --tests ReadFile \
+            --parallel 2 \
+            --json
+        env:
+          CI: true
+          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
+          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
+          NONINTERACTIVE_MODE: "true"
+
+  # Run regression tests on maintain challenges
+  regression-tests:
+    runs-on: ubuntu-latest
+    timeout-minutes: 45
+    if: github.ref == 'refs/heads/master' || github.ref == 'refs/heads/dev'
+    defaults:
+      run:
+        shell: bash
+        working-directory: classic
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+          submodules: true
+
+      - name: Set up Python ${{ env.min-python-version }}
+        uses: actions/setup-python@v5
+        with:
+          python-version: ${{ env.min-python-version }}
+
+      - name: Set up Python dependency cache
+        uses: actions/cache@v4
+        with:
+          path: ~/.cache/pypoetry
+          key: poetry-${{ runner.os }}-${{ hashFiles('classic/poetry.lock') }}
+
+      - name: Install Poetry
+        run: |
+          curl -sSL https://install.python-poetry.org | python3 -
+
+      - name: Install dependencies
+        run: poetry install
 
       - name: Run regression tests
-        working-directory: classic
         run: |
-          ./run agent start ${{ matrix.agent-name }}
-          cd ${{ matrix.agent-name }}
-
-          set +e # Ignore non-zero exit codes and continue execution
-          echo "Running the following command: poetry run agbenchmark --maintain --mock"
-          poetry run agbenchmark --maintain --mock
-          EXIT_CODE=$?
-          set -e  # Stop ignoring non-zero exit codes
-          # Check if the exit code was 5, and if so, exit with 0 instead
-          if [ $EXIT_CODE -eq 5 ]; then
-            echo "regression_tests.json is empty."
-          fi
-
-          echo "Running the following command: poetry run agbenchmark --mock"
-          poetry run agbenchmark --mock
-
-          echo "Running the following command: poetry run agbenchmark --mock --category=data"
-          poetry run agbenchmark --mock --category=data
-
-          echo "Running the following command: poetry run agbenchmark --mock --category=coding"
-          poetry run agbenchmark --mock --category=coding
-
-          # echo "Running the following command: poetry run agbenchmark --test=WriteFile"
-          # poetry run agbenchmark --test=WriteFile
-          cd ../benchmark
-          poetry install
-          echo "Adding the BUILD_SKILL_TREE environment variable. This will attempt to add new elements in the skill tree. If new elements are added, the CI fails because they should have been pushed"
-          export BUILD_SKILL_TREE=true
-
-          # poetry run agbenchmark --mock
-
-          # CHANGED=$(git diff --name-only | grep -E '(agbenchmark/challenges)|(../classic/frontend/assets)') || echo "No diffs"
-          # if [ ! -z "$CHANGED" ]; then
-          #   echo "There are unstaged changes please run agbenchmark and commit those changes since they are needed."
-          #   echo "$CHANGED"
-          #   exit 1
-          # else
-          #   echo "No unstaged changes."
-          # fi
+          echo "Running regression tests (previously beaten challenges)..."
+          poetry run direct-benchmark run \
+            --fresh \
+            --strategies one_shot \
+            --models claude \
+            --maintain \
+            --parallel 4 \
+            --json
         env:
+          CI: true
+          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
           OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
-          TELEMETRY_ENVIRONMENT: autogpt-benchmark-ci
-          TELEMETRY_OPT_IN: ${{ github.ref_name == 'master' }}
+          NONINTERACTIVE_MODE: "true"

.github/workflows/classic-forge-ci.yml

@@ -6,13 +6,15 @@ on:
     paths:
       - '.github/workflows/classic-forge-ci.yml'
       - 'classic/forge/**'
-      - '!classic/forge/tests/vcr_cassettes'
+      - 'classic/pyproject.toml'
+      - 'classic/poetry.lock'
   pull_request:
     branches: [ master, dev, release-* ]
     paths:
       - '.github/workflows/classic-forge-ci.yml'
       - 'classic/forge/**'
-      - '!classic/forge/tests/vcr_cassettes'
+      - 'classic/pyproject.toml'
+      - 'classic/poetry.lock'
 
 concurrency:
   group: ${{ format('forge-ci-{0}', github.head_ref && format('{0}-{1}', github.event_name, github.event.pull_request.number) || github.sha) }}
@@ -21,131 +23,60 @@ concurrency:
 defaults:
   run:
     shell: bash
-    working-directory: classic/forge
+    working-directory: classic
 
 jobs:
   test:
     permissions:
       contents: read
     timeout-minutes: 30
-    strategy:
-      fail-fast: false
-      matrix:
-        python-version: ["3.10"]
-        platform-os: [ubuntu, macos, macos-arm64, windows]
-    runs-on: ${{ matrix.platform-os != 'macos-arm64' && format('{0}-latest', matrix.platform-os) || 'macos-14' }}
+    runs-on: ubuntu-latest
 
     steps:
-      # Quite slow on macOS (2~4 minutes to set up Docker)
-      # - name: Set up Docker (macOS)
-      #   if: runner.os == 'macOS'
-      #   uses: crazy-max/ghaction-setup-docker@v3
-
-      - name: Start MinIO service (Linux)
-        if: runner.os == 'Linux'
+      - name: Start MinIO service
         working-directory: '.'
         run: |
           docker pull minio/minio:edge-cicd
           docker run -d -p 9000:9000 minio/minio:edge-cicd
 
-      - name: Start MinIO service (macOS)
-        if: runner.os == 'macOS'
-        working-directory: ${{ runner.temp }}
-        run: |
-          brew install minio/stable/minio
-          mkdir data
-          minio server ./data &
-
-      # No MinIO on Windows:
-      # - Windows doesn't support running Linux Docker containers
-      # - It doesn't seem possible to start background processes on Windows. They are
-      #   killed after the step returns.
-      #   See: https://github.com/actions/runner/issues/598#issuecomment-2011890429
-
       - name: Checkout repository
         uses: actions/checkout@v4
-        with:
-          fetch-depth: 0
-          submodules: true
 
-      - name: Checkout cassettes
-        if: ${{ startsWith(github.event_name, 'pull_request') }}
-        env:
-          PR_BASE: ${{ github.event.pull_request.base.ref }}
-          PR_BRANCH: ${{ github.event.pull_request.head.ref }}
-          PR_AUTHOR: ${{ github.event.pull_request.user.login }}
-        run: |
-          cassette_branch="${PR_AUTHOR}-${PR_BRANCH}"
-          cassette_base_branch="${PR_BASE}"
-          cd tests/vcr_cassettes
-
-          if ! git ls-remote --exit-code --heads origin $cassette_base_branch ; then
-            cassette_base_branch="master"
-          fi
-
-          if git ls-remote --exit-code --heads origin $cassette_branch ; then
-            git fetch origin $cassette_branch
-            git fetch origin $cassette_base_branch
-
-            git checkout $cassette_branch
-
-            # Pick non-conflicting cassette updates from the base branch
-            git merge --no-commit --strategy-option=ours origin/$cassette_base_branch
-            echo "Using cassettes from mirror branch '$cassette_branch'," \
-              "synced to upstream branch '$cassette_base_branch'."
-          else
-            git checkout -b $cassette_branch
-            echo "Branch '$cassette_branch' does not exist in cassette submodule." \
-              "Using cassettes from '$cassette_base_branch'."
-          fi
-
-      - name: Set up Python ${{ matrix.python-version }}
+      - name: Set up Python 3.12
         uses: actions/setup-python@v5
         with:
-          python-version: ${{ matrix.python-version }}
+          python-version: "3.12"
 
       - name: Set up Python dependency cache
-        # On Windows, unpacking cached dependencies takes longer than just installing them
-        if: runner.os != 'Windows'
         uses: actions/cache@v4
         with:
-          path: ${{ runner.os == 'macOS' && '~/Library/Caches/pypoetry' || '~/.cache/pypoetry' }}
-          key: poetry-${{ runner.os }}-${{ hashFiles('classic/forge/poetry.lock') }}
+          path: ~/.cache/pypoetry
+          key: poetry-${{ runner.os }}-${{ hashFiles('classic/poetry.lock') }}
 
-      - name: Install Poetry (Unix)
-        if: runner.os != 'Windows'
-        run: |
-          curl -sSL https://install.python-poetry.org | python3 -
-
-          if [ "${{ runner.os }}" = "macOS" ]; then
-            PATH="$HOME/.local/bin:$PATH"
-            echo "$HOME/.local/bin" >> $GITHUB_PATH
-          fi
-
-      - name: Install Poetry (Windows)
-        if: runner.os == 'Windows'
-        shell: pwsh
-        run: |
-          (Invoke-WebRequest -Uri https://install.python-poetry.org -UseBasicParsing).Content | python -
-
-          $env:PATH += ";$env:APPDATA\Python\Scripts"
-          echo "$env:APPDATA\Python\Scripts" >> $env:GITHUB_PATH
+      - name: Install Poetry
+        run: curl -sSL https://install.python-poetry.org | python3 -
 
       - name: Install Python dependencies
         run: poetry install
 
+      - name: Install Playwright browsers
+        run: poetry run playwright install chromium
+
       - name: Run pytest with coverage
         run: |
           poetry run pytest -vv \
             --cov=forge --cov-branch --cov-report term-missing --cov-report xml \
             --durations=10 \
             --junitxml=junit.xml -o junit_family=legacy \
-            forge
+            forge/forge forge/tests
         env:
           CI: true
           PLAIN_OUTPUT: True
+          # API keys - tests that need these will skip if not available
+          # Secrets are not available to fork PRs (GitHub security feature)
           OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
-          S3_ENDPOINT_URL: ${{ runner.os != 'Windows' && 'http://127.0.0.1:9000' || '' }}
+          ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY }}
+          S3_ENDPOINT_URL: http://127.0.0.1:9000
           AWS_ACCESS_KEY_ID: minioadmin
           AWS_SECRET_ACCESS_KEY: minioadmin
 
@@ -159,85 +90,11 @@ jobs:
         uses: codecov/codecov-action@v5
         with:
           token: ${{ secrets.CODECOV_TOKEN }}
-          flags: forge,${{ runner.os }}
-
-      - id: setup_git_auth
-        name: Set up git token authentication
-        # Cassettes may be pushed even when tests fail
-        if: success() || failure()
-        run: |
-          config_key="http.${{ github.server_url }}/.extraheader"
-          if [ "${{ runner.os }}" = 'macOS' ]; then
-            base64_pat=$(echo -n "pat:${{ secrets.PAT_REVIEW }}" | base64)
-          else
-            base64_pat=$(echo -n "pat:${{ secrets.PAT_REVIEW }}" | base64 -w0)
-          fi
-
-          git config "$config_key" \
-            "Authorization: Basic $base64_pat"
-
-          cd tests/vcr_cassettes
-          git config "$config_key" \
-            "Authorization: Basic $base64_pat"
-
-          echo "config_key=$config_key" >> $GITHUB_OUTPUT
-
-      - id: push_cassettes
-        name: Push updated cassettes
-        # For pull requests, push updated cassettes even when tests fail
-        if: github.event_name == 'push' || (! github.event.pull_request.head.repo.fork && (success() || failure()))
-        env:
-          PR_BRANCH: ${{ github.event.pull_request.head.ref }}
-          PR_AUTHOR: ${{ github.event.pull_request.user.login }}
-        run: |
-          if [ "${{ startsWith(github.event_name, 'pull_request') }}" = "true" ]; then
-            is_pull_request=true
-            cassette_branch="${PR_AUTHOR}-${PR_BRANCH}"
-          else
-            cassette_branch="${{ github.ref_name }}"
-          fi
-
-          cd tests/vcr_cassettes
-          # Commit & push changes to cassettes if any
-          if ! git diff --quiet; then
-            git add .
-            git commit -m "Auto-update cassettes"
-            git push origin HEAD:$cassette_branch
-            if [ ! $is_pull_request ]; then
-              cd ../..
-              git add tests/vcr_cassettes
-              git commit -m "Update cassette submodule"
-              git push origin HEAD:$cassette_branch
-            fi
-            echo "updated=true" >> $GITHUB_OUTPUT
-          else
-            echo "updated=false" >> $GITHUB_OUTPUT
-            echo "No cassette changes to commit"
-          fi
-
-      - name: Post Set up git token auth
-        if: steps.setup_git_auth.outcome == 'success'
-        run: |
-          git config --unset-all '${{ steps.setup_git_auth.outputs.config_key }}'
-          git submodule foreach git config --unset-all '${{ steps.setup_git_auth.outputs.config_key }}'
-
-      - name: Apply "behaviour change" label and comment on PR
-        if: ${{ startsWith(github.event_name, 'pull_request') }}
-        run: |
-          PR_NUMBER="${{ github.event.pull_request.number }}"
-          TOKEN="${{ secrets.PAT_REVIEW }}"
-          REPO="${{ github.repository }}"
-
-          if [[ "${{ steps.push_cassettes.outputs.updated }}" == "true" ]]; then
-            echo "Adding label and comment..."
-            echo $TOKEN | gh auth login --with-token
-            gh issue edit $PR_NUMBER --add-label "behaviour change"
-            gh issue comment $PR_NUMBER --body "You changed AutoGPT's behaviour on ${{ runner.os }}. The cassettes have been updated and will be merged to the submodule when this Pull Request gets merged."
-          fi
+          flags: forge
 
       - name: Upload logs to artifact
         if: always()
         uses: actions/upload-artifact@v4
         with:
           name: test-logs
-          path: classic/forge/logs/
+          path: classic/logs/

.github/workflows/classic-frontend-ci.yml

@@ -1,60 +0,0 @@
-name: Classic - Frontend CI/CD
-
-on:
-  push:
-    branches:
-      - master
-      - dev
-      - 'ci-test*' # This will match any branch that starts with "ci-test"
-    paths:
-      - 'classic/frontend/**'
-      - '.github/workflows/classic-frontend-ci.yml'
-  pull_request:
-    paths:
-      - 'classic/frontend/**'
-      - '.github/workflows/classic-frontend-ci.yml'
-
-jobs:
-  build:
-    permissions:
-      contents: write
-      pull-requests: write
-    runs-on: ubuntu-latest
-    env:
-      BUILD_BRANCH: ${{ format('classic-frontend-build/{0}', github.ref_name) }}
-
-    steps:
-      - name: Checkout Repo
-        uses: actions/checkout@v4
-
-      - name: Setup Flutter
-        uses: subosito/flutter-action@v2
-        with:
-          flutter-version: '3.13.2'
-
-      - name: Build Flutter to Web
-        run: |
-          cd classic/frontend
-          flutter build web --base-href /app/
-
-      # - name: Commit and Push to ${{ env.BUILD_BRANCH }}
-      #   if: github.event_name == 'push'
-      #   run: |
-      #     git config --local user.email "action@github.com"
-      #     git config --local user.name "GitHub Action"
-      #     git add classic/frontend/build/web
-      #     git checkout -B ${{ env.BUILD_BRANCH }}
-      #     git commit -m "Update frontend build to ${GITHUB_SHA:0:7}" -a
-      #     git push -f origin ${{ env.BUILD_BRANCH }}
-
-      - name: Create PR ${{ env.BUILD_BRANCH }} -> ${{ github.ref_name }}
-        if: github.event_name == 'push'
-        uses: peter-evans/create-pull-request@v8
-        with:
-          add-paths: classic/frontend/build/web
-          base: ${{ github.ref_name }}
-          branch: ${{ env.BUILD_BRANCH }}
-          delete-branch: true
-          title: "Update frontend build in `${{ github.ref_name }}`"
-          body: "This PR updates the frontend build based on commit ${{ github.sha }}."
-          commit-message: "Update frontend build based on commit ${{ github.sha }}"

.github/workflows/classic-python-checks.yml

@@ -7,7 +7,9 @@ on:
       - '.github/workflows/classic-python-checks-ci.yml'
       - 'classic/original_autogpt/**'
       - 'classic/forge/**'
-      - 'classic/benchmark/**'
+      - 'classic/direct_benchmark/**'
+      - 'classic/pyproject.toml'
+      - 'classic/poetry.lock'
       - '**.py'
       - '!classic/forge/tests/vcr_cassettes'
   pull_request:
@@ -16,7 +18,9 @@ on:
       - '.github/workflows/classic-python-checks-ci.yml'
       - 'classic/original_autogpt/**'
       - 'classic/forge/**'
-      - 'classic/benchmark/**'
+      - 'classic/direct_benchmark/**'
+      - 'classic/pyproject.toml'
+      - 'classic/poetry.lock'
       - '**.py'
       - '!classic/forge/tests/vcr_cassettes'
 
@@ -27,44 +31,13 @@ concurrency:
 defaults:
   run:
     shell: bash
+    working-directory: classic
 
 jobs:
-  get-changed-parts:
-    runs-on: ubuntu-latest
-    steps:
-      - name: Checkout repository
-        uses: actions/checkout@v4
-
-      - id: changes-in
-        name: Determine affected subprojects
-        uses: dorny/paths-filter@v3
-        with:
-          filters: |
-            original_autogpt:
-              - classic/original_autogpt/autogpt/**
-              - classic/original_autogpt/tests/**
-              - classic/original_autogpt/poetry.lock
-            forge:
-              - classic/forge/forge/**
-              - classic/forge/tests/**
-              - classic/forge/poetry.lock
-            benchmark:
-              - classic/benchmark/agbenchmark/**
-              - classic/benchmark/tests/**
-              - classic/benchmark/poetry.lock
-    outputs:
-      changed-parts: ${{ steps.changes-in.outputs.changes }}
-
   lint:
-    needs: get-changed-parts
     runs-on: ubuntu-latest
     env:
-      min-python-version: "3.10"
-
-    strategy:
-      matrix:
-        sub-package: ${{ fromJson(needs.get-changed-parts.outputs.changed-parts) }}
-      fail-fast: false
+      min-python-version: "3.12"
 
     steps:
       - name: Checkout repository
@@ -81,42 +54,31 @@ jobs:
         uses: actions/cache@v4
         with:
           path: ~/.cache/pypoetry
-          key: ${{ runner.os }}-poetry-${{ hashFiles(format('{0}/poetry.lock', matrix.sub-package)) }}
+          key: ${{ runner.os }}-poetry-${{ hashFiles('classic/poetry.lock') }}
 
       - name: Install Poetry
         run: curl -sSL https://install.python-poetry.org | python3 -
 
-      # Install dependencies
-
       - name: Install Python dependencies
-        run: poetry -C classic/${{ matrix.sub-package }} install
+        run: poetry install
 
       # Lint
 
       - name: Lint (isort)
         run: poetry run isort --check .
-        working-directory: classic/${{ matrix.sub-package }}
 
       - name: Lint (Black)
         if: success() || failure()
         run: poetry run black --check .
-        working-directory: classic/${{ matrix.sub-package }}
 
       - name: Lint (Flake8)
         if: success() || failure()
         run: poetry run flake8 .
-        working-directory: classic/${{ matrix.sub-package }}
 
   types:
-    needs: get-changed-parts
     runs-on: ubuntu-latest
     env:
-      min-python-version: "3.10"
-
-    strategy:
-      matrix:
-        sub-package: ${{ fromJson(needs.get-changed-parts.outputs.changed-parts) }}
-      fail-fast: false
+      min-python-version: "3.12"
 
     steps:
       - name: Checkout repository
@@ -133,19 +95,16 @@ jobs:
         uses: actions/cache@v4
         with:
           path: ~/.cache/pypoetry
-          key: ${{ runner.os }}-poetry-${{ hashFiles(format('{0}/poetry.lock', matrix.sub-package)) }}
+          key: ${{ runner.os }}-poetry-${{ hashFiles('classic/poetry.lock') }}
 
       - name: Install Poetry
         run: curl -sSL https://install.python-poetry.org | python3 -
 
-      # Install dependencies
-
       - name: Install Python dependencies
-        run: poetry -C classic/${{ matrix.sub-package }} install
+        run: poetry install
 
       # Typecheck
 
       - name: Typecheck
         if: success() || failure()
         run: poetry run pyright
-        working-directory: classic/${{ matrix.sub-package }}

.github/workflows/platform-backend-ci.yml

@@ -269,12 +269,14 @@ jobs:
           DATABASE_URL: ${{ steps.supabase.outputs.DB_URL }}
           DIRECT_URL: ${{ steps.supabase.outputs.DB_URL }}
 
-      - name: Run pytest
+      - name: Run pytest with coverage
         run: |
           if [[ "${{ runner.debug }}" == "1" ]]; then
-            poetry run pytest -s -vv -o log_cli=true -o log_cli_level=DEBUG
+            poetry run pytest -s -vv -o log_cli=true -o log_cli_level=DEBUG \
+              --cov=backend --cov-branch --cov-report term-missing --cov-report xml
           else
-            poetry run pytest -s -vv
+            poetry run pytest -s -vv \
+              --cov=backend --cov-branch --cov-report term-missing --cov-report xml
           fi
         env:
           LOG_LEVEL: ${{ runner.debug && 'DEBUG' || 'INFO' }}
@@ -287,11 +289,13 @@ jobs:
           REDIS_PORT: "6379"
           ENCRYPTION_KEY: "dvziYgz0KSK8FENhju0ZYi8-fRTfAdlz6YLhdB_jhNw=" # DO NOT USE IN PRODUCTION!!
 
-      # - name: Upload coverage reports to Codecov
-      #   uses: codecov/codecov-action@v4
-      #   with:
-      #     token: ${{ secrets.CODECOV_TOKEN }}
-      #     flags: backend,${{ runner.os }}
+      - name: Upload coverage reports to Codecov
+        if: ${{ !cancelled() }}
+        uses: codecov/codecov-action@v5
+        with:
+          token: ${{ secrets.CODECOV_TOKEN }}
+          flags: platform-backend
+          files: ./autogpt_platform/backend/coverage.xml
 
     env:
       CI: true

.github/workflows/platform-frontend-ci.yml

@@ -148,3 +148,11 @@ jobs:
 
       - name: Run Integration Tests
         run: pnpm test:unit
+
+      - name: Upload coverage reports to Codecov
+        if: ${{ !cancelled() }}
+        uses: codecov/codecov-action@v5
+        with:
+          token: ${{ secrets.CODECOV_TOKEN }}
+          flags: platform-frontend
+          files: ./autogpt_platform/frontend/coverage/cobertura-coverage.xml

.gitignore

@@ -3,6 +3,7 @@
 classic/original_autogpt/keys.py
 classic/original_autogpt/*.json
 auto_gpt_workspace/*
+.autogpt/
 *.mpeg
 .env
 # Root .env files
@@ -159,6 +160,10 @@ CURRENT_BULLETIN.md
 
 # AgBenchmark
 classic/benchmark/agbenchmark/reports/
+classic/reports/
+classic/direct_benchmark/reports/
+classic/.benchmark_workspaces/
+classic/direct_benchmark/.benchmark_workspaces/
 
 # Nodejs
 package-lock.json
@@ -177,9 +182,13 @@ autogpt_platform/backend/settings.py
 
 *.ign.*
 .test-contents
+**/.claude/settings.local.json
 .claude/settings.local.json
 CLAUDE.local.md
 /autogpt_platform/backend/logs
+
+# Test database
+test.db
 .next
 # Implementation plans (generated by AI agents)
 plans/

.gitmodules

@@ -1,3 +0,0 @@
-[submodule "classic/forge/tests/vcr_cassettes"]
-	path = classic/forge/tests/vcr_cassettes
-	url = https://github.com/Significant-Gravitas/Auto-GPT-test-cassettes

.pre-commit-config.yaml

@@ -84,51 +84,16 @@ repos:
         stages: [pre-commit, post-checkout]
 
       - id: poetry-install
-        name: Check & Install dependencies - Classic - AutoGPT
-        alias: poetry-install-classic-autogpt
+        name: Check & Install dependencies - Classic
+        alias: poetry-install-classic
         entry: >
           bash -c '
           if [ -n "$PRE_COMMIT_FROM_REF" ]; then
             git diff --name-only "$PRE_COMMIT_FROM_REF" "$PRE_COMMIT_TO_REF"
           else
             git diff --cached --name-only
-          fi | grep -qE "^classic/(original_autogpt|forge)/poetry\.lock$" || exit 0;
-          poetry -C classic/original_autogpt install
-          '
-        # include forge source (since it's a path dependency)
-        always_run: true
-        language: system
-        pass_filenames: false
-        stages: [pre-commit, post-checkout]
-
-      - id: poetry-install
-        name: Check & Install dependencies - Classic - Forge
-        alias: poetry-install-classic-forge
-        entry: >
-          bash -c '
-          if [ -n "$PRE_COMMIT_FROM_REF" ]; then
-            git diff --name-only "$PRE_COMMIT_FROM_REF" "$PRE_COMMIT_TO_REF"
-          else
-            git diff --cached --name-only
-          fi | grep -qE "^classic/forge/poetry\.lock$" || exit 0;
-          poetry -C classic/forge install
-          '
-        always_run: true
-        language: system
-        pass_filenames: false
-        stages: [pre-commit, post-checkout]
-
-      - id: poetry-install
-        name: Check & Install dependencies - Classic - Benchmark
-        alias: poetry-install-classic-benchmark
-        entry: >
-          bash -c '
-          if [ -n "$PRE_COMMIT_FROM_REF" ]; then
-            git diff --name-only "$PRE_COMMIT_FROM_REF" "$PRE_COMMIT_TO_REF"
-          else
-            git diff --cached --name-only
-          fi | grep -qE "^classic/benchmark/poetry\.lock$" || exit 0;
-          poetry -C classic/benchmark install
+          fi | grep -qE "^classic/poetry\.lock$" || exit 0;
+          poetry -C classic install
           '
         always_run: true
         language: system
@@ -223,26 +188,10 @@ repos:
         language: system
 
       - id: isort
-        name: Lint (isort) - Classic - AutoGPT
-        alias: isort-classic-autogpt
-        entry: poetry -P classic/original_autogpt run isort -p autogpt
-        files: ^classic/original_autogpt/
-        types: [file, python]
-        language: system
-
-      - id: isort
-        name: Lint (isort) - Classic - Forge
-        alias: isort-classic-forge
-        entry: poetry -P classic/forge run isort -p forge
-        files: ^classic/forge/
-        types: [file, python]
-        language: system
-
-      - id: isort
-        name: Lint (isort) - Classic - Benchmark
-        alias: isort-classic-benchmark
-        entry: poetry -P classic/benchmark run isort -p agbenchmark
-        files: ^classic/benchmark/
+        name: Lint (isort) - Classic
+        alias: isort-classic
+        entry: bash -c 'cd classic && poetry run isort $(echo "$@" | sed "s|classic/||g")' --
+        files: ^classic/(original_autogpt|forge|direct_benchmark)/
         types: [file, python]
         language: system
 
@@ -256,26 +205,13 @@ repos:
 
   - repo: https://github.com/PyCQA/flake8
     rev: 7.0.0
-    # To have flake8 load the config of the individual subprojects, we have to call
-    # them separately.
+    # Use consolidated flake8 config at classic/.flake8
     hooks:
       - id: flake8
-        name: Lint (Flake8) - Classic - AutoGPT
-        alias: flake8-classic-autogpt
-        files: ^classic/original_autogpt/(autogpt|scripts|tests)/
-        args: [--config=classic/original_autogpt/.flake8]
-
-      - id: flake8
-        name: Lint (Flake8) - Classic - Forge
-        alias: flake8-classic-forge
-        files: ^classic/forge/(forge|tests)/
-        args: [--config=classic/forge/.flake8]
-
-      - id: flake8
-        name: Lint (Flake8) - Classic - Benchmark
-        alias: flake8-classic-benchmark
-        files: ^classic/benchmark/(agbenchmark|tests)/((?!reports).)*[/.]
-        args: [--config=classic/benchmark/.flake8]
+        name: Lint (Flake8) - Classic
+        alias: flake8-classic
+        files: ^classic/(original_autogpt|forge|direct_benchmark)/
+        args: [--config=classic/.flake8]
 
   - repo: local
     hooks:
@@ -311,29 +247,10 @@ repos:
         pass_filenames: false
 
       - id: pyright
-        name: Typecheck - Classic - AutoGPT
-        alias: pyright-classic-autogpt
-        entry: poetry -C classic/original_autogpt run pyright
-        # include forge source (since it's a path dependency) but exclude *_test.py files:
-        files: ^(classic/original_autogpt/((autogpt|scripts|tests)/|poetry\.lock$)|classic/forge/(forge/.*(?<!_test)\.py|poetry\.lock)$)
-        types: [file]
-        language: system
-        pass_filenames: false
-
-      - id: pyright
-        name: Typecheck - Classic - Forge
-        alias: pyright-classic-forge
-        entry: poetry -C classic/forge run pyright
-        files: ^classic/forge/(forge/|poetry\.lock$)
-        types: [file]
-        language: system
-        pass_filenames: false
-
-      - id: pyright
-        name: Typecheck - Classic - Benchmark
-        alias: pyright-classic-benchmark
-        entry: poetry -C classic/benchmark run pyright
-        files: ^classic/benchmark/(agbenchmark/|tests/|poetry\.lock$)
+        name: Typecheck - Classic
+        alias: pyright-classic
+        entry: poetry -C classic run pyright
+        files: ^classic/(original_autogpt|forge|direct_benchmark)/.*\.py$|^classic/poetry\.lock$
         types: [file]
         language: system
         pass_filenames: false
@@ -360,26 +277,9 @@ repos:
   #       pass_filenames: false
 
   #     - id: pytest
-  #       name: Run tests - Classic - AutoGPT (excl. slow tests)
-  #       alias: pytest-classic-autogpt
-  #       entry: bash -c 'cd classic/original_autogpt && poetry run pytest --cov=autogpt -m "not slow" tests/unit tests/integration'
-  #       # include forge source (since it's a path dependency) but exclude *_test.py files:
-  #       files: ^(classic/original_autogpt/((autogpt|tests)/|poetry\.lock$)|classic/forge/(forge/.*(?<!_test)\.py|poetry\.lock)$)
-  #       language: system
-  #       pass_filenames: false
-
-  #     - id: pytest
-  #       name: Run tests - Classic - Forge (excl. slow tests)
-  #       alias: pytest-classic-forge
-  #       entry: bash -c 'cd classic/forge && poetry run pytest --cov=forge -m "not slow"'
-  #       files: ^classic/forge/(forge/|tests/|poetry\.lock$)
-  #       language: system
-  #       pass_filenames: false
-
-  #     - id: pytest
-  #       name: Run tests - Classic - Benchmark
-  #       alias: pytest-classic-benchmark
-  #       entry: bash -c 'cd classic/benchmark && poetry run pytest --cov=benchmark'
-  #       files: ^classic/benchmark/(agbenchmark/|tests/|poetry\.lock$)
+  #       name: Run tests - Classic (excl. slow tests)
+  #       alias: pytest-classic
+  #       entry: bash -c 'cd classic && poetry run pytest -m "not slow"'
+  #       files: ^classic/(original_autogpt|forge|direct_benchmark)/
   #       language: system
   #       pass_filenames: false

AGENTS.md

@@ -1,6 +1,6 @@
 # AutoGPT Platform Contribution Guide
 
-This guide provides context for Codex when updating the **autogpt_platform** folder.
+This guide provides context for coding agents when updating the **autogpt_platform** folder.
 
 ## Directory overview
 

CLAUDE.md

@@ -0,0 +1 @@
+@AGENTS.md

README.md

@@ -83,13 +83,13 @@ The AutoGPT frontend is where users interact with our powerful AI automation pla
 
    **Agent Builder:** For those who want to customize, our intuitive, low-code interface allows you to design and configure your own AI agents. 
    
-   **Workflow Management:** Build, modify, and optimize your automation workflows with ease. You build your agent by connecting blocks, where each block     performs a single action.
+   **Workflow Management:** Build, modify, and optimize your automation workflows with ease. You build your agent by connecting blocks, where each block performs a single action.
    
    **Deployment Controls:** Manage the lifecycle of your agents, from testing to production.
    
    **Ready-to-Use Agents:** Don't want to build? Simply select from our library of pre-configured agents and put them to work immediately.
    
-   **Agent Interaction:** Whether you've built your own or are using pre-configured agents, easily run and interact with them through our user-friendly      interface.
+   **Agent Interaction:** Whether you've built your own or are using pre-configured agents, easily run and interact with them through our user-friendly interface.
 
    **Monitoring and Analytics:** Keep track of your agents' performance and gain insights to continually improve your automation processes.
 

autogpt_platform/AGENTS.md

@@ -0,0 +1,120 @@
+# AutoGPT Platform
+
+This file provides guidance to coding agents when working with code in this repository.
+
+## Repository Overview
+
+AutoGPT Platform is a monorepo containing:
+
+- **Backend** (`backend`): Python FastAPI server with async support
+- **Frontend** (`frontend`): Next.js React application
+- **Shared Libraries** (`autogpt_libs`): Common Python utilities
+
+## Component Documentation
+
+- **Backend**: See @backend/AGENTS.md for backend-specific commands, architecture, and development tasks
+- **Frontend**: See @frontend/AGENTS.md for frontend-specific commands, architecture, and development patterns
+
+## Key Concepts
+
+1. **Agent Graphs**: Workflow definitions stored as JSON, executed by the backend
+2. **Blocks**: Reusable components in `backend/backend/blocks/` that perform specific tasks
+3. **Integrations**: OAuth and API connections stored per user
+4. **Store**: Marketplace for sharing agent templates
+5. **Virus Scanning**: ClamAV integration for file upload security
+
+### Environment Configuration
+
+#### Configuration Files
+
+- **Backend**: `backend/.env.default` (defaults) → `backend/.env` (user overrides)
+- **Frontend**: `frontend/.env.default` (defaults) → `frontend/.env` (user overrides)
+- **Platform**: `.env.default` (Supabase/shared defaults) → `.env` (user overrides)
+
+#### Docker Environment Loading Order
+
+1. `.env.default` files provide base configuration (tracked in git)
+2. `.env` files provide user-specific overrides (gitignored)
+3. Docker Compose `environment:` sections provide service-specific overrides
+4. Shell environment variables have highest precedence
+
+#### Key Points
+
+- All services use hardcoded defaults in docker-compose files (no `${VARIABLE}` substitutions)
+- The `env_file` directive loads variables INTO containers at runtime
+- Backend/Frontend services use YAML anchors for consistent configuration
+- Supabase services (`db/docker/docker-compose.yml`) follow the same pattern
+
+### Branching Strategy
+
+- **`dev`** is the main development branch. All PRs should target `dev`.
+- **`master`** is the production branch. Only used for production releases.
+
+### Creating Pull Requests
+
+- Create the PR against the `dev` branch of the repository.
+- **Split PRs by concern** — each PR should have a single clear purpose. For example, "usage tracking" and "credit charging" should be separate PRs even if related. Combining multiple concerns makes it harder for reviewers to understand what belongs to what.
+- Ensure the branch name is descriptive (e.g., `feature/add-new-block`)
+- Use conventional commit messages (see below)
+- **Structure the PR description with Why / What / How** — Why: the motivation (what problem it solves, what's broken/missing without it); What: high-level summary of changes; How: approach, key implementation details, or architecture decisions. Reviewers need all three to judge whether the approach fits the problem.
+- Fill out the .github/PULL_REQUEST_TEMPLATE.md template as the PR description
+- Always use `--body-file` to pass PR body — avoids shell interpretation of backticks and special characters:
+  ```bash
+  PR_BODY=$(mktemp)
+  cat > "$PR_BODY" << 'PREOF'
+  ## Summary
+  - use `backticks` freely here
+  PREOF
+  gh pr create --title "..." --body-file "$PR_BODY" --base dev
+  rm "$PR_BODY"
+  ```
+- Run the github pre-commit hooks to ensure code quality.
+
+### Test-Driven Development (TDD)
+
+When fixing a bug or adding a feature, follow a test-first approach:
+
+1. **Write a failing test first** — create a test that reproduces the bug or validates the new behavior, marked with `@pytest.mark.xfail` (backend) or `.fixme` (Playwright). Run it to confirm it fails for the right reason.
+2. **Implement the fix/feature** — write the minimal code to make the test pass.
+3. **Remove the xfail marker** — once the test passes, remove the `xfail`/`.fixme` annotation and run the full test suite to confirm nothing else broke.
+
+This ensures every change is covered by a test and that the test actually validates the intended behavior.
+
+### Reviewing/Revising Pull Requests
+
+Use `/pr-review` to review a PR or `/pr-address` to address comments.
+
+When fetching comments manually:
+- `gh api repos/Significant-Gravitas/AutoGPT/pulls/{N}/reviews --paginate` — top-level reviews
+- `gh api repos/Significant-Gravitas/AutoGPT/pulls/{N}/comments --paginate` — inline review comments (always paginate to avoid missing comments beyond page 1)
+- `gh api repos/Significant-Gravitas/AutoGPT/issues/{N}/comments` — PR conversation comments
+
+### Conventional Commits
+
+Use this format for commit messages and Pull Request titles:
+
+**Conventional Commit Types:**
+
+- `feat`: Introduces a new feature to the codebase
+- `fix`: Patches a bug in the codebase
+- `refactor`: Code change that neither fixes a bug nor adds a feature; also applies to removing features
+- `ci`: Changes to CI configuration
+- `docs`: Documentation-only changes
+- `dx`: Improvements to the developer experience
+
+**Recommended Base Scopes:**
+
+- `platform`: Changes affecting both frontend and backend
+- `frontend`
+- `backend`
+- `infra`
+- `blocks`: Modifications/additions of individual blocks
+
+**Subscope Examples:**
+
+- `backend/executor`
+- `backend/db`
+- `frontend/builder` (includes changes to the block UI component)
+- `infra/prod`
+
+Use these scopes and subscopes for clarity and consistency in commit messages.

autogpt_platform/CLAUDE.md

@@ -1,120 +1 @@
-# CLAUDE.md
-
-This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
-
-## Repository Overview
-
-AutoGPT Platform is a monorepo containing:
-
-- **Backend** (`backend`): Python FastAPI server with async support
-- **Frontend** (`frontend`): Next.js React application
-- **Shared Libraries** (`autogpt_libs`): Common Python utilities
-
-## Component Documentation
-
-- **Backend**: See @backend/CLAUDE.md for backend-specific commands, architecture, and development tasks
-- **Frontend**: See @frontend/CLAUDE.md for frontend-specific commands, architecture, and development patterns
-
-## Key Concepts
-
-1. **Agent Graphs**: Workflow definitions stored as JSON, executed by the backend
-2. **Blocks**: Reusable components in `backend/backend/blocks/` that perform specific tasks
-3. **Integrations**: OAuth and API connections stored per user
-4. **Store**: Marketplace for sharing agent templates
-5. **Virus Scanning**: ClamAV integration for file upload security
-
-### Environment Configuration
-
-#### Configuration Files
-
-- **Backend**: `backend/.env.default` (defaults) → `backend/.env` (user overrides)
-- **Frontend**: `frontend/.env.default` (defaults) → `frontend/.env` (user overrides)
-- **Platform**: `.env.default` (Supabase/shared defaults) → `.env` (user overrides)
-
-#### Docker Environment Loading Order
-
-1. `.env.default` files provide base configuration (tracked in git)
-2. `.env` files provide user-specific overrides (gitignored)
-3. Docker Compose `environment:` sections provide service-specific overrides
-4. Shell environment variables have highest precedence
-
-#### Key Points
-
-- All services use hardcoded defaults in docker-compose files (no `${VARIABLE}` substitutions)
-- The `env_file` directive loads variables INTO containers at runtime
-- Backend/Frontend services use YAML anchors for consistent configuration
-- Supabase services (`db/docker/docker-compose.yml`) follow the same pattern
-
-### Branching Strategy
-
-- **`dev`** is the main development branch. All PRs should target `dev`.
-- **`master`** is the production branch. Only used for production releases.
-
-### Creating Pull Requests
-
-- Create the PR against the `dev` branch of the repository.
-- **Split PRs by concern** — each PR should have a single clear purpose. For example, "usage tracking" and "credit charging" should be separate PRs even if related. Combining multiple concerns makes it harder for reviewers to understand what belongs to what.
-- Ensure the branch name is descriptive (e.g., `feature/add-new-block`)
-- Use conventional commit messages (see below)
-- **Structure the PR description with Why / What / How** — Why: the motivation (what problem it solves, what's broken/missing without it); What: high-level summary of changes; How: approach, key implementation details, or architecture decisions. Reviewers need all three to judge whether the approach fits the problem.
-- Fill out the .github/PULL_REQUEST_TEMPLATE.md template as the PR description
-- Always use `--body-file` to pass PR body — avoids shell interpretation of backticks and special characters:
-  ```bash
-  PR_BODY=$(mktemp)
-  cat > "$PR_BODY" << 'PREOF'
-  ## Summary
-  - use `backticks` freely here
-  PREOF
-  gh pr create --title "..." --body-file "$PR_BODY" --base dev
-  rm "$PR_BODY"
-  ```
-- Run the github pre-commit hooks to ensure code quality.
-
-### Test-Driven Development (TDD)
-
-When fixing a bug or adding a feature, follow a test-first approach:
-
-1. **Write a failing test first** — create a test that reproduces the bug or validates the new behavior, marked with `@pytest.mark.xfail` (backend) or `.fixme` (Playwright). Run it to confirm it fails for the right reason.
-2. **Implement the fix/feature** — write the minimal code to make the test pass.
-3. **Remove the xfail marker** — once the test passes, remove the `xfail`/`.fixme` annotation and run the full test suite to confirm nothing else broke.
-
-This ensures every change is covered by a test and that the test actually validates the intended behavior.
-
-### Reviewing/Revising Pull Requests
-
-Use `/pr-review` to review a PR or `/pr-address` to address comments.
-
-When fetching comments manually:
-- `gh api repos/Significant-Gravitas/AutoGPT/pulls/{N}/reviews --paginate` — top-level reviews
-- `gh api repos/Significant-Gravitas/AutoGPT/pulls/{N}/comments --paginate` — inline review comments (always paginate to avoid missing comments beyond page 1)
-- `gh api repos/Significant-Gravitas/AutoGPT/issues/{N}/comments` — PR conversation comments
-
-### Conventional Commits
-
-Use this format for commit messages and Pull Request titles:
-
-**Conventional Commit Types:**
-
-- `feat`: Introduces a new feature to the codebase
-- `fix`: Patches a bug in the codebase
-- `refactor`: Code change that neither fixes a bug nor adds a feature; also applies to removing features
-- `ci`: Changes to CI configuration
-- `docs`: Documentation-only changes
-- `dx`: Improvements to the developer experience
-
-**Recommended Base Scopes:**
-
-- `platform`: Changes affecting both frontend and backend
-- `frontend`
-- `backend`
-- `infra`
-- `blocks`: Modifications/additions of individual blocks
-
-**Subscope Examples:**
-
-- `backend/executor`
-- `backend/db`
-- `frontend/builder` (includes changes to the block UI component)
-- `infra/prod`
-
-Use these scopes and subscopes for clarity and consistency in commit messages.
+@AGENTS.md

autogpt_platform/backend/.env.default

@@ -178,6 +178,7 @@ SMTP_USERNAME=
 SMTP_PASSWORD=
 
 # Business & Marketing Tools
+AGENTMAIL_API_KEY=
 APOLLO_API_KEY=
 ENRICHLAYER_API_KEY=
 AYRSHARE_API_KEY=

autogpt_platform/backend/AGENTS.md

@@ -0,0 +1,227 @@
+# Backend
+
+This file provides guidance to coding agents when working with the backend.
+
+## Essential Commands
+
+To run something with Python package dependencies you MUST use `poetry run ...`.
+
+```bash
+# Install dependencies
+poetry install
+
+# Run database migrations
+poetry run prisma migrate dev
+
+# Start all services (database, redis, rabbitmq, clamav)
+docker compose up -d
+
+# Run the backend as a whole
+poetry run app
+
+# Run tests
+poetry run test
+
+# Run specific test
+poetry run pytest path/to/test_file.py::test_function_name
+
+# Run block tests (tests that validate all blocks work correctly)
+poetry run pytest backend/blocks/test/test_block.py -xvs
+
+# Run tests for a specific block (e.g., GetCurrentTimeBlock)
+poetry run pytest 'backend/blocks/test/test_block.py::test_available_blocks[GetCurrentTimeBlock]' -xvs
+
+# Lint and format
+# prefer format if you want to just "fix" it and only get the errors that can't be autofixed
+poetry run format  # Black + isort
+poetry run lint    # ruff
+```
+
+More details can be found in @TESTING.md
+
+### Creating/Updating Snapshots
+
+When you first write a test or when the expected output changes:
+
+```bash
+poetry run pytest path/to/test.py --snapshot-update
+```
+
+⚠️ **Important**: Always review snapshot changes before committing! Use `git diff` to verify the changes are expected.
+
+## Architecture
+
+- **API Layer**: FastAPI with REST and WebSocket endpoints
+- **Database**: PostgreSQL with Prisma ORM, includes pgvector for embeddings
+- **Queue System**: RabbitMQ for async task processing
+- **Execution Engine**: Separate executor service processes agent workflows
+- **Authentication**: JWT-based with Supabase integration
+- **Security**: Cache protection middleware prevents sensitive data caching in browsers/proxies
+
+## Code Style
+
+- **Top-level imports only** — no local/inner imports (lazy imports only for heavy optional deps like `openpyxl`)
+- **Absolute imports** — use `from backend.module import ...` for cross-package imports. Single-dot relative (`from .sibling import ...`) is acceptable for sibling modules within the same package (e.g., blocks). Avoid double-dot relative imports (`from ..parent import ...`) — use the absolute path instead
+- **No duck typing** — no `hasattr`/`getattr`/`isinstance` for type dispatch; use typed interfaces/unions/protocols
+- **Pydantic models** over dataclass/namedtuple/dict for structured data
+- **No linter suppressors** — no `# type: ignore`, `# noqa`, `# pyright: ignore`; fix the type/code
+- **List comprehensions** over manual loop-and-append
+- **Early return** — guard clauses first, avoid deep nesting
+- **f-strings vs printf syntax in log statements** — Use `%s` for deferred interpolation in `debug` statements, f-strings elsewhere for readability: `logger.debug("Processing %s items", count)`, `logger.info(f"Processing {count} items")`
+- **Sanitize error paths** — `os.path.basename()` in error messages to avoid leaking directory structure
+- **TOCTOU awareness** — avoid check-then-act patterns for file access and credit charging
+- **`Security()` vs `Depends()`** — use `Security()` for auth deps to get proper OpenAPI security spec
+- **Redis pipelines** — `transaction=True` for atomicity on multi-step operations
+- **`max(0, value)` guards** — for computed values that should never be negative
+- **SSE protocol** — `data:` lines for frontend-parsed events (must match Zod schema), `: comment` lines for heartbeats/status
+- **File length** — keep files under ~300 lines; if a file grows beyond this, split by responsibility (e.g. extract helpers, models, or a sub-module into a new file). Never keep appending to a long file.
+- **Function length** — keep functions under ~40 lines; extract named helpers when a function grows longer. Long functions are a sign of mixed concerns, not complexity.
+- **Top-down ordering** — define the main/public function or class first, then the helpers it uses below. A reader should encounter high-level logic before implementation details.
+
+## Testing Approach
+
+- Uses pytest with snapshot testing for API responses
+- Test files are colocated with source files (`*_test.py`)
+- Mock at boundaries — mock where the symbol is **used**, not where it's **defined**
+- After refactoring, update mock targets to match new module paths
+- Use `AsyncMock` for async functions (`from unittest.mock import AsyncMock`)
+
+### Test-Driven Development (TDD)
+
+When fixing a bug or adding a feature, write the test **before** the implementation:
+
+```python
+# 1. Write a failing test marked xfail
+@pytest.mark.xfail(reason="Bug #1234: widget crashes on empty input")
+def test_widget_handles_empty_input():
+    result = widget.process("")
+    assert result == Widget.EMPTY_RESULT
+
+# 2. Run it — confirm it fails (XFAIL)
+# poetry run pytest path/to/test.py::test_widget_handles_empty_input -xvs
+
+# 3. Implement the fix
+
+# 4. Remove xfail, run again — confirm it passes
+def test_widget_handles_empty_input():
+    result = widget.process("")
+    assert result == Widget.EMPTY_RESULT
+```
+
+This catches regressions and proves the fix actually works. **Every bug fix should include a test that would have caught it.**
+
+## Database Schema
+
+Key models (defined in `schema.prisma`):
+
+- `User`: Authentication and profile data
+- `AgentGraph`: Workflow definitions with version control
+- `AgentGraphExecution`: Execution history and results
+- `AgentNode`: Individual nodes in a workflow
+- `StoreListing`: Marketplace listings for sharing agents
+
+## Environment Configuration
+
+- **Backend**: `.env.default` (defaults) → `.env` (user overrides)
+
+## Common Development Tasks
+
+### Adding a new block
+
+Follow the comprehensive [Block SDK Guide](@../../docs/platform/block-sdk-guide.md) which covers:
+
+- Provider configuration with `ProviderBuilder`
+- Block schema definition
+- Authentication (API keys, OAuth, webhooks)
+- Testing and validation
+- File organization
+
+Quick steps:
+
+1. Create new file in `backend/blocks/`
+2. Configure provider using `ProviderBuilder` in `_config.py`
+3. Inherit from `Block` base class
+4. Define input/output schemas using `BlockSchema`
+5. Implement async `run` method
+6. Generate unique block ID using `uuid.uuid4()`
+7. Test with `poetry run pytest backend/blocks/test/test_block.py`
+
+Note: when making many new blocks analyze the interfaces for each of these blocks and picture if they would go well together in a graph-based editor or would they struggle to connect productively?
+ex: do the inputs and outputs tie well together?
+
+If you get any pushback or hit complex block conditions check the new_blocks guide in the docs.
+
+#### Handling files in blocks with `store_media_file()`
+
+When blocks need to work with files (images, videos, documents), use `store_media_file()` from `backend.util.file`. The `return_format` parameter determines what you get back:
+
+| Format | Use When | Returns |
+|--------|----------|---------|
+| `"for_local_processing"` | Processing with local tools (ffmpeg, MoviePy, PIL) | Local file path (e.g., `"image.png"`) |
+| `"for_external_api"` | Sending content to external APIs (Replicate, OpenAI) | Data URI (e.g., `"data:image/png;base64,..."`) |
+| `"for_block_output"` | Returning output from your block | Smart: `workspace://` in CoPilot, data URI in graphs |
+
+**Examples:**
+
+```python
+# INPUT: Need to process file locally with ffmpeg
+local_path = await store_media_file(
+    file=input_data.video,
+    execution_context=execution_context,
+    return_format="for_local_processing",
+)
+# local_path = "video.mp4" - use with Path/ffmpeg/etc
+
+# INPUT: Need to send to external API like Replicate
+image_b64 = await store_media_file(
+    file=input_data.image,
+    execution_context=execution_context,
+    return_format="for_external_api",
+)
+# image_b64 = "data:image/png;base64,iVBORw0..." - send to API
+
+# OUTPUT: Returning result from block
+result_url = await store_media_file(
+    file=generated_image_url,
+    execution_context=execution_context,
+    return_format="for_block_output",
+)
+yield "image_url", result_url
+# In CoPilot: result_url = "workspace://abc123"
+# In graphs:  result_url = "data:image/png;base64,..."
+```
+
+**Key points:**
+
+- `for_block_output` is the ONLY format that auto-adapts to execution context
+- Always use `for_block_output` for block outputs unless you have a specific reason not to
+- Never hardcode workspace checks - let `for_block_output` handle it
+
+### Modifying the API
+
+1. Update route in `backend/api/features/`
+2. Add/update Pydantic models in same directory
+3. Write tests alongside the route file
+4. Run `poetry run test` to verify
+
+## Workspace & Media Files
+
+**Read [Workspace & Media Architecture](../../docs/platform/workspace-media-architecture.md) when:**
+- Working on CoPilot file upload/download features
+- Building blocks that handle `MediaFileType` inputs/outputs
+- Modifying `WorkspaceManager` or `store_media_file()`
+- Debugging file persistence or virus scanning issues
+
+Covers: `WorkspaceManager` (persistent storage with session scoping), `store_media_file()` (media normalization pipeline), and responsibility boundaries for virus scanning and persistence.
+
+## Security Implementation
+
+### Cache Protection Middleware
+
+- Located in `backend/api/middleware/security.py`
+- Default behavior: Disables caching for ALL endpoints with `Cache-Control: no-store, no-cache, must-revalidate, private`
+- Uses an allow list approach - only explicitly permitted paths can be cached
+- Cacheable paths include: static assets (`static/*`, `_next/static/*`), health checks, public store pages, documentation
+- Prevents sensitive data (auth tokens, API keys, user data) from being cached by browsers/proxies
+- To allow caching for a new endpoint, add it to `CACHEABLE_PATHS` in the middleware
+- Applied to both main API server and external API applications

autogpt_platform/backend/CLAUDE.md

@@ -1,227 +1 @@
-# CLAUDE.md - Backend
-
-This file provides guidance to Claude Code when working with the backend.
-
-## Essential Commands
-
-To run something with Python package dependencies you MUST use `poetry run ...`.
-
-```bash
-# Install dependencies
-poetry install
-
-# Run database migrations
-poetry run prisma migrate dev
-
-# Start all services (database, redis, rabbitmq, clamav)
-docker compose up -d
-
-# Run the backend as a whole
-poetry run app
-
-# Run tests
-poetry run test
-
-# Run specific test
-poetry run pytest path/to/test_file.py::test_function_name
-
-# Run block tests (tests that validate all blocks work correctly)
-poetry run pytest backend/blocks/test/test_block.py -xvs
-
-# Run tests for a specific block (e.g., GetCurrentTimeBlock)
-poetry run pytest 'backend/blocks/test/test_block.py::test_available_blocks[GetCurrentTimeBlock]' -xvs
-
-# Lint and format
-# prefer format if you want to just "fix" it and only get the errors that can't be autofixed
-poetry run format  # Black + isort
-poetry run lint    # ruff
-```
-
-More details can be found in @TESTING.md
-
-### Creating/Updating Snapshots
-
-When you first write a test or when the expected output changes:
-
-```bash
-poetry run pytest path/to/test.py --snapshot-update
-```
-
-⚠️ **Important**: Always review snapshot changes before committing! Use `git diff` to verify the changes are expected.
-
-## Architecture
-
-- **API Layer**: FastAPI with REST and WebSocket endpoints
-- **Database**: PostgreSQL with Prisma ORM, includes pgvector for embeddings
-- **Queue System**: RabbitMQ for async task processing
-- **Execution Engine**: Separate executor service processes agent workflows
-- **Authentication**: JWT-based with Supabase integration
-- **Security**: Cache protection middleware prevents sensitive data caching in browsers/proxies
-
-## Code Style
-
-- **Top-level imports only** — no local/inner imports (lazy imports only for heavy optional deps like `openpyxl`)
-- **Absolute imports** — use `from backend.module import ...` for cross-package imports. Single-dot relative (`from .sibling import ...`) is acceptable for sibling modules within the same package (e.g., blocks). Avoid double-dot relative imports (`from ..parent import ...`) — use the absolute path instead
-- **No duck typing** — no `hasattr`/`getattr`/`isinstance` for type dispatch; use typed interfaces/unions/protocols
-- **Pydantic models** over dataclass/namedtuple/dict for structured data
-- **No linter suppressors** — no `# type: ignore`, `# noqa`, `# pyright: ignore`; fix the type/code
-- **List comprehensions** over manual loop-and-append
-- **Early return** — guard clauses first, avoid deep nesting
-- **f-strings vs printf syntax in log statements** — Use `%s` for deferred interpolation in `debug` statements, f-strings elsewhere for readability: `logger.debug("Processing %s items", count)`, `logger.info(f"Processing {count} items")`
-- **Sanitize error paths** — `os.path.basename()` in error messages to avoid leaking directory structure
-- **TOCTOU awareness** — avoid check-then-act patterns for file access and credit charging
-- **`Security()` vs `Depends()`** — use `Security()` for auth deps to get proper OpenAPI security spec
-- **Redis pipelines** — `transaction=True` for atomicity on multi-step operations
-- **`max(0, value)` guards** — for computed values that should never be negative
-- **SSE protocol** — `data:` lines for frontend-parsed events (must match Zod schema), `: comment` lines for heartbeats/status
-- **File length** — keep files under ~300 lines; if a file grows beyond this, split by responsibility (e.g. extract helpers, models, or a sub-module into a new file). Never keep appending to a long file.
-- **Function length** — keep functions under ~40 lines; extract named helpers when a function grows longer. Long functions are a sign of mixed concerns, not complexity.
-- **Top-down ordering** — define the main/public function or class first, then the helpers it uses below. A reader should encounter high-level logic before implementation details.
-
-## Testing Approach
-
-- Uses pytest with snapshot testing for API responses
-- Test files are colocated with source files (`*_test.py`)
-- Mock at boundaries — mock where the symbol is **used**, not where it's **defined**
-- After refactoring, update mock targets to match new module paths
-- Use `AsyncMock` for async functions (`from unittest.mock import AsyncMock`)
-
-### Test-Driven Development (TDD)
-
-When fixing a bug or adding a feature, write the test **before** the implementation:
-
-```python
-# 1. Write a failing test marked xfail
-@pytest.mark.xfail(reason="Bug #1234: widget crashes on empty input")
-def test_widget_handles_empty_input():
-    result = widget.process("")
-    assert result == Widget.EMPTY_RESULT
-
-# 2. Run it — confirm it fails (XFAIL)
-# poetry run pytest path/to/test.py::test_widget_handles_empty_input -xvs
-
-# 3. Implement the fix
-
-# 4. Remove xfail, run again — confirm it passes
-def test_widget_handles_empty_input():
-    result = widget.process("")
-    assert result == Widget.EMPTY_RESULT
-```
-
-This catches regressions and proves the fix actually works. **Every bug fix should include a test that would have caught it.**
-
-## Database Schema
-
-Key models (defined in `schema.prisma`):
-
-- `User`: Authentication and profile data
-- `AgentGraph`: Workflow definitions with version control
-- `AgentGraphExecution`: Execution history and results
-- `AgentNode`: Individual nodes in a workflow
-- `StoreListing`: Marketplace listings for sharing agents
-
-## Environment Configuration
-
-- **Backend**: `.env.default` (defaults) → `.env` (user overrides)
-
-## Common Development Tasks
-
-### Adding a new block
-
-Follow the comprehensive [Block SDK Guide](@../../docs/content/platform/block-sdk-guide.md) which covers:
-
-- Provider configuration with `ProviderBuilder`
-- Block schema definition
-- Authentication (API keys, OAuth, webhooks)
-- Testing and validation
-- File organization
-
-Quick steps:
-
-1. Create new file in `backend/blocks/`
-2. Configure provider using `ProviderBuilder` in `_config.py`
-3. Inherit from `Block` base class
-4. Define input/output schemas using `BlockSchema`
-5. Implement async `run` method
-6. Generate unique block ID using `uuid.uuid4()`
-7. Test with `poetry run pytest backend/blocks/test/test_block.py`
-
-Note: when making many new blocks analyze the interfaces for each of these blocks and picture if they would go well together in a graph-based editor or would they struggle to connect productively?
-ex: do the inputs and outputs tie well together?
-
-If you get any pushback or hit complex block conditions check the new_blocks guide in the docs.
-
-#### Handling files in blocks with `store_media_file()`
-
-When blocks need to work with files (images, videos, documents), use `store_media_file()` from `backend.util.file`. The `return_format` parameter determines what you get back:
-
-| Format | Use When | Returns |
-|--------|----------|---------|
-| `"for_local_processing"` | Processing with local tools (ffmpeg, MoviePy, PIL) | Local file path (e.g., `"image.png"`) |
-| `"for_external_api"` | Sending content to external APIs (Replicate, OpenAI) | Data URI (e.g., `"data:image/png;base64,..."`) |
-| `"for_block_output"` | Returning output from your block | Smart: `workspace://` in CoPilot, data URI in graphs |
-
-**Examples:**
-
-```python
-# INPUT: Need to process file locally with ffmpeg
-local_path = await store_media_file(
-    file=input_data.video,
-    execution_context=execution_context,
-    return_format="for_local_processing",
-)
-# local_path = "video.mp4" - use with Path/ffmpeg/etc
-
-# INPUT: Need to send to external API like Replicate
-image_b64 = await store_media_file(
-    file=input_data.image,
-    execution_context=execution_context,
-    return_format="for_external_api",
-)
-# image_b64 = "data:image/png;base64,iVBORw0..." - send to API
-
-# OUTPUT: Returning result from block
-result_url = await store_media_file(
-    file=generated_image_url,
-    execution_context=execution_context,
-    return_format="for_block_output",
-)
-yield "image_url", result_url
-# In CoPilot: result_url = "workspace://abc123"
-# In graphs:  result_url = "data:image/png;base64,..."
-```
-
-**Key points:**
-
-- `for_block_output` is the ONLY format that auto-adapts to execution context
-- Always use `for_block_output` for block outputs unless you have a specific reason not to
-- Never hardcode workspace checks - let `for_block_output` handle it
-
-### Modifying the API
-
-1. Update route in `backend/api/features/`
-2. Add/update Pydantic models in same directory
-3. Write tests alongside the route file
-4. Run `poetry run test` to verify
-
-## Workspace & Media Files
-
-**Read [Workspace & Media Architecture](../../docs/platform/workspace-media-architecture.md) when:**
-- Working on CoPilot file upload/download features
-- Building blocks that handle `MediaFileType` inputs/outputs
-- Modifying `WorkspaceManager` or `store_media_file()`
-- Debugging file persistence or virus scanning issues
-
-Covers: `WorkspaceManager` (persistent storage with session scoping), `store_media_file()` (media normalization pipeline), and responsibility boundaries for virus scanning and persistence.
-
-## Security Implementation
-
-### Cache Protection Middleware
-
-- Located in `backend/api/middleware/security.py`
-- Default behavior: Disables caching for ALL endpoints with `Cache-Control: no-store, no-cache, must-revalidate, private`
-- Uses an allow list approach - only explicitly permitted paths can be cached
-- Cacheable paths include: static assets (`static/*`, `_next/static/*`), health checks, public store pages, documentation
-- Prevents sensitive data (auth tokens, API keys, user data) from being cached by browsers/proxies
-- To allow caching for a new endpoint, add it to `CACHEABLE_PATHS` in the middleware
-- Applied to both main API server and external API applications
+@AGENTS.md

autogpt_platform/backend/backend/api/external/v1/integrations.py

@@ -31,7 +31,10 @@ from backend.data.model import (
     UserPasswordCredentials,
     is_sdk_default,
 )
-from backend.integrations.credentials_store import provider_matches
+from backend.integrations.credentials_store import (
+    is_system_credential,
+    provider_matches,
+)
 from backend.integrations.creds_manager import IntegrationCredentialsManager
 from backend.integrations.oauth import CREDENTIALS_BY_PROVIDER, HANDLERS_BY_NAME
 from backend.integrations.providers import ProviderName
@@ -618,6 +621,11 @@ async def delete_credential(
         raise HTTPException(
             status_code=status.HTTP_404_NOT_FOUND, detail="Credentials not found"
         )
+    if is_system_credential(cred_id):
+        raise HTTPException(
+            status_code=status.HTTP_403_FORBIDDEN,
+            detail="System-managed credentials cannot be deleted",
+        )
     creds = await creds_manager.store.get_creds_by_id(auth.user_id, cred_id)
     if not creds:
         raise HTTPException(

autogpt_platform/backend/backend/api/external/v1/tools.py

@@ -72,7 +72,7 @@ class RunAgentRequest(BaseModel):
 
 def _create_ephemeral_session(user_id: str) -> ChatSession:
     """Create an ephemeral session for stateless API requests."""
-    return ChatSession.new(user_id)
+    return ChatSession.new(user_id, dry_run=False)
 
 
 @tools_router.post(

autogpt_platform/backend/backend/api/features/admin/rate_limit_admin_routes.py

@@ -0,0 +1,146 @@
+"""Admin endpoints for checking and resetting user CoPilot rate limit usage."""
+
+import logging
+from typing import Optional
+
+from autogpt_libs.auth import get_user_id, requires_admin_user
+from fastapi import APIRouter, Body, HTTPException, Security
+from pydantic import BaseModel
+
+from backend.copilot.config import ChatConfig
+from backend.copilot.rate_limit import (
+    get_global_rate_limits,
+    get_usage_status,
+    reset_user_usage,
+)
+from backend.data.user import get_user_by_email, get_user_email_by_id
+
+logger = logging.getLogger(__name__)
+
+config = ChatConfig()
+
+router = APIRouter(
+    prefix="/admin",
+    tags=["copilot", "admin"],
+    dependencies=[Security(requires_admin_user)],
+)
+
+
+class UserRateLimitResponse(BaseModel):
+    user_id: str
+    user_email: Optional[str] = None
+    daily_token_limit: int
+    weekly_token_limit: int
+    daily_tokens_used: int
+    weekly_tokens_used: int
+
+
+async def _resolve_user_id(
+    user_id: Optional[str], email: Optional[str]
+) -> tuple[str, Optional[str]]:
+    """Resolve a user_id and email from the provided parameters.
+
+    Returns (user_id, email). Accepts either user_id or email; at least one
+    must be provided.  When both are provided, ``email`` takes precedence.
+    """
+    if email:
+        user = await get_user_by_email(email)
+        if not user:
+            raise HTTPException(
+                status_code=404, detail="No user found with the provided email."
+            )
+        return user.id, email
+
+    if not user_id:
+        raise HTTPException(
+            status_code=400,
+            detail="Either user_id or email query parameter is required.",
+        )
+
+    # We have a user_id; try to look up their email for display purposes.
+    # This is non-critical -- a failure should not block the response.
+    try:
+        resolved_email = await get_user_email_by_id(user_id)
+    except Exception:
+        logger.warning("Failed to resolve email for user %s", user_id, exc_info=True)
+        resolved_email = None
+    return user_id, resolved_email
+
+
+@router.get(
+    "/rate_limit",
+    response_model=UserRateLimitResponse,
+    summary="Get User Rate Limit",
+)
+async def get_user_rate_limit(
+    user_id: Optional[str] = None,
+    email: Optional[str] = None,
+    admin_user_id: str = Security(get_user_id),
+) -> UserRateLimitResponse:
+    """Get a user's current usage and effective rate limits. Admin-only.
+
+    Accepts either ``user_id`` or ``email`` as a query parameter.
+    When ``email`` is provided the user is looked up by email first.
+    """
+    resolved_id, resolved_email = await _resolve_user_id(user_id, email)
+
+    logger.info("Admin %s checking rate limit for user %s", admin_user_id, resolved_id)
+
+    daily_limit, weekly_limit = await get_global_rate_limits(
+        resolved_id, config.daily_token_limit, config.weekly_token_limit
+    )
+    usage = await get_usage_status(resolved_id, daily_limit, weekly_limit)
+
+    return UserRateLimitResponse(
+        user_id=resolved_id,
+        user_email=resolved_email,
+        daily_token_limit=daily_limit,
+        weekly_token_limit=weekly_limit,
+        daily_tokens_used=usage.daily.used,
+        weekly_tokens_used=usage.weekly.used,
+    )
+
+
+@router.post(
+    "/rate_limit/reset",
+    response_model=UserRateLimitResponse,
+    summary="Reset User Rate Limit Usage",
+)
+async def reset_user_rate_limit(
+    user_id: str = Body(embed=True),
+    reset_weekly: bool = Body(False, embed=True),
+    admin_user_id: str = Security(get_user_id),
+) -> UserRateLimitResponse:
+    """Reset a user's daily usage counter (and optionally weekly). Admin-only."""
+    logger.info(
+        "Admin %s resetting rate limit for user %s (reset_weekly=%s)",
+        admin_user_id,
+        user_id,
+        reset_weekly,
+    )
+
+    try:
+        await reset_user_usage(user_id, reset_weekly=reset_weekly)
+    except Exception as e:
+        logger.exception("Failed to reset user usage")
+        raise HTTPException(status_code=500, detail="Failed to reset usage") from e
+
+    daily_limit, weekly_limit = await get_global_rate_limits(
+        user_id, config.daily_token_limit, config.weekly_token_limit
+    )
+    usage = await get_usage_status(user_id, daily_limit, weekly_limit)
+
+    try:
+        resolved_email = await get_user_email_by_id(user_id)
+    except Exception:
+        logger.warning("Failed to resolve email for user %s", user_id, exc_info=True)
+        resolved_email = None
+
+    return UserRateLimitResponse(
+        user_id=user_id,
+        user_email=resolved_email,
+        daily_token_limit=daily_limit,
+        weekly_token_limit=weekly_limit,
+        daily_tokens_used=usage.daily.used,
+        weekly_tokens_used=usage.weekly.used,
+    )

autogpt_platform/backend/backend/api/features/admin/rate_limit_admin_routes_test.py

@@ -0,0 +1,263 @@
+import json
+from types import SimpleNamespace
+from unittest.mock import AsyncMock
+
+import fastapi
+import fastapi.testclient
+import pytest
+import pytest_mock
+from autogpt_libs.auth.jwt_utils import get_jwt_payload
+from pytest_snapshot.plugin import Snapshot
+
+from backend.copilot.rate_limit import CoPilotUsageStatus, UsageWindow
+
+from .rate_limit_admin_routes import router as rate_limit_admin_router
+
+app = fastapi.FastAPI()
+app.include_router(rate_limit_admin_router)
+
+client = fastapi.testclient.TestClient(app)
+
+_MOCK_MODULE = "backend.api.features.admin.rate_limit_admin_routes"
+
+_TARGET_EMAIL = "target@example.com"
+
+
+@pytest.fixture(autouse=True)
+def setup_app_admin_auth(mock_jwt_admin):
+    """Setup admin auth overrides for all tests in this module"""
+    app.dependency_overrides[get_jwt_payload] = mock_jwt_admin["get_jwt_payload"]
+    yield
+    app.dependency_overrides.clear()
+
+
+def _mock_usage_status(
+    daily_used: int = 500_000, weekly_used: int = 3_000_000
+) -> CoPilotUsageStatus:
+    from datetime import UTC, datetime, timedelta
+
+    now = datetime.now(UTC)
+    return CoPilotUsageStatus(
+        daily=UsageWindow(
+            used=daily_used, limit=2_500_000, resets_at=now + timedelta(hours=6)
+        ),
+        weekly=UsageWindow(
+            used=weekly_used, limit=12_500_000, resets_at=now + timedelta(days=3)
+        ),
+    )
+
+
+def _patch_rate_limit_deps(
+    mocker: pytest_mock.MockerFixture,
+    target_user_id: str,
+    daily_used: int = 500_000,
+    weekly_used: int = 3_000_000,
+):
+    """Patch the common rate-limit + user-lookup dependencies."""
+    mocker.patch(
+        f"{_MOCK_MODULE}.get_global_rate_limits",
+        new_callable=AsyncMock,
+        return_value=(2_500_000, 12_500_000),
+    )
+    mocker.patch(
+        f"{_MOCK_MODULE}.get_usage_status",
+        new_callable=AsyncMock,
+        return_value=_mock_usage_status(daily_used=daily_used, weekly_used=weekly_used),
+    )
+    mocker.patch(
+        f"{_MOCK_MODULE}.get_user_email_by_id",
+        new_callable=AsyncMock,
+        return_value=_TARGET_EMAIL,
+    )
+
+
+def test_get_rate_limit(
+    mocker: pytest_mock.MockerFixture,
+    configured_snapshot: Snapshot,
+    target_user_id: str,
+) -> None:
+    """Test getting rate limit and usage for a user."""
+    _patch_rate_limit_deps(mocker, target_user_id)
+
+    response = client.get("/admin/rate_limit", params={"user_id": target_user_id})
+
+    assert response.status_code == 200
+    data = response.json()
+    assert data["user_id"] == target_user_id
+    assert data["user_email"] == _TARGET_EMAIL
+    assert data["daily_token_limit"] == 2_500_000
+    assert data["weekly_token_limit"] == 12_500_000
+    assert data["daily_tokens_used"] == 500_000
+    assert data["weekly_tokens_used"] == 3_000_000
+
+    configured_snapshot.assert_match(
+        json.dumps(data, indent=2, sort_keys=True) + "\n",
+        "get_rate_limit",
+    )
+
+
+def test_get_rate_limit_by_email(
+    mocker: pytest_mock.MockerFixture,
+    target_user_id: str,
+) -> None:
+    """Test looking up rate limits via email instead of user_id."""
+    _patch_rate_limit_deps(mocker, target_user_id)
+
+    mock_user = SimpleNamespace(id=target_user_id, email=_TARGET_EMAIL)
+    mocker.patch(
+        f"{_MOCK_MODULE}.get_user_by_email",
+        new_callable=AsyncMock,
+        return_value=mock_user,
+    )
+
+    response = client.get("/admin/rate_limit", params={"email": _TARGET_EMAIL})
+
+    assert response.status_code == 200
+    data = response.json()
+    assert data["user_id"] == target_user_id
+    assert data["user_email"] == _TARGET_EMAIL
+    assert data["daily_token_limit"] == 2_500_000
+
+
+def test_get_rate_limit_by_email_not_found(
+    mocker: pytest_mock.MockerFixture,
+) -> None:
+    """Test that looking up a non-existent email returns 404."""
+    mocker.patch(
+        f"{_MOCK_MODULE}.get_user_by_email",
+        new_callable=AsyncMock,
+        return_value=None,
+    )
+
+    response = client.get("/admin/rate_limit", params={"email": "nobody@example.com"})
+
+    assert response.status_code == 404
+
+
+def test_get_rate_limit_no_params() -> None:
+    """Test that omitting both user_id and email returns 400."""
+    response = client.get("/admin/rate_limit")
+    assert response.status_code == 400
+
+
+def test_reset_user_usage_daily_only(
+    mocker: pytest_mock.MockerFixture,
+    configured_snapshot: Snapshot,
+    target_user_id: str,
+) -> None:
+    """Test resetting only daily usage (default behaviour)."""
+    mock_reset = mocker.patch(
+        f"{_MOCK_MODULE}.reset_user_usage",
+        new_callable=AsyncMock,
+    )
+    _patch_rate_limit_deps(mocker, target_user_id, daily_used=0, weekly_used=3_000_000)
+
+    response = client.post(
+        "/admin/rate_limit/reset",
+        json={"user_id": target_user_id},
+    )
+
+    assert response.status_code == 200
+    data = response.json()
+    assert data["daily_tokens_used"] == 0
+    # Weekly is untouched
+    assert data["weekly_tokens_used"] == 3_000_000
+
+    mock_reset.assert_awaited_once_with(target_user_id, reset_weekly=False)
+
+    configured_snapshot.assert_match(
+        json.dumps(data, indent=2, sort_keys=True) + "\n",
+        "reset_user_usage_daily_only",
+    )
+
+
+def test_reset_user_usage_daily_and_weekly(
+    mocker: pytest_mock.MockerFixture,
+    configured_snapshot: Snapshot,
+    target_user_id: str,
+) -> None:
+    """Test resetting both daily and weekly usage."""
+    mock_reset = mocker.patch(
+        f"{_MOCK_MODULE}.reset_user_usage",
+        new_callable=AsyncMock,
+    )
+    _patch_rate_limit_deps(mocker, target_user_id, daily_used=0, weekly_used=0)
+
+    response = client.post(
+        "/admin/rate_limit/reset",
+        json={"user_id": target_user_id, "reset_weekly": True},
+    )
+
+    assert response.status_code == 200
+    data = response.json()
+    assert data["daily_tokens_used"] == 0
+    assert data["weekly_tokens_used"] == 0
+
+    mock_reset.assert_awaited_once_with(target_user_id, reset_weekly=True)
+
+    configured_snapshot.assert_match(
+        json.dumps(data, indent=2, sort_keys=True) + "\n",
+        "reset_user_usage_daily_and_weekly",
+    )
+
+
+def test_reset_user_usage_redis_failure(
+    mocker: pytest_mock.MockerFixture,
+    target_user_id: str,
+) -> None:
+    """Test that Redis failure on reset returns 500."""
+    mocker.patch(
+        f"{_MOCK_MODULE}.reset_user_usage",
+        new_callable=AsyncMock,
+        side_effect=Exception("Redis connection refused"),
+    )
+
+    response = client.post(
+        "/admin/rate_limit/reset",
+        json={"user_id": target_user_id},
+    )
+
+    assert response.status_code == 500
+
+
+def test_get_rate_limit_email_lookup_failure(
+    mocker: pytest_mock.MockerFixture,
+    target_user_id: str,
+) -> None:
+    """Test that failing to resolve a user email degrades gracefully."""
+    mocker.patch(
+        f"{_MOCK_MODULE}.get_global_rate_limits",
+        new_callable=AsyncMock,
+        return_value=(2_500_000, 12_500_000),
+    )
+    mocker.patch(
+        f"{_MOCK_MODULE}.get_usage_status",
+        new_callable=AsyncMock,
+        return_value=_mock_usage_status(),
+    )
+    mocker.patch(
+        f"{_MOCK_MODULE}.get_user_email_by_id",
+        new_callable=AsyncMock,
+        side_effect=Exception("DB connection lost"),
+    )
+
+    response = client.get("/admin/rate_limit", params={"user_id": target_user_id})
+
+    assert response.status_code == 200
+    data = response.json()
+    assert data["user_id"] == target_user_id
+    assert data["user_email"] is None
+
+
+def test_admin_endpoints_require_admin_role(mock_jwt_user) -> None:
+    """Test that rate limit admin endpoints require admin role."""
+    app.dependency_overrides[get_jwt_payload] = mock_jwt_user["get_jwt_payload"]
+
+    response = client.get("/admin/rate_limit", params={"user_id": "test"})
+    assert response.status_code == 403
+
+    response = client.post(
+        "/admin/rate_limit/reset",
+        json={"user_id": "test"},
+    )
+    assert response.status_code == 403

autogpt_platform/backend/backend/api/features/chat/routes.py

@@ -11,7 +11,7 @@ from autogpt_libs import auth
 from fastapi import APIRouter, HTTPException, Query, Response, Security
 from fastapi.responses import StreamingResponse
 from prisma.models import UserWorkspaceFile
-from pydantic import BaseModel, Field, field_validator
+from pydantic import BaseModel, ConfigDict, Field, field_validator
 
 from backend.copilot import service as chat_service
 from backend.copilot import stream_registry
@@ -20,6 +20,7 @@ from backend.copilot.executor.utils import enqueue_cancel_task, enqueue_copilot_
 from backend.copilot.model import (
     ChatMessage,
     ChatSession,
+    ChatSessionMetadata,
     append_and_save_message,
     create_chat_session,
     delete_chat_session,
@@ -30,8 +31,14 @@ from backend.copilot.model import (
 from backend.copilot.rate_limit import (
     CoPilotUsageStatus,
     RateLimitExceeded,
+    acquire_reset_lock,
     check_rate_limit,
+    get_daily_reset_count,
+    get_global_rate_limits,
     get_usage_status,
+    increment_daily_reset_count,
+    release_reset_lock,
+    reset_daily_usage,
 )
 from backend.copilot.response_model import StreamError, StreamFinish, StreamHeartbeat
 from backend.copilot.tools.e2b_sandbox import kill_sandbox
@@ -59,9 +66,16 @@ from backend.copilot.tools.models import (
     UnderstandingUpdatedResponse,
 )
 from backend.copilot.tracking import track_user_message
+from backend.data.credit import UsageTransactionMetadata, get_user_credit_model
 from backend.data.redis_client import get_redis_async
+from backend.data.understanding import get_business_understanding
 from backend.data.workspace import get_or_create_workspace
-from backend.util.exceptions import NotFoundError
+from backend.util.exceptions import InsufficientBalanceError, NotFoundError
+from backend.util.settings import Settings
+
+settings = Settings()
+
+logger = logging.getLogger(__name__)
 
 config = ChatConfig()
 
@@ -69,8 +83,6 @@ _UUID_RE = re.compile(
     r"^[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$", re.I
 )
 
-logger = logging.getLogger(__name__)
-
 
 async def _validate_and_get_session(
     session_id: str,
@@ -101,12 +113,25 @@ class StreamChatRequest(BaseModel):
     )  # Workspace file IDs attached to this message
 
 
+class CreateSessionRequest(BaseModel):
+    """Request model for creating a new chat session.
+
+    ``dry_run`` is a **top-level** field — do not nest it inside ``metadata``.
+    Extra/unknown fields are rejected (422) to prevent silent mis-use.
+    """
+
+    model_config = ConfigDict(extra="forbid")
+
+    dry_run: bool = False
+
+
 class CreateSessionResponse(BaseModel):
     """Response model containing information on a newly created chat session."""
 
     id: str
     created_at: str
     user_id: str | None
+    metadata: ChatSessionMetadata = ChatSessionMetadata()
 
 
 class ActiveStreamInfo(BaseModel):
@@ -127,6 +152,7 @@ class SessionDetailResponse(BaseModel):
     active_stream: ActiveStreamInfo | None = None  # Present if stream is still active
     total_prompt_tokens: int = 0
     total_completion_tokens: int = 0
+    metadata: ChatSessionMetadata = ChatSessionMetadata()
 
 
 class SessionSummaryResponse(BaseModel):
@@ -237,6 +263,7 @@ async def list_sessions(
 )
 async def create_session(
     user_id: Annotated[str, Security(auth.get_user_id)],
+    request: CreateSessionRequest | None = None,
 ) -> CreateSessionResponse:
     """
     Create a new chat session.
@@ -245,22 +272,28 @@ async def create_session(
 
     Args:
         user_id: The authenticated user ID parsed from the JWT (required).
+        request: Optional request body. When provided, ``dry_run=True``
+            forces run_block and run_agent calls to use dry-run simulation.
 
     Returns:
         CreateSessionResponse: Details of the created session.
 
     """
+    dry_run = request.dry_run if request else False
+
     logger.info(
         f"Creating session with user_id: "
         f"...{user_id[-8:] if len(user_id) > 8 else '<redacted>'}"
+        f"{', dry_run=True' if dry_run else ''}"
     )
 
-    session = await create_chat_session(user_id)
+    session = await create_chat_session(user_id, dry_run=dry_run)
 
     return CreateSessionResponse(
         id=session.session_id,
         created_at=session.started_at.isoformat(),
         user_id=session.user_id,
+        metadata=session.metadata,
     )
 
 
@@ -409,6 +442,7 @@ async def get_session(
         active_stream=active_stream_info,
         total_prompt_tokens=total_prompt,
         total_completion_tokens=total_completion,
+        metadata=session.metadata,
     )
 
 
@@ -421,11 +455,189 @@ async def get_copilot_usage(
     """Get CoPilot usage status for the authenticated user.
 
     Returns current token usage vs limits for daily and weekly windows.
+    Global defaults sourced from LaunchDarkly (falling back to config).
     """
+    daily_limit, weekly_limit = await get_global_rate_limits(
+        user_id, config.daily_token_limit, config.weekly_token_limit
+    )
     return await get_usage_status(
         user_id=user_id,
-        daily_token_limit=config.daily_token_limit,
-        weekly_token_limit=config.weekly_token_limit,
+        daily_token_limit=daily_limit,
+        weekly_token_limit=weekly_limit,
+        rate_limit_reset_cost=config.rate_limit_reset_cost,
+    )
+
+
+class RateLimitResetResponse(BaseModel):
+    """Response from resetting the daily rate limit."""
+
+    success: bool
+    credits_charged: int = Field(description="Credits charged (in cents)")
+    remaining_balance: int = Field(description="Credit balance after charge (in cents)")
+    usage: CoPilotUsageStatus = Field(description="Updated usage status after reset")
+
+
+@router.post(
+    "/usage/reset",
+    status_code=200,
+    responses={
+        400: {
+            "description": "Bad Request (feature disabled or daily limit not reached)"
+        },
+        402: {"description": "Payment Required (insufficient credits)"},
+        429: {
+            "description": "Too Many Requests (max daily resets exceeded or reset in progress)"
+        },
+        503: {
+            "description": "Service Unavailable (Redis reset failed; credits refunded or support needed)"
+        },
+    },
+)
+async def reset_copilot_usage(
+    user_id: Annotated[str, Security(auth.get_user_id)],
+) -> RateLimitResetResponse:
+    """Reset the daily CoPilot rate limit by spending credits.
+
+    Allows users who have hit their daily token limit to spend credits
+    to reset their daily usage counter and continue working.
+    Returns 400 if the feature is disabled or the user is not over the limit.
+    Returns 402 if the user has insufficient credits.
+    """
+    cost = config.rate_limit_reset_cost
+    if cost <= 0:
+        raise HTTPException(
+            status_code=400,
+            detail="Rate limit reset is not available.",
+        )
+
+    if not settings.config.enable_credit:
+        raise HTTPException(
+            status_code=400,
+            detail="Rate limit reset is not available (credit system is disabled).",
+        )
+
+    daily_limit, weekly_limit = await get_global_rate_limits(
+        user_id, config.daily_token_limit, config.weekly_token_limit
+    )
+
+    if daily_limit <= 0:
+        raise HTTPException(
+            status_code=400,
+            detail="No daily limit is configured — nothing to reset.",
+        )
+
+    # Check max daily resets.  get_daily_reset_count returns None when Redis
+    # is unavailable; reject the reset in that case to prevent unlimited
+    # free resets when the counter store is down.
+    reset_count = await get_daily_reset_count(user_id)
+    if reset_count is None:
+        raise HTTPException(
+            status_code=503,
+            detail="Unable to verify reset eligibility — please try again later.",
+        )
+    if config.max_daily_resets > 0 and reset_count >= config.max_daily_resets:
+        raise HTTPException(
+            status_code=429,
+            detail=f"You've used all {config.max_daily_resets} resets for today.",
+        )
+
+    # Acquire a per-user lock to prevent TOCTOU races (concurrent resets).
+    if not await acquire_reset_lock(user_id):
+        raise HTTPException(
+            status_code=429,
+            detail="A reset is already in progress. Please try again.",
+        )
+
+    try:
+        # Verify the user is actually at or over their daily limit.
+        # (rate_limit_reset_cost intentionally omitted — this object is only
+        # used for limit checks, not returned to the client.)
+        usage_status = await get_usage_status(
+            user_id=user_id,
+            daily_token_limit=daily_limit,
+            weekly_token_limit=weekly_limit,
+        )
+        if daily_limit > 0 and usage_status.daily.used < daily_limit:
+            raise HTTPException(
+                status_code=400,
+                detail="You have not reached your daily limit yet.",
+            )
+
+        # If the weekly limit is also exhausted, resetting the daily counter
+        # won't help — the user would still be blocked by the weekly limit.
+        if weekly_limit > 0 and usage_status.weekly.used >= weekly_limit:
+            raise HTTPException(
+                status_code=400,
+                detail="Your weekly limit is also reached. Resetting the daily limit won't help.",
+            )
+
+        # Charge credits.
+        credit_model = await get_user_credit_model(user_id)
+        try:
+            remaining = await credit_model.spend_credits(
+                user_id=user_id,
+                cost=cost,
+                metadata=UsageTransactionMetadata(
+                    reason="CoPilot daily rate limit reset",
+                ),
+            )
+        except InsufficientBalanceError as e:
+            raise HTTPException(
+                status_code=402,
+                detail="Insufficient credits to reset your rate limit.",
+            ) from e
+
+        # Reset daily usage in Redis.  If this fails, refund the credits
+        # so the user is not charged for a service they did not receive.
+        if not await reset_daily_usage(user_id, daily_token_limit=daily_limit):
+            # Compensate: refund the charged credits.
+            refunded = False
+            try:
+                await credit_model.top_up_credits(user_id, cost)
+                refunded = True
+                logger.warning(
+                    "Refunded %d credits to user %s after Redis reset failure",
+                    cost,
+                    user_id[:8],
+                )
+            except Exception:
+                logger.error(
+                    "CRITICAL: Failed to refund %d credits to user %s "
+                    "after Redis reset failure — manual intervention required",
+                    cost,
+                    user_id[:8],
+                    exc_info=True,
+                )
+            if refunded:
+                raise HTTPException(
+                    status_code=503,
+                    detail="Rate limit reset failed — please try again later. "
+                    "Your credits have not been charged.",
+                )
+            raise HTTPException(
+                status_code=503,
+                detail="Rate limit reset failed and the automatic refund "
+                "also failed. Please contact support for assistance.",
+            )
+
+        # Track the reset count for daily cap enforcement.
+        await increment_daily_reset_count(user_id)
+    finally:
+        await release_reset_lock(user_id)
+
+    # Return updated usage status.
+    updated_usage = await get_usage_status(
+        user_id=user_id,
+        daily_token_limit=daily_limit,
+        weekly_token_limit=weekly_limit,
+        rate_limit_reset_cost=config.rate_limit_reset_cost,
+    )
+
+    return RateLimitResetResponse(
+        success=True,
+        credits_charged=cost,
+        remaining_balance=remaining,
+        usage=updated_usage,
     )
 
 
@@ -526,12 +738,16 @@ async def stream_chat_post(
 
     # Pre-turn rate limit check (token-based).
     # check_rate_limit short-circuits internally when both limits are 0.
+    # Global defaults sourced from LaunchDarkly, falling back to config.
     if user_id:
         try:
+            daily_limit, weekly_limit = await get_global_rate_limits(
+                user_id, config.daily_token_limit, config.weekly_token_limit
+            )
             await check_rate_limit(
                 user_id=user_id,
-                daily_token_limit=config.daily_token_limit,
-                weekly_token_limit=config.weekly_token_limit,
+                daily_token_limit=daily_limit,
+                weekly_token_limit=weekly_limit,
             )
         except RateLimitExceeded as e:
             raise HTTPException(status_code=429, detail=str(e)) from e
@@ -894,6 +1110,47 @@ async def session_assign_user(
     return {"status": "ok"}
 
 
+# ========== Suggested Prompts ==========
+
+
+class SuggestedTheme(BaseModel):
+    """A themed group of suggested prompts."""
+
+    name: str
+    prompts: list[str]
+
+
+class SuggestedPromptsResponse(BaseModel):
+    """Response model for user-specific suggested prompts grouped by theme."""
+
+    themes: list[SuggestedTheme]
+
+
+@router.get(
+    "/suggested-prompts",
+    dependencies=[Security(auth.requires_user)],
+)
+async def get_suggested_prompts(
+    user_id: Annotated[str, Security(auth.get_user_id)],
+) -> SuggestedPromptsResponse:
+    """
+    Get LLM-generated suggested prompts grouped by theme.
+
+    Returns personalized quick-action prompts based on the user's
+    business understanding. Returns empty themes list if no custom
+    prompts are available.
+    """
+    understanding = await get_business_understanding(user_id)
+    if understanding is None or not understanding.suggested_prompts:
+        return SuggestedPromptsResponse(themes=[])
+
+    themes = [
+        SuggestedTheme(name=name, prompts=prompts)
+        for name, prompts in understanding.suggested_prompts.items()
+    ]
+    return SuggestedPromptsResponse(themes=themes)
+
+
 # ========== Configuration ==========
 
 
@@ -942,7 +1199,7 @@ async def health_check() -> dict:
     )
 
     # Create and retrieve session to verify full data layer
-    session = await create_chat_session(health_check_user_id)
+    session = await create_chat_session(health_check_user_id, dry_run=False)
     await get_chat_session(session.session_id, health_check_user_id)
 
     return {

autogpt_platform/backend/backend/api/features/chat/routes_test.py

@@ -1,7 +1,7 @@
 """Tests for chat API routes: session title update, file attachment validation, usage, and rate limiting."""
 
 from datetime import UTC, datetime, timedelta
-from unittest.mock import AsyncMock
+from unittest.mock import AsyncMock, MagicMock
 
 import fastapi
 import fastapi.testclient
@@ -368,6 +368,7 @@ def test_usage_returns_daily_and_weekly(
         user_id=test_user_id,
         daily_token_limit=10000,
         weekly_token_limit=50000,
+        rate_limit_reset_cost=chat_routes.config.rate_limit_reset_cost,
     )
 
 
@@ -380,6 +381,7 @@ def test_usage_uses_config_limits(
 
     mocker.patch.object(chat_routes.config, "daily_token_limit", 99999)
     mocker.patch.object(chat_routes.config, "weekly_token_limit", 77777)
+    mocker.patch.object(chat_routes.config, "rate_limit_reset_cost", 500)
 
     response = client.get("/usage")
 
@@ -388,6 +390,7 @@ def test_usage_uses_config_limits(
         user_id=test_user_id,
         daily_token_limit=99999,
         weekly_token_limit=77777,
+        rate_limit_reset_cost=500,
     )
 
 
@@ -400,3 +403,126 @@ def test_usage_rejects_unauthenticated_request() -> None:
     response = unauthenticated_client.get("/usage")
 
     assert response.status_code == 401
+
+
+# ─── Suggested prompts endpoint ──────────────────────────────────────
+
+
+def _mock_get_business_understanding(
+    mocker: pytest_mock.MockerFixture,
+    *,
+    return_value=None,
+):
+    """Mock get_business_understanding."""
+    return mocker.patch(
+        "backend.api.features.chat.routes.get_business_understanding",
+        new_callable=AsyncMock,
+        return_value=return_value,
+    )
+
+
+def test_suggested_prompts_returns_themes(
+    mocker: pytest_mock.MockerFixture,
+    test_user_id: str,
+) -> None:
+    """User with themed prompts gets them back as themes list."""
+    mock_understanding = MagicMock()
+    mock_understanding.suggested_prompts = {
+        "Learn": ["L1", "L2"],
+        "Create": ["C1"],
+    }
+    _mock_get_business_understanding(mocker, return_value=mock_understanding)
+
+    response = client.get("/suggested-prompts")
+
+    assert response.status_code == 200
+    data = response.json()
+    assert "themes" in data
+    themes_by_name = {t["name"]: t["prompts"] for t in data["themes"]}
+    assert themes_by_name["Learn"] == ["L1", "L2"]
+    assert themes_by_name["Create"] == ["C1"]
+
+
+def test_suggested_prompts_no_understanding(
+    mocker: pytest_mock.MockerFixture,
+    test_user_id: str,
+) -> None:
+    """User with no understanding gets empty themes list."""
+    _mock_get_business_understanding(mocker, return_value=None)
+
+    response = client.get("/suggested-prompts")
+
+    assert response.status_code == 200
+    assert response.json() == {"themes": []}
+
+
+def test_suggested_prompts_empty_prompts(
+    mocker: pytest_mock.MockerFixture,
+    test_user_id: str,
+) -> None:
+    """User with understanding but empty prompts gets empty themes list."""
+    mock_understanding = MagicMock()
+    mock_understanding.suggested_prompts = {}
+    _mock_get_business_understanding(mocker, return_value=mock_understanding)
+
+    response = client.get("/suggested-prompts")
+
+    assert response.status_code == 200
+    assert response.json() == {"themes": []}
+
+
+# ─── Create session: dry_run contract ─────────────────────────────────
+
+
+def _mock_create_chat_session(mocker: pytest_mock.MockerFixture):
+    """Mock create_chat_session to return a fake session."""
+    from backend.copilot.model import ChatSession
+
+    async def _fake_create(user_id: str, *, dry_run: bool):
+        return ChatSession.new(user_id, dry_run=dry_run)
+
+    return mocker.patch(
+        "backend.api.features.chat.routes.create_chat_session",
+        new_callable=AsyncMock,
+        side_effect=_fake_create,
+    )
+
+
+def test_create_session_dry_run_true(
+    mocker: pytest_mock.MockerFixture,
+    test_user_id: str,
+) -> None:
+    """Sending ``{"dry_run": true}`` sets metadata.dry_run to True."""
+    _mock_create_chat_session(mocker)
+
+    response = client.post("/sessions", json={"dry_run": True})
+
+    assert response.status_code == 200
+    assert response.json()["metadata"]["dry_run"] is True
+
+
+def test_create_session_dry_run_default_false(
+    mocker: pytest_mock.MockerFixture,
+    test_user_id: str,
+) -> None:
+    """Empty body defaults dry_run to False."""
+    _mock_create_chat_session(mocker)
+
+    response = client.post("/sessions")
+
+    assert response.status_code == 200
+    assert response.json()["metadata"]["dry_run"] is False
+
+
+def test_create_session_rejects_nested_metadata(
+    test_user_id: str,
+) -> None:
+    """Sending ``{"metadata": {"dry_run": true}}`` must return 422, not silently
+    default to ``dry_run=False``. This guards against the common mistake of
+    nesting dry_run inside metadata instead of providing it at the top level."""
+    response = client.post(
+        "/sessions",
+        json={"metadata": {"dry_run": True}},
+    )
+
+    assert response.status_code == 422

autogpt_platform/backend/backend/api/features/integrations/router.py

@@ -40,11 +40,15 @@ from backend.data.onboarding import OnboardingStep, complete_onboarding_step
 from backend.data.user import get_user_integrations
 from backend.executor.utils import add_graph_execution
 from backend.integrations.ayrshare import AyrshareClient, SocialPlatform
-from backend.integrations.credentials_store import provider_matches
+from backend.integrations.credentials_store import (
+    is_system_credential,
+    provider_matches,
+)
 from backend.integrations.creds_manager import (
     IntegrationCredentialsManager,
     create_mcp_oauth_handler,
 )
+from backend.integrations.managed_credentials import ensure_managed_credentials
 from backend.integrations.oauth import CREDENTIALS_BY_PROVIDER, HANDLERS_BY_NAME
 from backend.integrations.providers import ProviderName
 from backend.integrations.webhooks import get_webhook_manager
@@ -110,6 +114,7 @@ class CredentialsMetaResponse(BaseModel):
         default=None,
         description="Host pattern for host-scoped or MCP server URL for MCP credentials",
     )
+    is_managed: bool = False
 
     @model_validator(mode="before")
     @classmethod
@@ -148,6 +153,7 @@ def to_meta_response(cred: Credentials) -> CredentialsMetaResponse:
         scopes=cred.scopes if isinstance(cred, OAuth2Credentials) else None,
         username=cred.username if isinstance(cred, OAuth2Credentials) else None,
         host=CredentialsMetaResponse.get_host(cred),
+        is_managed=cred.is_managed,
     )
 
 
@@ -224,6 +230,9 @@ async def callback(
 async def list_credentials(
     user_id: Annotated[str, Security(get_user_id)],
 ) -> list[CredentialsMetaResponse]:
+    # Fire-and-forget: provision missing managed credentials in the background.
+    # The credential appears on the next page load; listing is never blocked.
+    asyncio.create_task(ensure_managed_credentials(user_id, creds_manager.store))
     credentials = await creds_manager.store.get_all_creds(user_id)
 
     return [
@@ -238,6 +247,7 @@ async def list_credentials_by_provider(
     ],
     user_id: Annotated[str, Security(get_user_id)],
 ) -> list[CredentialsMetaResponse]:
+    asyncio.create_task(ensure_managed_credentials(user_id, creds_manager.store))
     credentials = await creds_manager.store.get_creds_by_provider(user_id, provider)
 
     return [
@@ -332,6 +342,11 @@ async def delete_credentials(
         raise HTTPException(
             status_code=status.HTTP_404_NOT_FOUND, detail="Credentials not found"
         )
+    if is_system_credential(cred_id):
+        raise HTTPException(
+            status_code=status.HTTP_403_FORBIDDEN,
+            detail="System-managed credentials cannot be deleted",
+        )
     creds = await creds_manager.store.get_creds_by_id(user_id, cred_id)
     if not creds:
         raise HTTPException(
@@ -342,6 +357,11 @@ async def delete_credentials(
             status_code=status.HTTP_404_NOT_FOUND,
             detail="Credentials not found",
         )
+    if creds.is_managed:
+        raise HTTPException(
+            status_code=status.HTTP_403_FORBIDDEN,
+            detail="AutoGPT-managed credentials cannot be deleted",
+        )
 
     try:
         await remove_all_webhooks_for_credentials(user_id, creds, force)

autogpt_platform/backend/backend/api/features/integrations/router_test.py

@@ -1,6 +1,7 @@
 """Tests for credentials API security: no secret leakage, SDK defaults filtered."""
 
-from unittest.mock import AsyncMock, patch
+from contextlib import asynccontextmanager
+from unittest.mock import AsyncMock, MagicMock, patch
 
 import fastapi
 import fastapi.testclient
@@ -276,3 +277,294 @@ class TestCreateCredentialNoSecretInResponse:
 
         assert resp.status_code == 403
         mock_mgr.create.assert_not_called()
+
+
+class TestManagedCredentials:
+    """AutoGPT-managed credentials cannot be deleted by users."""
+
+    def test_delete_is_managed_returns_403(self):
+        cred = APIKeyCredentials(
+            id="managed-cred-1",
+            provider="agent_mail",
+            title="AgentMail (managed by AutoGPT)",
+            api_key=SecretStr("sk-managed-key"),
+            is_managed=True,
+        )
+        with patch(
+            "backend.api.features.integrations.router.creds_manager"
+        ) as mock_mgr:
+            mock_mgr.store.get_creds_by_id = AsyncMock(return_value=cred)
+            resp = client.request("DELETE", "/agent_mail/credentials/managed-cred-1")
+
+        assert resp.status_code == 403
+        assert "AutoGPT-managed" in resp.json()["detail"]
+
+    def test_list_credentials_includes_is_managed_field(self):
+        managed = APIKeyCredentials(
+            id="managed-1",
+            provider="agent_mail",
+            title="AgentMail (managed)",
+            api_key=SecretStr("sk-key"),
+            is_managed=True,
+        )
+        regular = APIKeyCredentials(
+            id="regular-1",
+            provider="openai",
+            title="My Key",
+            api_key=SecretStr("sk-key"),
+        )
+        with patch(
+            "backend.api.features.integrations.router.creds_manager"
+        ) as mock_mgr:
+            mock_mgr.store.get_all_creds = AsyncMock(return_value=[managed, regular])
+            resp = client.get("/credentials")
+
+        assert resp.status_code == 200
+        data = resp.json()
+        managed_cred = next(c for c in data if c["id"] == "managed-1")
+        regular_cred = next(c for c in data if c["id"] == "regular-1")
+        assert managed_cred["is_managed"] is True
+        assert regular_cred["is_managed"] is False
+
+
+# ---------------------------------------------------------------------------
+# Managed credential provisioning infrastructure
+# ---------------------------------------------------------------------------
+
+
+def _make_managed_cred(
+    provider: str = "agent_mail", pod_id: str = "pod-abc"
+) -> APIKeyCredentials:
+    return APIKeyCredentials(
+        id="managed-auto",
+        provider=provider,
+        title="AgentMail (managed by AutoGPT)",
+        api_key=SecretStr("sk-pod-key"),
+        is_managed=True,
+        metadata={"pod_id": pod_id},
+    )
+
+
+def _make_store_mock(**kwargs) -> MagicMock:
+    """Create a store mock with a working async ``locks()`` context manager."""
+
+    @asynccontextmanager
+    async def _noop_locked(key):
+        yield
+
+    locks_obj = MagicMock()
+    locks_obj.locked = _noop_locked
+
+    store = MagicMock(**kwargs)
+    store.locks = AsyncMock(return_value=locks_obj)
+    return store
+
+
+class TestEnsureManagedCredentials:
+    """Unit tests for the ensure/cleanup helpers in managed_credentials.py."""
+
+    @pytest.mark.asyncio
+    async def test_provisions_when_missing(self):
+        """Provider.provision() is called when no managed credential exists."""
+        from backend.integrations.managed_credentials import (
+            _PROVIDERS,
+            _provisioned_users,
+            ensure_managed_credentials,
+        )
+
+        cred = _make_managed_cred()
+        provider = MagicMock()
+        provider.provider_name = "test_provider"
+        provider.is_available = AsyncMock(return_value=True)
+        provider.provision = AsyncMock(return_value=cred)
+
+        store = _make_store_mock()
+        store.has_managed_credential = AsyncMock(return_value=False)
+        store.add_managed_credential = AsyncMock()
+
+        saved = dict(_PROVIDERS)
+        _PROVIDERS.clear()
+        _PROVIDERS["test_provider"] = provider
+        _provisioned_users.pop("user-1", None)
+        try:
+            await ensure_managed_credentials("user-1", store)
+        finally:
+            _PROVIDERS.clear()
+            _PROVIDERS.update(saved)
+            _provisioned_users.pop("user-1", None)
+
+        provider.provision.assert_awaited_once_with("user-1")
+        store.add_managed_credential.assert_awaited_once_with("user-1", cred)
+
+    @pytest.mark.asyncio
+    async def test_skips_when_already_exists(self):
+        """Provider.provision() is NOT called when managed credential exists."""
+        from backend.integrations.managed_credentials import (
+            _PROVIDERS,
+            _provisioned_users,
+            ensure_managed_credentials,
+        )
+
+        provider = MagicMock()
+        provider.provider_name = "test_provider"
+        provider.is_available = AsyncMock(return_value=True)
+        provider.provision = AsyncMock()
+
+        store = _make_store_mock()
+        store.has_managed_credential = AsyncMock(return_value=True)
+
+        saved = dict(_PROVIDERS)
+        _PROVIDERS.clear()
+        _PROVIDERS["test_provider"] = provider
+        _provisioned_users.pop("user-1", None)
+        try:
+            await ensure_managed_credentials("user-1", store)
+        finally:
+            _PROVIDERS.clear()
+            _PROVIDERS.update(saved)
+            _provisioned_users.pop("user-1", None)
+
+        provider.provision.assert_not_awaited()
+
+    @pytest.mark.asyncio
+    async def test_skips_when_unavailable(self):
+        """Provider.provision() is NOT called when provider is not available."""
+        from backend.integrations.managed_credentials import (
+            _PROVIDERS,
+            _provisioned_users,
+            ensure_managed_credentials,
+        )
+
+        provider = MagicMock()
+        provider.provider_name = "test_provider"
+        provider.is_available = AsyncMock(return_value=False)
+        provider.provision = AsyncMock()
+
+        store = _make_store_mock()
+        store.has_managed_credential = AsyncMock()
+
+        saved = dict(_PROVIDERS)
+        _PROVIDERS.clear()
+        _PROVIDERS["test_provider"] = provider
+        _provisioned_users.pop("user-1", None)
+        try:
+            await ensure_managed_credentials("user-1", store)
+        finally:
+            _PROVIDERS.clear()
+            _PROVIDERS.update(saved)
+            _provisioned_users.pop("user-1", None)
+
+        provider.provision.assert_not_awaited()
+        store.has_managed_credential.assert_not_awaited()
+
+    @pytest.mark.asyncio
+    async def test_provision_failure_does_not_propagate(self):
+        """A failed provision is logged but does not raise."""
+        from backend.integrations.managed_credentials import (
+            _PROVIDERS,
+            _provisioned_users,
+            ensure_managed_credentials,
+        )
+
+        provider = MagicMock()
+        provider.provider_name = "test_provider"
+        provider.is_available = AsyncMock(return_value=True)
+        provider.provision = AsyncMock(side_effect=RuntimeError("boom"))
+
+        store = _make_store_mock()
+        store.has_managed_credential = AsyncMock(return_value=False)
+
+        saved = dict(_PROVIDERS)
+        _PROVIDERS.clear()
+        _PROVIDERS["test_provider"] = provider
+        _provisioned_users.pop("user-1", None)
+        try:
+            await ensure_managed_credentials("user-1", store)
+        finally:
+            _PROVIDERS.clear()
+            _PROVIDERS.update(saved)
+            _provisioned_users.pop("user-1", None)
+
+        # No exception raised — provisioning failure is swallowed.
+
+
+class TestCleanupManagedCredentials:
+    """Unit tests for cleanup_managed_credentials."""
+
+    @pytest.mark.asyncio
+    async def test_calls_deprovision_for_managed_creds(self):
+        from backend.integrations.managed_credentials import (
+            _PROVIDERS,
+            cleanup_managed_credentials,
+        )
+
+        cred = _make_managed_cred()
+        provider = MagicMock()
+        provider.provider_name = "agent_mail"
+        provider.deprovision = AsyncMock()
+
+        store = MagicMock()
+        store.get_all_creds = AsyncMock(return_value=[cred])
+
+        saved = dict(_PROVIDERS)
+        _PROVIDERS.clear()
+        _PROVIDERS["agent_mail"] = provider
+        try:
+            await cleanup_managed_credentials("user-1", store)
+        finally:
+            _PROVIDERS.clear()
+            _PROVIDERS.update(saved)
+
+        provider.deprovision.assert_awaited_once_with("user-1", cred)
+
+    @pytest.mark.asyncio
+    async def test_skips_non_managed_creds(self):
+        from backend.integrations.managed_credentials import (
+            _PROVIDERS,
+            cleanup_managed_credentials,
+        )
+
+        regular = _make_api_key_cred()
+        provider = MagicMock()
+        provider.provider_name = "openai"
+        provider.deprovision = AsyncMock()
+
+        store = MagicMock()
+        store.get_all_creds = AsyncMock(return_value=[regular])
+
+        saved = dict(_PROVIDERS)
+        _PROVIDERS.clear()
+        _PROVIDERS["openai"] = provider
+        try:
+            await cleanup_managed_credentials("user-1", store)
+        finally:
+            _PROVIDERS.clear()
+            _PROVIDERS.update(saved)
+
+        provider.deprovision.assert_not_awaited()
+
+    @pytest.mark.asyncio
+    async def test_deprovision_failure_does_not_propagate(self):
+        from backend.integrations.managed_credentials import (
+            _PROVIDERS,
+            cleanup_managed_credentials,
+        )
+
+        cred = _make_managed_cred()
+        provider = MagicMock()
+        provider.provider_name = "agent_mail"
+        provider.deprovision = AsyncMock(side_effect=RuntimeError("boom"))
+
+        store = MagicMock()
+        store.get_all_creds = AsyncMock(return_value=[cred])
+
+        saved = dict(_PROVIDERS)
+        _PROVIDERS.clear()
+        _PROVIDERS["agent_mail"] = provider
+        try:
+            await cleanup_managed_credentials("user-1", store)
+        finally:
+            _PROVIDERS.clear()
+            _PROVIDERS.update(saved)
+
+        # No exception raised — cleanup failure is swallowed.

autogpt_platform/backend/backend/api/features/library/_add_to_library.py

@@ -17,8 +17,6 @@ from backend.data.includes import library_agent_include
 from backend.util.exceptions import NotFoundError
 from backend.util.json import SafeJson
 
-from .db import get_library_agent_by_graph_id, update_library_agent
-
 logger = logging.getLogger(__name__)
 
 
@@ -61,28 +59,17 @@ async def add_graph_to_library(
     graph_model: GraphModel,
     user_id: str,
 ) -> library_model.LibraryAgent:
-    """Check existing / restore soft-deleted / create new LibraryAgent."""
-    if existing := await get_library_agent_by_graph_id(
-        user_id, graph_model.id, graph_model.version
-    ):
-        return existing
+    """Check existing / restore soft-deleted / create new LibraryAgent.
 
-    deleted_agent = await prisma.models.LibraryAgent.prisma().find_unique(
-        where={
-            "userId_agentGraphId_agentGraphVersion": {
-                "userId": user_id,
-                "agentGraphId": graph_model.id,
-                "agentGraphVersion": graph_model.version,
-            }
-        },
+    Uses a create-then-catch-UniqueViolationError-then-update pattern on
+    the (userId, agentGraphId, agentGraphVersion) composite unique constraint.
+    This is more robust than ``upsert`` because Prisma's upsert atomicity
+    guarantees are not well-documented for all versions.
+    """
+    settings_json = SafeJson(GraphSettings.from_graph(graph_model).model_dump())
+    _include = library_agent_include(
+        user_id, include_nodes=False, include_executions=False
     )
-    if deleted_agent and (deleted_agent.isDeleted or deleted_agent.isArchived):
-        return await update_library_agent(
-            deleted_agent.id,
-            user_id,
-            is_deleted=False,
-            is_archived=False,
-        )
 
     try:
         added_agent = await prisma.models.LibraryAgent.prisma().create(
@@ -98,23 +85,32 @@ async def add_graph_to_library(
                 },
                 "isCreatedByUser": False,
                 "useGraphIsActiveVersion": False,
-                "settings": SafeJson(
-                    GraphSettings.from_graph(graph_model).model_dump()
-                ),
+                "settings": settings_json,
             },
-            include=library_agent_include(
-                user_id, include_nodes=False, include_executions=False
-            ),
+            include=_include,
         )
     except prisma.errors.UniqueViolationError:
-        # Race condition: concurrent request created the row between our
-        # check and create.  Re-read instead of crashing.
-        existing = await get_library_agent_by_graph_id(
-            user_id, graph_model.id, graph_model.version
+        # Already exists — update to restore if previously soft-deleted/archived
+        added_agent = await prisma.models.LibraryAgent.prisma().update(
+            where={
+                "userId_agentGraphId_agentGraphVersion": {
+                    "userId": user_id,
+                    "agentGraphId": graph_model.id,
+                    "agentGraphVersion": graph_model.version,
+                }
+            },
+            data={
+                "isDeleted": False,
+                "isArchived": False,
+                "settings": settings_json,
+            },
+            include=_include,
         )
-        if existing:
-            return existing
-        raise  # Shouldn't happen, but don't swallow unexpected errors
+        if added_agent is None:
+            raise NotFoundError(
+                f"LibraryAgent for graph #{graph_model.id} "
+                f"v{graph_model.version} not found after UniqueViolationError"
+            )
 
     logger.debug(
         f"Added graph #{graph_model.id} v{graph_model.version} "

autogpt_platform/backend/backend/api/features/library/_add_to_library_test.py

@@ -1,71 +1,80 @@
 from unittest.mock import AsyncMock, MagicMock, patch
 
+import prisma.errors
 import pytest
 
 from ._add_to_library import add_graph_to_library
 
 
 @pytest.mark.asyncio
-async def test_add_graph_to_library_restores_archived_agent() -> None:
-    graph_model = MagicMock(id="graph-id", version=2)
-    archived_agent = MagicMock(id="library-agent-id", isDeleted=False, isArchived=True)
-    restored_agent = MagicMock(name="LibraryAgentModel")
+async def test_add_graph_to_library_create_new_agent() -> None:
+    """When no matching LibraryAgent exists, create inserts a new one."""
+    graph_model = MagicMock(id="graph-id", version=2, nodes=[])
+    created_agent = MagicMock(name="CreatedLibraryAgent")
+    converted_agent = MagicMock(name="ConvertedLibraryAgent")
 
     with (
-        patch(
-            "backend.api.features.library._add_to_library.get_library_agent_by_graph_id",
-            new=AsyncMock(return_value=None),
-        ),
         patch(
             "backend.api.features.library._add_to_library.prisma.models.LibraryAgent.prisma"
         ) as mock_prisma,
         patch(
-            "backend.api.features.library._add_to_library.update_library_agent",
-            new=AsyncMock(return_value=restored_agent),
-        ) as mock_update,
+            "backend.api.features.library._add_to_library.library_model.LibraryAgent.from_db",
+            return_value=converted_agent,
+        ) as mock_from_db,
     ):
-        mock_prisma.return_value.find_unique = AsyncMock(return_value=archived_agent)
+        mock_prisma.return_value.create = AsyncMock(return_value=created_agent)
 
         result = await add_graph_to_library("slv-id", graph_model, "user-id")
 
-    assert result is restored_agent
-    mock_update.assert_awaited_once_with(
-        "library-agent-id",
-        "user-id",
-        is_deleted=False,
-        is_archived=False,
-    )
-    mock_prisma.return_value.create.assert_not_called()
+    assert result is converted_agent
+    mock_from_db.assert_called_once_with(created_agent)
+    # Verify create was called with correct data
+    create_call = mock_prisma.return_value.create.call_args
+    create_data = create_call.kwargs["data"]
+    assert create_data["User"] == {"connect": {"id": "user-id"}}
+    assert create_data["AgentGraph"] == {
+        "connect": {"graphVersionId": {"id": "graph-id", "version": 2}}
+    }
+    assert create_data["isCreatedByUser"] is False
+    assert create_data["useGraphIsActiveVersion"] is False
 
 
 @pytest.mark.asyncio
-async def test_add_graph_to_library_restores_deleted_agent() -> None:
-    graph_model = MagicMock(id="graph-id", version=2)
-    deleted_agent = MagicMock(id="library-agent-id", isDeleted=True, isArchived=False)
-    restored_agent = MagicMock(name="LibraryAgentModel")
+async def test_add_graph_to_library_unique_violation_updates_existing() -> None:
+    """UniqueViolationError on create falls back to update."""
+    graph_model = MagicMock(id="graph-id", version=2, nodes=[])
+    updated_agent = MagicMock(name="UpdatedLibraryAgent")
+    converted_agent = MagicMock(name="ConvertedLibraryAgent")
 
     with (
-        patch(
-            "backend.api.features.library._add_to_library.get_library_agent_by_graph_id",
-            new=AsyncMock(return_value=None),
-        ),
         patch(
             "backend.api.features.library._add_to_library.prisma.models.LibraryAgent.prisma"
         ) as mock_prisma,
         patch(
-            "backend.api.features.library._add_to_library.update_library_agent",
-            new=AsyncMock(return_value=restored_agent),
-        ) as mock_update,
+            "backend.api.features.library._add_to_library.library_model.LibraryAgent.from_db",
+            return_value=converted_agent,
+        ) as mock_from_db,
     ):
-        mock_prisma.return_value.find_unique = AsyncMock(return_value=deleted_agent)
+        mock_prisma.return_value.create = AsyncMock(
+            side_effect=prisma.errors.UniqueViolationError(
+                MagicMock(), message="unique constraint"
+            )
+        )
+        mock_prisma.return_value.update = AsyncMock(return_value=updated_agent)
 
         result = await add_graph_to_library("slv-id", graph_model, "user-id")
 
-    assert result is restored_agent
-    mock_update.assert_awaited_once_with(
-        "library-agent-id",
-        "user-id",
-        is_deleted=False,
-        is_archived=False,
-    )
-    mock_prisma.return_value.create.assert_not_called()
+    assert result is converted_agent
+    mock_from_db.assert_called_once_with(updated_agent)
+    # Verify update was called with correct where and data
+    update_call = mock_prisma.return_value.update.call_args
+    assert update_call.kwargs["where"] == {
+        "userId_agentGraphId_agentGraphVersion": {
+            "userId": "user-id",
+            "agentGraphId": "graph-id",
+            "agentGraphVersion": 2,
+        }
+    }
+    update_data = update_call.kwargs["data"]
+    assert update_data["isDeleted"] is False
+    assert update_data["isArchived"] is False

autogpt_platform/backend/backend/api/features/library/db.py

@@ -436,32 +436,58 @@ async def create_library_agent(
     async with transaction() as tx:
         library_agents = await asyncio.gather(
             *(
-                prisma.models.LibraryAgent.prisma(tx).create(
-                    data=prisma.types.LibraryAgentCreateInput(
-                        isCreatedByUser=(user_id == user_id),
-                        useGraphIsActiveVersion=True,
-                        User={"connect": {"id": user_id}},
-                        AgentGraph={
-                            "connect": {
-                                "graphVersionId": {
-                                    "id": graph_entry.id,
-                                    "version": graph_entry.version,
+                prisma.models.LibraryAgent.prisma(tx).upsert(
+                    where={
+                        "userId_agentGraphId_agentGraphVersion": {
+                            "userId": user_id,
+                            "agentGraphId": graph_entry.id,
+                            "agentGraphVersion": graph_entry.version,
+                        }
+                    },
+                    data={
+                        "create": prisma.types.LibraryAgentCreateInput(
+                            isCreatedByUser=(user_id == graph.user_id),
+                            useGraphIsActiveVersion=True,
+                            User={"connect": {"id": user_id}},
+                            AgentGraph={
+                                "connect": {
+                                    "graphVersionId": {
+                                        "id": graph_entry.id,
+                                        "version": graph_entry.version,
+                                    }
                                 }
-                            }
+                            },
+                            settings=SafeJson(
+                                GraphSettings.from_graph(
+                                    graph_entry,
+                                    hitl_safe_mode=hitl_safe_mode,
+                                    sensitive_action_safe_mode=sensitive_action_safe_mode,
+                                ).model_dump()
+                            ),
+                            **(
+                                {"Folder": {"connect": {"id": folder_id}}}
+                                if folder_id and graph_entry is graph
+                                else {}
+                            ),
+                        ),
+                        "update": {
+                            "isDeleted": False,
+                            "isArchived": False,
+                            "useGraphIsActiveVersion": True,
+                            "settings": SafeJson(
+                                GraphSettings.from_graph(
+                                    graph_entry,
+                                    hitl_safe_mode=hitl_safe_mode,
+                                    sensitive_action_safe_mode=sensitive_action_safe_mode,
+                                ).model_dump()
+                            ),
+                            **(
+                                {"Folder": {"connect": {"id": folder_id}}}
+                                if folder_id and graph_entry is graph
+                                else {}
+                            ),
                         },
-                        settings=SafeJson(
-                            GraphSettings.from_graph(
-                                graph_entry,
-                                hitl_safe_mode=hitl_safe_mode,
-                                sensitive_action_safe_mode=sensitive_action_safe_mode,
-                            ).model_dump()
-                        ),
-                        **(
-                            {"Folder": {"connect": {"id": folder_id}}}
-                            if folder_id and graph_entry is graph
-                            else {}
-                        ),
-                    ),
+                    },
                     include=library_agent_include(
                         user_id, include_nodes=False, include_executions=False
                     ),

autogpt_platform/backend/backend/api/features/library/db_test.py

@@ -1,4 +1,6 @@
+from contextlib import asynccontextmanager
 from datetime import datetime
+from unittest.mock import AsyncMock, MagicMock, patch
 
 import prisma.enums
 import prisma.models
@@ -85,10 +87,6 @@ async def test_get_library_agents(mocker):
 async def test_add_agent_to_library(mocker):
     await connect()
 
-    # Mock the transaction context
-    mock_transaction = mocker.patch("backend.api.features.library.db.transaction")
-    mock_transaction.return_value.__aenter__ = mocker.AsyncMock(return_value=None)
-    mock_transaction.return_value.__aexit__ = mocker.AsyncMock(return_value=None)
     # Mock data
     mock_store_listing_data = prisma.models.StoreListingVersion(
         id="version123",
@@ -143,13 +141,11 @@ async def test_add_agent_to_library(mocker):
     )
 
     mock_library_agent = mocker.patch("prisma.models.LibraryAgent.prisma")
-    mock_library_agent.return_value.find_first = mocker.AsyncMock(return_value=None)
-    mock_library_agent.return_value.find_unique = mocker.AsyncMock(return_value=None)
     mock_library_agent.return_value.create = mocker.AsyncMock(
         return_value=mock_library_agent_data
     )
 
-    # Mock graph_db.get_graph function that's called to check for HITL blocks
+    # Mock graph_db.get_graph function that's called in resolve_graph_for_library
     # (lives in _add_to_library.py after refactor, not db.py)
     mock_graph_db = mocker.patch(
         "backend.api.features.library._add_to_library.graph_db"
@@ -175,37 +171,27 @@ async def test_add_agent_to_library(mocker):
     mock_store_listing_version.return_value.find_unique.assert_called_once_with(
         where={"id": "version123"}, include={"AgentGraph": True}
     )
-    mock_library_agent.return_value.find_unique.assert_called_once_with(
-        where={
-            "userId_agentGraphId_agentGraphVersion": {
-                "userId": "test-user",
-                "agentGraphId": "agent1",
-                "agentGraphVersion": 1,
-            }
-        },
-    )
     # Check that create was called with the expected data including settings
     create_call_args = mock_library_agent.return_value.create.call_args
     assert create_call_args is not None
 
-    # Verify the main structure
-    expected_data = {
+    # Verify the create data structure
+    create_data = create_call_args.kwargs["data"]
+    expected_create = {
         "User": {"connect": {"id": "test-user"}},
         "AgentGraph": {"connect": {"graphVersionId": {"id": "agent1", "version": 1}}},
         "isCreatedByUser": False,
+        "useGraphIsActiveVersion": False,
     }
-
-    actual_data = create_call_args[1]["data"]
-    # Check that all expected fields are present
-    for key, value in expected_data.items():
-        assert actual_data[key] == value
+    for key, value in expected_create.items():
+        assert create_data[key] == value
 
     # Check that settings field is present and is a SafeJson object
-    assert "settings" in actual_data
-    assert hasattr(actual_data["settings"], "__class__")  # Should be a SafeJson object
+    assert "settings" in create_data
+    assert hasattr(create_data["settings"], "__class__")  # Should be a SafeJson object
 
     # Check include parameter
-    assert create_call_args[1]["include"] == library_agent_include(
+    assert create_call_args.kwargs["include"] == library_agent_include(
         "test-user", include_nodes=False, include_executions=False
     )
 
@@ -320,3 +306,50 @@ async def test_update_graph_in_library_allows_archived_library_agent(mocker):
         include_archived=True,
     )
     mock_update_library_agent.assert_awaited_once_with("test-user", created_graph)
+
+
+@pytest.mark.asyncio
+async def test_create_library_agent_uses_upsert():
+    """create_library_agent should use upsert (not create) to handle duplicates."""
+    mock_graph = MagicMock()
+    mock_graph.id = "graph-1"
+    mock_graph.version = 1
+    mock_graph.user_id = "user-1"
+    mock_graph.nodes = []
+    mock_graph.sub_graphs = []
+
+    mock_upserted = MagicMock(name="UpsertedLibraryAgent")
+
+    @asynccontextmanager
+    async def fake_tx():
+        yield None
+
+    with (
+        patch("backend.api.features.library.db.transaction", fake_tx),
+        patch("prisma.models.LibraryAgent.prisma") as mock_prisma,
+        patch(
+            "backend.api.features.library.db.add_generated_agent_image",
+            new=AsyncMock(),
+        ),
+        patch(
+            "backend.api.features.library.model.LibraryAgent.from_db",
+            return_value=MagicMock(),
+        ),
+    ):
+        mock_prisma.return_value.upsert = AsyncMock(return_value=mock_upserted)
+
+        result = await db.create_library_agent(mock_graph, "user-1")
+
+    assert len(result) == 1
+    upsert_call = mock_prisma.return_value.upsert.call_args
+    assert upsert_call is not None
+    # Verify the upsert where clause uses the composite unique key
+    where = upsert_call.kwargs["where"]
+    assert "userId_agentGraphId_agentGraphVersion" in where
+    # Verify the upsert data has both create and update branches
+    data = upsert_call.kwargs["data"]
+    assert "create" in data
+    assert "update" in data
+    # Verify update branch restores soft-deleted/archived agents
+    assert data["update"]["isDeleted"] is False
+    assert data["update"]["isArchived"] is False

autogpt_platform/backend/backend/api/features/oauth_test.py

@@ -12,6 +12,7 @@ Tests cover:
 5. Complete OAuth flow end-to-end
 """
 
+import asyncio
 import base64
 import hashlib
 import secrets
@@ -58,14 +59,27 @@ async def test_user(server, test_user_id: str):
 
     yield test_user_id
 
-    # Cleanup - delete in correct order due to foreign key constraints
-    await PrismaOAuthAccessToken.prisma().delete_many(where={"userId": test_user_id})
-    await PrismaOAuthRefreshToken.prisma().delete_many(where={"userId": test_user_id})
-    await PrismaOAuthAuthorizationCode.prisma().delete_many(
-        where={"userId": test_user_id}
-    )
-    await PrismaOAuthApplication.prisma().delete_many(where={"ownerId": test_user_id})
-    await PrismaUser.prisma().delete(where={"id": test_user_id})
+    # Cleanup - delete in correct order due to foreign key constraints.
+    # Wrap in try/except because the event loop or Prisma engine may already
+    # be closed during session teardown on Python 3.12+.
+    try:
+        await asyncio.gather(
+            PrismaOAuthAccessToken.prisma().delete_many(where={"userId": test_user_id}),
+            PrismaOAuthRefreshToken.prisma().delete_many(
+                where={"userId": test_user_id}
+            ),
+            PrismaOAuthAuthorizationCode.prisma().delete_many(
+                where={"userId": test_user_id}
+            ),
+        )
+        await asyncio.gather(
+            PrismaOAuthApplication.prisma().delete_many(
+                where={"ownerId": test_user_id}
+            ),
+            PrismaUser.prisma().delete(where={"id": test_user_id}),
+        )
+    except RuntimeError:
+        pass
 
 
 @pytest_asyncio.fixture

autogpt_platform/backend/backend/api/rest_api.py

@@ -18,6 +18,7 @@ from prisma.errors import PrismaError
 
 import backend.api.features.admin.credit_admin_routes
 import backend.api.features.admin.execution_analytics_routes
+import backend.api.features.admin.rate_limit_admin_routes
 import backend.api.features.admin.store_admin_routes
 import backend.api.features.builder
 import backend.api.features.builder.routes
@@ -117,6 +118,11 @@ async def lifespan_context(app: fastapi.FastAPI):
 
     AutoRegistry.patch_integrations()
 
+    # Register managed credential providers (e.g. AgentMail)
+    from backend.integrations.managed_providers import register_all
+
+    register_all()
+
     await backend.data.block.initialize_blocks()
 
     await backend.data.user.migrate_and_encrypt_user_integrations()
@@ -318,6 +324,11 @@ app.include_router(
     tags=["v2", "admin"],
     prefix="/api/executions",
 )
+app.include_router(
+    backend.api.features.admin.rate_limit_admin_routes.router,
+    tags=["v2", "admin"],
+    prefix="/api/copilot",
+)
 app.include_router(
     backend.api.features.executions.review.routes.router,
     tags=["v2", "executions", "review"],

autogpt_platform/backend/backend/blocks/_base.py

@@ -698,13 +698,30 @@ class Block(ABC, Generic[BlockSchemaInputType, BlockSchemaOutputType]):
             if should_pause:
                 return
 
-        # Validate the input data (original or reviewer-modified) once
-        if error := self.input_schema.validate_data(input_data):
-            raise BlockInputError(
-                message=f"Unable to execute block with invalid input data: {error}",
-                block_name=self.name,
-                block_id=self.id,
-            )
+        # Validate the input data (original or reviewer-modified) once.
+        # In dry-run mode, credential fields may contain sentinel None values
+        # that would fail JSON schema required checks.  We still validate the
+        # non-credential fields so blocks that execute for real during dry-run
+        # (e.g. AgentExecutorBlock) get proper input validation.
+        is_dry_run = getattr(kwargs.get("execution_context"), "dry_run", False)
+        if is_dry_run:
+            cred_field_names = set(self.input_schema.get_credentials_fields().keys())
+            non_cred_data = {
+                k: v for k, v in input_data.items() if k not in cred_field_names
+            }
+            if error := self.input_schema.validate_data(non_cred_data):
+                raise BlockInputError(
+                    message=f"Unable to execute block with invalid input data: {error}",
+                    block_name=self.name,
+                    block_id=self.id,
+                )
+        else:
+            if error := self.input_schema.validate_data(input_data):
+                raise BlockInputError(
+                    message=f"Unable to execute block with invalid input data: {error}",
+                    block_name=self.name,
+                    block_id=self.id,
+                )
 
         # Use the validated input data
         async for output_name, output_data in self.run(

autogpt_platform/backend/backend/blocks/agent.py

@@ -49,11 +49,17 @@ class AgentExecutorBlock(Block):
         @classmethod
         def get_missing_input(cls, data: BlockInput) -> set[str]:
             required_fields = cls.get_input_schema(data).get("required", [])
-            return set(required_fields) - set(data)
+            # Check against the nested `inputs` dict, not the top-level node
+            # data — required fields like "topic" live inside data["inputs"],
+            # not at data["topic"].
+            provided = data.get("inputs", {})
+            return set(required_fields) - set(provided)
 
         @classmethod
         def get_mismatch_error(cls, data: BlockInput) -> str | None:
-            return validate_with_jsonschema(cls.get_input_schema(data), data)
+            return validate_with_jsonschema(
+                cls.get_input_schema(data), data.get("inputs", {})
+            )
 
     class Output(BlockSchema):
         # Use BlockSchema to avoid automatic error field that could clash with graph outputs
@@ -88,6 +94,7 @@ class AgentExecutorBlock(Block):
             execution_context=execution_context.model_copy(
                 update={"parent_execution_id": graph_exec_id},
             ),
+            dry_run=execution_context.dry_run,
         )
 
         logger = execution_utils.LogMetadata(
@@ -149,14 +156,19 @@ class AgentExecutorBlock(Block):
                 ExecutionStatus.TERMINATED,
                 ExecutionStatus.FAILED,
             ]:
-                logger.debug(
-                    f"Execution {log_id} received event {event.event_type} with status {event.status}"
+                logger.info(
+                    f"Execution {log_id} skipping event {event.event_type} status={event.status} "
+                    f"node={getattr(event, 'node_exec_id', '?')}"
                 )
                 continue
 
             if event.event_type == ExecutionEventType.GRAPH_EXEC_UPDATE:
                 # If the graph execution is COMPLETED, TERMINATED, or FAILED,
                 # we can stop listening for further events.
+                logger.info(
+                    f"Execution {log_id} graph completed with status {event.status}, "
+                    f"yielded {len(yielded_node_exec_ids)} outputs"
+                )
                 self.merge_stats(
                     NodeExecutionStats(
                         extra_cost=event.stats.cost if event.stats else 0,

autogpt_platform/backend/backend/blocks/ai_condition.py

@@ -1,3 +1,4 @@
+import re
 from typing import Any
 
 from backend.blocks._base import (
@@ -19,6 +20,33 @@ from backend.blocks.llm import (
 )
 from backend.data.model import APIKeyCredentials, NodeExecutionStats, SchemaField
 
+# Minimum max_output_tokens accepted by OpenAI-compatible APIs.
+# A true/false answer fits comfortably within this budget.
+MIN_LLM_OUTPUT_TOKENS = 16
+
+
+def _parse_boolean_response(response_text: str) -> tuple[bool, str | None]:
+    """Parse an LLM response into a boolean result.
+
+    Returns a ``(result, error)`` tuple.  *error* is ``None`` when the
+    response is unambiguous; otherwise it contains a diagnostic message
+    and *result* defaults to ``False``.
+    """
+    text = response_text.strip().lower()
+    if text == "true":
+        return True, None
+    if text == "false":
+        return False, None
+
+    # Fuzzy match – use word boundaries to avoid false positives like "untrue".
+    tokens = set(re.findall(r"\b(true|false|yes|no|1|0)\b", text))
+    if tokens == {"true"} or tokens == {"yes"} or tokens == {"1"}:
+        return True, None
+    if tokens == {"false"} or tokens == {"no"} or tokens == {"0"}:
+        return False, None
+
+    return False, f"Unclear AI response: '{response_text}'"
+
 
 class AIConditionBlock(AIBlockBase):
     """
@@ -162,54 +190,26 @@ class AIConditionBlock(AIBlockBase):
         ]
 
         # Call the LLM
-        try:
-            response = await self.llm_call(
-                credentials=credentials,
-                llm_model=input_data.model,
-                prompt=prompt,
-                max_tokens=10,  # We only expect a true/false response
+        response = await self.llm_call(
+            credentials=credentials,
+            llm_model=input_data.model,
+            prompt=prompt,
+            max_tokens=MIN_LLM_OUTPUT_TOKENS,
+        )
+
+        # Extract the boolean result from the response
+        result, error = _parse_boolean_response(response.response)
+        if error:
+            yield "error", error
+
+        # Update internal stats
+        self.merge_stats(
+            NodeExecutionStats(
+                input_token_count=response.prompt_tokens,
+                output_token_count=response.completion_tokens,
             )
-
-            # Extract the boolean result from the response
-            response_text = response.response.strip().lower()
-            if response_text == "true":
-                result = True
-            elif response_text == "false":
-                result = False
-            else:
-                # If the response is not clear, try to interpret it using word boundaries
-                import re
-
-                # Use word boundaries to avoid false positives like 'untrue' or '10'
-                tokens = set(re.findall(r"\b(true|false|yes|no|1|0)\b", response_text))
-
-                if tokens == {"true"} or tokens == {"yes"} or tokens == {"1"}:
-                    result = True
-                elif tokens == {"false"} or tokens == {"no"} or tokens == {"0"}:
-                    result = False
-                else:
-                    # Unclear or conflicting response - default to False and yield error
-                    result = False
-                    yield "error", f"Unclear AI response: '{response.response}'"
-
-            # Update internal stats
-            self.merge_stats(
-                NodeExecutionStats(
-                    input_token_count=response.prompt_tokens,
-                    output_token_count=response.completion_tokens,
-                )
-            )
-            self.prompt = response.prompt
-
-        except Exception as e:
-            # In case of any error, default to False to be safe
-            result = False
-            # Log the error but don't fail the block execution
-            import logging
-
-            logger = logging.getLogger(__name__)
-            logger.error(f"AI condition evaluation failed: {str(e)}")
-            yield "error", f"AI evaluation failed: {str(e)}"
+        )
+        self.prompt = response.prompt
 
         # Yield results
         yield "result", result

autogpt_platform/backend/backend/blocks/ai_condition_test.py

@@ -0,0 +1,147 @@
+"""Tests for AIConditionBlock – regression coverage for max_tokens and error propagation."""
+
+from __future__ import annotations
+
+from typing import cast
+
+import pytest
+
+from backend.blocks.ai_condition import (
+    MIN_LLM_OUTPUT_TOKENS,
+    AIConditionBlock,
+    _parse_boolean_response,
+)
+from backend.blocks.llm import (
+    DEFAULT_LLM_MODEL,
+    TEST_CREDENTIALS,
+    TEST_CREDENTIALS_INPUT,
+    AICredentials,
+    LLMResponse,
+)
+
+_TEST_AI_CREDENTIALS = cast(AICredentials, TEST_CREDENTIALS_INPUT)
+
+
+# ---------------------------------------------------------------------------
+# Helper to collect all yields from the async generator
+# ---------------------------------------------------------------------------
+
+
+async def _collect_outputs(block: AIConditionBlock, input_data, credentials):
+    outputs: dict[str, object] = {}
+    async for name, value in block.run(input_data, credentials=credentials):
+        outputs[name] = value
+    return outputs
+
+
+def _make_input(**overrides) -> AIConditionBlock.Input:
+    defaults: dict = {
+        "input_value": "hello@example.com",
+        "condition": "the input is an email address",
+        "yes_value": "yes!",
+        "no_value": "no!",
+        "model": DEFAULT_LLM_MODEL,
+        "credentials": TEST_CREDENTIALS_INPUT,
+    }
+    defaults.update(overrides)
+    return AIConditionBlock.Input(**defaults)
+
+
+def _mock_llm_response(response_text: str) -> LLMResponse:
+    return LLMResponse(
+        raw_response="",
+        prompt=[],
+        response=response_text,
+        tool_calls=None,
+        prompt_tokens=10,
+        completion_tokens=5,
+        reasoning=None,
+    )
+
+
+# ---------------------------------------------------------------------------
+# _parse_boolean_response unit tests
+# ---------------------------------------------------------------------------
+
+
+class TestParseBooleanResponse:
+    def test_true_exact(self):
+        assert _parse_boolean_response("true") == (True, None)
+
+    def test_false_exact(self):
+        assert _parse_boolean_response("false") == (False, None)
+
+    def test_true_with_whitespace(self):
+        assert _parse_boolean_response("  True  ") == (True, None)
+
+    def test_yes_fuzzy(self):
+        assert _parse_boolean_response("Yes") == (True, None)
+
+    def test_no_fuzzy(self):
+        assert _parse_boolean_response("no") == (False, None)
+
+    def test_one_fuzzy(self):
+        assert _parse_boolean_response("1") == (True, None)
+
+    def test_zero_fuzzy(self):
+        assert _parse_boolean_response("0") == (False, None)
+
+    def test_unclear_response(self):
+        result, error = _parse_boolean_response("I'm not sure")
+        assert result is False
+        assert error is not None
+        assert "Unclear" in error
+
+    def test_conflicting_tokens(self):
+        result, error = _parse_boolean_response("true and false")
+        assert result is False
+        assert error is not None
+
+
+# ---------------------------------------------------------------------------
+# Regression: max_tokens is set to MIN_LLM_OUTPUT_TOKENS
+# ---------------------------------------------------------------------------
+
+
+class TestMaxTokensRegression:
+    @pytest.mark.asyncio
+    async def test_llm_call_receives_min_output_tokens(self):
+        """max_tokens must be MIN_LLM_OUTPUT_TOKENS (16) – the previous value
+        of 1 was too low and caused OpenAI to reject the request."""
+        block = AIConditionBlock()
+        captured_kwargs: dict = {}
+
+        async def spy_llm_call(**kwargs):
+            captured_kwargs.update(kwargs)
+            return _mock_llm_response("true")
+
+        block.llm_call = spy_llm_call  # type: ignore[assignment]
+
+        input_data = _make_input()
+        await _collect_outputs(block, input_data, credentials=TEST_CREDENTIALS)
+
+        assert captured_kwargs["max_tokens"] == MIN_LLM_OUTPUT_TOKENS
+        assert captured_kwargs["max_tokens"] == 16
+
+
+# ---------------------------------------------------------------------------
+# Regression: exceptions from llm_call must propagate
+# ---------------------------------------------------------------------------
+
+
+class TestExceptionPropagation:
+    @pytest.mark.asyncio
+    async def test_llm_call_exception_propagates(self):
+        """If llm_call raises, the exception must NOT be swallowed.
+        Previously the block caught all exceptions and silently returned
+        result=False."""
+        block = AIConditionBlock()
+
+        async def boom(**kwargs):
+            raise RuntimeError("LLM provider error")
+
+        block.llm_call = boom  # type: ignore[assignment]
+
+        input_data = _make_input()
+        with pytest.raises(RuntimeError, match="LLM provider error"):
+            await _collect_outputs(block, input_data, credentials=TEST_CREDENTIALS)

autogpt_platform/backend/backend/blocks/autopilot.py

@@ -146,6 +146,21 @@ class AutoPilotBlock(Block):
             advanced=True,
         )
 
+        dry_run: bool = SchemaField(
+            description=(
+                "When enabled, run_block and run_agent tool calls in this "
+                "autopilot session are forced to use dry-run simulation mode. "
+                "No real API calls, side effects, or credits are consumed "
+                "by those tools. Useful for testing agent wiring and "
+                "previewing outputs. "
+                "Only applies when creating a new session (session_id is empty). "
+                "When reusing an existing session_id, the session's original "
+                "dry_run setting is preserved."
+            ),
+            default=False,
+            advanced=True,
+        )
+
         # timeout_seconds removed: the SDK manages its own heartbeat-based
         # timeouts internally; wrapping with asyncio.timeout corrupts the
         # SDK's internal stream (see service.py CRITICAL comment).
@@ -232,11 +247,11 @@ class AutoPilotBlock(Block):
             },
         )
 
-    async def create_session(self, user_id: str) -> str:
+    async def create_session(self, user_id: str, *, dry_run: bool) -> str:
         """Create a new chat session and return its ID (mockable for tests)."""
         from backend.copilot.model import create_chat_session  # avoid circular import
 
-        session = await create_chat_session(user_id)
+        session = await create_chat_session(user_id, dry_run=dry_run)
         return session.session_id
 
     async def execute_copilot(
@@ -367,7 +382,9 @@ class AutoPilotBlock(Block):
         # even if the downstream stream fails (avoids orphaned sessions).
         sid = input_data.session_id
         if not sid:
-            sid = await self.create_session(execution_context.user_id)
+            sid = await self.create_session(
+                execution_context.user_id, dry_run=input_data.dry_run
+            )
 
         # NOTE: No asyncio.timeout() here — the SDK manages its own
         # heartbeat-based timeouts internally.  Wrapping with asyncio.timeout

autogpt_platform/backend/backend/blocks/discord/bot_blocks.py

@@ -73,7 +73,7 @@ class ReadDiscordMessagesBlock(Block):
             id="df06086a-d5ac-4abb-9996-2ad0acb2eff7",
             input_schema=ReadDiscordMessagesBlock.Input,  # Assign input schema
             output_schema=ReadDiscordMessagesBlock.Output,  # Assign output schema
-            description="Reads messages from a Discord channel using a bot token.",
+            description="Reads new messages from a Discord channel using a bot token and triggers when a new message is posted",
             categories={BlockCategory.SOCIAL},
             test_input={
                 "continuous_read": False,

autogpt_platform/backend/backend/blocks/google/gmail.py

@@ -1,5 +1,6 @@
 import asyncio
 import base64
+import re
 from abc import ABC
 from email import encoders
 from email.mime.base import MIMEBase
@@ -8,7 +9,7 @@ from email.mime.text import MIMEText
 from email.policy import SMTP
 from email.utils import getaddresses, parseaddr
 from pathlib import Path
-from typing import List, Literal, Optional
+from typing import List, Literal, Optional, Protocol, runtime_checkable
 
 from google.oauth2.credentials import Credentials
 from googleapiclient.discovery import build
@@ -42,8 +43,52 @@ NO_WRAP_POLICY = SMTP.clone(max_line_length=0)
 
 
 def serialize_email_recipients(recipients: list[str]) -> str:
-    """Serialize recipients list to comma-separated string."""
-    return ", ".join(recipients)
+    """Serialize recipients list to comma-separated string.
+
+    Strips leading/trailing whitespace from each address to keep MIME
+    headers clean (mirrors the strip done in ``validate_email_recipients``).
+    """
+    return ", ".join(addr.strip() for addr in recipients)
+
+
+# RFC 5322 simplified pattern: local@domain where domain has at least one dot
+_EMAIL_RE = re.compile(r"^[^@\s]+@[^@\s]+\.[^@\s]+$")
+
+
+def validate_email_recipients(recipients: list[str], field_name: str = "to") -> None:
+    """Validate that all recipients are plausible email addresses.
+
+    Raises ``ValueError`` with a user-friendly message listing every
+    invalid entry so the caller (or LLM) can correct them in one pass.
+    """
+    invalid = [addr for addr in recipients if not _EMAIL_RE.match(addr.strip())]
+    if invalid:
+        formatted = ", ".join(f"'{a}'" for a in invalid)
+        raise ValueError(
+            f"Invalid email address(es) in '{field_name}': {formatted}. "
+            f"Each entry must be a valid email address (e.g. user@example.com)."
+        )
+
+
+@runtime_checkable
+class HasRecipients(Protocol):
+    to: list[str]
+    cc: list[str]
+    bcc: list[str]
+
+
+def validate_all_recipients(input_data: HasRecipients) -> None:
+    """Validate to/cc/bcc recipient fields on an input namespace.
+
+    Calls ``validate_email_recipients`` for ``to`` (required) and
+    ``cc``/``bcc`` (when non-empty), raising ``ValueError`` on the
+    first field that contains an invalid address.
+    """
+    validate_email_recipients(input_data.to, "to")
+    if input_data.cc:
+        validate_email_recipients(input_data.cc, "cc")
+    if input_data.bcc:
+        validate_email_recipients(input_data.bcc, "bcc")
 
 
 def _make_mime_text(
@@ -100,14 +145,16 @@ async def create_mime_message(
 ) -> str:
     """Create a MIME message with attachments and return base64-encoded raw message."""
 
+    validate_all_recipients(input_data)
+
     message = MIMEMultipart()
     message["to"] = serialize_email_recipients(input_data.to)
     message["subject"] = input_data.subject
 
     if input_data.cc:
-        message["cc"] = ", ".join(input_data.cc)
+        message["cc"] = serialize_email_recipients(input_data.cc)
     if input_data.bcc:
-        message["bcc"] = ", ".join(input_data.bcc)
+        message["bcc"] = serialize_email_recipients(input_data.bcc)
 
     # Use the new helper function with content_type if available
     content_type = getattr(input_data, "content_type", None)
@@ -1167,13 +1214,15 @@ async def _build_reply_message(
         references.append(headers["message-id"])
 
     # Create MIME message
+    validate_all_recipients(input_data)
+
     msg = MIMEMultipart()
     if input_data.to:
-        msg["To"] = ", ".join(input_data.to)
+        msg["To"] = serialize_email_recipients(input_data.to)
     if input_data.cc:
-        msg["Cc"] = ", ".join(input_data.cc)
+        msg["Cc"] = serialize_email_recipients(input_data.cc)
     if input_data.bcc:
-        msg["Bcc"] = ", ".join(input_data.bcc)
+        msg["Bcc"] = serialize_email_recipients(input_data.bcc)
     msg["Subject"] = subject
     if headers.get("message-id"):
         msg["In-Reply-To"] = headers["message-id"]
@@ -1685,13 +1734,16 @@ To: {original_to}
         else:
             body = f"{forward_header}\n\n{original_body}"
 
+        # Validate all recipient lists before building the MIME message
+        validate_all_recipients(input_data)
+
         # Create MIME message
         msg = MIMEMultipart()
-        msg["To"] = ", ".join(input_data.to)
+        msg["To"] = serialize_email_recipients(input_data.to)
         if input_data.cc:
-            msg["Cc"] = ", ".join(input_data.cc)
+            msg["Cc"] = serialize_email_recipients(input_data.cc)
         if input_data.bcc:
-            msg["Bcc"] = ", ".join(input_data.bcc)
+            msg["Bcc"] = serialize_email_recipients(input_data.bcc)
         msg["Subject"] = subject
 
         # Add body with proper content type

autogpt_platform/backend/backend/blocks/io.py

@@ -2,6 +2,8 @@ import copy
 from datetime import date, time
 from typing import Any, Optional
 
+from pydantic import AliasChoices, Field
+
 from backend.blocks._base import (
     Block,
     BlockCategory,
@@ -28,9 +30,9 @@ class AgentInputBlock(Block):
     """
     This block is used to provide input to the graph.
 
-    It takes in a value, name, description, default values list and bool to limit selection to default values.
+    It takes in a value, name, and description.
 
-    It Outputs the value passed as input.
+    It outputs the value passed as input.
     """
 
     class Input(BlockSchemaInput):
@@ -47,12 +49,6 @@ class AgentInputBlock(Block):
             default=None,
             advanced=True,
         )
-        placeholder_values: list = SchemaField(
-            description="The placeholder values to be passed as input.",
-            default_factory=list,
-            advanced=True,
-            hidden=True,
-        )
         advanced: bool = SchemaField(
             description="Whether to show the input in the advanced section, if the field is not required.",
             default=False,
@@ -65,10 +61,7 @@ class AgentInputBlock(Block):
         )
 
         def generate_schema(self):
-            schema = copy.deepcopy(self.get_field_schema("value"))
-            if possible_values := self.placeholder_values:
-                schema["enum"] = possible_values
-            return schema
+            return copy.deepcopy(self.get_field_schema("value"))
 
     class Output(BlockSchema):
         # Use BlockSchema to avoid automatic error field for interface definition
@@ -86,18 +79,16 @@ class AgentInputBlock(Block):
                         "value": "Hello, World!",
                         "name": "input_1",
                         "description": "Example test input.",
-                        "placeholder_values": [],
                     },
                     {
-                        "value": "Hello, World!",
+                        "value": 42,
                         "name": "input_2",
-                        "description": "Example test input with placeholders.",
-                        "placeholder_values": ["Hello, World!"],
+                        "description": "Example numeric input.",
                     },
                 ],
                 "test_output": [
                     ("result", "Hello, World!"),
-                    ("result", "Hello, World!"),
+                    ("result", 42),
                 ],
                 "categories": {BlockCategory.INPUT, BlockCategory.BASIC},
                 "block_type": BlockType.INPUT,
@@ -245,13 +236,11 @@ class AgentShortTextInputBlock(AgentInputBlock):
                     "value": "Hello",
                     "name": "short_text_1",
                     "description": "Short text example 1",
-                    "placeholder_values": [],
                 },
                 {
                     "value": "Quick test",
                     "name": "short_text_2",
                     "description": "Short text example 2",
-                    "placeholder_values": ["Quick test", "Another option"],
                 },
             ],
             test_output=[
@@ -285,13 +274,11 @@ class AgentLongTextInputBlock(AgentInputBlock):
                     "value": "Lorem ipsum dolor sit amet...",
                     "name": "long_text_1",
                     "description": "Long text example 1",
-                    "placeholder_values": [],
                 },
                 {
                     "value": "Another multiline text input.",
                     "name": "long_text_2",
                     "description": "Long text example 2",
-                    "placeholder_values": ["Another multiline text input."],
                 },
             ],
             test_output=[
@@ -325,13 +312,11 @@ class AgentNumberInputBlock(AgentInputBlock):
                     "value": 42,
                     "name": "number_input_1",
                     "description": "Number example 1",
-                    "placeholder_values": [],
                 },
                 {
                     "value": 314,
                     "name": "number_input_2",
                     "description": "Number example 2",
-                    "placeholder_values": [314, 2718],
                 },
             ],
             test_output=[
@@ -484,7 +469,8 @@ class AgentFileInputBlock(AgentInputBlock):
 
 class AgentDropdownInputBlock(AgentInputBlock):
     """
-    A specialized text input block that relies on placeholder_values to present a dropdown.
+    A specialized text input block that presents a dropdown selector
+    restricted to a fixed set of values.
     """
 
     class Input(AgentInputBlock.Input):
@@ -494,13 +480,26 @@ class AgentDropdownInputBlock(AgentInputBlock):
             advanced=False,
             title="Default Value",
         )
-        placeholder_values: list = SchemaField(
-            description="Possible values for the dropdown.",
+        # Use Field() directly (not SchemaField) to pass validation_alias,
+        # which handles backward compat for legacy "placeholder_values" across
+        # all construction paths (model_construct, __init__, model_validate).
+        options: list = Field(
             default_factory=list,
-            advanced=False,
             title="Dropdown Options",
+            description=(
+                "If provided, renders the input as a dropdown selector "
+                "restricted to these values. Leave empty for free-text input."
+            ),
+            validation_alias=AliasChoices("options", "placeholder_values"),
+            json_schema_extra={"advanced": False, "secret": False},
         )
 
+        def generate_schema(self):
+            schema = super().generate_schema()
+            if possible_values := self.options:
+                schema["enum"] = possible_values
+            return schema
+
     class Output(AgentInputBlock.Output):
         result: str = SchemaField(description="Selected dropdown value.")
 
@@ -515,13 +514,13 @@ class AgentDropdownInputBlock(AgentInputBlock):
                 {
                     "value": "Option A",
                     "name": "dropdown_1",
-                    "placeholder_values": ["Option A", "Option B", "Option C"],
+                    "options": ["Option A", "Option B", "Option C"],
                     "description": "Dropdown example 1",
                 },
                 {
                     "value": "Option C",
                     "name": "dropdown_2",
-                    "placeholder_values": ["Option A", "Option B", "Option C"],
+                    "options": ["Option A", "Option B", "Option C"],
                     "description": "Dropdown example 2",
                 },
             ],

autogpt_platform/backend/backend/blocks/llm.py

@@ -104,6 +104,18 @@ class LlmModelMeta(EnumMeta):
 
 
 class LlmModel(str, Enum, metaclass=LlmModelMeta):
+
+    @classmethod
+    def _missing_(cls, value: object) -> "LlmModel | None":
+        """Handle provider-prefixed model names like 'anthropic/claude-sonnet-4-6'."""
+        if isinstance(value, str) and "/" in value:
+            stripped = value.split("/", 1)[1]
+            try:
+                return cls(stripped)
+            except ValueError:
+                return None
+        return None
+
     # OpenAI models
     O3_MINI = "o3-mini"
     O3 = "o3-2025-04-16"
@@ -712,6 +724,9 @@ def convert_openai_tool_fmt_to_anthropic(
 def extract_openai_reasoning(response) -> str | None:
     """Extract reasoning from OpenAI-compatible response if available."""
     """Note: This will likely not working since the reasoning is not present in another Response API"""
+    if not response.choices:
+        logger.warning("LLM response has empty choices in extract_openai_reasoning")
+        return None
     reasoning = None
     choice = response.choices[0]
     if hasattr(choice, "reasoning") and getattr(choice, "reasoning", None):
@@ -727,6 +742,9 @@ def extract_openai_reasoning(response) -> str | None:
 
 def extract_openai_tool_calls(response) -> list[ToolContentBlock] | None:
     """Extract tool calls from OpenAI-compatible response."""
+    if not response.choices:
+        logger.warning("LLM response has empty choices in extract_openai_tool_calls")
+        return None
     if response.choices[0].message.tool_calls:
         return [
             ToolContentBlock(
@@ -960,6 +978,8 @@ async def llm_call(
             response_format=response_format,  # type: ignore
             max_tokens=max_tokens,
         )
+        if not response.choices:
+            raise ValueError("Groq returned empty choices in response")
         return LLMResponse(
             raw_response=response.choices[0].message,
             prompt=prompt,
@@ -1019,12 +1039,8 @@ async def llm_call(
             parallel_tool_calls=parallel_tool_calls_param,
         )
 
-        # If there's no response, raise an error
         if not response.choices:
-            if response:
-                raise ValueError(f"OpenRouter error: {response}")
-            else:
-                raise ValueError("No response from OpenRouter.")
+            raise ValueError(f"OpenRouter returned empty choices: {response}")
 
         tool_calls = extract_openai_tool_calls(response)
         reasoning = extract_openai_reasoning(response)
@@ -1061,12 +1077,8 @@ async def llm_call(
             parallel_tool_calls=parallel_tool_calls_param,
         )
 
-        # If there's no response, raise an error
         if not response.choices:
-            if response:
-                raise ValueError(f"Llama API error: {response}")
-            else:
-                raise ValueError("No response from Llama API.")
+            raise ValueError(f"Llama API returned empty choices: {response}")
 
         tool_calls = extract_openai_tool_calls(response)
         reasoning = extract_openai_reasoning(response)
@@ -1096,6 +1108,8 @@ async def llm_call(
             messages=prompt,  # type: ignore
             max_tokens=max_tokens,
         )
+        if not completion.choices:
+            raise ValueError("AI/ML API returned empty choices in response")
 
         return LLMResponse(
             raw_response=completion.choices[0].message,
@@ -1132,6 +1146,9 @@ async def llm_call(
             parallel_tool_calls=parallel_tool_calls_param,
         )
 
+        if not response.choices:
+            raise ValueError(f"v0 API returned empty choices: {response}")
+
         tool_calls = extract_openai_tool_calls(response)
         reasoning = extract_openai_reasoning(response)
 
@@ -1999,6 +2016,19 @@ class AIConversationBlock(AIBlockBase):
     async def run(
         self, input_data: Input, *, credentials: APIKeyCredentials, **kwargs
     ) -> BlockOutput:
+        has_messages = any(
+            isinstance(m, dict)
+            and isinstance(m.get("content"), str)
+            and bool(m["content"].strip())
+            for m in (input_data.messages or [])
+        )
+        has_prompt = bool(input_data.prompt and input_data.prompt.strip())
+        if not has_messages and not has_prompt:
+            raise ValueError(
+                "Cannot call LLM with no messages and no prompt. "
+                "Provide at least one message or a non-empty prompt."
+            )
+
         response = await self.llm_call(
             AIStructuredResponseGeneratorBlock.Input(
                 prompt=input_data.prompt,

autogpt_platform/backend/backend/blocks/mcp/block.py

@@ -89,6 +89,12 @@ class MCPToolBlock(Block):
             default={},
             hidden=True,
         )
+        tool_description: str = SchemaField(
+            description="Description of the selected MCP tool. "
+            "Populated automatically when a tool is selected.",
+            default="",
+            hidden=True,
+        )
 
         tool_arguments: dict[str, Any] = SchemaField(
             description="Arguments to pass to the selected MCP tool. "

autogpt_platform/backend/backend/blocks/sql_query_block.py

@@ -0,0 +1,323 @@
+import asyncio
+from typing import Any, Literal
+
+from pydantic import SecretStr
+from sqlalchemy.engine.url import URL
+from sqlalchemy.exc import DBAPIError, OperationalError, ProgrammingError
+
+from backend.blocks._base import (
+    Block,
+    BlockCategory,
+    BlockOutput,
+    BlockSchemaInput,
+    BlockSchemaOutput,
+)
+from backend.blocks.sql_query_helpers import (
+    _DATABASE_TYPE_DEFAULT_PORT,
+    _DATABASE_TYPE_TO_DRIVER,
+    DatabaseType,
+    _execute_query,
+    _sanitize_error,
+    _validate_query_is_read_only,
+    _validate_single_statement,
+)
+from backend.data.model import (
+    CredentialsField,
+    CredentialsMetaInput,
+    SchemaField,
+    UserPasswordCredentials,
+)
+from backend.integrations.providers import ProviderName
+from backend.util.request import resolve_and_check_blocked
+
+TEST_CREDENTIALS = UserPasswordCredentials(
+    id="01234567-89ab-cdef-0123-456789abcdef",
+    provider="database",
+    username=SecretStr("test_user"),
+    password=SecretStr("test_pass"),
+    title="Mock Database credentials",
+)
+
+TEST_CREDENTIALS_INPUT = {
+    "provider": TEST_CREDENTIALS.provider,
+    "id": TEST_CREDENTIALS.id,
+    "type": TEST_CREDENTIALS.type,
+    "title": TEST_CREDENTIALS.title,
+}
+
+DatabaseCredentials = UserPasswordCredentials
+DatabaseCredentialsInput = CredentialsMetaInput[
+    Literal[ProviderName.DATABASE],
+    Literal["user_password"],
+]
+
+
+def DatabaseCredentialsField() -> DatabaseCredentialsInput:
+    return CredentialsField(
+        description="Database username and password",
+    )
+
+
+class SQLQueryBlock(Block):
+    class Input(BlockSchemaInput):
+        database_type: DatabaseType = SchemaField(
+            default=DatabaseType.POSTGRES,
+            description="Database engine",
+            advanced=False,
+        )
+        host: SecretStr = SchemaField(
+            description=(
+                "Database hostname or IP address. "
+                "Treated as a secret to avoid leaking infrastructure details. "
+                "Private/internal IPs are blocked (SSRF protection)."
+            ),
+            placeholder="db.example.com",
+            secret=True,
+        )
+        port: int | None = SchemaField(
+            default=None,
+            description=(
+                "Database port (leave empty for default: "
+                "PostgreSQL: 5432, MySQL: 3306, MSSQL: 1433)"
+            ),
+            ge=1,
+            le=65535,
+        )
+        database: str = SchemaField(
+            description="Name of the database to connect to",
+            placeholder="my_database",
+        )
+        query: str = SchemaField(
+            description="SQL query to execute",
+            placeholder="SELECT * FROM analytics.daily_active_users LIMIT 10",
+        )
+        read_only: bool = SchemaField(
+            default=True,
+            description=(
+                "When enabled (default), only SELECT queries are allowed "
+                "and the database session is set to read-only mode. "
+                "Disable to allow write operations (INSERT, UPDATE, DELETE, etc.)."
+            ),
+        )
+        timeout: int = SchemaField(
+            default=30,
+            description="Query timeout in seconds (max 120)",
+            ge=1,
+            le=120,
+        )
+        max_rows: int = SchemaField(
+            default=1000,
+            description="Maximum number of rows to return (max 10000)",
+            ge=1,
+            le=10000,
+        )
+        credentials: DatabaseCredentialsInput = DatabaseCredentialsField()
+
+    class Output(BlockSchemaOutput):
+        results: list[dict[str, Any]] = SchemaField(
+            description="Query results as a list of row dictionaries"
+        )
+        columns: list[str] = SchemaField(
+            description="Column names from the query result"
+        )
+        row_count: int = SchemaField(description="Number of rows returned")
+        truncated: bool = SchemaField(
+            description=(
+                "True when the result set was capped by max_rows, "
+                "indicating additional rows exist in the database"
+            )
+        )
+        affected_rows: int = SchemaField(
+            description="Number of rows affected by a write query (INSERT/UPDATE/DELETE)"
+        )
+        error: str = SchemaField(description="Error message if the query failed")
+
+    def __init__(self):
+        super().__init__(
+            id="4dc35c0f-4fd8-465e-9616-5a216f1ba2bc",
+            description=(
+                "Execute a SQL query. Read-only by default for safety "
+                "-- disable to allow write operations. "
+                "Supports PostgreSQL, MySQL, and MSSQL via SQLAlchemy."
+            ),
+            categories={BlockCategory.DATA},
+            input_schema=SQLQueryBlock.Input,
+            output_schema=SQLQueryBlock.Output,
+            test_input={
+                "query": "SELECT 1 AS test_col",
+                "database_type": DatabaseType.POSTGRES,
+                "host": "localhost",
+                "database": "test_db",
+                "timeout": 30,
+                "max_rows": 1000,
+                "credentials": TEST_CREDENTIALS_INPUT,
+            },
+            test_credentials=TEST_CREDENTIALS,
+            test_output=[
+                ("results", [{"test_col": 1}]),
+                ("columns", ["test_col"]),
+                ("row_count", 1),
+                ("truncated", False),
+            ],
+            test_mock={
+                "execute_query": lambda *_args, **_kwargs: (
+                    [{"test_col": 1}],
+                    ["test_col"],
+                    -1,
+                    False,
+                ),
+                "check_host_allowed": lambda *_args, **_kwargs: ["127.0.0.1"],
+            },
+        )
+
+    @staticmethod
+    async def check_host_allowed(host: str) -> list[str]:
+        """Validate that the given host is not a private/blocked address.
+
+        Returns the list of resolved IP addresses so the caller can pin the
+        connection to the validated IP (preventing DNS rebinding / TOCTOU).
+        Raises ValueError or OSError if the host is blocked.
+        Extracted as a method so it can be mocked during block tests.
+        """
+        return await resolve_and_check_blocked(host)
+
+    @staticmethod
+    def execute_query(
+        connection_url: URL | str,
+        query: str,
+        timeout: int,
+        max_rows: int,
+        read_only: bool = True,
+        database_type: DatabaseType = DatabaseType.POSTGRES,
+    ) -> tuple[list[dict[str, Any]], list[str], int, bool]:
+        """Execute a SQL query and return (rows, columns, affected_rows, truncated).
+
+        Delegates to ``_execute_query`` in ``sql_query_helpers``.
+        Extracted as a method so it can be mocked during block tests.
+        """
+        return _execute_query(
+            connection_url=connection_url,
+            query=query,
+            timeout=timeout,
+            max_rows=max_rows,
+            read_only=read_only,
+            database_type=database_type,
+        )
+
+    async def run(
+        self,
+        input_data: Input,
+        *,
+        credentials: DatabaseCredentials,
+        **_kwargs: Any,
+    ) -> BlockOutput:
+        # Validate query structure and read-only constraints.
+        error = self._validate_query(input_data)
+        if error:
+            yield "error", error
+            return
+
+        # Validate host and resolve for SSRF protection.
+        host, pinned_host, error = await self._resolve_host(input_data)
+        if error:
+            yield "error", error
+            return
+
+        # Build connection URL and execute.
+        port = input_data.port or _DATABASE_TYPE_DEFAULT_PORT[input_data.database_type]
+        username = credentials.username.get_secret_value()
+        connection_url = URL.create(
+            drivername=_DATABASE_TYPE_TO_DRIVER[input_data.database_type],
+            username=username,
+            password=credentials.password.get_secret_value(),
+            host=pinned_host,
+            port=port,
+            database=input_data.database,
+        )
+        conn_str = connection_url.render_as_string(hide_password=True)
+        db_name = input_data.database
+
+        def _sanitize(err: Exception) -> str:
+            return _sanitize_error(
+                str(err).strip(),
+                conn_str,
+                host=pinned_host,
+                original_host=host,
+                username=username,
+                port=port,
+                database=db_name,
+            )
+
+        try:
+            results, columns, affected, truncated = await asyncio.to_thread(
+                self.execute_query,
+                connection_url=connection_url,
+                query=input_data.query,
+                timeout=input_data.timeout,
+                max_rows=input_data.max_rows,
+                read_only=input_data.read_only,
+                database_type=input_data.database_type,
+            )
+            yield "results", results
+            yield "columns", columns
+            yield "row_count", len(results)
+            yield "truncated", truncated
+            if affected >= 0:
+                yield "affected_rows", affected
+        except OperationalError as e:
+            yield (
+                "error",
+                self._classify_operational_error(
+                    _sanitize(e),
+                    input_data.timeout,
+                ),
+            )
+        except ProgrammingError as e:
+            yield "error", f"SQL error: {_sanitize(e)}"
+        except DBAPIError as e:
+            yield "error", f"Database error: {_sanitize(e)}"
+        except ModuleNotFoundError:
+            yield (
+                "error",
+                (
+                    f"Database driver not available for "
+                    f"{input_data.database_type.value}. "
+                    f"Please contact the platform administrator."
+                ),
+            )
+
+    @staticmethod
+    def _validate_query(input_data: "SQLQueryBlock.Input") -> str | None:
+        """Validate query structure and read-only constraints."""
+        stmt_error, parsed_stmt = _validate_single_statement(input_data.query)
+        if stmt_error:
+            return stmt_error
+        assert parsed_stmt is not None
+        if input_data.read_only:
+            return _validate_query_is_read_only(parsed_stmt)
+        return None
+
+    async def _resolve_host(
+        self, input_data: "SQLQueryBlock.Input"
+    ) -> tuple[str, str, str | None]:
+        """Validate and resolve the database host. Returns (host, pinned_ip, error)."""
+        host = input_data.host.get_secret_value().strip()
+        if not host:
+            return "", "", "Database host is required."
+        if host.startswith("/"):
+            return host, "", "Unix socket connections are not allowed."
+        try:
+            resolved_ips = await self.check_host_allowed(host)
+        except (ValueError, OSError) as e:
+            return host, "", f"Blocked host: {str(e).strip()}"
+        return host, resolved_ips[0], None
+
+    @staticmethod
+    def _classify_operational_error(sanitized_msg: str, timeout: int) -> str:
+        """Classify an already-sanitized OperationalError for user display."""
+        lower = sanitized_msg.lower()
+        if "timeout" in lower or "cancel" in lower:
+            return f"Query timed out after {timeout}s."
+        if "connect" in lower:
+            return f"Failed to connect to database: {sanitized_msg}"
+        return f"Database error: {sanitized_msg}"

autogpt_platform/backend/backend/blocks/sql_query_helpers.py

@@ -0,0 +1,430 @@
+import re
+from datetime import date, datetime, time
+from decimal import Decimal
+from enum import Enum
+from typing import Any
+
+import sqlparse
+from sqlalchemy import create_engine, text
+from sqlalchemy.engine.url import URL
+
+
+class DatabaseType(str, Enum):
+    POSTGRES = "postgres"
+    MYSQL = "mysql"
+    MSSQL = "mssql"
+
+
+# Defense-in-depth: reject queries containing data-modifying keywords.
+# These are checked against parsed SQL tokens (not raw text) so column names
+# and string literals do not cause false positives.
+_DISALLOWED_KEYWORDS = {
+    "INSERT",
+    "UPDATE",
+    "DELETE",
+    "DROP",
+    "ALTER",
+    "CREATE",
+    "TRUNCATE",
+    "GRANT",
+    "REVOKE",
+    "COPY",
+    "EXECUTE",
+    "CALL",
+    "SET",
+    "RESET",
+    "DISCARD",
+    "NOTIFY",
+    "DO",
+    # MySQL file exfiltration: LOAD DATA LOCAL INFILE reads server/client files
+    "LOAD",
+    # MySQL REPLACE is INSERT-or-UPDATE; data modification
+    "REPLACE",
+    # ANSI MERGE (UPSERT) modifies data
+    "MERGE",
+    # MSSQL BULK INSERT loads external files into tables
+    "BULK",
+    # MSSQL EXEC / EXEC sp_name runs stored procedures (arbitrary code)
+    "EXEC",
+}
+
+# Map DatabaseType enum values to the expected SQLAlchemy driver prefix.
+_DATABASE_TYPE_TO_DRIVER = {
+    DatabaseType.POSTGRES: "postgresql",
+    DatabaseType.MYSQL: "mysql+pymysql",
+    DatabaseType.MSSQL: "mssql+pymssql",
+}
+
+# Connection timeout in seconds passed to the DBAPI driver (connect_timeout /
+# login_timeout).  This bounds how long the driver waits to establish a TCP
+# connection to the database server.  It is separate from the per-statement
+# timeout configured via SET commands inside _configure_session().
+_CONNECT_TIMEOUT_SECONDS = 10
+
+# Default ports for each database type.
+_DATABASE_TYPE_DEFAULT_PORT = {
+    DatabaseType.POSTGRES: 5432,
+    DatabaseType.MYSQL: 3306,
+    DatabaseType.MSSQL: 1433,
+}
+
+
+def _sanitize_error(
+    error_msg: str,
+    connection_string: str,
+    *,
+    host: str = "",
+    original_host: str = "",
+    username: str = "",
+    port: int = 0,
+    database: str = "",
+) -> str:
+    """Remove connection string, credentials, and infrastructure details
+    from error messages so they are safe to expose to the LLM.
+
+    Scrubs:
+    - The full connection string
+    - URL-embedded credentials (``://user:pass@``)
+    - ``password=<value>`` key-value pairs
+    - The database hostname / IP used for the connection
+    - The original (pre-resolution) hostname provided by the user
+    - Any IPv4 addresses that appear in the message
+    - Any bracketed IPv6 addresses (e.g. ``[::1]``, ``[fe80::1%eth0]``)
+    - The database username
+    - The database port number
+    - The database name
+    """
+    sanitized = error_msg.replace(connection_string, "<connection_string>")
+    sanitized = re.sub(r"password=[^\s&]+", "password=***", sanitized)
+    sanitized = re.sub(r"://[^@]+@", "://***:***@", sanitized)
+
+    # Replace the known host (may be an IP already) before the generic IP pass.
+    # Also replace the original (pre-DNS-resolution) hostname if it differs.
+    if original_host and original_host != host:
+        sanitized = sanitized.replace(original_host, "<host>")
+    if host:
+        sanitized = sanitized.replace(host, "<host>")
+
+    # Replace any remaining IPv4 addresses (e.g. resolved IPs the driver logs)
+    sanitized = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", "<ip>", sanitized)
+
+    # Replace bracketed IPv6 addresses (e.g. "[::1]", "[fe80::1%eth0]")
+    sanitized = re.sub(r"\[[0-9a-fA-F:]+(?:%[^\]]+)?\]", "<ip>", sanitized)
+
+    # Replace the database username (handles double-quoted, single-quoted,
+    # and unquoted formats across PostgreSQL, MySQL, and MSSQL error messages).
+    if username:
+        sanitized = re.sub(
+            r"""for user ["']?""" + re.escape(username) + r"""["']?""",
+            "for user <user>",
+            sanitized,
+        )
+        # Catch remaining bare occurrences in various quote styles:
+        # - PostgreSQL: "FATAL:  role "myuser" does not exist"
+        # - MySQL: "Access denied for user 'myuser'@'host'"
+        # - MSSQL: "Login failed for user 'myuser'"
+        sanitized = sanitized.replace(f'"{username}"', "<user>")
+        sanitized = sanitized.replace(f"'{username}'", "<user>")
+
+    # Replace the port number (handles "port 5432" and ":5432" formats)
+    if port:
+        port_str = re.escape(str(port))
+        sanitized = re.sub(
+            r"(?:port |:)" + port_str + r"(?![0-9])",
+            lambda m: ("port " if m.group().startswith("p") else ":") + "<port>",
+            sanitized,
+        )
+
+    # Replace the database name to avoid leaking internal infrastructure names.
+    # Use word-boundary regex to prevent mangling when the database name is a
+    # common substring (e.g. "test", "data", "on").
+    if database:
+        sanitized = re.sub(r"\b" + re.escape(database) + r"\b", "<database>", sanitized)
+
+    return sanitized
+
+
+def _extract_keyword_tokens(parsed: sqlparse.sql.Statement) -> list[str]:
+    """Extract keyword tokens from a parsed SQL statement.
+
+    Uses sqlparse token type classification to collect Keyword/DML/DDL/DCL
+    tokens. String literals and identifiers have different token types, so
+    they are naturally excluded from the result.
+    """
+    return [
+        token.normalized.upper()
+        for token in parsed.flatten()
+        if token.ttype
+        in (
+            sqlparse.tokens.Keyword,
+            sqlparse.tokens.Keyword.DML,
+            sqlparse.tokens.Keyword.DDL,
+            sqlparse.tokens.Keyword.DCL,
+        )
+    ]
+
+
+def _has_disallowed_into(stmt: sqlparse.sql.Statement) -> bool:
+    """Check if a statement contains a disallowed ``INTO`` clause.
+
+    ``SELECT ... INTO @variable`` is a valid read-only MySQL syntax that stores
+    a query result into a session-scoped user variable.  All other forms of
+    ``INTO`` are data-modifying or file-writing and must be blocked:
+
+    * ``SELECT ... INTO new_table``  (PostgreSQL / MSSQL – creates a table)
+    * ``SELECT ... INTO OUTFILE``    (MySQL – writes to the filesystem)
+    * ``SELECT ... INTO DUMPFILE``   (MySQL – writes to the filesystem)
+    * ``INSERT INTO ...``            (already blocked by INSERT being in the
+      disallowed set, but we reject INTO as well for defense-in-depth)
+
+    Returns ``True`` if the statement contains a disallowed ``INTO``.
+    """
+    flat = list(stmt.flatten())
+    for i, token in enumerate(flat):
+        if not (
+            token.ttype in (sqlparse.tokens.Keyword,)
+            and token.normalized.upper() == "INTO"
+        ):
+            continue
+
+        # Look at the first non-whitespace token after INTO.
+        j = i + 1
+        while j < len(flat) and flat[j].ttype is sqlparse.tokens.Text.Whitespace:
+            j += 1
+
+        if j >= len(flat):
+            # INTO at the very end – malformed, block it.
+            return True
+
+        next_token = flat[j]
+        # MySQL user variable: either a single Name starting with "@"
+        # (e.g. ``@total``) or a bare ``@`` Operator token followed by a Name.
+        if next_token.ttype is sqlparse.tokens.Name and next_token.value.startswith(
+            "@"
+        ):
+            continue
+        if next_token.ttype is sqlparse.tokens.Operator and next_token.value == "@":
+            continue
+
+        # Everything else (table name, OUTFILE, DUMPFILE, etc.) is disallowed.
+        return True
+
+    return False
+
+
+def _validate_query_is_read_only(stmt: sqlparse.sql.Statement) -> str | None:
+    """Validate that a parsed SQL statement is read-only (SELECT/WITH only).
+
+    Accepts an already-parsed statement from ``_validate_single_statement``
+    to avoid re-parsing. Checks:
+    1. Statement type must be SELECT (sqlparse classifies WITH...SELECT as SELECT)
+    2. No disallowed keywords (INSERT, UPDATE, DELETE, DROP, etc.)
+    3. No disallowed INTO clauses (allows MySQL ``SELECT ... INTO @variable``)
+
+    Returns an error message if the query is not read-only, None otherwise.
+    """
+    # sqlparse returns 'SELECT' for SELECT and WITH...SELECT queries
+    if stmt.get_type() != "SELECT":
+        return "Only SELECT queries are allowed."
+
+    # Defense-in-depth: check parsed keyword tokens for disallowed keywords
+    for kw in _extract_keyword_tokens(stmt):
+        # Normalize multi-word tokens (e.g. "SET LOCAL" -> "SET")
+        base_kw = kw.split()[0] if " " in kw else kw
+        if base_kw in _DISALLOWED_KEYWORDS:
+            return f"Disallowed SQL keyword: {kw}"
+
+    # Contextual check for INTO: allow MySQL @variable syntax, block everything else
+    if _has_disallowed_into(stmt):
+        return "Disallowed SQL keyword: INTO"
+
+    return None
+
+
+def _validate_single_statement(
+    query: str,
+) -> tuple[str | None, sqlparse.sql.Statement | None]:
+    """Validate that the query contains exactly one non-empty SQL statement.
+
+    Returns (error_message, parsed_statement). If error_message is not None,
+    the query is invalid and parsed_statement will be None.
+    """
+    stripped = query.strip().rstrip(";").strip()
+    if not stripped:
+        return "Query is empty.", None
+
+    # Parse the SQL using sqlparse for proper tokenization
+    statements = sqlparse.parse(stripped)
+
+    # Filter out empty statements and comment-only statements
+    statements = [
+        s
+        for s in statements
+        if s.tokens
+        and str(s).strip()
+        and not all(
+            t.is_whitespace or t.ttype in sqlparse.tokens.Comment for t in s.flatten()
+        )
+    ]
+
+    if not statements:
+        return "Query is empty.", None
+
+    # Reject multiple statements -- prevents injection via semicolons
+    if len(statements) > 1:
+        return "Only single statements are allowed.", None
+
+    return None, statements[0]
+
+
+def _serialize_value(value: Any) -> Any:
+    """Convert database-specific types to JSON-serializable Python types."""
+    if isinstance(value, Decimal):
+        # NaN / Infinity are not valid JSON numbers; serialize as strings.
+        if value.is_nan() or value.is_infinite():
+            return str(value)
+        # Use int for whole numbers; use str for fractional to preserve exact
+        # precision (float would silently round high-precision analytics values).
+        if value == value.to_integral_value():
+            return int(value)
+        return str(value)
+    if isinstance(value, (datetime, date, time)):
+        return value.isoformat()
+    if isinstance(value, memoryview):
+        return bytes(value).hex()
+    if isinstance(value, bytes):
+        return value.hex()
+    return value
+
+
+def _configure_session(
+    conn: Any,
+    dialect_name: str,
+    timeout_ms: str,
+    read_only: bool,
+) -> None:
+    """Set session-level timeout and read-only mode for the given dialect.
+
+    Timeout limitations by database:
+
+    * **PostgreSQL** – ``statement_timeout`` reliably cancels any running
+      statement (SELECT or DML) after the configured duration.
+    * **MySQL** – ``MAX_EXECUTION_TIME`` only applies to **read-only SELECT**
+      statements.  DML (INSERT/UPDATE/DELETE) and DDL are *not* bounded by
+      this hint; they rely on the server's ``wait_timeout`` /
+      ``interactive_timeout`` instead.  There is no session-level setting in
+      MySQL that reliably cancels long-running writes.
+    * **MSSQL** – ``SET LOCK_TIMEOUT`` only limits how long the server waits
+      to acquire a **lock**.  CPU-bound queries (e.g. large scans, hash
+      joins) that do not block on locks will *not* be cancelled.  MSSQL has
+      no session-level ``statement_timeout`` equivalent; the closest
+      mechanism is Resource Governor (requires sysadmin configuration) or
+      ``CONTEXT_INFO``-based external monitoring.
+
+    Note: SQLite is not supported by this block.  The ``_configure_session``
+    function is a no-op for unrecognised dialect names, so an SQLite engine
+    would skip all SET commands silently.  The block's ``DatabaseType`` enum
+    intentionally excludes SQLite.
+    """
+    if dialect_name == "postgresql":
+        conn.execute(text("SET statement_timeout = " + timeout_ms))
+        if read_only:
+            conn.execute(text("SET default_transaction_read_only = ON"))
+    elif dialect_name == "mysql":
+        # NOTE: MAX_EXECUTION_TIME only applies to SELECT statements.
+        # Write queries (INSERT/UPDATE/DELETE) are not bounded by this
+        # setting; they rely on the database's wait_timeout instead.
+        # See docstring above for full limitations.
+        conn.execute(text("SET SESSION MAX_EXECUTION_TIME = " + timeout_ms))
+        if read_only:
+            conn.execute(text("SET SESSION TRANSACTION READ ONLY"))
+    elif dialect_name == "mssql":
+        # MSSQL: SET LOCK_TIMEOUT limits lock-wait time (ms) only.
+        # CPU-bound queries without lock contention are NOT cancelled.
+        # See docstring above for full limitations.
+        conn.execute(text("SET LOCK_TIMEOUT " + timeout_ms))
+        # MSSQL lacks a session-level read-only mode like
+        # PostgreSQL/MySQL.  Read-only enforcement is handled by
+        # the SQL validation layer (_validate_query_is_read_only)
+        # and the ROLLBACK in the finally block.
+
+
+def _run_in_transaction(
+    conn: Any,
+    dialect_name: str,
+    query: str,
+    max_rows: int,
+    read_only: bool,
+) -> tuple[list[dict[str, Any]], list[str], int, bool]:
+    """Execute a query inside an explicit transaction, returning results.
+
+    Returns ``(rows, columns, affected_rows, truncated)`` where *truncated*
+    is ``True`` when ``fetchmany`` returned exactly ``max_rows`` rows,
+    indicating that additional rows may exist in the result set.
+    """
+    # MSSQL uses T-SQL "BEGIN TRANSACTION"; others use "BEGIN".
+    begin_stmt = "BEGIN TRANSACTION" if dialect_name == "mssql" else "BEGIN"
+    conn.execute(text(begin_stmt))
+    try:
+        result = conn.execute(text(query))
+        affected = result.rowcount if not result.returns_rows else -1
+        columns = list(result.keys()) if result.returns_rows else []
+        rows = result.fetchmany(max_rows) if result.returns_rows else []
+        truncated = len(rows) == max_rows
+        results = [
+            {col: _serialize_value(val) for col, val in zip(columns, row)}
+            for row in rows
+        ]
+    except Exception:
+        try:
+            conn.execute(text("ROLLBACK"))
+        except Exception:
+            pass
+        raise
+    else:
+        conn.execute(text("ROLLBACK" if read_only else "COMMIT"))
+    return results, columns, affected, truncated
+
+
+def _execute_query(
+    connection_url: URL | str,
+    query: str,
+    timeout: int,
+    max_rows: int,
+    read_only: bool = True,
+    database_type: DatabaseType = DatabaseType.POSTGRES,
+) -> tuple[list[dict[str, Any]], list[str], int, bool]:
+    """Execute a SQL query and return (rows, columns, affected_rows, truncated).
+
+    Uses SQLAlchemy to connect to any supported database.
+    For SELECT queries, rows are limited to ``max_rows`` via DBAPI fetchmany.
+    ``truncated`` is ``True`` when the result set was capped by ``max_rows``.
+    For write queries, affected_rows contains the rowcount from the driver.
+    When ``read_only`` is True, the database session is set to read-only
+    mode and the transaction is always rolled back.
+    """
+    # Determine driver-specific connection timeout argument.
+    # pymssql uses "login_timeout", while PostgreSQL/MySQL use "connect_timeout".
+    timeout_key = (
+        "login_timeout" if database_type == DatabaseType.MSSQL else "connect_timeout"
+    )
+    engine = create_engine(
+        connection_url, connect_args={timeout_key: _CONNECT_TIMEOUT_SECONDS}
+    )
+    try:
+        with engine.connect() as conn:
+            # Use AUTOCOMMIT so SET commands take effect immediately.
+            conn = conn.execution_options(isolation_level="AUTOCOMMIT")
+
+            # Compute timeout in milliseconds. The value is Pydantic-validated
+            # (ge=1, le=120), but we use int() as defense-in-depth.
+            # NOTE: SET commands do not support bind parameters in most
+            # databases, so we use str(int(...)) for safe interpolation.
+            timeout_ms = str(int(timeout * 1000))
+
+            _configure_session(conn, engine.dialect.name, timeout_ms, read_only)
+            return _run_in_transaction(
+                conn, engine.dialect.name, query, max_rows, read_only
+            )
+    finally:
+        engine.dispose()

autogpt_platform/backend/backend/blocks/test/test_block.py

@@ -4,6 +4,8 @@ import pytest
 
 from backend.blocks import get_blocks
 from backend.blocks._base import Block, BlockSchemaInput
+from backend.blocks.io import AgentDropdownInputBlock, AgentInputBlock
+from backend.data.graph import BaseGraph
 from backend.data.model import SchemaField
 from backend.util.test import execute_block_test
 
@@ -279,3 +281,113 @@ class TestAutoCredentialsFieldsValidation:
         assert "Duplicate auto_credentials kwarg_name 'credentials'" in str(
             exc_info.value
         )
+
+
+def test_agent_input_block_ignores_legacy_placeholder_values():
+    """Verify AgentInputBlock.Input.model_construct tolerates extra placeholder_values
+    for backward compatibility with existing agent JSON."""
+    legacy_data = {
+        "name": "url",
+        "value": "",
+        "description": "Enter a URL",
+        "placeholder_values": ["https://example.com"],
+    }
+    instance = AgentInputBlock.Input.model_construct(**legacy_data)
+    schema = instance.generate_schema()
+    assert (
+        "enum" not in schema
+    ), "AgentInputBlock should not produce enum from legacy placeholder_values"
+
+
+def test_dropdown_input_block_produces_enum():
+    """Verify AgentDropdownInputBlock.Input.generate_schema() produces enum
+    using the canonical 'options' field name."""
+    opts = ["Option A", "Option B"]
+    instance = AgentDropdownInputBlock.Input.model_construct(
+        name="choice", value=None, options=opts
+    )
+    schema = instance.generate_schema()
+    assert schema.get("enum") == opts
+
+
+def test_dropdown_input_block_legacy_placeholder_values_produces_enum():
+    """Verify backward compat: passing legacy 'placeholder_values' to
+    AgentDropdownInputBlock still produces enum via model_construct remap."""
+    opts = ["Option A", "Option B"]
+    instance = AgentDropdownInputBlock.Input.model_construct(
+        name="choice", value=None, placeholder_values=opts
+    )
+    schema = instance.generate_schema()
+    assert (
+        schema.get("enum") == opts
+    ), "Legacy placeholder_values should be remapped to options"
+
+
+def test_generate_schema_integration_legacy_placeholder_values():
+    """Test the full Graph._generate_schema path with legacy placeholder_values
+    on AgentInputBlock — verifies no enum leaks through the graph loading path."""
+    legacy_input_default = {
+        "name": "url",
+        "value": "",
+        "description": "Enter a URL",
+        "placeholder_values": ["https://example.com"],
+    }
+    result = BaseGraph._generate_schema(
+        (AgentInputBlock.Input, legacy_input_default),
+    )
+    url_props = result["properties"]["url"]
+    assert (
+        "enum" not in url_props
+    ), "Graph schema should not contain enum from AgentInputBlock placeholder_values"
+
+
+def test_generate_schema_integration_dropdown_produces_enum():
+    """Test the full Graph._generate_schema path with AgentDropdownInputBlock
+    — verifies enum IS produced for dropdown blocks using canonical field name."""
+    dropdown_input_default = {
+        "name": "color",
+        "value": None,
+        "options": ["Red", "Green", "Blue"],
+    }
+    result = BaseGraph._generate_schema(
+        (AgentDropdownInputBlock.Input, dropdown_input_default),
+    )
+    color_props = result["properties"]["color"]
+    assert color_props.get("enum") == [
+        "Red",
+        "Green",
+        "Blue",
+    ], "Graph schema should contain enum from AgentDropdownInputBlock"
+
+
+def test_generate_schema_integration_dropdown_legacy_placeholder_values():
+    """Test the full Graph._generate_schema path with AgentDropdownInputBlock
+    using legacy 'placeholder_values' — verifies backward compat produces enum."""
+    legacy_dropdown_input_default = {
+        "name": "color",
+        "value": None,
+        "placeholder_values": ["Red", "Green", "Blue"],
+    }
+    result = BaseGraph._generate_schema(
+        (AgentDropdownInputBlock.Input, legacy_dropdown_input_default),
+    )
+    color_props = result["properties"]["color"]
+    assert color_props.get("enum") == [
+        "Red",
+        "Green",
+        "Blue",
+    ], "Legacy placeholder_values should still produce enum via model_construct remap"
+
+
+def test_dropdown_input_block_init_legacy_placeholder_values():
+    """Verify backward compat: constructing AgentDropdownInputBlock.Input via
+    model_validate with legacy 'placeholder_values' correctly maps to 'options'."""
+    opts = ["Option A", "Option B"]
+    instance = AgentDropdownInputBlock.Input.model_validate(
+        {"name": "choice", "value": None, "placeholder_values": opts}
+    )
+    assert (
+        instance.options == opts
+    ), "Legacy placeholder_values should be remapped to options via model_validate"
+    schema = instance.generate_schema()
+    assert schema.get("enum") == opts

autogpt_platform/backend/backend/blocks/test/test_blocks_dos_vulnerability.py

@@ -207,6 +207,51 @@ class TestXMLParserBlockSecurity:
                 pass
 
 
+class TestXMLParserBlockSyntaxErrors:
+    """XML syntax errors should raise ValueError (not SyntaxError).
+
+    This ensures the base Block.execute() wraps them as BlockExecutionError
+    (expected / user-caused) instead of BlockUnknownError (unexpected / alerts
+    Sentry).
+    """
+
+    async def test_unclosed_tag_raises_value_error(self):
+        """Unclosed tags should raise ValueError, not SyntaxError."""
+        block = XMLParserBlock()
+        bad_xml = "<root><unclosed>"
+
+        with pytest.raises(ValueError, match="Unclosed tag"):
+            async for _ in block.run(XMLParserBlock.Input(input_xml=bad_xml)):
+                pass
+
+    async def test_unexpected_closing_tag_raises_value_error(self):
+        """Extra closing tags should raise ValueError, not SyntaxError."""
+        block = XMLParserBlock()
+        bad_xml = "</unexpected>"
+
+        with pytest.raises(ValueError):
+            async for _ in block.run(XMLParserBlock.Input(input_xml=bad_xml)):
+                pass
+
+    async def test_empty_xml_raises_value_error(self):
+        """Empty XML input should raise ValueError."""
+        block = XMLParserBlock()
+
+        with pytest.raises(ValueError, match="XML input is empty"):
+            async for _ in block.run(XMLParserBlock.Input(input_xml="")):
+                pass
+
+    async def test_syntax_error_from_parser_becomes_value_error(self):
+        """SyntaxErrors from gravitasml library become ValueError (BlockExecutionError)."""
+        block = XMLParserBlock()
+        # Malformed XML that might trigger a SyntaxError from the parser
+        bad_xml = "<root><child>no closing"
+
+        with pytest.raises(ValueError):
+            async for _ in block.run(XMLParserBlock.Input(input_xml=bad_xml)):
+                pass
+
+
 class TestStoreMediaFileSecurity:
     """Test file storage security limits."""
 

autogpt_platform/backend/backend/blocks/test/test_llm.py

@@ -488,6 +488,154 @@ class TestLLMStatsTracking:
         assert outputs["response"] == {"result": "test"}
 
 
+class TestAIConversationBlockValidation:
+    """Test that AIConversationBlock validates inputs before calling the LLM."""
+
+    @pytest.mark.asyncio
+    async def test_empty_messages_and_empty_prompt_raises_error(self):
+        """Empty messages with no prompt should raise ValueError, not a cryptic API error."""
+        block = llm.AIConversationBlock()
+
+        input_data = llm.AIConversationBlock.Input(
+            messages=[],
+            prompt="",
+            model=llm.DEFAULT_LLM_MODEL,
+            credentials=_TEST_AI_CREDENTIALS,
+        )
+
+        with pytest.raises(ValueError, match="no messages and no prompt"):
+            async for _ in block.run(input_data, credentials=llm.TEST_CREDENTIALS):
+                pass
+
+    @pytest.mark.asyncio
+    async def test_empty_messages_with_prompt_succeeds(self):
+        """Empty messages but a non-empty prompt should proceed without error."""
+        block = llm.AIConversationBlock()
+
+        async def mock_llm_call(input_data, credentials):
+            return {"response": "OK"}
+
+        with patch.object(block, "llm_call", new=AsyncMock(side_effect=mock_llm_call)):
+            input_data = llm.AIConversationBlock.Input(
+                messages=[],
+                prompt="Hello, how are you?",
+                model=llm.DEFAULT_LLM_MODEL,
+                credentials=_TEST_AI_CREDENTIALS,
+            )
+
+            outputs = {}
+            async for name, data in block.run(
+                input_data, credentials=llm.TEST_CREDENTIALS
+            ):
+                outputs[name] = data
+
+        assert outputs["response"] == "OK"
+
+    @pytest.mark.asyncio
+    async def test_nonempty_messages_with_empty_prompt_succeeds(self):
+        """Non-empty messages with no prompt should proceed without error."""
+        block = llm.AIConversationBlock()
+
+        async def mock_llm_call(input_data, credentials):
+            return {"response": "response from conversation"}
+
+        with patch.object(block, "llm_call", new=AsyncMock(side_effect=mock_llm_call)):
+            input_data = llm.AIConversationBlock.Input(
+                messages=[{"role": "user", "content": "Hello"}],
+                prompt="",
+                model=llm.DEFAULT_LLM_MODEL,
+                credentials=_TEST_AI_CREDENTIALS,
+            )
+
+            outputs = {}
+            async for name, data in block.run(
+                input_data, credentials=llm.TEST_CREDENTIALS
+            ):
+                outputs[name] = data
+
+        assert outputs["response"] == "response from conversation"
+
+    @pytest.mark.asyncio
+    async def test_messages_with_empty_content_raises_error(self):
+        """Messages with empty content strings should be treated as no messages."""
+        block = llm.AIConversationBlock()
+
+        input_data = llm.AIConversationBlock.Input(
+            messages=[{"role": "user", "content": ""}],
+            prompt="",
+            model=llm.DEFAULT_LLM_MODEL,
+            credentials=_TEST_AI_CREDENTIALS,
+        )
+
+        with pytest.raises(ValueError, match="no messages and no prompt"):
+            async for _ in block.run(input_data, credentials=llm.TEST_CREDENTIALS):
+                pass
+
+    @pytest.mark.asyncio
+    async def test_messages_with_whitespace_content_raises_error(self):
+        """Messages with whitespace-only content should be treated as no messages."""
+        block = llm.AIConversationBlock()
+
+        input_data = llm.AIConversationBlock.Input(
+            messages=[{"role": "user", "content": "   "}],
+            prompt="",
+            model=llm.DEFAULT_LLM_MODEL,
+            credentials=_TEST_AI_CREDENTIALS,
+        )
+
+        with pytest.raises(ValueError, match="no messages and no prompt"):
+            async for _ in block.run(input_data, credentials=llm.TEST_CREDENTIALS):
+                pass
+
+    @pytest.mark.asyncio
+    async def test_messages_with_none_entry_raises_error(self):
+        """Messages list containing None should be treated as no messages."""
+        block = llm.AIConversationBlock()
+
+        input_data = llm.AIConversationBlock.Input(
+            messages=[None],
+            prompt="",
+            model=llm.DEFAULT_LLM_MODEL,
+            credentials=_TEST_AI_CREDENTIALS,
+        )
+
+        with pytest.raises(ValueError, match="no messages and no prompt"):
+            async for _ in block.run(input_data, credentials=llm.TEST_CREDENTIALS):
+                pass
+
+    @pytest.mark.asyncio
+    async def test_messages_with_empty_dict_raises_error(self):
+        """Messages list containing empty dict should be treated as no messages."""
+        block = llm.AIConversationBlock()
+
+        input_data = llm.AIConversationBlock.Input(
+            messages=[{}],
+            prompt="",
+            model=llm.DEFAULT_LLM_MODEL,
+            credentials=_TEST_AI_CREDENTIALS,
+        )
+
+        with pytest.raises(ValueError, match="no messages and no prompt"):
+            async for _ in block.run(input_data, credentials=llm.TEST_CREDENTIALS):
+                pass
+
+    @pytest.mark.asyncio
+    async def test_messages_with_none_content_raises_error(self):
+        """Messages with content=None should not crash with AttributeError."""
+        block = llm.AIConversationBlock()
+
+        input_data = llm.AIConversationBlock.Input(
+            messages=[{"role": "user", "content": None}],
+            prompt="",
+            model=llm.DEFAULT_LLM_MODEL,
+            credentials=_TEST_AI_CREDENTIALS,
+        )
+
+        with pytest.raises(ValueError, match="no messages and no prompt"):
+            async for _ in block.run(input_data, credentials=llm.TEST_CREDENTIALS):
+                pass
+
+
 class TestAITextSummarizerValidation:
     """Test that AITextSummarizerBlock validates LLM responses are strings."""
 
@@ -809,3 +957,33 @@ class TestUserErrorStatusCodeHandling:
 
         mock_warning.assert_called_once()
         mock_exception.assert_not_called()
+
+
+class TestLlmModelMissing:
+    """Test that LlmModel handles provider-prefixed model names."""
+
+    def test_provider_prefixed_model_resolves(self):
+        """Provider-prefixed model string should resolve to the correct enum member."""
+        assert (
+            llm.LlmModel("anthropic/claude-sonnet-4-6")
+            == llm.LlmModel.CLAUDE_4_6_SONNET
+        )
+
+    def test_bare_model_still_works(self):
+        """Bare (non-prefixed) model string should still resolve correctly."""
+        assert llm.LlmModel("claude-sonnet-4-6") == llm.LlmModel.CLAUDE_4_6_SONNET
+
+    def test_invalid_prefixed_model_raises(self):
+        """Unknown provider-prefixed model string should raise ValueError."""
+        with pytest.raises(ValueError):
+            llm.LlmModel("invalid/nonexistent-model")
+
+    def test_slash_containing_value_direct_lookup(self):
+        """Enum values with '/' (e.g., OpenRouter models) should resolve via direct lookup, not _missing_."""
+        assert llm.LlmModel("google/gemini-2.5-pro") == llm.LlmModel.GEMINI_2_5_PRO
+
+    def test_double_prefixed_slash_model(self):
+        """Double-prefixed value should still resolve by stripping first prefix."""
+        assert (
+            llm.LlmModel("extra/google/gemini-2.5-pro") == llm.LlmModel.GEMINI_2_5_PRO
+        )

autogpt_platform/backend/backend/blocks/test/test_llm_empty_choices.py

@@ -0,0 +1,87 @@
+"""Tests for empty-choices guard in extract_openai_tool_calls() and extract_openai_reasoning()."""
+
+from unittest.mock import MagicMock
+
+from backend.blocks.llm import extract_openai_reasoning, extract_openai_tool_calls
+
+
+class TestExtractOpenaiToolCallsEmptyChoices:
+    """extract_openai_tool_calls() must return None when choices is empty."""
+
+    def test_returns_none_for_empty_choices(self):
+        response = MagicMock()
+        response.choices = []
+        assert extract_openai_tool_calls(response) is None
+
+    def test_returns_none_for_none_choices(self):
+        response = MagicMock()
+        response.choices = None
+        assert extract_openai_tool_calls(response) is None
+
+    def test_returns_tool_calls_when_choices_present(self):
+        tool = MagicMock()
+        tool.id = "call_1"
+        tool.type = "function"
+        tool.function.name = "my_func"
+        tool.function.arguments = '{"a": 1}'
+
+        message = MagicMock()
+        message.tool_calls = [tool]
+
+        choice = MagicMock()
+        choice.message = message
+
+        response = MagicMock()
+        response.choices = [choice]
+
+        result = extract_openai_tool_calls(response)
+        assert result is not None
+        assert len(result) == 1
+        assert result[0].function.name == "my_func"
+
+    def test_returns_none_when_no_tool_calls(self):
+        message = MagicMock()
+        message.tool_calls = None
+
+        choice = MagicMock()
+        choice.message = message
+
+        response = MagicMock()
+        response.choices = [choice]
+
+        assert extract_openai_tool_calls(response) is None
+
+
+class TestExtractOpenaiReasoningEmptyChoices:
+    """extract_openai_reasoning() must return None when choices is empty."""
+
+    def test_returns_none_for_empty_choices(self):
+        response = MagicMock()
+        response.choices = []
+        assert extract_openai_reasoning(response) is None
+
+    def test_returns_none_for_none_choices(self):
+        response = MagicMock()
+        response.choices = None
+        assert extract_openai_reasoning(response) is None
+
+    def test_returns_reasoning_from_choice(self):
+        choice = MagicMock()
+        choice.reasoning = "Step-by-step reasoning"
+        choice.message = MagicMock(spec=[])  # no 'reasoning' attr on message
+
+        response = MagicMock(spec=[])  # no 'reasoning' attr on response
+        response.choices = [choice]
+
+        result = extract_openai_reasoning(response)
+        assert result == "Step-by-step reasoning"
+
+    def test_returns_none_when_no_reasoning(self):
+        choice = MagicMock(spec=[])  # no 'reasoning' attr
+        choice.message = MagicMock(spec=[])  # no 'reasoning' attr
+
+        response = MagicMock(spec=[])  # no 'reasoning' attr
+        response.choices = [choice]
+
+        result = extract_openai_reasoning(response)
+        assert result is None

autogpt_platform/backend/backend/blocks/test/test_orchestrator.py

@@ -1074,6 +1074,7 @@ async def test_orchestrator_uses_customized_name_for_blocks():
     mock_node.block_id = StoreValueBlock().id
     mock_node.metadata = {"customized_name": "My Custom Tool Name"}
     mock_node.block = StoreValueBlock()
+    mock_node.input_default = {}
 
     # Create a mock link
     mock_link = MagicMock(spec=Link)
@@ -1105,6 +1106,7 @@ async def test_orchestrator_falls_back_to_block_name():
     mock_node.block_id = StoreValueBlock().id
     mock_node.metadata = {}  # No customized_name
     mock_node.block = StoreValueBlock()
+    mock_node.input_default = {}
 
     # Create a mock link
     mock_link = MagicMock(spec=Link)

autogpt_platform/backend/backend/blocks/test/test_orchestrator_execution_mode.py

@@ -0,0 +1,202 @@
+"""Tests for ExecutionMode enum and provider validation in the orchestrator.
+
+Covers:
+- ExecutionMode enum members exist and have stable values
+- EXTENDED_THINKING provider validation (anthropic/open_router allowed, others rejected)
+- EXTENDED_THINKING model-name validation (must start with "claude")
+"""
+
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+
+from backend.blocks.llm import LlmModel
+from backend.blocks.orchestrator import ExecutionMode, OrchestratorBlock
+
+# ---------------------------------------------------------------------------
+# ExecutionMode enum integrity
+# ---------------------------------------------------------------------------
+
+
+class TestExecutionModeEnum:
+    """Guard against accidental renames or removals of enum members."""
+
+    def test_built_in_exists(self):
+        assert hasattr(ExecutionMode, "BUILT_IN")
+        assert ExecutionMode.BUILT_IN.value == "built_in"
+
+    def test_extended_thinking_exists(self):
+        assert hasattr(ExecutionMode, "EXTENDED_THINKING")
+        assert ExecutionMode.EXTENDED_THINKING.value == "extended_thinking"
+
+    def test_exactly_two_members(self):
+        """If a new mode is added, this test should be updated intentionally."""
+        assert set(ExecutionMode.__members__.keys()) == {
+            "BUILT_IN",
+            "EXTENDED_THINKING",
+        }
+
+    def test_string_enum(self):
+        """ExecutionMode is a str enum so it serialises cleanly to JSON."""
+        assert isinstance(ExecutionMode.BUILT_IN, str)
+        assert isinstance(ExecutionMode.EXTENDED_THINKING, str)
+
+    def test_round_trip_from_value(self):
+        """Constructing from the string value should return the same member."""
+        assert ExecutionMode("built_in") is ExecutionMode.BUILT_IN
+        assert ExecutionMode("extended_thinking") is ExecutionMode.EXTENDED_THINKING
+
+
+# ---------------------------------------------------------------------------
+# Provider validation (inline in OrchestratorBlock.run)
+# ---------------------------------------------------------------------------
+
+
+def _make_model_stub(provider: str, value: str):
+    """Create a lightweight stub that behaves like LlmModel for validation."""
+    metadata = MagicMock()
+    metadata.provider = provider
+    stub = MagicMock()
+    stub.metadata = metadata
+    stub.value = value
+    return stub
+
+
+class TestExtendedThinkingProviderValidation:
+    """The orchestrator rejects EXTENDED_THINKING for non-Anthropic providers."""
+
+    def test_anthropic_provider_accepted(self):
+        """provider='anthropic' + claude model should not raise."""
+        model = _make_model_stub("anthropic", "claude-opus-4-6")
+        provider = model.metadata.provider
+        model_name = model.value
+        assert provider in ("anthropic", "open_router")
+        assert model_name.startswith("claude")
+
+    def test_open_router_provider_accepted(self):
+        """provider='open_router' + claude model should not raise."""
+        model = _make_model_stub("open_router", "claude-sonnet-4-6")
+        provider = model.metadata.provider
+        model_name = model.value
+        assert provider in ("anthropic", "open_router")
+        assert model_name.startswith("claude")
+
+    def test_openai_provider_rejected(self):
+        """provider='openai' should be rejected for EXTENDED_THINKING."""
+        model = _make_model_stub("openai", "gpt-4o")
+        provider = model.metadata.provider
+        assert provider not in ("anthropic", "open_router")
+
+    def test_groq_provider_rejected(self):
+        model = _make_model_stub("groq", "llama-3.3-70b-versatile")
+        provider = model.metadata.provider
+        assert provider not in ("anthropic", "open_router")
+
+    def test_non_claude_model_rejected_even_if_anthropic_provider(self):
+        """A hypothetical non-Claude model with provider='anthropic' is rejected."""
+        model = _make_model_stub("anthropic", "not-a-claude-model")
+        model_name = model.value
+        assert not model_name.startswith("claude")
+
+    def test_real_gpt4o_model_rejected(self):
+        """Verify a real LlmModel enum member (GPT4O) fails the provider check."""
+        model = LlmModel.GPT4O
+        provider = model.metadata.provider
+        assert provider not in ("anthropic", "open_router")
+
+    def test_real_claude_model_passes(self):
+        """Verify a real LlmModel enum member (CLAUDE_4_6_SONNET) passes."""
+        model = LlmModel.CLAUDE_4_6_SONNET
+        provider = model.metadata.provider
+        model_name = model.value
+        assert provider in ("anthropic", "open_router")
+        assert model_name.startswith("claude")
+
+
+# ---------------------------------------------------------------------------
+# Integration-style: exercise the validation branch via OrchestratorBlock.run
+# ---------------------------------------------------------------------------
+
+
+def _make_input_data(model, execution_mode=ExecutionMode.EXTENDED_THINKING):
+    """Build a minimal MagicMock that satisfies OrchestratorBlock.run's early path."""
+    inp = MagicMock()
+    inp.execution_mode = execution_mode
+    inp.model = model
+    inp.prompt = "test"
+    inp.sys_prompt = ""
+    inp.conversation_history = []
+    inp.last_tool_output = None
+    inp.prompt_values = {}
+    return inp
+
+
+async def _collect_run_outputs(block, input_data, **kwargs):
+    """Exhaust the OrchestratorBlock.run async generator, collecting outputs."""
+    outputs = []
+    async for item in block.run(input_data, **kwargs):
+        outputs.append(item)
+    return outputs
+
+
+class TestExtendedThinkingValidationRaisesInBlock:
+    """Call OrchestratorBlock.run far enough to trigger the ValueError."""
+
+    @pytest.mark.asyncio
+    async def test_non_anthropic_provider_raises_valueerror(self):
+        """EXTENDED_THINKING + openai provider raises ValueError."""
+        block = OrchestratorBlock()
+        input_data = _make_input_data(model=LlmModel.GPT4O)
+
+        with (
+            patch.object(
+                block,
+                "_create_tool_node_signatures",
+                new_callable=AsyncMock,
+                return_value=[],
+            ),
+            pytest.raises(ValueError, match="Anthropic-compatible"),
+        ):
+            await _collect_run_outputs(
+                block,
+                input_data,
+                credentials=MagicMock(),
+                graph_id="g",
+                node_id="n",
+                graph_exec_id="ge",
+                node_exec_id="ne",
+                user_id="u",
+                graph_version=1,
+                execution_context=MagicMock(),
+                execution_processor=MagicMock(),
+            )
+
+    @pytest.mark.asyncio
+    async def test_non_claude_model_with_anthropic_provider_raises(self):
+        """A model with anthropic provider but non-claude name raises ValueError."""
+        block = OrchestratorBlock()
+        fake_model = _make_model_stub("anthropic", "not-a-claude-model")
+        input_data = _make_input_data(model=fake_model)
+
+        with (
+            patch.object(
+                block,
+                "_create_tool_node_signatures",
+                new_callable=AsyncMock,
+                return_value=[],
+            ),
+            pytest.raises(ValueError, match="only supports Claude models"),
+        ):
+            await _collect_run_outputs(
+                block,
+                input_data,
+                credentials=MagicMock(),
+                graph_id="g",
+                node_id="n",
+                graph_exec_id="ge",
+                node_exec_id="ne",
+                user_id="u",
+                graph_version=1,
+                execution_context=MagicMock(),
+                execution_processor=MagicMock(),
+            )

autogpt_platform/backend/backend/blocks/xml_parser.py

@@ -44,7 +44,7 @@ class XMLParserBlock(Block):
             elif token.type == "TAG_CLOSE":
                 depth -= 1
                 if depth < 0:
-                    raise SyntaxError("Unexpected closing tag in XML input.")
+                    raise ValueError("Unexpected closing tag in XML input.")
             elif token.type in {"TEXT", "ESCAPE"}:
                 if depth == 0 and token.value:
                     raise ValueError(
@@ -53,7 +53,7 @@ class XMLParserBlock(Block):
                     )
 
         if depth != 0:
-            raise SyntaxError("Unclosed tag detected in XML input.")
+            raise ValueError("Unclosed tag detected in XML input.")
         if not root_seen:
             raise ValueError("XML must include a root element.")
 
@@ -76,4 +76,7 @@ class XMLParserBlock(Block):
         except ValueError as val_e:
             raise ValueError(f"Validation error for dict:{val_e}") from val_e
         except SyntaxError as syn_e:
-            raise SyntaxError(f"Error in input xml syntax: {syn_e}") from syn_e
+            # Raise as ValueError so the base Block.execute() wraps it as
+            # BlockExecutionError (expected user-caused failure) instead of
+            # BlockUnknownError (unexpected platform error that alerts Sentry).
+            raise ValueError(f"Error in input xml syntax: {syn_e}") from syn_e

autogpt_platform/backend/backend/copilot/baseline/service.py

@@ -9,12 +9,16 @@ shared tool registry as the SDK path.
 import asyncio
 import logging
 import uuid
-from collections.abc import AsyncGenerator
-from typing import Any
+from collections.abc import AsyncGenerator, Sequence
+from dataclasses import dataclass, field
+from functools import partial
+from typing import Any, cast
 
 import orjson
 from langfuse import propagate_attributes
+from openai.types.chat import ChatCompletionMessageParam, ChatCompletionToolParam
 
+from backend.copilot.context import set_execution_context
 from backend.copilot.model import (
     ChatMessage,
     ChatSession,
@@ -48,7 +52,17 @@ from backend.copilot.token_tracking import persist_and_record_usage
 from backend.copilot.tools import execute_tool, get_available_tools
 from backend.copilot.tracking import track_user_message
 from backend.util.exceptions import NotFoundError
-from backend.util.prompt import compress_context
+from backend.util.prompt import (
+    compress_context,
+    estimate_token_count,
+    estimate_token_count_str,
+)
+from backend.util.tool_call_loop import (
+    LLMLoopResponse,
+    LLMToolCall,
+    ToolCallResult,
+    tool_call_loop,
+)
 
 logger = logging.getLogger(__name__)
 
@@ -59,6 +73,247 @@ _background_tasks: set[asyncio.Task[Any]] = set()
 _MAX_TOOL_ROUNDS = 30
 
 
+@dataclass
+class _BaselineStreamState:
+    """Mutable state shared between the tool-call loop callbacks.
+
+    Extracted from ``stream_chat_completion_baseline`` so that the callbacks
+    can be module-level functions instead of deeply nested closures.
+    """
+
+    pending_events: list[StreamBaseResponse] = field(default_factory=list)
+    assistant_text: str = ""
+    text_block_id: str = field(default_factory=lambda: str(uuid.uuid4()))
+    text_started: bool = False
+    turn_prompt_tokens: int = 0
+    turn_completion_tokens: int = 0
+
+
+async def _baseline_llm_caller(
+    messages: list[dict[str, Any]],
+    tools: Sequence[Any],
+    *,
+    state: _BaselineStreamState,
+) -> LLMLoopResponse:
+    """Stream an OpenAI-compatible response and collect results.
+
+    Extracted from ``stream_chat_completion_baseline`` for readability.
+    """
+    state.pending_events.append(StreamStartStep())
+
+    round_text = ""
+    try:
+        client = _get_openai_client()
+        typed_messages = cast(list[ChatCompletionMessageParam], messages)
+        if tools:
+            typed_tools = cast(list[ChatCompletionToolParam], tools)
+            response = await client.chat.completions.create(
+                model=config.model,
+                messages=typed_messages,
+                tools=typed_tools,
+                stream=True,
+                stream_options={"include_usage": True},
+            )
+        else:
+            response = await client.chat.completions.create(
+                model=config.model,
+                messages=typed_messages,
+                stream=True,
+                stream_options={"include_usage": True},
+            )
+        tool_calls_by_index: dict[int, dict[str, str]] = {}
+
+        async for chunk in response:
+            if chunk.usage:
+                state.turn_prompt_tokens += chunk.usage.prompt_tokens or 0
+                state.turn_completion_tokens += chunk.usage.completion_tokens or 0
+
+            delta = chunk.choices[0].delta if chunk.choices else None
+            if not delta:
+                continue
+
+            if delta.content:
+                if not state.text_started:
+                    state.pending_events.append(StreamTextStart(id=state.text_block_id))
+                    state.text_started = True
+                round_text += delta.content
+                state.pending_events.append(
+                    StreamTextDelta(id=state.text_block_id, delta=delta.content)
+                )
+
+            if delta.tool_calls:
+                for tc in delta.tool_calls:
+                    idx = tc.index
+                    if idx not in tool_calls_by_index:
+                        tool_calls_by_index[idx] = {
+                            "id": "",
+                            "name": "",
+                            "arguments": "",
+                        }
+                    entry = tool_calls_by_index[idx]
+                    if tc.id:
+                        entry["id"] = tc.id
+                    if tc.function and tc.function.name:
+                        entry["name"] = tc.function.name
+                    if tc.function and tc.function.arguments:
+                        entry["arguments"] += tc.function.arguments
+
+        # Close text block
+        if state.text_started:
+            state.pending_events.append(StreamTextEnd(id=state.text_block_id))
+            state.text_started = False
+            state.text_block_id = str(uuid.uuid4())
+    finally:
+        # Always persist partial text so the session history stays consistent,
+        # even when the stream is interrupted by an exception.
+        state.assistant_text += round_text
+        # Always emit StreamFinishStep to match the StreamStartStep,
+        # even if an exception occurred during streaming.
+        state.pending_events.append(StreamFinishStep())
+
+    # Convert to shared format
+    llm_tool_calls = [
+        LLMToolCall(
+            id=tc["id"],
+            name=tc["name"],
+            arguments=tc["arguments"] or "{}",
+        )
+        for tc in tool_calls_by_index.values()
+    ]
+
+    return LLMLoopResponse(
+        response_text=round_text or None,
+        tool_calls=llm_tool_calls,
+        raw_response=None,  # Not needed for baseline conversation updater
+        prompt_tokens=0,  # Tracked via state accumulators
+        completion_tokens=0,
+    )
+
+
+async def _baseline_tool_executor(
+    tool_call: LLMToolCall,
+    tools: Sequence[Any],
+    *,
+    state: _BaselineStreamState,
+    user_id: str | None,
+    session: ChatSession,
+) -> ToolCallResult:
+    """Execute a tool via the copilot tool registry.
+
+    Extracted from ``stream_chat_completion_baseline`` for readability.
+    """
+    tool_call_id = tool_call.id
+    tool_name = tool_call.name
+    raw_args = tool_call.arguments or "{}"
+
+    try:
+        tool_args = orjson.loads(raw_args)
+    except orjson.JSONDecodeError as parse_err:
+        parse_error = f"Invalid JSON arguments for tool '{tool_name}': {parse_err}"
+        logger.warning("[Baseline] %s", parse_error)
+        state.pending_events.append(
+            StreamToolOutputAvailable(
+                toolCallId=tool_call_id,
+                toolName=tool_name,
+                output=parse_error,
+                success=False,
+            )
+        )
+        return ToolCallResult(
+            tool_call_id=tool_call_id,
+            tool_name=tool_name,
+            content=parse_error,
+            is_error=True,
+        )
+
+    state.pending_events.append(
+        StreamToolInputStart(toolCallId=tool_call_id, toolName=tool_name)
+    )
+    state.pending_events.append(
+        StreamToolInputAvailable(
+            toolCallId=tool_call_id,
+            toolName=tool_name,
+            input=tool_args,
+        )
+    )
+
+    try:
+        result: StreamToolOutputAvailable = await execute_tool(
+            tool_name=tool_name,
+            parameters=tool_args,
+            user_id=user_id,
+            session=session,
+            tool_call_id=tool_call_id,
+        )
+        state.pending_events.append(result)
+        tool_output = (
+            result.output if isinstance(result.output, str) else str(result.output)
+        )
+        return ToolCallResult(
+            tool_call_id=tool_call_id,
+            tool_name=tool_name,
+            content=tool_output,
+        )
+    except Exception as e:
+        error_output = f"Tool execution error: {e}"
+        logger.error(
+            "[Baseline] Tool %s failed: %s",
+            tool_name,
+            error_output,
+            exc_info=True,
+        )
+        state.pending_events.append(
+            StreamToolOutputAvailable(
+                toolCallId=tool_call_id,
+                toolName=tool_name,
+                output=error_output,
+                success=False,
+            )
+        )
+        return ToolCallResult(
+            tool_call_id=tool_call_id,
+            tool_name=tool_name,
+            content=error_output,
+            is_error=True,
+        )
+
+
+def _baseline_conversation_updater(
+    messages: list[dict[str, Any]],
+    response: LLMLoopResponse,
+    tool_results: list[ToolCallResult] | None = None,
+) -> None:
+    """Update OpenAI message list with assistant response + tool results.
+
+    Extracted from ``stream_chat_completion_baseline`` for readability.
+    """
+    if tool_results:
+        # Build assistant message with tool_calls
+        assistant_msg: dict[str, Any] = {"role": "assistant"}
+        if response.response_text:
+            assistant_msg["content"] = response.response_text
+        assistant_msg["tool_calls"] = [
+            {
+                "id": tc.id,
+                "type": "function",
+                "function": {"name": tc.name, "arguments": tc.arguments},
+            }
+            for tc in response.tool_calls
+        ]
+        messages.append(assistant_msg)
+        for tr in tool_results:
+            messages.append(
+                {
+                    "role": "tool",
+                    "tool_call_id": tr.tool_call_id,
+                    "content": tr.content,
+                }
+            )
+    else:
+        if response.response_text:
+            messages.append({"role": "assistant", "content": response.response_text})
+
+
 async def _update_title_async(
     session_id: str, message: str, user_id: str | None
 ) -> None:
@@ -203,6 +458,9 @@ async def stream_chat_completion_baseline(
 
     tools = get_available_tools()
 
+    # Propagate execution context so tool handlers can read session-level flags.
+    set_execution_context(user_id, session)
+
     yield StreamStart(messageId=message_id, sessionId=session_id)
 
     # Propagate user/session context to Langfuse so all LLM calls within
@@ -219,191 +477,32 @@ async def stream_chat_completion_baseline(
     except Exception:
         logger.warning("[Baseline] Langfuse trace context setup failed")
 
-    assistant_text = ""
-    text_block_id = str(uuid.uuid4())
-    text_started = False
-    step_open = False
-    # Token usage accumulators — populated from streaming chunks
-    turn_prompt_tokens = 0
-    turn_completion_tokens = 0
     _stream_error = False  # Track whether an error occurred during streaming
+    state = _BaselineStreamState()
+
+    # Bind extracted module-level callbacks to this request's state/session
+    # using functools.partial so they satisfy the Protocol signatures.
+    _bound_llm_caller = partial(_baseline_llm_caller, state=state)
+    _bound_tool_executor = partial(
+        _baseline_tool_executor, state=state, user_id=user_id, session=session
+    )
+
     try:
-        for _round in range(_MAX_TOOL_ROUNDS):
-            # Open a new step for each LLM round
-            yield StreamStartStep()
-            step_open = True
+        loop_result = None
+        async for loop_result in tool_call_loop(
+            messages=openai_messages,
+            tools=tools,
+            llm_call=_bound_llm_caller,
+            execute_tool=_bound_tool_executor,
+            update_conversation=_baseline_conversation_updater,
+            max_iterations=_MAX_TOOL_ROUNDS,
+        ):
+            # Drain buffered events after each iteration (real-time streaming)
+            for evt in state.pending_events:
+                yield evt
+            state.pending_events.clear()
 
-            # Stream a response from the model
-            create_kwargs: dict[str, Any] = dict(
-                model=config.model,
-                messages=openai_messages,
-                stream=True,
-                stream_options={"include_usage": True},
-            )
-            if tools:
-                create_kwargs["tools"] = tools
-            response = await _get_openai_client().chat.completions.create(**create_kwargs)  # type: ignore[arg-type]  # dynamic kwargs
-
-            # Accumulate streamed response (text + tool calls)
-            round_text = ""
-            tool_calls_by_index: dict[int, dict[str, str]] = {}
-
-            async for chunk in response:
-                # Capture token usage from the streaming chunk.
-                # OpenRouter normalises all providers into OpenAI format
-                # where prompt_tokens already includes cached tokens
-                # (unlike Anthropic's native API). Use += to sum all
-                # tool-call rounds since each API call is independent.
-                # NOTE: stream_options={"include_usage": True} is not
-                # universally supported — some providers (Mistral, Llama
-                # via OpenRouter) always return chunk.usage=None. When
-                # that happens, tokens stay 0 and the tiktoken fallback
-                # below activates. Fail-open: one round is estimated.
-                if chunk.usage:
-                    turn_prompt_tokens += chunk.usage.prompt_tokens or 0
-                    turn_completion_tokens += chunk.usage.completion_tokens or 0
-
-                delta = chunk.choices[0].delta if chunk.choices else None
-                if not delta:
-                    continue
-
-                # Text content
-                if delta.content:
-                    if not text_started:
-                        yield StreamTextStart(id=text_block_id)
-                        text_started = True
-                    round_text += delta.content
-                    yield StreamTextDelta(id=text_block_id, delta=delta.content)
-
-                # Tool call fragments (streamed incrementally)
-                if delta.tool_calls:
-                    for tc in delta.tool_calls:
-                        idx = tc.index
-                        if idx not in tool_calls_by_index:
-                            tool_calls_by_index[idx] = {
-                                "id": "",
-                                "name": "",
-                                "arguments": "",
-                            }
-                        entry = tool_calls_by_index[idx]
-                        if tc.id:
-                            entry["id"] = tc.id
-                        if tc.function and tc.function.name:
-                            entry["name"] = tc.function.name
-                        if tc.function and tc.function.arguments:
-                            entry["arguments"] += tc.function.arguments
-
-            # Close text block if we had one this round
-            if text_started:
-                yield StreamTextEnd(id=text_block_id)
-                text_started = False
-                text_block_id = str(uuid.uuid4())
-
-            # Accumulate text for session persistence
-            assistant_text += round_text
-
-            # No tool calls -> model is done
-            if not tool_calls_by_index:
-                yield StreamFinishStep()
-                step_open = False
-                break
-
-            # Close step before tool execution
-            yield StreamFinishStep()
-            step_open = False
-
-            # Append the assistant message with tool_calls to context.
-            assistant_msg: dict[str, Any] = {"role": "assistant"}
-            if round_text:
-                assistant_msg["content"] = round_text
-            assistant_msg["tool_calls"] = [
-                {
-                    "id": tc["id"],
-                    "type": "function",
-                    "function": {
-                        "name": tc["name"],
-                        "arguments": tc["arguments"] or "{}",
-                    },
-                }
-                for tc in tool_calls_by_index.values()
-            ]
-            openai_messages.append(assistant_msg)
-
-            # Execute each tool call and stream events
-            for tc in tool_calls_by_index.values():
-                tool_call_id = tc["id"]
-                tool_name = tc["name"]
-                raw_args = tc["arguments"] or "{}"
-                try:
-                    tool_args = orjson.loads(raw_args)
-                except orjson.JSONDecodeError as parse_err:
-                    parse_error = (
-                        f"Invalid JSON arguments for tool '{tool_name}': {parse_err}"
-                    )
-                    logger.warning("[Baseline] %s", parse_error)
-                    yield StreamToolOutputAvailable(
-                        toolCallId=tool_call_id,
-                        toolName=tool_name,
-                        output=parse_error,
-                        success=False,
-                    )
-                    openai_messages.append(
-                        {
-                            "role": "tool",
-                            "tool_call_id": tool_call_id,
-                            "content": parse_error,
-                        }
-                    )
-                    continue
-
-                yield StreamToolInputStart(toolCallId=tool_call_id, toolName=tool_name)
-                yield StreamToolInputAvailable(
-                    toolCallId=tool_call_id,
-                    toolName=tool_name,
-                    input=tool_args,
-                )
-
-                # Execute via shared tool registry
-                try:
-                    result: StreamToolOutputAvailable = await execute_tool(
-                        tool_name=tool_name,
-                        parameters=tool_args,
-                        user_id=user_id,
-                        session=session,
-                        tool_call_id=tool_call_id,
-                    )
-                    yield result
-                    tool_output = (
-                        result.output
-                        if isinstance(result.output, str)
-                        else str(result.output)
-                    )
-                except Exception as e:
-                    error_output = f"Tool execution error: {e}"
-                    logger.error(
-                        "[Baseline] Tool %s failed: %s",
-                        tool_name,
-                        error_output,
-                        exc_info=True,
-                    )
-                    yield StreamToolOutputAvailable(
-                        toolCallId=tool_call_id,
-                        toolName=tool_name,
-                        output=error_output,
-                        success=False,
-                    )
-                    tool_output = error_output
-
-                # Append tool result to context for next round
-                openai_messages.append(
-                    {
-                        "role": "tool",
-                        "tool_call_id": tool_call_id,
-                        "content": tool_output,
-                    }
-                )
-        else:
-            # for-loop exhausted without break -> tool-round limit hit
+        if loop_result and not loop_result.finished_naturally:
             limit_msg = (
                 f"Exceeded {_MAX_TOOL_ROUNDS} tool-call rounds "
                 "without a final response."
@@ -418,11 +517,28 @@ async def stream_chat_completion_baseline(
         _stream_error = True
         error_msg = str(e) or type(e).__name__
         logger.error("[Baseline] Streaming error: %s", error_msg, exc_info=True)
-        # Close any open text/step before emitting error
-        if text_started:
-            yield StreamTextEnd(id=text_block_id)
-        if step_open:
-            yield StreamFinishStep()
+        # Close any open text block.  The llm_caller's finally block
+        # already appended StreamFinishStep to pending_events, so we must
+        # insert StreamTextEnd *before* StreamFinishStep to preserve the
+        # protocol ordering:
+        #   StreamStartStep -> StreamTextStart -> ...deltas... ->
+        #   StreamTextEnd -> StreamFinishStep
+        # Appending (or yielding directly) would place it after
+        # StreamFinishStep, violating the protocol.
+        if state.text_started:
+            # Find the last StreamFinishStep and insert before it.
+            insert_pos = len(state.pending_events)
+            for i in range(len(state.pending_events) - 1, -1, -1):
+                if isinstance(state.pending_events[i], StreamFinishStep):
+                    insert_pos = i
+                    break
+            state.pending_events.insert(
+                insert_pos, StreamTextEnd(id=state.text_block_id)
+            )
+        # Drain pending events in correct order
+        for evt in state.pending_events:
+            yield evt
+        state.pending_events.clear()
         yield StreamError(errorText=error_msg, code="baseline_error")
         # Still persist whatever we got
     finally:
@@ -442,26 +558,21 @@ async def stream_chat_completion_baseline(
         # Skip fallback when an error occurred and no output was produced —
         # charging rate-limit tokens for completely failed requests is unfair.
         if (
-            turn_prompt_tokens == 0
-            and turn_completion_tokens == 0
-            and not (_stream_error and not assistant_text)
+            state.turn_prompt_tokens == 0
+            and state.turn_completion_tokens == 0
+            and not (_stream_error and not state.assistant_text)
         ):
-            from backend.util.prompt import (
-                estimate_token_count,
-                estimate_token_count_str,
-            )
-
-            turn_prompt_tokens = max(
+            state.turn_prompt_tokens = max(
                 estimate_token_count(openai_messages, model=config.model), 1
             )
-            turn_completion_tokens = estimate_token_count_str(
-                assistant_text, model=config.model
+            state.turn_completion_tokens = estimate_token_count_str(
+                state.assistant_text, model=config.model
             )
             logger.info(
                 "[Baseline] No streaming usage reported; estimated tokens: "
                 "prompt=%d, completion=%d",
-                turn_prompt_tokens,
-                turn_completion_tokens,
+                state.turn_prompt_tokens,
+                state.turn_completion_tokens,
             )
 
         # Persist token usage to session and record for rate limiting.
@@ -471,15 +582,15 @@ async def stream_chat_completion_baseline(
         await persist_and_record_usage(
             session=session,
             user_id=user_id,
-            prompt_tokens=turn_prompt_tokens,
-            completion_tokens=turn_completion_tokens,
+            prompt_tokens=state.turn_prompt_tokens,
+            completion_tokens=state.turn_completion_tokens,
             log_prefix="[Baseline]",
         )
 
         # Persist assistant response
-        if assistant_text:
+        if state.assistant_text:
             session.messages.append(
-                ChatMessage(role="assistant", content=assistant_text)
+                ChatMessage(role="assistant", content=state.assistant_text)
             )
         try:
             await upsert_chat_session(session)
@@ -491,11 +602,11 @@ async def stream_chat_completion_baseline(
     # aclose() — doing so raises RuntimeError on client disconnect.
     # On GeneratorExit the client is already gone, so unreachable yields
     # are harmless; on normal completion they reach the SSE stream.
-    if turn_prompt_tokens > 0 or turn_completion_tokens > 0:
+    if state.turn_prompt_tokens > 0 or state.turn_completion_tokens > 0:
         yield StreamUsage(
-            prompt_tokens=turn_prompt_tokens,
-            completion_tokens=turn_completion_tokens,
-            total_tokens=turn_prompt_tokens + turn_completion_tokens,
+            prompt_tokens=state.turn_prompt_tokens,
+            completion_tokens=state.turn_completion_tokens,
+            total_tokens=state.turn_prompt_tokens + state.turn_completion_tokens,
         )
 
     yield StreamFinish()

autogpt_platform/backend/backend/copilot/baseline/service_test.py

@@ -31,7 +31,7 @@ async def test_baseline_multi_turn(setup_test_user, test_user_id):
     if not api_key:
         return pytest.skip("OPEN_ROUTER_API_KEY is not set, skipping test")
 
-    session = await create_chat_session(test_user_id)
+    session = await create_chat_session(test_user_id, dry_run=False)
     session = await upsert_chat_session(session)
 
     # --- Turn 1: send a message with a unique keyword ---

autogpt_platform/backend/backend/copilot/config.py

@@ -20,6 +20,10 @@ class ChatConfig(BaseSettings):
         default="openai/gpt-4o-mini",
         description="Model to use for generating session titles (should be fast/cheap)",
     )
+    simulation_model: str = Field(
+        default="google/gemini-2.5-flash",
+        description="Model for dry-run block simulation (should be fast/cheap with good JSON output)",
+    )
     api_key: str | None = Field(default=None, description="OpenAI API key")
     base_url: str | None = Field(
         default=OPENROUTER_BASE_URL,
@@ -91,6 +95,20 @@ class ChatConfig(BaseSettings):
         description="Max tokens per week, resets Monday 00:00 UTC (0 = unlimited)",
     )
 
+    # Cost (in credits / cents) to reset the daily rate limit using credits.
+    # When a user hits their daily limit, they can spend this amount to reset
+    # the daily counter and keep working.  Set to 0 to disable the feature.
+    rate_limit_reset_cost: int = Field(
+        default=500,
+        ge=0,
+        description="Credit cost (in cents) for resetting the daily rate limit. 0 = disabled.",
+    )
+    max_daily_resets: int = Field(
+        default=5,
+        ge=0,
+        description="Maximum number of credit-based rate limit resets per user per day. 0 = unlimited.",
+    )
+
     # Claude Agent SDK Configuration
     use_claude_agent_sdk: bool = Field(
         default=True,
@@ -164,7 +182,7 @@ class ChatConfig(BaseSettings):
 
         Single source of truth for "will the SDK route through OpenRouter?".
         Checks the flag *and* that ``api_key`` + a valid ``base_url`` are
-        present — mirrors the fallback logic in ``_build_sdk_env``.
+        present — mirrors the fallback logic in ``build_sdk_env``.
         """
         if not self.use_openrouter:
             return False

autogpt_platform/backend/backend/copilot/context.py

@@ -149,7 +149,8 @@ def is_allowed_local_path(path: str, sdk_cwd: str | None = None) -> bool:
 
     Allowed:
     - Files under *sdk_cwd* (``/tmp/copilot-<session>/``)
-    - Files under ``~/.claude/projects/<encoded-cwd>/<uuid>/tool-results/...``.
+    - Files under ``~/.claude/projects/<encoded-cwd>/<uuid>/tool-results/...``
+      or ``tool-outputs/...``.
       The SDK nests tool-results under a conversation UUID directory;
       the UUID segment is validated with ``_UUID_RE``.
     """
@@ -174,17 +175,20 @@ def is_allowed_local_path(path: str, sdk_cwd: str | None = None) -> bool:
         # Defence-in-depth: ensure project_dir didn't escape the base.
         if not project_dir.startswith(SDK_PROJECTS_DIR + os.sep):
             return False
-        # Only allow: <encoded-cwd>/<uuid>/tool-results/<file>
+        # Only allow: <encoded-cwd>/<uuid>/<tool-dir>/<file>
         # The SDK always creates a conversation UUID directory between
-        # the project dir and tool-results/.
+        # the project dir and the tool directory.
+        # Accept both "tool-results" (SDK's persisted outputs) and
+        # "tool-outputs" (the model sometimes confuses workspace paths
+        # with filesystem paths and generates this variant).
         if resolved.startswith(project_dir + os.sep):
             relative = resolved[len(project_dir) + 1 :]
             parts = relative.split(os.sep)
-            # Require exactly: [<uuid>, "tool-results", <file>, ...]
+            # Require exactly: [<uuid>, "tool-results"|"tool-outputs", <file>, ...]
             if (
                 len(parts) >= 3
                 and _UUID_RE.match(parts[0])
-                and parts[1] == "tool-results"
+                and parts[1] in ("tool-results", "tool-outputs")
             ):
                 return True
 

autogpt_platform/backend/backend/copilot/context_test.py

@@ -134,6 +134,21 @@ def test_is_allowed_local_path_tool_results_with_uuid():
         _current_project_dir.set("")
 
 
+def test_is_allowed_local_path_tool_outputs_with_uuid():
+    """Files under <encoded-cwd>/<uuid>/tool-outputs/ are also allowed."""
+    encoded = "test-encoded-dir"
+    conv_uuid = "a1b2c3d4-e5f6-7890-abcd-ef1234567890"
+    path = os.path.join(
+        SDK_PROJECTS_DIR, encoded, conv_uuid, "tool-outputs", "output.json"
+    )
+
+    _current_project_dir.set(encoded)
+    try:
+        assert is_allowed_local_path(path, sdk_cwd=None)
+    finally:
+        _current_project_dir.set("")
+
+
 def test_is_allowed_local_path_tool_results_without_uuid_rejected():
     """Direct <encoded-cwd>/tool-results/ (no UUID) is rejected."""
     encoded = "test-encoded-dir"
@@ -159,7 +174,7 @@ def test_is_allowed_local_path_sibling_of_tool_results_is_rejected():
 
 
 def test_is_allowed_local_path_valid_uuid_wrong_segment_name_rejected():
-    """A valid UUID dir but non-'tool-results' second segment is rejected."""
+    """A valid UUID dir but non-'tool-results'/'tool-outputs' second segment is rejected."""
     encoded = "test-encoded-dir"
     uuid_str = "12345678-1234-5678-9abc-def012345678"
     path = os.path.join(

autogpt_platform/backend/backend/copilot/db.py

@@ -18,7 +18,13 @@ from prisma.types import (
 from backend.data import db
 from backend.util.json import SafeJson, sanitize_string
 
-from .model import ChatMessage, ChatSession, ChatSessionInfo
+from .model import (
+    ChatMessage,
+    ChatSession,
+    ChatSessionInfo,
+    ChatSessionMetadata,
+    invalidate_session_cache,
+)
 
 logger = logging.getLogger(__name__)
 
@@ -35,6 +41,7 @@ async def get_chat_session(session_id: str) -> ChatSession | None:
 async def create_chat_session(
     session_id: str,
     user_id: str,
+    metadata: ChatSessionMetadata | None = None,
 ) -> ChatSessionInfo:
     """Create a new chat session in the database."""
     data = ChatSessionCreateInput(
@@ -43,6 +50,7 @@ async def create_chat_session(
         credentials=SafeJson({}),
         successfulAgentRuns=SafeJson({}),
         successfulAgentSchedules=SafeJson({}),
+        metadata=SafeJson((metadata or ChatSessionMetadata()).model_dump()),
     )
     prisma_session = await PrismaChatSession.prisma().create(data=data)
     return ChatSessionInfo.from_db(prisma_session)
@@ -57,7 +65,12 @@ async def update_chat_session(
     total_completion_tokens: int | None = None,
     title: str | None = None,
 ) -> ChatSession | None:
-    """Update a chat session's metadata."""
+    """Update a chat session's mutable fields.
+
+    Note: ``metadata`` (which includes ``dry_run``) is intentionally omitted —
+    it is set once at creation time and treated as immutable for the lifetime
+    of the session.
+    """
     data: ChatSessionUpdateInput = {"updatedAt": datetime.now(UTC)}
 
     if credentials is not None:
@@ -217,6 +230,9 @@ async def add_chat_messages_batch(
                     if msg.get("function_call") is not None:
                         data["functionCall"] = SafeJson(msg["function_call"])
 
+                    if msg.get("duration_ms") is not None:
+                        data["durationMs"] = msg["duration_ms"]
+
                     messages_data.append(data)
 
                 # Run create_many and session update in parallel within transaction
@@ -359,3 +375,22 @@ async def update_tool_message_content(
             f"tool_call_id {tool_call_id}: {e}"
         )
         return False
+
+
+async def set_turn_duration(session_id: str, duration_ms: int) -> None:
+    """Set durationMs on the last assistant message in a session.
+
+    Also invalidates the Redis session cache so the next GET returns
+    the updated duration.
+    """
+    last_msg = await PrismaChatMessage.prisma().find_first(
+        where={"sessionId": session_id, "role": "assistant"},
+        order={"sequence": "desc"},
+    )
+    if last_msg:
+        await PrismaChatMessage.prisma().update(
+            where={"id": last_msg.id},
+            data={"durationMs": duration_ms},
+        )
+        # Invalidate cache so the session is re-fetched from DB with durationMs
+        await invalidate_session_cache(session_id)

autogpt_platform/backend/backend/copilot/integration_creds.py

@@ -59,6 +59,16 @@ _null_cache: TTLCache[tuple[str, str], bool] = TTLCache(
     maxsize=_CACHE_MAX_SIZE, ttl=_NULL_CACHE_TTL
 )
 
+# GitHub user identity caches (keyed by user_id only, not provider tuple).
+# Declared here so invalidate_user_provider_cache() can reference them.
+_GH_IDENTITY_CACHE_TTL = 600.0  # 10 min — profile data rarely changes
+_gh_identity_cache: TTLCache[str, dict[str, str]] = TTLCache(
+    maxsize=_CACHE_MAX_SIZE, ttl=_GH_IDENTITY_CACHE_TTL
+)
+_gh_identity_null_cache: TTLCache[str, bool] = TTLCache(
+    maxsize=_CACHE_MAX_SIZE, ttl=_NULL_CACHE_TTL
+)
+
 
 def invalidate_user_provider_cache(user_id: str, provider: str) -> None:
     """Remove the cached entry for *user_id*/*provider* from both caches.
@@ -66,11 +76,19 @@ def invalidate_user_provider_cache(user_id: str, provider: str) -> None:
     Call this after storing new credentials so that the next
     ``get_provider_token()`` call performs a fresh DB lookup instead of
     serving a stale TTL-cached result.
+
+    For GitHub specifically, also clears the git-identity caches so that
+    ``get_github_user_git_identity()`` re-fetches the user's profile on
+    the next call instead of serving stale identity data.
     """
     key = (user_id, provider)
     _token_cache.pop(key, None)
     _null_cache.pop(key, None)
 
+    if provider == "github":
+        _gh_identity_cache.pop(user_id, None)
+        _gh_identity_null_cache.pop(user_id, None)
+
 
 # Register this module's cache-bust function with the credentials manager so
 # that any create/update/delete operation immediately evicts stale cache
@@ -123,6 +141,7 @@ async def get_provider_token(user_id: str, provider: str) -> str | None:
         [c for c in creds_list if c.type == "oauth2"],
         key=lambda c: 0 if "repo" in (cast(OAuth2Credentials, c).scopes or []) else 1,
     )
+    refresh_failed = False
     for creds in oauth2_creds:
         if creds.type == "oauth2":
             try:
@@ -141,6 +160,7 @@ async def get_provider_token(user_id: str, provider: str) -> str | None:
                 # Do NOT fall back to the stale token — it is likely expired
                 # or revoked.  Returning None forces the caller to re-auth,
                 # preventing the LLM from receiving a non-functional token.
+                refresh_failed = True
                 continue
             _token_cache[cache_key] = token
             return token
@@ -152,8 +172,12 @@ async def get_provider_token(user_id: str, provider: str) -> str | None:
             _token_cache[cache_key] = token
             return token
 
-    # No credentials found — cache to avoid repeated DB hits.
-    _null_cache[cache_key] = True
+    # Only cache "not connected" when the user truly has no credentials for this
+    # provider.  If we had OAuth credentials but refresh failed (e.g. transient
+    # network error, event-loop mismatch), do NOT cache the negative result —
+    # the next call should retry the refresh instead of being blocked for 60 s.
+    if not refresh_failed:
+        _null_cache[cache_key] = True
     return None
 
 
@@ -171,3 +195,76 @@ async def get_integration_env_vars(user_id: str) -> dict[str, str]:
             for var in var_names:
                 env[var] = token
     return env
+
+
+# ---------------------------------------------------------------------------
+# GitHub user identity (for git committer env vars)
+# ---------------------------------------------------------------------------
+
+
+async def get_github_user_git_identity(user_id: str) -> dict[str, str] | None:
+    """Fetch the GitHub user's name and email for git committer env vars.
+
+    Uses the ``/user`` GitHub API endpoint with the user's stored token.
+    Returns a dict with ``GIT_AUTHOR_NAME``, ``GIT_AUTHOR_EMAIL``,
+    ``GIT_COMMITTER_NAME``, and ``GIT_COMMITTER_EMAIL`` if the user has a
+    connected GitHub account.  Returns ``None`` otherwise.
+
+    Results are cached for 10 minutes; "not connected" results are cached for
+    60 s (same as null-token cache).
+    """
+    if user_id in _gh_identity_null_cache:
+        return None
+    if cached := _gh_identity_cache.get(user_id):
+        return cached
+
+    token = await get_provider_token(user_id, "github")
+    if not token:
+        _gh_identity_null_cache[user_id] = True
+        return None
+
+    import aiohttp
+
+    try:
+        async with aiohttp.ClientSession() as session:
+            async with session.get(
+                "https://api.github.com/user",
+                headers={
+                    "Authorization": f"token {token}",
+                    "Accept": "application/vnd.github+json",
+                },
+                timeout=aiohttp.ClientTimeout(total=5),
+            ) as resp:
+                if resp.status != 200:
+                    logger.warning(
+                        "[git-identity] GitHub /user returned %s for user %s",
+                        resp.status,
+                        user_id,
+                    )
+                    return None
+                data = await resp.json()
+    except Exception as exc:
+        logger.warning(
+            "[git-identity] Failed to fetch GitHub profile for user %s: %s",
+            user_id,
+            exc,
+        )
+        return None
+
+    name = data.get("name") or data.get("login") or "AutoGPT User"
+    # GitHub may return email=null if the user has set their email to private.
+    # Fall back to the noreply address GitHub generates for every account.
+    email = data.get("email")
+    if not email:
+        gh_id = data.get("id", "")
+        login = data.get("login", "user")
+        email = f"{gh_id}+{login}@users.noreply.github.com"
+
+    identity = {
+        "GIT_AUTHOR_NAME": name,
+        "GIT_AUTHOR_EMAIL": email,
+        "GIT_COMMITTER_NAME": name,
+        "GIT_COMMITTER_EMAIL": email,
+    }
+    _gh_identity_cache[user_id] = identity
+    return identity

autogpt_platform/backend/backend/copilot/integration_creds_test.py

@@ -9,6 +9,8 @@ from backend.copilot.integration_creds import (
     _NULL_CACHE_TTL,
     _TOKEN_CACHE_TTL,
     PROVIDER_ENV_VARS,
+    _gh_identity_cache,
+    _gh_identity_null_cache,
     _null_cache,
     _token_cache,
     get_integration_env_vars,
@@ -49,9 +51,13 @@ def clear_caches():
     """Ensure clean caches before and after every test."""
     _token_cache.clear()
     _null_cache.clear()
+    _gh_identity_cache.clear()
+    _gh_identity_null_cache.clear()
     yield
     _token_cache.clear()
     _null_cache.clear()
+    _gh_identity_cache.clear()
+    _gh_identity_null_cache.clear()
 
 
 class TestInvalidateUserProviderCache:
@@ -77,6 +83,34 @@ class TestInvalidateUserProviderCache:
         invalidate_user_provider_cache(_USER, _PROVIDER)
         assert other_key in _token_cache
 
+    def test_clears_gh_identity_cache_for_github_provider(self):
+        """When provider is 'github', identity caches must also be cleared."""
+        _gh_identity_cache[_USER] = {
+            "GIT_AUTHOR_NAME": "Old Name",
+            "GIT_AUTHOR_EMAIL": "old@example.com",
+            "GIT_COMMITTER_NAME": "Old Name",
+            "GIT_COMMITTER_EMAIL": "old@example.com",
+        }
+        invalidate_user_provider_cache(_USER, "github")
+        assert _USER not in _gh_identity_cache
+
+    def test_clears_gh_identity_null_cache_for_github_provider(self):
+        """When provider is 'github', the identity null-cache must also be cleared."""
+        _gh_identity_null_cache[_USER] = True
+        invalidate_user_provider_cache(_USER, "github")
+        assert _USER not in _gh_identity_null_cache
+
+    def test_does_not_clear_gh_identity_cache_for_other_providers(self):
+        """When provider is NOT 'github', identity caches must be left alone."""
+        _gh_identity_cache[_USER] = {
+            "GIT_AUTHOR_NAME": "Some Name",
+            "GIT_AUTHOR_EMAIL": "some@example.com",
+            "GIT_COMMITTER_NAME": "Some Name",
+            "GIT_COMMITTER_EMAIL": "some@example.com",
+        }
+        invalidate_user_provider_cache(_USER, "some-other-provider")
+        assert _USER in _gh_identity_cache
+
 
 class TestGetProviderToken:
     @pytest.mark.asyncio(loop_scope="session")
@@ -129,8 +163,15 @@ class TestGetProviderToken:
         assert result == "oauth-tok"
 
     @pytest.mark.asyncio(loop_scope="session")
-    async def test_oauth2_refresh_failure_returns_none(self):
-        """On refresh failure, return None instead of caching a stale token."""
+    async def test_oauth2_refresh_failure_returns_none_without_null_cache(self):
+        """On refresh failure, return None but do NOT cache in null_cache.
+
+        The user has credentials — they just couldn't be refreshed right now
+        (e.g. transient network error or event-loop mismatch in the copilot
+        executor).  Caching a negative result would block all credential
+        lookups for 60 s even though the creds exist and may refresh fine
+        on the next attempt.
+        """
         oauth_creds = _make_oauth2_creds("stale-oauth-tok")
         mock_manager = MagicMock()
         mock_manager.store.get_creds_by_provider = AsyncMock(return_value=[oauth_creds])
@@ -141,6 +182,8 @@ class TestGetProviderToken:
 
         # Stale tokens must NOT be returned — forces re-auth.
         assert result is None
+        # Must NOT cache negative result when refresh failed — next call retries.
+        assert (_USER, _PROVIDER) not in _null_cache
 
     @pytest.mark.asyncio(loop_scope="session")
     async def test_no_credentials_caches_null_entry(self):
@@ -176,6 +219,96 @@ class TestGetProviderToken:
         assert _NULL_CACHE_TTL < _TOKEN_CACHE_TTL
 
 
+class TestThreadSafetyLocks:
+    """Bug reproduction: shared AsyncRedisKeyedMutex across threads caused
+    'Future attached to a different loop' when copilot workers accessed
+    credentials from different event loops."""
+
+    @pytest.mark.asyncio(loop_scope="session")
+    async def test_store_locks_returns_per_thread_instance(self):
+        """IntegrationCredentialsStore.locks() must return different instances
+        for different threads (via @thread_cached)."""
+        import asyncio
+        import concurrent.futures
+
+        from backend.integrations.credentials_store import IntegrationCredentialsStore
+
+        store = IntegrationCredentialsStore()
+
+        async def get_locks_id():
+            mock_redis = AsyncMock()
+            with patch(
+                "backend.integrations.credentials_store.get_redis_async",
+                return_value=mock_redis,
+            ):
+                locks = await store.locks()
+                return id(locks)
+
+        # Get locks from main thread
+        main_id = await get_locks_id()
+
+        # Get locks from a worker thread
+        def run_in_thread():
+            loop = asyncio.new_event_loop()
+            try:
+                return loop.run_until_complete(get_locks_id())
+            finally:
+                loop.close()
+
+        with concurrent.futures.ThreadPoolExecutor(max_workers=1) as pool:
+            worker_id = await asyncio.get_event_loop().run_in_executor(
+                pool, run_in_thread
+            )
+
+        assert main_id != worker_id, (
+            "Store.locks() returned the same instance across threads. "
+            "This would cause 'Future attached to a different loop' errors."
+        )
+
+    @pytest.mark.asyncio(loop_scope="session")
+    async def test_manager_delegates_to_store_locks(self):
+        """IntegrationCredentialsManager.locks() should delegate to store."""
+        from backend.integrations.creds_manager import IntegrationCredentialsManager
+
+        manager = IntegrationCredentialsManager()
+        mock_redis = AsyncMock()
+
+        with patch(
+            "backend.integrations.credentials_store.get_redis_async",
+            return_value=mock_redis,
+        ):
+            locks = await manager.locks()
+
+        # Should have gotten it from the store
+        assert locks is not None
+
+
+class TestRefreshUnlockedPath:
+    """Bug reproduction: copilot worker threads need lock-free refresh because
+    Redis-backed asyncio.Lock created on one event loop can't be used on another."""
+
+    @pytest.mark.asyncio(loop_scope="session")
+    async def test_refresh_if_needed_lock_false_skips_redis(self):
+        """refresh_if_needed(lock=False) must not touch Redis locks at all."""
+        from backend.integrations.creds_manager import IntegrationCredentialsManager
+
+        manager = IntegrationCredentialsManager()
+        creds = _make_oauth2_creds()
+
+        mock_handler = MagicMock()
+        mock_handler.needs_refresh = MagicMock(return_value=False)
+
+        with patch(
+            "backend.integrations.creds_manager._get_provider_oauth_handler",
+            new_callable=AsyncMock,
+            return_value=mock_handler,
+        ):
+            result = await manager.refresh_if_needed(_USER, creds, lock=False)
+
+        # Should return credentials without touching locks
+        assert result.id == creds.id
+
+
 class TestGetIntegrationEnvVars:
     @pytest.mark.asyncio(loop_scope="session")
     async def test_injects_all_env_vars_for_provider(self):

autogpt_platform/backend/backend/copilot/model.py

@@ -46,6 +46,16 @@ def _get_session_cache_key(session_id: str) -> str:
 # ===================== Chat data models ===================== #
 
 
+class ChatSessionMetadata(BaseModel):
+    """Typed metadata stored in the ``metadata`` JSON column of ChatSession.
+
+    Add new session-level flags here instead of adding DB columns —
+    no migration required for new fields as long as a default is provided.
+    """
+
+    dry_run: bool = False
+
+
 class ChatMessage(BaseModel):
     role: str
     content: str | None = None
@@ -54,6 +64,7 @@ class ChatMessage(BaseModel):
     refusal: str | None = None
     tool_calls: list[dict] | None = None
     function_call: dict | None = None
+    duration_ms: int | None = None
 
     @staticmethod
     def from_db(prisma_message: PrismaChatMessage) -> "ChatMessage":
@@ -66,6 +77,7 @@ class ChatMessage(BaseModel):
             refusal=prisma_message.refusal,
             tool_calls=_parse_json_field(prisma_message.toolCalls),
             function_call=_parse_json_field(prisma_message.functionCall),
+            duration_ms=prisma_message.durationMs,
         )
 
 
@@ -88,6 +100,12 @@ class ChatSessionInfo(BaseModel):
     updated_at: datetime
     successful_agent_runs: dict[str, int] = {}
     successful_agent_schedules: dict[str, int] = {}
+    metadata: ChatSessionMetadata = ChatSessionMetadata()
+
+    @property
+    def dry_run(self) -> bool:
+        """Convenience accessor for ``metadata.dry_run``."""
+        return self.metadata.dry_run
 
     @classmethod
     def from_db(cls, prisma_session: PrismaChatSession) -> Self:
@@ -101,6 +119,10 @@ class ChatSessionInfo(BaseModel):
             prisma_session.successfulAgentSchedules, default={}
         )
 
+        # Parse typed metadata from the JSON column.
+        raw_metadata = _parse_json_field(prisma_session.metadata, default={})
+        metadata = ChatSessionMetadata.model_validate(raw_metadata)
+
         # Calculate usage from token counts.
         # NOTE: Per-turn cache_read_tokens / cache_creation_tokens breakdown
         # is lost after persistence — the DB only stores aggregate prompt and
@@ -126,6 +148,7 @@ class ChatSessionInfo(BaseModel):
             updated_at=prisma_session.updatedAt,
             successful_agent_runs=successful_agent_runs,
             successful_agent_schedules=successful_agent_schedules,
+            metadata=metadata,
         )
 
 
@@ -133,7 +156,7 @@ class ChatSession(ChatSessionInfo):
     messages: list[ChatMessage]
 
     @classmethod
-    def new(cls, user_id: str) -> Self:
+    def new(cls, user_id: str, *, dry_run: bool) -> Self:
         return cls(
             session_id=str(uuid.uuid4()),
             user_id=user_id,
@@ -143,6 +166,7 @@ class ChatSession(ChatSessionInfo):
             credentials={},
             started_at=datetime.now(UTC),
             updated_at=datetime.now(UTC),
+            metadata=ChatSessionMetadata(dry_run=dry_run),
         )
 
     @classmethod
@@ -530,6 +554,7 @@ async def _save_session_to_db(
             await db.create_chat_session(
                 session_id=session.session_id,
                 user_id=session.user_id,
+                metadata=session.metadata,
             )
             existing_message_count = 0
 
@@ -607,21 +632,27 @@ async def append_and_save_message(session_id: str, message: ChatMessage) -> Chat
         return session
 
 
-async def create_chat_session(user_id: str) -> ChatSession:
+async def create_chat_session(user_id: str, *, dry_run: bool) -> ChatSession:
     """Create a new chat session and persist it.
 
+    Args:
+        user_id: The authenticated user ID.
+        dry_run: When True, run_block and run_agent tool calls in this
+            session are forced to use dry-run simulation mode.
+
     Raises:
         DatabaseError: If the database write fails. We fail fast to ensure
             callers never receive a non-persisted session that only exists
             in cache (which would be lost when the cache expires).
     """
-    session = ChatSession.new(user_id)
+    session = ChatSession.new(user_id, dry_run=dry_run)
 
     # Create in database first - fail fast if this fails
     try:
         await chat_db().create_chat_session(
             session_id=session.session_id,
             user_id=user_id,
+            metadata=session.metadata,
         )
     except Exception as e:
         logger.error(f"Failed to create session {session.session_id} in database: {e}")

autogpt_platform/backend/backend/copilot/model_test.py

@@ -46,7 +46,7 @@ messages = [
 
 @pytest.mark.asyncio(loop_scope="session")
 async def test_chatsession_serialization_deserialization():
-    s = ChatSession.new(user_id="abc123")
+    s = ChatSession.new(user_id="abc123", dry_run=False)
     s.messages = messages
     s.usage = [Usage(prompt_tokens=100, completion_tokens=200, total_tokens=300)]
     serialized = s.model_dump_json()
@@ -57,7 +57,7 @@ async def test_chatsession_serialization_deserialization():
 @pytest.mark.asyncio(loop_scope="session")
 async def test_chatsession_redis_storage(setup_test_user, test_user_id):
 
-    s = ChatSession.new(user_id=test_user_id)
+    s = ChatSession.new(user_id=test_user_id, dry_run=False)
     s.messages = messages
 
     s = await upsert_chat_session(s)
@@ -75,7 +75,7 @@ async def test_chatsession_redis_storage_user_id_mismatch(
     setup_test_user, test_user_id
 ):
 
-    s = ChatSession.new(user_id=test_user_id)
+    s = ChatSession.new(user_id=test_user_id, dry_run=False)
     s.messages = messages
     s = await upsert_chat_session(s)
 
@@ -90,7 +90,7 @@ async def test_chatsession_db_storage(setup_test_user, test_user_id):
     from backend.data.redis_client import get_redis_async
 
     # Create session with messages including assistant message
-    s = ChatSession.new(user_id=test_user_id)
+    s = ChatSession.new(user_id=test_user_id, dry_run=False)
     s.messages = messages  # Contains user, assistant, and tool messages
     assert s.session_id is not None, "Session id is not set"
     # Upsert to save to both cache and DB
@@ -241,7 +241,7 @@ _raw_tc2 = {
 
 def test_add_tool_call_appends_to_existing_assistant():
     """When the last assistant is from the current turn, tool_call is added to it."""
-    session = ChatSession.new(user_id="u")
+    session = ChatSession.new(user_id="u", dry_run=False)
     session.messages = [
         ChatMessage(role="user", content="hi"),
         ChatMessage(role="assistant", content="working on it"),
@@ -254,7 +254,7 @@ def test_add_tool_call_appends_to_existing_assistant():
 
 def test_add_tool_call_creates_assistant_when_none_exists():
     """When there's no current-turn assistant, a new one is created."""
-    session = ChatSession.new(user_id="u")
+    session = ChatSession.new(user_id="u", dry_run=False)
     session.messages = [
         ChatMessage(role="user", content="hi"),
     ]
@@ -267,7 +267,7 @@ def test_add_tool_call_creates_assistant_when_none_exists():
 
 def test_add_tool_call_does_not_cross_user_boundary():
     """A user message acts as a boundary — previous assistant is not modified."""
-    session = ChatSession.new(user_id="u")
+    session = ChatSession.new(user_id="u", dry_run=False)
     session.messages = [
         ChatMessage(role="assistant", content="old turn"),
         ChatMessage(role="user", content="new message"),
@@ -282,7 +282,7 @@ def test_add_tool_call_does_not_cross_user_boundary():
 
 def test_add_tool_call_multiple_times():
     """Multiple long-running tool calls accumulate on the same assistant."""
-    session = ChatSession.new(user_id="u")
+    session = ChatSession.new(user_id="u", dry_run=False)
     session.messages = [
         ChatMessage(role="user", content="hi"),
         ChatMessage(role="assistant", content="doing stuff"),
@@ -300,7 +300,7 @@ def test_add_tool_call_multiple_times():
 
 def test_to_openai_messages_merges_split_assistants():
     """End-to-end: session with split assistants produces valid OpenAI messages."""
-    session = ChatSession.new(user_id="u")
+    session = ChatSession.new(user_id="u", dry_run=False)
     session.messages = [
         ChatMessage(role="user", content="build agent"),
         ChatMessage(role="assistant", content="Let me build that"),
@@ -352,7 +352,7 @@ async def test_concurrent_saves_collision_detection(setup_test_user, test_user_i
     import asyncio
 
     # Create a session with initial messages
-    session = ChatSession.new(user_id=test_user_id)
+    session = ChatSession.new(user_id=test_user_id, dry_run=False)
     for i in range(3):
         session.messages.append(
             ChatMessage(

autogpt_platform/backend/backend/copilot/permissions.py

@@ -66,6 +66,7 @@ from pydantic import BaseModel, PrivateAttr
 ToolName = Literal[
     # Platform tools (must match keys in TOOL_REGISTRY)
     "add_understanding",
+    "ask_question",
     "bash_exec",
     "browser_act",
     "browser_navigate",
@@ -102,6 +103,7 @@ ToolName = Literal[
     "web_fetch",
     "write_workspace_file",
     # SDK built-ins
+    "Agent",
     "Edit",
     "Glob",
     "Grep",

autogpt_platform/backend/backend/copilot/permissions_test.py

@@ -544,6 +544,7 @@ class TestApplyToolPermissions:
 class TestSdkBuiltinToolNames:
     def test_expected_builtins_present(self):
         expected = {
+            "Agent",
             "Read",
             "Write",
             "Edit",

autogpt_platform/backend/backend/copilot/prompting.py

@@ -18,6 +18,18 @@ After `write_workspace_file`, embed the `download_url` in Markdown:
 - Image: `![chart](workspace://file_id#image/png)`
 - Video: `![recording](workspace://file_id#video/mp4)`
 
+### Handling binary/image data in tool outputs — CRITICAL
+When a tool output contains base64-encoded binary data (images, PDFs, etc.):
+1. **NEVER** try to inline or render the base64 content in your response.
+2. **Save** the data to workspace using `write_workspace_file` (pass the base64 data URI as content).
+3. **Show** the result via the workspace download URL in Markdown: `![image](workspace://file_id#image/png)`.
+
+### Passing large data between tools — CRITICAL
+When tool outputs produce large text that you need to feed into another tool:
+- **NEVER** copy-paste the full text into the next tool call argument.
+- **Save** the output to a file (workspace or local), then use `@@agptfile:` references.
+- This avoids token limits and ensures data integrity.
+
 ### File references — @@agptfile:
 Pass large file content to tools by reference: `@@agptfile:<uri>[<start>-<end>]`
 - `workspace://<file_id>` or `workspace:///<path>` — workspace files
@@ -107,6 +119,13 @@ Do not re-fetch or re-generate data you already have from prior tool calls.
 After building the file, reference it with `@@agptfile:` in other tools:
 `@@agptfile:/home/user/report.md`
 
+### Web search best practices
+- If 3 similar web searches don't return the specific data you need, conclude
+  it isn't publicly available and work with what you have.
+- Prefer fewer, well-targeted searches over many variations of the same query.
+- When spawning sub-agents for research, ensure each has a distinct
+  non-overlapping scope to avoid redundant searches.
+
 ### Sub-agent tasks
 - When using the Task tool, NEVER set `run_in_background` to true.
   All tasks must run in the foreground.
@@ -131,6 +150,11 @@ parent autopilot handles orchestration.
 # E2B-only notes — E2B has full internet access so gh CLI works there.
 # Not shown in local (bubblewrap) mode: --unshare-net blocks all network.
 _E2B_TOOL_NOTES = """
+### SDK tool-result files in E2B
+When you `Read` an SDK tool-result file, it is automatically copied into the
+sandbox so `bash_exec` can access it for further processing.
+The exact sandbox path is shown in the `[Sandbox copy available at ...]` note.
+
 ### GitHub CLI (`gh`) and git
 - If the user has connected their GitHub account, both `gh` and `git` are
   pre-authenticated — use them directly without any manual login step.
@@ -196,18 +220,22 @@ def _build_storage_supplement(
    - Files here **survive across sessions indefinitely**
 
 ### Moving files between storages
-- **{file_move_name_1_to_2}**: Copy to persistent workspace
-- **{file_move_name_2_to_1}**: Download for processing
+- **{file_move_name_1_to_2}**: `write_workspace_file(filename="output.json", source_path="/path/to/local/file")`
+- **{file_move_name_2_to_1}**: `read_workspace_file(path="tool-outputs/data.json", save_to_path="{working_dir}/data.json")`
 
 ### File persistence
 Important files (code, configs, outputs) should be saved to workspace to ensure they persist.
 
 ### SDK tool-result files
 When tool outputs are large, the SDK truncates them and saves the full output to
-a local file under `~/.claude/projects/.../tool-results/`. To read these files,
-always use `read_file` or `Read` (NOT `read_workspace_file`).
-`read_workspace_file` reads from cloud workspace storage, where SDK
-tool-results are NOT stored.
+a local file under `~/.claude/projects/.../tool-results/` (or `tool-outputs/`).
+To read these files, use `Read` — it reads from the host filesystem.
+
+### Large tool outputs saved to workspace
+When a tool output contains `<tool-output-truncated workspace_path="...">`, the
+full output is in workspace storage (NOT on the local filesystem). To access it:
+- Use `read_workspace_file(path="...", offset=..., length=50000)` for reading sections.
+- To process in the sandbox, use `read_workspace_file(path="...", save_to_path="{working_dir}/file.json")` first, then use `bash_exec` on the local copy.
 {_SHARED_TOOL_NOTES}{extra_notes}"""
 
 

autogpt_platform/backend/backend/copilot/prompting_test.py

@@ -0,0 +1,28 @@
+"""Tests for agent generation guide — verifies clarification section."""
+
+from pathlib import Path
+
+
+class TestAgentGenerationGuideContainsClarifySection:
+    """The agent generation guide must include the clarification section."""
+
+    def test_guide_includes_clarify_section(self):
+        guide_path = Path(__file__).parent / "sdk" / "agent_generation_guide.md"
+        content = guide_path.read_text(encoding="utf-8")
+        assert "Before or During Building" in content
+
+    def test_guide_mentions_find_block_for_clarification(self):
+        guide_path = Path(__file__).parent / "sdk" / "agent_generation_guide.md"
+        content = guide_path.read_text(encoding="utf-8")
+        clarify_section = content.split("Before or During Building")[1].split(
+            "### Workflow"
+        )[0]
+        assert "find_block" in clarify_section
+
+    def test_guide_mentions_ask_question_tool(self):
+        guide_path = Path(__file__).parent / "sdk" / "agent_generation_guide.md"
+        content = guide_path.read_text(encoding="utf-8")
+        clarify_section = content.split("Before or During Building")[1].split(
+            "### Workflow"
+        )[0]
+        assert "ask_question" in clarify_section

autogpt_platform/backend/backend/copilot/rate_limit.py

@@ -36,6 +36,10 @@ class CoPilotUsageStatus(BaseModel):
 
     daily: UsageWindow
     weekly: UsageWindow
+    reset_cost: int = Field(
+        default=0,
+        description="Credit cost (in cents) to reset the daily limit. 0 = feature disabled.",
+    )
 
 
 class RateLimitExceeded(Exception):
@@ -61,6 +65,7 @@ async def get_usage_status(
     user_id: str,
     daily_token_limit: int,
     weekly_token_limit: int,
+    rate_limit_reset_cost: int = 0,
 ) -> CoPilotUsageStatus:
     """Get current usage status for a user.
 
@@ -68,6 +73,7 @@ async def get_usage_status(
         user_id: The user's ID.
         daily_token_limit: Max tokens per day (0 = unlimited).
         weekly_token_limit: Max tokens per week (0 = unlimited).
+        rate_limit_reset_cost: Credit cost (cents) to reset daily limit (0 = disabled).
 
     Returns:
         CoPilotUsageStatus with current usage and limits.
@@ -97,6 +103,7 @@ async def get_usage_status(
             limit=weekly_token_limit,
             resets_at=_weekly_reset_time(now=now),
         ),
+        reset_cost=rate_limit_reset_cost,
     )
 
 
@@ -141,6 +148,111 @@ async def check_rate_limit(
         raise RateLimitExceeded("weekly", _weekly_reset_time(now=now))
 
 
+async def reset_daily_usage(user_id: str, daily_token_limit: int = 0) -> bool:
+    """Reset a user's daily token usage counter in Redis.
+
+    Called after a user pays credits to extend their daily limit.
+    Also reduces the weekly usage counter by ``daily_token_limit`` tokens
+    (clamped to 0) so the user effectively gets one extra day's worth of
+    weekly capacity.
+
+    Args:
+        user_id: The user's ID.
+        daily_token_limit: The configured daily token limit. When positive,
+            the weekly counter is reduced by this amount.
+
+    Returns False if Redis is unavailable so the caller can handle
+    compensation (fail-closed for billed operations, unlike the read-only
+    rate-limit checks which fail-open).
+    """
+    now = datetime.now(UTC)
+    try:
+        redis = await get_redis_async()
+
+        # Use a MULTI/EXEC transaction so that DELETE (daily) and DECRBY
+        # (weekly) either both execute or neither does.  This prevents the
+        # scenario where the daily counter is cleared but the weekly
+        # counter is not decremented — which would let the caller refund
+        # credits even though the daily limit was already reset.
+        d_key = _daily_key(user_id, now=now)
+        w_key = _weekly_key(user_id, now=now) if daily_token_limit > 0 else None
+
+        pipe = redis.pipeline(transaction=True)
+        pipe.delete(d_key)
+        if w_key is not None:
+            pipe.decrby(w_key, daily_token_limit)
+        results = await pipe.execute()
+
+        # Clamp negative weekly counter to 0 (best-effort; not critical).
+        if w_key is not None:
+            new_val = results[1]  # DECRBY result
+            if new_val < 0:
+                await redis.set(w_key, 0, keepttl=True)
+
+        logger.info("Reset daily usage for user %s", user_id[:8])
+        return True
+    except (RedisError, ConnectionError, OSError):
+        logger.warning("Redis unavailable for resetting daily usage")
+        return False
+
+
+_RESET_LOCK_PREFIX = "copilot:reset_lock"
+_RESET_COUNT_PREFIX = "copilot:reset_count"
+
+
+async def acquire_reset_lock(user_id: str, ttl_seconds: int = 10) -> bool:
+    """Acquire a short-lived lock to serialize rate limit resets per user."""
+    try:
+        redis = await get_redis_async()
+        key = f"{_RESET_LOCK_PREFIX}:{user_id}"
+        return bool(await redis.set(key, "1", nx=True, ex=ttl_seconds))
+    except (RedisError, ConnectionError, OSError) as exc:
+        logger.warning("Redis unavailable for reset lock, rejecting reset: %s", exc)
+        return False
+
+
+async def release_reset_lock(user_id: str) -> None:
+    """Release the per-user reset lock."""
+    try:
+        redis = await get_redis_async()
+        await redis.delete(f"{_RESET_LOCK_PREFIX}:{user_id}")
+    except (RedisError, ConnectionError, OSError):
+        pass  # Lock will expire via TTL
+
+
+async def get_daily_reset_count(user_id: str) -> int | None:
+    """Get how many times the user has reset today.
+
+    Returns None when Redis is unavailable so callers can fail-closed
+    for billed operations (as opposed to failing open for read-only
+    rate-limit checks).
+    """
+    now = datetime.now(UTC)
+    try:
+        redis = await get_redis_async()
+        key = f"{_RESET_COUNT_PREFIX}:{user_id}:{now.strftime('%Y-%m-%d')}"
+        val = await redis.get(key)
+        return int(val or 0)
+    except (RedisError, ConnectionError, OSError):
+        logger.warning("Redis unavailable for reading daily reset count")
+        return None
+
+
+async def increment_daily_reset_count(user_id: str) -> None:
+    """Increment and track how many resets this user has done today."""
+    now = datetime.now(UTC)
+    try:
+        redis = await get_redis_async()
+        key = f"{_RESET_COUNT_PREFIX}:{user_id}:{now.strftime('%Y-%m-%d')}"
+        pipe = redis.pipeline(transaction=True)
+        pipe.incr(key)
+        seconds_until_reset = int((_daily_reset_time(now=now) - now).total_seconds())
+        pipe.expire(key, max(seconds_until_reset, 1))
+        await pipe.execute()
+    except (RedisError, ConnectionError, OSError):
+        logger.warning("Redis unavailable for tracking reset count")
+
+
 async def record_token_usage(
     user_id: str,
     prompt_tokens: int,
@@ -231,6 +343,67 @@ async def record_token_usage(
         )
 
 
+async def get_global_rate_limits(
+    user_id: str,
+    config_daily: int,
+    config_weekly: int,
+) -> tuple[int, int]:
+    """Resolve global rate limits from LaunchDarkly, falling back to config.
+
+    Args:
+        user_id: User ID for LD flag evaluation context.
+        config_daily: Fallback daily limit from ChatConfig.
+        config_weekly: Fallback weekly limit from ChatConfig.
+
+    Returns:
+        (daily_token_limit, weekly_token_limit) tuple.
+    """
+    # Lazy import to avoid circular dependency:
+    # rate_limit -> feature_flag -> settings -> ... -> rate_limit
+    from backend.util.feature_flag import Flag, get_feature_flag_value
+
+    daily_raw = await get_feature_flag_value(
+        Flag.COPILOT_DAILY_TOKEN_LIMIT.value, user_id, config_daily
+    )
+    weekly_raw = await get_feature_flag_value(
+        Flag.COPILOT_WEEKLY_TOKEN_LIMIT.value, user_id, config_weekly
+    )
+    try:
+        daily = max(0, int(daily_raw))
+    except (TypeError, ValueError):
+        logger.warning("Invalid LD value for daily token limit: %r", daily_raw)
+        daily = config_daily
+    try:
+        weekly = max(0, int(weekly_raw))
+    except (TypeError, ValueError):
+        logger.warning("Invalid LD value for weekly token limit: %r", weekly_raw)
+        weekly = config_weekly
+    return daily, weekly
+
+
+async def reset_user_usage(user_id: str, *, reset_weekly: bool = False) -> None:
+    """Reset a user's usage counters.
+
+    Always deletes the daily Redis key.  When *reset_weekly* is ``True``,
+    the weekly key is deleted as well.
+
+    Unlike read paths (``get_usage_status``, ``check_rate_limit``) which
+    fail-open on Redis errors, resets intentionally re-raise so the caller
+    knows the operation did not succeed.  A silent failure here would leave
+    the admin believing the counters were zeroed when they were not.
+    """
+    now = datetime.now(UTC)
+    keys_to_delete = [_daily_key(user_id, now=now)]
+    if reset_weekly:
+        keys_to_delete.append(_weekly_key(user_id, now=now))
+    try:
+        redis = await get_redis_async()
+        await redis.delete(*keys_to_delete)
+    except (RedisError, ConnectionError, OSError):
+        logger.warning("Redis unavailable for resetting user usage")
+        raise
+
+
 # ---------------------------------------------------------------------------
 # Private helpers
 # ---------------------------------------------------------------------------

autogpt_platform/backend/backend/copilot/rate_limit_test.py

@@ -12,6 +12,7 @@ from .rate_limit import (
     check_rate_limit,
     get_usage_status,
     record_token_usage,
+    reset_daily_usage,
 )
 
 _USER = "test-user-rl"
@@ -332,3 +333,91 @@ class TestRecordTokenUsage:
         ):
             # Should not raise — fail-open
             await record_token_usage(_USER, prompt_tokens=100, completion_tokens=50)
+
+
+# ---------------------------------------------------------------------------
+# reset_daily_usage
+# ---------------------------------------------------------------------------
+
+
+class TestResetDailyUsage:
+    @staticmethod
+    def _make_pipeline_mock(decrby_result: int = 0) -> MagicMock:
+        """Create a pipeline mock that returns [delete_result, decrby_result]."""
+        pipe = MagicMock()
+        pipe.execute = AsyncMock(return_value=[1, decrby_result])
+        return pipe
+
+    @pytest.mark.asyncio
+    async def test_deletes_daily_key(self):
+        mock_pipe = self._make_pipeline_mock(decrby_result=0)
+        mock_redis = AsyncMock()
+        mock_redis.pipeline = lambda **_kw: mock_pipe
+
+        with patch(
+            "backend.copilot.rate_limit.get_redis_async",
+            return_value=mock_redis,
+        ):
+            result = await reset_daily_usage(_USER, daily_token_limit=10000)
+
+        assert result is True
+        mock_pipe.delete.assert_called_once()
+
+    @pytest.mark.asyncio
+    async def test_reduces_weekly_usage_via_decrby(self):
+        """Weekly counter should be reduced via DECRBY in the pipeline."""
+        mock_pipe = self._make_pipeline_mock(decrby_result=35000)
+        mock_redis = AsyncMock()
+        mock_redis.pipeline = lambda **_kw: mock_pipe
+
+        with patch(
+            "backend.copilot.rate_limit.get_redis_async",
+            return_value=mock_redis,
+        ):
+            await reset_daily_usage(_USER, daily_token_limit=10000)
+
+        mock_pipe.decrby.assert_called_once()
+        mock_redis.set.assert_not_called()  # 35000 > 0, no clamp needed
+
+    @pytest.mark.asyncio
+    async def test_clamps_negative_weekly_to_zero(self):
+        """If DECRBY goes negative, SET to 0 (outside the pipeline)."""
+        mock_pipe = self._make_pipeline_mock(decrby_result=-5000)
+        mock_redis = AsyncMock()
+        mock_redis.pipeline = lambda **_kw: mock_pipe
+
+        with patch(
+            "backend.copilot.rate_limit.get_redis_async",
+            return_value=mock_redis,
+        ):
+            await reset_daily_usage(_USER, daily_token_limit=10000)
+
+        mock_pipe.decrby.assert_called_once()
+        mock_redis.set.assert_called_once()
+
+    @pytest.mark.asyncio
+    async def test_no_weekly_reduction_when_daily_limit_zero(self):
+        """When daily_token_limit is 0, weekly counter should not be touched."""
+        mock_pipe = self._make_pipeline_mock()
+        mock_pipe.execute = AsyncMock(return_value=[1])  # only delete result
+        mock_redis = AsyncMock()
+        mock_redis.pipeline = lambda **_kw: mock_pipe
+
+        with patch(
+            "backend.copilot.rate_limit.get_redis_async",
+            return_value=mock_redis,
+        ):
+            await reset_daily_usage(_USER, daily_token_limit=0)
+
+        mock_pipe.delete.assert_called_once()
+        mock_pipe.decrby.assert_not_called()
+
+    @pytest.mark.asyncio
+    async def test_returns_false_when_redis_unavailable(self):
+        with patch(
+            "backend.copilot.rate_limit.get_redis_async",
+            side_effect=ConnectionError("Redis down"),
+        ):
+            result = await reset_daily_usage(_USER, daily_token_limit=10000)
+
+        assert result is False

autogpt_platform/backend/backend/copilot/reset_usage_test.py

@@ -0,0 +1,330 @@
+"""Unit tests for the POST /usage/reset endpoint."""
+
+from __future__ import annotations
+
+from datetime import UTC, datetime, timedelta
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+from fastapi import HTTPException
+
+from backend.api.features.chat.routes import reset_copilot_usage
+from backend.copilot.rate_limit import CoPilotUsageStatus, UsageWindow
+from backend.util.exceptions import InsufficientBalanceError
+
+
+# Minimal config mock matching ChatConfig fields used by the endpoint.
+def _make_config(
+    rate_limit_reset_cost: int = 500,
+    daily_token_limit: int = 2_500_000,
+    weekly_token_limit: int = 12_500_000,
+    max_daily_resets: int = 5,
+):
+    cfg = MagicMock()
+    cfg.rate_limit_reset_cost = rate_limit_reset_cost
+    cfg.daily_token_limit = daily_token_limit
+    cfg.weekly_token_limit = weekly_token_limit
+    cfg.max_daily_resets = max_daily_resets
+    return cfg
+
+
+def _usage(daily_used: int = 3_000_000, daily_limit: int = 2_500_000):
+    return CoPilotUsageStatus(
+        daily=UsageWindow(
+            used=daily_used,
+            limit=daily_limit,
+            resets_at=datetime.now(UTC) + timedelta(hours=6),
+        ),
+        weekly=UsageWindow(
+            used=5_000_000,
+            limit=12_500_000,
+            resets_at=datetime.now(UTC) + timedelta(days=3),
+        ),
+    )
+
+
+_MODULE = "backend.api.features.chat.routes"
+
+
+def _mock_settings(enable_credit: bool = True):
+    """Return a mock Settings object with the given enable_credit flag."""
+    mock = MagicMock()
+    mock.config.enable_credit = enable_credit
+    return mock
+
+
+@pytest.mark.asyncio
+class TestResetCopilotUsage:
+    async def test_feature_disabled_returns_400(self):
+        """When rate_limit_reset_cost=0, endpoint returns 400."""
+
+        with patch(f"{_MODULE}.config", _make_config(rate_limit_reset_cost=0)):
+            with pytest.raises(HTTPException) as exc_info:
+                await reset_copilot_usage(user_id="user-1")
+            assert exc_info.value.status_code == 400
+            assert "not available" in exc_info.value.detail
+
+    async def test_no_daily_limit_returns_400(self):
+        """When daily_token_limit=0 (unlimited), endpoint returns 400."""
+
+        with (
+            patch(f"{_MODULE}.config", _make_config(daily_token_limit=0)),
+            patch(f"{_MODULE}.settings", _mock_settings()),
+            patch(
+                f"{_MODULE}.get_global_rate_limits",
+                AsyncMock(return_value=(0, 12_500_000)),
+            ),
+        ):
+            with pytest.raises(HTTPException) as exc_info:
+                await reset_copilot_usage(user_id="user-1")
+            assert exc_info.value.status_code == 400
+            assert "nothing to reset" in exc_info.value.detail.lower()
+
+    async def test_not_at_limit_returns_400(self):
+        """When user hasn't hit their daily limit, returns 400."""
+
+        cfg = _make_config()
+        with (
+            patch(f"{_MODULE}.config", cfg),
+            patch(f"{_MODULE}.settings", _mock_settings()),
+            patch(
+                f"{_MODULE}.get_global_rate_limits",
+                AsyncMock(return_value=(2_500_000, 12_500_000)),
+            ),
+            patch(f"{_MODULE}.get_daily_reset_count", AsyncMock(return_value=0)),
+            patch(f"{_MODULE}.acquire_reset_lock", AsyncMock(return_value=True)),
+            patch(f"{_MODULE}.release_reset_lock", AsyncMock()) as mock_release,
+            patch(
+                f"{_MODULE}.get_usage_status",
+                AsyncMock(return_value=_usage(daily_used=1_000_000)),
+            ),
+        ):
+            with pytest.raises(HTTPException) as exc_info:
+                await reset_copilot_usage(user_id="user-1")
+            assert exc_info.value.status_code == 400
+            assert "not reached" in exc_info.value.detail
+            mock_release.assert_awaited_once()
+
+    async def test_insufficient_credits_returns_402(self):
+        """When user doesn't have enough credits, returns 402."""
+
+        mock_credit_model = AsyncMock()
+        mock_credit_model.spend_credits.side_effect = InsufficientBalanceError(
+            message="Insufficient balance",
+            user_id="user-1",
+            balance=50,
+            amount=200,
+        )
+
+        cfg = _make_config()
+        with (
+            patch(f"{_MODULE}.config", cfg),
+            patch(f"{_MODULE}.settings", _mock_settings()),
+            patch(
+                f"{_MODULE}.get_global_rate_limits",
+                AsyncMock(return_value=(2_500_000, 12_500_000)),
+            ),
+            patch(f"{_MODULE}.get_daily_reset_count", AsyncMock(return_value=0)),
+            patch(f"{_MODULE}.acquire_reset_lock", AsyncMock(return_value=True)),
+            patch(f"{_MODULE}.release_reset_lock", AsyncMock()) as mock_release,
+            patch(
+                f"{_MODULE}.get_usage_status",
+                AsyncMock(return_value=_usage()),
+            ),
+            patch(
+                f"{_MODULE}.get_user_credit_model",
+                AsyncMock(return_value=mock_credit_model),
+            ),
+        ):
+            with pytest.raises(HTTPException) as exc_info:
+                await reset_copilot_usage(user_id="user-1")
+            assert exc_info.value.status_code == 402
+            mock_release.assert_awaited_once()
+
+    async def test_happy_path(self):
+        """Successful reset: charges credits, resets usage, returns response."""
+
+        mock_credit_model = AsyncMock()
+        mock_credit_model.spend_credits.return_value = 1500  # remaining balance
+
+        cfg = _make_config()
+        updated_usage = _usage(daily_used=0)
+
+        with (
+            patch(f"{_MODULE}.config", cfg),
+            patch(f"{_MODULE}.settings", _mock_settings()),
+            patch(
+                f"{_MODULE}.get_global_rate_limits",
+                AsyncMock(return_value=(2_500_000, 12_500_000)),
+            ),
+            patch(f"{_MODULE}.get_daily_reset_count", AsyncMock(return_value=0)),
+            patch(f"{_MODULE}.acquire_reset_lock", AsyncMock(return_value=True)),
+            patch(f"{_MODULE}.release_reset_lock", AsyncMock()),
+            patch(
+                f"{_MODULE}.get_usage_status",
+                AsyncMock(side_effect=[_usage(), updated_usage]),
+            ),
+            patch(
+                f"{_MODULE}.get_user_credit_model",
+                AsyncMock(return_value=mock_credit_model),
+            ),
+            patch(
+                f"{_MODULE}.reset_daily_usage", AsyncMock(return_value=True)
+            ) as mock_reset,
+            patch(f"{_MODULE}.increment_daily_reset_count", AsyncMock()) as mock_incr,
+        ):
+            result = await reset_copilot_usage(user_id="user-1")
+            assert result.success is True
+            assert result.credits_charged == 500
+            assert result.remaining_balance == 1500
+            mock_reset.assert_awaited_once()
+            mock_incr.assert_awaited_once()
+
+    async def test_max_daily_resets_exceeded(self):
+        """When user has exhausted daily resets, returns 429."""
+
+        cfg = _make_config(max_daily_resets=3)
+        with (
+            patch(f"{_MODULE}.config", cfg),
+            patch(f"{_MODULE}.settings", _mock_settings()),
+            patch(
+                f"{_MODULE}.get_global_rate_limits",
+                AsyncMock(return_value=(2_500_000, 12_500_000)),
+            ),
+            patch(f"{_MODULE}.get_daily_reset_count", AsyncMock(return_value=3)),
+        ):
+            with pytest.raises(HTTPException) as exc_info:
+                await reset_copilot_usage(user_id="user-1")
+            assert exc_info.value.status_code == 429
+
+    async def test_credit_system_disabled_returns_400(self):
+        """When enable_credit=False, endpoint returns 400."""
+
+        with (
+            patch(f"{_MODULE}.config", _make_config()),
+            patch(f"{_MODULE}.settings", _mock_settings(enable_credit=False)),
+        ):
+            with pytest.raises(HTTPException) as exc_info:
+                await reset_copilot_usage(user_id="user-1")
+            assert exc_info.value.status_code == 400
+            assert "credit system is disabled" in exc_info.value.detail.lower()
+
+    async def test_weekly_limit_exhausted_returns_400(self):
+        """When the weekly limit is also exhausted, resetting daily won't help."""
+
+        cfg = _make_config()
+        weekly_exhausted = CoPilotUsageStatus(
+            daily=UsageWindow(
+                used=3_000_000,
+                limit=2_500_000,
+                resets_at=datetime.now(UTC) + timedelta(hours=6),
+            ),
+            weekly=UsageWindow(
+                used=12_500_000,
+                limit=12_500_000,
+                resets_at=datetime.now(UTC) + timedelta(days=3),
+            ),
+        )
+        with (
+            patch(f"{_MODULE}.config", cfg),
+            patch(f"{_MODULE}.settings", _mock_settings()),
+            patch(
+                f"{_MODULE}.get_global_rate_limits",
+                AsyncMock(return_value=(2_500_000, 12_500_000)),
+            ),
+            patch(f"{_MODULE}.get_daily_reset_count", AsyncMock(return_value=0)),
+            patch(f"{_MODULE}.acquire_reset_lock", AsyncMock(return_value=True)),
+            patch(f"{_MODULE}.release_reset_lock", AsyncMock()) as mock_release,
+            patch(
+                f"{_MODULE}.get_usage_status",
+                AsyncMock(return_value=weekly_exhausted),
+            ),
+        ):
+            with pytest.raises(HTTPException) as exc_info:
+                await reset_copilot_usage(user_id="user-1")
+            assert exc_info.value.status_code == 400
+            assert "weekly" in exc_info.value.detail.lower()
+            mock_release.assert_awaited_once()
+
+    async def test_redis_failure_for_reset_count_returns_503(self):
+        """When Redis is unavailable for get_daily_reset_count, returns 503."""
+
+        with (
+            patch(f"{_MODULE}.config", _make_config()),
+            patch(f"{_MODULE}.settings", _mock_settings()),
+            patch(
+                f"{_MODULE}.get_global_rate_limits",
+                AsyncMock(return_value=(2_500_000, 12_500_000)),
+            ),
+            patch(f"{_MODULE}.get_daily_reset_count", AsyncMock(return_value=None)),
+        ):
+            with pytest.raises(HTTPException) as exc_info:
+                await reset_copilot_usage(user_id="user-1")
+            assert exc_info.value.status_code == 503
+            assert "verify" in exc_info.value.detail.lower()
+
+    async def test_redis_reset_failure_refunds_credits(self):
+        """When reset_daily_usage fails, credits are refunded and 503 returned."""
+
+        mock_credit_model = AsyncMock()
+        mock_credit_model.spend_credits.return_value = 1500
+
+        cfg = _make_config()
+        with (
+            patch(f"{_MODULE}.config", cfg),
+            patch(f"{_MODULE}.settings", _mock_settings()),
+            patch(
+                f"{_MODULE}.get_global_rate_limits",
+                AsyncMock(return_value=(2_500_000, 12_500_000)),
+            ),
+            patch(f"{_MODULE}.get_daily_reset_count", AsyncMock(return_value=0)),
+            patch(f"{_MODULE}.acquire_reset_lock", AsyncMock(return_value=True)),
+            patch(f"{_MODULE}.release_reset_lock", AsyncMock()),
+            patch(
+                f"{_MODULE}.get_usage_status",
+                AsyncMock(return_value=_usage()),
+            ),
+            patch(
+                f"{_MODULE}.get_user_credit_model",
+                AsyncMock(return_value=mock_credit_model),
+            ),
+            patch(f"{_MODULE}.reset_daily_usage", AsyncMock(return_value=False)),
+        ):
+            with pytest.raises(HTTPException) as exc_info:
+                await reset_copilot_usage(user_id="user-1")
+            assert exc_info.value.status_code == 503
+            assert "not been charged" in exc_info.value.detail
+            mock_credit_model.top_up_credits.assert_awaited_once()
+
+    async def test_redis_reset_failure_refund_also_fails(self):
+        """When both reset and refund fail, error message reflects the truth."""
+
+        mock_credit_model = AsyncMock()
+        mock_credit_model.spend_credits.return_value = 1500
+        mock_credit_model.top_up_credits.side_effect = RuntimeError("db down")
+
+        cfg = _make_config()
+        with (
+            patch(f"{_MODULE}.config", cfg),
+            patch(f"{_MODULE}.settings", _mock_settings()),
+            patch(
+                f"{_MODULE}.get_global_rate_limits",
+                AsyncMock(return_value=(2_500_000, 12_500_000)),
+            ),
+            patch(f"{_MODULE}.get_daily_reset_count", AsyncMock(return_value=0)),
+            patch(f"{_MODULE}.acquire_reset_lock", AsyncMock(return_value=True)),
+            patch(f"{_MODULE}.release_reset_lock", AsyncMock()),
+            patch(
+                f"{_MODULE}.get_usage_status",
+                AsyncMock(return_value=_usage()),
+            ),
+            patch(
+                f"{_MODULE}.get_user_credit_model",
+                AsyncMock(return_value=mock_credit_model),
+            ),
+            patch(f"{_MODULE}.reset_daily_usage", AsyncMock(return_value=False)),
+        ):
+            with pytest.raises(HTTPException) as exc_info:
+                await reset_copilot_usage(user_id="user-1")
+            assert exc_info.value.status_code == 503
+            assert "contact support" in exc_info.value.detail.lower()

autogpt_platform/backend/backend/copilot/sdk/agent_generation_guide.md

@@ -3,6 +3,29 @@
 You can create, edit, and customize agents directly. You ARE the brain —
 generate the agent JSON yourself using block schemas, then validate and save.
 
+### Clarifying — Before or During Building
+
+Use `ask_question` whenever the user's intent is ambiguous — whether
+that's before starting or midway through the workflow. Common moments:
+
+- **Before building**: output format, delivery channel, data source, or
+  trigger is unspecified.
+- **During block discovery**: multiple blocks could fit and the user
+  should choose.
+- **During JSON generation**: a wiring decision depends on user
+  preference.
+
+Steps:
+1. Call `find_block` (or another discovery tool) to learn what the
+   platform actually supports for the ambiguous dimension.
+2. Call `ask_question` with a concrete question listing the discovered
+   options (e.g. "The platform supports Gmail, Slack, and Google Docs —
+   which should the agent use for delivery?").
+3. **Wait for the user's answer** before continuing.
+
+**Skip this** when the goal already specifies all dimensions (e.g.
+"scrape prices from Amazon and email me daily").
+
 ### Workflow for Creating/Editing Agents
 
 1. **Discover blocks**: Call `find_block(query, include_schemas=true)` to
@@ -67,9 +90,17 @@ These define the agent's interface — what it accepts and what it produces.
 **AgentInputBlock** (ID: `c0a8e994-ebf1-4a9c-a4d8-89d09c86741b`):
 - Defines a user-facing input field on the agent
 - Required `input_default` fields: `name` (str), `value` (default: null)
-- Optional: `title`, `description`, `placeholder_values` (for dropdowns)
+- Optional: `title`, `description`
 - Output: `result` — the user-provided value at runtime
 - Create one AgentInputBlock per distinct input the agent needs
+- For dropdown/select inputs, use **AgentDropdownInputBlock** instead (see below)
+
+**AgentDropdownInputBlock** (ID: `655d6fdf-a334-421c-b733-520549c07cd1`):
+- Specialized input block that presents a dropdown/select to the user
+- Required `input_default` fields: `name` (str)
+- Optional: `options` (list of dropdown values; when omitted/empty, input behaves as free-text), `title`, `description`, `value` (default selection)
+- Output: `result` — the user-selected value at runtime
+- Use this instead of AgentInputBlock when the user should pick from a fixed set of options
 
 **AgentOutputBlock** (ID: `363ae599-353e-4804-937e-b2ee3cef3da4`):
 - Defines a user-facing output displayed after the agent runs
@@ -222,6 +253,17 @@ real API calls, credentials, or credits:
 3. **Iterate**: If the dry run reveals wiring issues or missing inputs, fix
    the agent JSON and re-save before suggesting a real execution.
 
+**Special block behaviour in dry-run mode:**
+- **OrchestratorBlock** and **AgentExecutorBlock** execute for real so the
+  orchestrator can make LLM calls and agent executors can spawn child graphs.
+  Their downstream tool blocks and child-graph blocks are still simulated.
+  Note: real LLM inference calls are made (consuming API quota), even though
+  platform credits are not charged. Agent-mode iterations are capped at 1 in
+  dry-run to keep it fast.
+- **MCPToolBlock** is simulated using the selected tool's name and JSON Schema
+  so the LLM can produce a realistic mock response without connecting to the
+  MCP server.
+
 ### Example: Simple AI Text Processor
 
 A minimal agent with input, processing, and output:

autogpt_platform/backend/backend/copilot/sdk/compaction_test.py

@@ -25,7 +25,7 @@ from backend.copilot.sdk.compaction import (
 
 
 def _make_session() -> ChatSession:
-    return ChatSession.new(user_id="test-user")
+    return ChatSession.new(user_id="test-user", dry_run=False)
 
 
 # ---------------------------------------------------------------------------

autogpt_platform/backend/backend/copilot/sdk/conftest.py

@@ -2,14 +2,30 @@
 
 from __future__ import annotations
 
+from collections.abc import AsyncIterator
 from unittest.mock import patch
 from uuid import uuid4
 
 import pytest
+import pytest_asyncio
 
 from backend.util import json
 
 
+@pytest_asyncio.fixture(scope="session", loop_scope="session", name="server")
+async def _server_noop() -> None:
+    """No-op server stub — SDK tests don't need the full backend."""
+    return None
+
+
+@pytest_asyncio.fixture(
+    scope="session", loop_scope="session", autouse=True, name="graph_cleanup"
+)
+async def _graph_cleanup_noop() -> AsyncIterator[None]:
+    """No-op graph cleanup stub."""
+    yield
+
+
 @pytest.fixture()
 def mock_chat_config():
     """Mock ChatConfig so compact_transcript tests skip real config lookup."""
@@ -25,24 +41,64 @@ def build_test_transcript(pairs: list[tuple[str, str]]) -> str:
 
     Use this helper in any copilot SDK test that needs a well-formed
     transcript without hitting the real storage layer.
+
+    Delegates to ``build_structured_transcript`` — plain content strings
+    are automatically wrapped in ``[{"type": "text", "text": ...}]`` for
+    assistant messages.
+    """
+    # Cast widening: tuple[str, str] is structurally compatible with
+    # tuple[str, str | list[dict]] but list invariance requires explicit
+    # annotation.
+    widened: list[tuple[str, str | list[dict]]] = list(pairs)
+    return build_structured_transcript(widened)
+
+
+def build_structured_transcript(
+    entries: list[tuple[str, str | list[dict]]],
+) -> str:
+    """Build a JSONL transcript with structured content blocks.
+
+    Each entry is (role, content) where content is either a plain string
+    (for user messages) or a list of content block dicts (for assistant
+    messages with thinking/tool_use/text blocks).
+
+    Example::
+
+        build_structured_transcript([
+            ("user", "Hello"),
+            ("assistant", [
+                {"type": "thinking", "thinking": "...", "signature": "sig1"},
+                {"type": "text", "text": "Hi there"},
+            ]),
+        ])
     """
     lines: list[str] = []
     last_uuid: str | None = None
-    for role, content in pairs:
+    for role, content in entries:
         uid = str(uuid4())
         entry_type = "assistant" if role == "assistant" else "user"
-        msg: dict = {"role": role, "content": content}
-        if role == "assistant":
-            msg.update(
-                {
-                    "model": "",
-                    "id": f"msg_{uid[:8]}",
-                    "type": "message",
-                    "content": [{"type": "text", "text": content}],
-                    "stop_reason": "end_turn",
-                    "stop_sequence": None,
-                }
-            )
+        if role == "assistant" and isinstance(content, list):
+            msg: dict = {
+                "role": "assistant",
+                "model": "claude-test",
+                "id": f"msg_{uid[:8]}",
+                "type": "message",
+                "content": content,
+                "stop_reason": "end_turn",
+                "stop_sequence": None,
+            }
+        elif role == "assistant":
+            msg = {
+                "role": "assistant",
+                "model": "claude-test",
+                "id": f"msg_{uid[:8]}",
+                "type": "message",
+                "content": [{"type": "text", "text": content}],
+                "stop_reason": "end_turn",
+                "stop_sequence": None,
+            }
+        else:
+            msg = {"role": role, "content": content}
         entry = {
             "type": entry_type,
             "uuid": uid,

autogpt_platform/backend/backend/copilot/sdk/e2b_file_tools.py

@@ -8,6 +8,9 @@ SDK-internal paths (``~/.claude/projects/…/tool-results/``) are handled
 by the separate ``Read`` MCP tool registered in ``tool_adapter.py``.
 """
 
+import asyncio
+import base64
+import hashlib
 import itertools
 import json
 import logging
@@ -28,6 +31,12 @@ from backend.copilot.context import (
 
 logger = logging.getLogger(__name__)
 
+# Default number of lines returned by ``read_file`` when the caller does not
+# specify a limit.  Also used as the threshold in ``bridge_to_sandbox`` to
+# decide whether the model is requesting the full file (and thus whether the
+# bridge copy is worthwhile).
+_DEFAULT_READ_LIMIT = 2000
+
 
 async def _check_sandbox_symlink_escape(
     sandbox: Any,
@@ -89,7 +98,7 @@ def _get_sandbox_and_path(
     return sandbox, remote
 
 
-async def _sandbox_write(sandbox: Any, path: str, content: str) -> None:
+async def _sandbox_write(sandbox: Any, path: str, content: str | bytes) -> None:
     """Write *content* to *path* inside the sandbox.
 
     The E2B filesystem API (``sandbox.files.write``) and the command API
@@ -102,11 +111,14 @@ async def _sandbox_write(sandbox: Any, path: str, content: str) -> None:
     To work around this, writes targeting ``/tmp`` are performed via
     ``tee`` through the command API, which runs as the sandbox ``user``
     and can therefore always overwrite user-owned files.
+
+    *content* may be ``str`` (text) or ``bytes`` (binary).  Both paths
+    are handled correctly: text is encoded to bytes for the base64 shell
+    pipe, and raw bytes are passed through without any encoding.
     """
     if path == "/tmp" or path.startswith("/tmp/"):
-        import base64 as _b64
-
-        encoded = _b64.b64encode(content.encode()).decode()
+        raw = content.encode() if isinstance(content, str) else content
+        encoded = base64.b64encode(raw).decode()
         result = await sandbox.commands.run(
             f"echo {shlex.quote(encoded)} | base64 -d > {shlex.quote(path)}",
             cwd=E2B_WORKDIR,
@@ -128,14 +140,25 @@ async def _handle_read_file(args: dict[str, Any]) -> dict[str, Any]:
     """Read lines from a sandbox file, falling back to the local host for SDK-internal paths."""
     file_path: str = args.get("file_path", "")
     offset: int = max(0, int(args.get("offset", 0)))
-    limit: int = max(1, int(args.get("limit", 2000)))
+    limit: int = max(1, int(args.get("limit", _DEFAULT_READ_LIMIT)))
 
     if not file_path:
         return _mcp("file_path is required", error=True)
 
-    # SDK-internal paths (tool-results, ephemeral working dir) stay on the host.
+    # SDK-internal paths (tool-results/tool-outputs, ephemeral working dir)
+    # stay on the host.  When E2B is active, also copy the file into the
+    # sandbox so bash_exec can access it for further processing.
     if _is_allowed_local(file_path):
-        return _read_local(file_path, offset, limit)
+        result = _read_local(file_path, offset, limit)
+        if not result.get("isError"):
+            sandbox = _get_sandbox()
+            if sandbox is not None:
+                annotation = await bridge_and_annotate(
+                    sandbox, file_path, offset, limit
+                )
+                if annotation:
+                    result["content"][0]["text"] += annotation
+        return result
 
     result = _get_sandbox_and_path(file_path)
     if isinstance(result, dict):
@@ -302,6 +325,103 @@ async def _handle_grep(args: dict[str, Any]) -> dict[str, Any]:
     return _mcp(output if output else "No matches found.")
 
 
+# Bridging: copy SDK-internal files into E2B sandbox
+
+# Files larger than this are written to /home/user/ via sandbox.files.write()
+# instead of /tmp/ via shell base64, to avoid shell argument length limits
+# and E2B command timeouts.  Base64 expands content by ~33%, so keep this
+# well under the typical Linux ARG_MAX (128 KB).
+_BRIDGE_SHELL_MAX_BYTES = 32 * 1024  # 32 KB
+# Files larger than this are skipped entirely to avoid excessive transfer times.
+_BRIDGE_SKIP_BYTES = 50 * 1024 * 1024  # 50 MB
+
+
+async def bridge_to_sandbox(
+    sandbox: Any, file_path: str, offset: int, limit: int
+) -> str | None:
+    """Best-effort copy of a host-side SDK file into the E2B sandbox.
+
+    When the model reads an SDK-internal file (e.g. tool-results), it often
+    wants to process the data with bash.  Copying the file into the sandbox
+    under a stable name lets ``bash_exec`` access it without extra steps.
+
+    Only copies when offset=0 and limit is large enough to indicate the model
+    wants the full file.  Errors are logged but never propagated.
+
+    Returns the sandbox path on success, or ``None`` on skip/failure.
+
+    Size handling:
+    - <= 32 KB: written to ``/tmp/<hash>-<basename>`` via shell base64
+      (``_sandbox_write``).  Kept small to stay within ARG_MAX.
+    - 32 KB - 50 MB: written to ``/home/user/<hash>-<basename>`` via
+      ``sandbox.files.write()`` to avoid shell argument length limits.
+    - > 50 MB: skipped entirely with a warning.
+
+    The sandbox filename is prefixed with a short hash of the full source
+    path to avoid collisions when different source files share the same
+    basename (e.g. multiple ``result.json`` files).
+    """
+    if offset != 0 or limit < _DEFAULT_READ_LIMIT:
+        return None
+    try:
+        expanded = os.path.realpath(os.path.expanduser(file_path))
+        basename = os.path.basename(expanded)
+        source_id = hashlib.sha256(expanded.encode()).hexdigest()[:12]
+        unique_name = f"{source_id}-{basename}"
+        file_size = os.path.getsize(expanded)
+        if file_size > _BRIDGE_SKIP_BYTES:
+            logger.warning(
+                "[E2B] Skipping bridge for large file (%d bytes): %s",
+                file_size,
+                basename,
+            )
+            return None
+
+        def _read_bytes() -> bytes:
+            with open(expanded, "rb") as fh:
+                return fh.read()
+
+        raw_content = await asyncio.to_thread(_read_bytes)
+        try:
+            text_content: str | None = raw_content.decode("utf-8")
+        except UnicodeDecodeError:
+            text_content = None
+        data: str | bytes = text_content if text_content is not None else raw_content
+        if file_size <= _BRIDGE_SHELL_MAX_BYTES:
+            sandbox_path = f"/tmp/{unique_name}"
+            await _sandbox_write(sandbox, sandbox_path, data)
+        else:
+            sandbox_path = f"/home/user/{unique_name}"
+            await sandbox.files.write(sandbox_path, data)
+        logger.info(
+            "[E2B] Bridged SDK file to sandbox: %s -> %s", basename, sandbox_path
+        )
+        return sandbox_path
+    except Exception:
+        logger.warning(
+            "[E2B] Failed to bridge SDK file to sandbox: %s",
+            file_path,
+            exc_info=True,
+        )
+        return None
+
+
+async def bridge_and_annotate(
+    sandbox: Any, file_path: str, offset: int, limit: int
+) -> str | None:
+    """Bridge a host file to the sandbox and return a newline-prefixed annotation.
+
+    Combines ``bridge_to_sandbox`` with the standard annotation suffix so
+    callers don't need to duplicate the pattern.  Returns a string like
+    ``"\\n[Sandbox copy available at /tmp/abc-file.txt]"`` on success, or
+    ``None`` if bridging was skipped or failed.
+    """
+    sandbox_path = await bridge_to_sandbox(sandbox, file_path, offset, limit)
+    if sandbox_path is None:
+        return None
+    return f"\n[Sandbox copy available at {sandbox_path}]"
+
+
 # Local read (for SDK-internal paths)
 
 

autogpt_platform/backend/backend/copilot/sdk/e2b_file_tools_test.py

@@ -3,6 +3,7 @@
 Pure unit tests with no external dependencies (no E2B, no sandbox).
 """
 
+import hashlib
 import os
 import shutil
 from types import SimpleNamespace
@@ -13,12 +14,26 @@ import pytest
 from backend.copilot.context import E2B_WORKDIR, SDK_PROJECTS_DIR, _current_project_dir
 
 from .e2b_file_tools import (
+    _BRIDGE_SHELL_MAX_BYTES,
+    _BRIDGE_SKIP_BYTES,
+    _DEFAULT_READ_LIMIT,
     _check_sandbox_symlink_escape,
     _read_local,
     _sandbox_write,
+    bridge_and_annotate,
+    bridge_to_sandbox,
     resolve_sandbox_path,
 )
 
+
+def _expected_bridge_path(file_path: str, prefix: str = "/tmp") -> str:
+    """Compute the expected sandbox path for a bridged file."""
+    expanded = os.path.realpath(os.path.expanduser(file_path))
+    basename = os.path.basename(expanded)
+    source_id = hashlib.sha256(expanded.encode()).hexdigest()[:12]
+    return f"{prefix}/{source_id}-{basename}"
+
+
 # ---------------------------------------------------------------------------
 # resolve_sandbox_path — sandbox path normalisation & boundary enforcement
 # ---------------------------------------------------------------------------
@@ -91,9 +106,9 @@ class TestResolveSandboxPath:
 # ---------------------------------------------------------------------------
 # _read_local — host filesystem reads with allowlist enforcement
 #
-# In E2B mode, _read_local only allows tool-results paths (via
-# is_allowed_local_path without sdk_cwd).  Regular files live on the
-# sandbox, not the host.
+# In E2B mode, _read_local only allows tool-results/tool-outputs paths
+# (via is_allowed_local_path without sdk_cwd).  Regular files live on
+# the sandbox, not the host.
 # ---------------------------------------------------------------------------
 
 
@@ -119,7 +134,7 @@ class TestReadLocal:
         )
         token = _current_project_dir.set(encoded)
         try:
-            result = _read_local(filepath, offset=0, limit=2000)
+            result = _read_local(filepath, offset=0, limit=_DEFAULT_READ_LIMIT)
             assert result["isError"] is False
             assert "line 1" in result["content"][0]["text"]
             assert "line 2" in result["content"][0]["text"]
@@ -127,6 +142,25 @@ class TestReadLocal:
             _current_project_dir.reset(token)
             os.unlink(filepath)
 
+    def test_read_tool_outputs_file(self):
+        """Reading a tool-outputs file should also succeed."""
+        encoded = "-tmp-copilot-e2b-test-read-outputs"
+        tool_outputs_dir = os.path.join(
+            SDK_PROJECTS_DIR, encoded, self._CONV_UUID, "tool-outputs"
+        )
+        os.makedirs(tool_outputs_dir, exist_ok=True)
+        filepath = os.path.join(tool_outputs_dir, "sdk-abc123.json")
+        with open(filepath, "w") as f:
+            f.write('{"data": "test"}\n')
+        token = _current_project_dir.set(encoded)
+        try:
+            result = _read_local(filepath, offset=0, limit=_DEFAULT_READ_LIMIT)
+            assert result["isError"] is False
+            assert "test" in result["content"][0]["text"]
+        finally:
+            _current_project_dir.reset(token)
+            shutil.rmtree(os.path.join(SDK_PROJECTS_DIR, encoded), ignore_errors=True)
+
     def test_read_disallowed_path_blocked(self):
         """Reading /etc/passwd should be blocked by the allowlist."""
         result = _read_local("/etc/passwd", offset=0, limit=10)
@@ -335,3 +369,199 @@ class TestSandboxWrite:
         encoded_in_cmd = call_args.split("echo ")[1].split(" |")[0].strip("'")
         decoded = base64.b64decode(encoded_in_cmd).decode()
         assert decoded == content
+
+
+# ---------------------------------------------------------------------------
+# bridge_to_sandbox — copy SDK-internal files into E2B sandbox
+# ---------------------------------------------------------------------------
+
+
+def _make_bridge_sandbox() -> SimpleNamespace:
+    """Build a sandbox mock suitable for bridge_to_sandbox tests."""
+    run_result = SimpleNamespace(stdout="", stderr="", exit_code=0)
+    commands = SimpleNamespace(run=AsyncMock(return_value=run_result))
+    files = SimpleNamespace(write=AsyncMock())
+    return SimpleNamespace(commands=commands, files=files)
+
+
+class TestBridgeToSandbox:
+    @pytest.mark.asyncio
+    async def test_happy_path_small_file(self, tmp_path):
+        """A small file is bridged to /tmp/<hash>-<basename> via _sandbox_write."""
+        f = tmp_path / "result.json"
+        f.write_text('{"ok": true}')
+        sandbox = _make_bridge_sandbox()
+
+        result = await bridge_to_sandbox(
+            sandbox, str(f), offset=0, limit=_DEFAULT_READ_LIMIT
+        )
+
+        expected = _expected_bridge_path(str(f))
+        assert result == expected
+        sandbox.commands.run.assert_called_once()
+        cmd = sandbox.commands.run.call_args[0][0]
+        assert "result.json" in cmd
+        sandbox.files.write.assert_not_called()
+
+    @pytest.mark.asyncio
+    async def test_skip_when_offset_nonzero(self, tmp_path):
+        """Bridging is skipped when offset != 0 (partial read)."""
+        f = tmp_path / "data.txt"
+        f.write_text("content")
+        sandbox = _make_bridge_sandbox()
+
+        result = await bridge_to_sandbox(
+            sandbox, str(f), offset=10, limit=_DEFAULT_READ_LIMIT
+        )
+
+        assert result is None
+        sandbox.commands.run.assert_not_called()
+        sandbox.files.write.assert_not_called()
+
+    @pytest.mark.asyncio
+    async def test_skip_when_limit_too_small(self, tmp_path):
+        """Bridging is skipped when limit < _DEFAULT_READ_LIMIT (partial read)."""
+        f = tmp_path / "data.txt"
+        f.write_text("content")
+        sandbox = _make_bridge_sandbox()
+
+        await bridge_to_sandbox(sandbox, str(f), offset=0, limit=100)
+
+        sandbox.commands.run.assert_not_called()
+        sandbox.files.write.assert_not_called()
+
+    @pytest.mark.asyncio
+    async def test_nonexistent_file_does_not_raise(self, tmp_path):
+        """Bridging a non-existent file logs but does not propagate errors."""
+        sandbox = _make_bridge_sandbox()
+
+        await bridge_to_sandbox(
+            sandbox, str(tmp_path / "ghost.txt"), offset=0, limit=_DEFAULT_READ_LIMIT
+        )
+
+        sandbox.commands.run.assert_not_called()
+        sandbox.files.write.assert_not_called()
+
+    @pytest.mark.asyncio
+    async def test_sandbox_write_failure_returns_none(self, tmp_path):
+        """If sandbox write fails, returns None (best-effort)."""
+        f = tmp_path / "data.txt"
+        f.write_text("content")
+        sandbox = _make_bridge_sandbox()
+        sandbox.commands.run.side_effect = RuntimeError("E2B timeout")
+
+        result = await bridge_to_sandbox(
+            sandbox, str(f), offset=0, limit=_DEFAULT_READ_LIMIT
+        )
+
+        assert result is None
+
+    @pytest.mark.asyncio
+    async def test_large_file_uses_files_api(self, tmp_path):
+        """Files > 32 KB but <= 50 MB are written to /home/user/ via files.write."""
+        f = tmp_path / "big.json"
+        f.write_bytes(b"x" * (_BRIDGE_SHELL_MAX_BYTES + 1))
+        sandbox = _make_bridge_sandbox()
+
+        result = await bridge_to_sandbox(
+            sandbox, str(f), offset=0, limit=_DEFAULT_READ_LIMIT
+        )
+
+        expected = _expected_bridge_path(str(f), prefix="/home/user")
+        assert result == expected
+        sandbox.files.write.assert_called_once()
+        call_args = sandbox.files.write.call_args[0]
+        assert call_args[0] == expected
+        sandbox.commands.run.assert_not_called()
+
+    @pytest.mark.asyncio
+    async def test_small_binary_file_preserves_bytes(self, tmp_path):
+        """A small binary file is bridged to /tmp via base64 without corruption."""
+        binary_data = bytes(range(256))
+        f = tmp_path / "image.png"
+        f.write_bytes(binary_data)
+        sandbox = _make_bridge_sandbox()
+
+        result = await bridge_to_sandbox(
+            sandbox, str(f), offset=0, limit=_DEFAULT_READ_LIMIT
+        )
+
+        expected = _expected_bridge_path(str(f))
+        assert result == expected
+        sandbox.commands.run.assert_called_once()
+        cmd = sandbox.commands.run.call_args[0][0]
+        assert "base64" in cmd
+        sandbox.files.write.assert_not_called()
+
+    @pytest.mark.asyncio
+    async def test_large_binary_file_writes_raw_bytes(self, tmp_path):
+        """A large binary file is bridged to /home/user/ as raw bytes."""
+        binary_data = bytes(range(256)) * 200
+        f = tmp_path / "photo.jpg"
+        f.write_bytes(binary_data)
+        sandbox = _make_bridge_sandbox()
+
+        result = await bridge_to_sandbox(
+            sandbox, str(f), offset=0, limit=_DEFAULT_READ_LIMIT
+        )
+
+        expected = _expected_bridge_path(str(f), prefix="/home/user")
+        assert result == expected
+        sandbox.files.write.assert_called_once()
+        call_args = sandbox.files.write.call_args[0]
+        assert call_args[0] == expected
+        assert call_args[1] == binary_data
+        sandbox.commands.run.assert_not_called()
+
+    @pytest.mark.asyncio
+    async def test_very_large_file_skipped(self, tmp_path):
+        """Files > 50 MB are skipped entirely."""
+        f = tmp_path / "huge.bin"
+        # Create a sparse file to avoid actually writing 50 MB
+        with open(f, "wb") as fh:
+            fh.seek(_BRIDGE_SKIP_BYTES + 1)
+            fh.write(b"\0")
+        sandbox = _make_bridge_sandbox()
+
+        result = await bridge_to_sandbox(
+            sandbox, str(f), offset=0, limit=_DEFAULT_READ_LIMIT
+        )
+
+        assert result is None
+
+        sandbox.commands.run.assert_not_called()
+        sandbox.files.write.assert_not_called()
+
+
+# ---------------------------------------------------------------------------
+# bridge_and_annotate — shared helper wrapping bridge_to_sandbox + annotation
+# ---------------------------------------------------------------------------
+
+
+class TestBridgeAndAnnotate:
+    @pytest.mark.asyncio
+    async def test_returns_annotation_on_success(self, tmp_path):
+        """On success, returns a newline-prefixed annotation with the sandbox path."""
+        f = tmp_path / "data.json"
+        f.write_text('{"ok": true}')
+        sandbox = _make_bridge_sandbox()
+
+        annotation = await bridge_and_annotate(
+            sandbox, str(f), offset=0, limit=_DEFAULT_READ_LIMIT
+        )
+
+        expected_path = _expected_bridge_path(str(f))
+        assert annotation == f"\n[Sandbox copy available at {expected_path}]"
+
+    @pytest.mark.asyncio
+    async def test_returns_none_when_skipped(self, tmp_path):
+        """When bridging is skipped (e.g. offset != 0), returns None."""
+        f = tmp_path / "data.json"
+        f.write_text("content")
+        sandbox = _make_bridge_sandbox()
+
+        annotation = await bridge_and_annotate(
+            sandbox, str(f), offset=10, limit=_DEFAULT_READ_LIMIT
+        )
+
+        assert annotation is None

autogpt_platform/backend/backend/copilot/sdk/e2e_compaction_test.py

@@ -275,7 +275,7 @@ class TestCompactionE2E:
 
         # --- Step 7: CompactionTracker receives PreCompact hook ---
         tracker = CompactionTracker()
-        session = ChatSession.new(user_id="test-user")
+        session = ChatSession.new(user_id="test-user", dry_run=False)
         tracker.on_compact(str(session_file))
 
         # --- Step 8: Next SDK message arrives → emit_start ---
@@ -376,7 +376,7 @@ class TestCompactionE2E:
         monkeypatch.setenv("CLAUDE_CONFIG_DIR", str(config_dir))
 
         tracker = CompactionTracker()
-        session = ChatSession.new(user_id="test")
+        session = ChatSession.new(user_id="test", dry_run=False)
         builder = TranscriptBuilder()
 
         # --- First query with compaction ---

autogpt_platform/backend/backend/copilot/sdk/env.py

@@ -0,0 +1,82 @@
+"""SDK environment variable builder — importable without circular deps.
+
+Extracted from ``service.py`` so that ``backend.blocks.orchestrator``
+can reuse the same subscription / OpenRouter / direct-Anthropic logic
+without pulling in the full copilot service module (which would create a
+circular import through ``executor`` → ``credit`` → ``block_cost_config``).
+"""
+
+from __future__ import annotations
+
+from backend.copilot.config import ChatConfig
+from backend.copilot.sdk.subscription import validate_subscription
+
+# ChatConfig is stateless (reads env vars) — a separate instance is fine.
+# A singleton would require importing service.py which causes the circular dep
+# this module was created to avoid.
+config = ChatConfig()
+
+
+def build_sdk_env(
+    session_id: str | None = None,
+    user_id: str | None = None,
+    sdk_cwd: str | None = None,
+) -> dict[str, str]:
+    """Build env vars for the SDK CLI subprocess.
+
+    Three modes (checked in order):
+    1. **Subscription** — clears all keys; CLI uses ``claude login`` auth.
+    2. **Direct Anthropic** — returns ``{}``; subprocess inherits
+       ``ANTHROPIC_API_KEY`` from the parent environment.
+    3. **OpenRouter** (default) — overrides base URL and auth token to
+       route through the proxy, with Langfuse trace headers.
+
+    When *sdk_cwd* is provided, ``CLAUDE_CODE_TMPDIR`` is set so that
+    the CLI writes temp/sub-agent output inside the per-session workspace
+    directory rather than an inaccessible system temp path.
+    """
+    # --- Mode 1: Claude Code subscription auth ---
+    if config.use_claude_code_subscription:
+        validate_subscription()
+        env: dict[str, str] = {
+            "ANTHROPIC_API_KEY": "",
+            "ANTHROPIC_AUTH_TOKEN": "",
+            "ANTHROPIC_BASE_URL": "",
+        }
+        if sdk_cwd:
+            env["CLAUDE_CODE_TMPDIR"] = sdk_cwd
+        return env
+
+    # --- Mode 2: Direct Anthropic (no proxy hop) ---
+    if not config.openrouter_active:
+        env = {}
+        if sdk_cwd:
+            env["CLAUDE_CODE_TMPDIR"] = sdk_cwd
+        return env
+
+    # --- Mode 3: OpenRouter proxy ---
+    base = (config.base_url or "").rstrip("/")
+    if base.endswith("/v1"):
+        base = base[:-3]
+    env = {
+        "ANTHROPIC_BASE_URL": base,
+        "ANTHROPIC_AUTH_TOKEN": config.api_key or "",
+        "ANTHROPIC_API_KEY": "",  # force CLI to use AUTH_TOKEN
+    }
+
+    # Inject broadcast headers so OpenRouter forwards traces to Langfuse.
+    def _safe(v: str) -> str:
+        return v.replace("\r", "").replace("\n", "").strip()[:128]
+
+    parts = []
+    if session_id:
+        parts.append(f"x-session-id: {_safe(session_id)}")
+    if user_id:
+        parts.append(f"x-user-id: {_safe(user_id)}")
+    if parts:
+        env["ANTHROPIC_CUSTOM_HEADERS"] = "\n".join(parts)
+
+    if sdk_cwd:
+        env["CLAUDE_CODE_TMPDIR"] = sdk_cwd
+
+    return env

autogpt_platform/backend/backend/copilot/sdk/env_test.py

@@ -0,0 +1,293 @@
+"""Tests for build_sdk_env() — the SDK subprocess environment builder."""
+
+from unittest.mock import patch
+
+import pytest
+
+from backend.copilot.config import ChatConfig
+
+# ---------------------------------------------------------------------------
+# Helpers — build a ChatConfig with explicit field values so tests don't
+# depend on real environment variables.
+# ---------------------------------------------------------------------------
+
+
+def _make_config(**overrides) -> ChatConfig:
+    """Create a ChatConfig with safe defaults, applying *overrides*."""
+    defaults = {
+        "use_claude_code_subscription": False,
+        "use_openrouter": False,
+        "api_key": None,
+        "base_url": None,
+    }
+    defaults.update(overrides)
+    return ChatConfig(**defaults)
+
+
+# ---------------------------------------------------------------------------
+# Mode 1 — Subscription auth
+# ---------------------------------------------------------------------------
+
+
+class TestBuildSdkEnvSubscription:
+    """When ``use_claude_code_subscription`` is True, keys are blanked."""
+
+    @patch("backend.copilot.sdk.env.validate_subscription")
+    def test_returns_blanked_keys(self, mock_validate):
+        """Subscription mode clears API_KEY, AUTH_TOKEN, and BASE_URL."""
+        cfg = _make_config(use_claude_code_subscription=True)
+        with patch("backend.copilot.sdk.env.config", cfg):
+            from backend.copilot.sdk.env import build_sdk_env
+
+            result = build_sdk_env()
+
+        assert result == {
+            "ANTHROPIC_API_KEY": "",
+            "ANTHROPIC_AUTH_TOKEN": "",
+            "ANTHROPIC_BASE_URL": "",
+        }
+        mock_validate.assert_called_once()
+
+    @patch(
+        "backend.copilot.sdk.env.validate_subscription",
+        side_effect=RuntimeError("CLI not found"),
+    )
+    def test_propagates_validation_error(self, mock_validate):
+        """If validate_subscription fails, the error bubbles up."""
+        cfg = _make_config(use_claude_code_subscription=True)
+        with patch("backend.copilot.sdk.env.config", cfg):
+            from backend.copilot.sdk.env import build_sdk_env
+
+            with pytest.raises(RuntimeError, match="CLI not found"):
+                build_sdk_env()
+
+
+# ---------------------------------------------------------------------------
+# Mode 2 — Direct Anthropic (no OpenRouter)
+# ---------------------------------------------------------------------------
+
+
+class TestBuildSdkEnvDirectAnthropic:
+    """When OpenRouter is inactive, return empty dict (inherit parent env)."""
+
+    def test_returns_empty_dict_when_openrouter_inactive(self):
+        cfg = _make_config(use_openrouter=False)
+        with patch("backend.copilot.sdk.env.config", cfg):
+            from backend.copilot.sdk.env import build_sdk_env
+
+            result = build_sdk_env()
+
+        assert result == {}
+
+    def test_returns_empty_dict_when_openrouter_flag_true_but_no_key(self):
+        """OpenRouter flag is True but no api_key => openrouter_active is False."""
+        cfg = _make_config(use_openrouter=True, base_url="https://openrouter.ai/api/v1")
+        # Force api_key to None after construction (field_validator may pick up env vars)
+        object.__setattr__(cfg, "api_key", None)
+        assert not cfg.openrouter_active
+        with patch("backend.copilot.sdk.env.config", cfg):
+            from backend.copilot.sdk.env import build_sdk_env
+
+            result = build_sdk_env()
+
+        assert result == {}
+
+
+# ---------------------------------------------------------------------------
+# Mode 3 — OpenRouter proxy
+# ---------------------------------------------------------------------------
+
+
+class TestBuildSdkEnvOpenRouter:
+    """When OpenRouter is active, return proxy env vars."""
+
+    def _openrouter_config(self, **overrides):
+        defaults = {
+            "use_openrouter": True,
+            "api_key": "sk-or-test-key",
+            "base_url": "https://openrouter.ai/api/v1",
+        }
+        defaults.update(overrides)
+        return _make_config(**defaults)
+
+    def test_basic_openrouter_env(self):
+        cfg = self._openrouter_config()
+        with patch("backend.copilot.sdk.env.config", cfg):
+            from backend.copilot.sdk.env import build_sdk_env
+
+            result = build_sdk_env()
+
+        assert result["ANTHROPIC_BASE_URL"] == "https://openrouter.ai/api"
+        assert result["ANTHROPIC_AUTH_TOKEN"] == "sk-or-test-key"
+        assert result["ANTHROPIC_API_KEY"] == ""
+        assert "ANTHROPIC_CUSTOM_HEADERS" not in result
+
+    def test_strips_trailing_v1(self):
+        """The /v1 suffix is stripped from the base URL."""
+        cfg = self._openrouter_config(base_url="https://openrouter.ai/api/v1")
+        with patch("backend.copilot.sdk.env.config", cfg):
+            from backend.copilot.sdk.env import build_sdk_env
+
+            result = build_sdk_env()
+
+        assert result["ANTHROPIC_BASE_URL"] == "https://openrouter.ai/api"
+
+    def test_strips_trailing_v1_and_slash(self):
+        """Trailing slash before /v1 strip is handled."""
+        cfg = self._openrouter_config(base_url="https://openrouter.ai/api/v1/")
+        with patch("backend.copilot.sdk.env.config", cfg):
+            from backend.copilot.sdk.env import build_sdk_env
+
+            result = build_sdk_env()
+
+        # rstrip("/") first, then remove /v1
+        assert result["ANTHROPIC_BASE_URL"] == "https://openrouter.ai/api"
+
+    def test_no_v1_suffix_left_alone(self):
+        """A base URL without /v1 is used as-is."""
+        cfg = self._openrouter_config(base_url="https://custom-proxy.example.com")
+        with patch("backend.copilot.sdk.env.config", cfg):
+            from backend.copilot.sdk.env import build_sdk_env
+
+            result = build_sdk_env()
+
+        assert result["ANTHROPIC_BASE_URL"] == "https://custom-proxy.example.com"
+
+    def test_session_id_header(self):
+        cfg = self._openrouter_config()
+        with patch("backend.copilot.sdk.env.config", cfg):
+            from backend.copilot.sdk.env import build_sdk_env
+
+            result = build_sdk_env(session_id="sess-123")
+
+        assert "ANTHROPIC_CUSTOM_HEADERS" in result
+        assert "x-session-id: sess-123" in result["ANTHROPIC_CUSTOM_HEADERS"]
+
+    def test_user_id_header(self):
+        cfg = self._openrouter_config()
+        with patch("backend.copilot.sdk.env.config", cfg):
+            from backend.copilot.sdk.env import build_sdk_env
+
+            result = build_sdk_env(user_id="user-456")
+
+        assert "x-user-id: user-456" in result["ANTHROPIC_CUSTOM_HEADERS"]
+
+    def test_both_headers(self):
+        cfg = self._openrouter_config()
+        with patch("backend.copilot.sdk.env.config", cfg):
+            from backend.copilot.sdk.env import build_sdk_env
+
+            result = build_sdk_env(session_id="s1", user_id="u2")
+
+        headers = result["ANTHROPIC_CUSTOM_HEADERS"]
+        assert "x-session-id: s1" in headers
+        assert "x-user-id: u2" in headers
+        # They should be newline-separated
+        assert "\n" in headers
+
+    def test_header_sanitisation_strips_newlines(self):
+        """Newlines/carriage-returns in header values are stripped."""
+        cfg = self._openrouter_config()
+        with patch("backend.copilot.sdk.env.config", cfg):
+            from backend.copilot.sdk.env import build_sdk_env
+
+            result = build_sdk_env(session_id="bad\r\nvalue")
+
+        header_val = result["ANTHROPIC_CUSTOM_HEADERS"]
+        # The _safe helper removes \r and \n
+        assert "\r" not in header_val.split(": ", 1)[1]
+        assert "badvalue" in header_val
+
+    def test_header_value_truncated_to_128_chars(self):
+        """Header values are truncated to 128 characters."""
+        cfg = self._openrouter_config()
+        with patch("backend.copilot.sdk.env.config", cfg):
+            from backend.copilot.sdk.env import build_sdk_env
+
+            long_id = "x" * 200
+            result = build_sdk_env(session_id=long_id)
+
+        # The value after "x-session-id: " should be at most 128 chars
+        header_line = result["ANTHROPIC_CUSTOM_HEADERS"]
+        value = header_line.split(": ", 1)[1]
+        assert len(value) == 128
+
+
+# ---------------------------------------------------------------------------
+# Mode priority
+# ---------------------------------------------------------------------------
+
+
+class TestBuildSdkEnvModePriority:
+    """Subscription mode takes precedence over OpenRouter."""
+
+    @patch("backend.copilot.sdk.env.validate_subscription")
+    def test_subscription_overrides_openrouter(self, mock_validate):
+        cfg = _make_config(
+            use_claude_code_subscription=True,
+            use_openrouter=True,
+            api_key="sk-or-key",
+            base_url="https://openrouter.ai/api/v1",
+        )
+        with patch("backend.copilot.sdk.env.config", cfg):
+            from backend.copilot.sdk.env import build_sdk_env
+
+            result = build_sdk_env()
+
+        # Should get subscription result, not OpenRouter
+        assert result == {
+            "ANTHROPIC_API_KEY": "",
+            "ANTHROPIC_AUTH_TOKEN": "",
+            "ANTHROPIC_BASE_URL": "",
+        }
+
+
+# ---------------------------------------------------------------------------
+# CLAUDE_CODE_TMPDIR integration
+# ---------------------------------------------------------------------------
+
+
+class TestClaudeCodeTmpdir:
+    """Verify build_sdk_env() sets CLAUDE_CODE_TMPDIR from *sdk_cwd*."""
+
+    def test_tmpdir_set_when_sdk_cwd_is_truthy(self):
+        """CLAUDE_CODE_TMPDIR is set to sdk_cwd when sdk_cwd is truthy."""
+        cfg = _make_config(use_openrouter=False)
+        with patch("backend.copilot.sdk.env.config", cfg):
+            from backend.copilot.sdk.env import build_sdk_env
+
+            result = build_sdk_env(sdk_cwd="/tmp/copilot-workspace")
+
+        assert result["CLAUDE_CODE_TMPDIR"] == "/tmp/copilot-workspace"
+
+    def test_tmpdir_not_set_when_sdk_cwd_is_none(self):
+        """CLAUDE_CODE_TMPDIR is NOT in the env when sdk_cwd is None."""
+        cfg = _make_config(use_openrouter=False)
+        with patch("backend.copilot.sdk.env.config", cfg):
+            from backend.copilot.sdk.env import build_sdk_env
+
+            result = build_sdk_env(sdk_cwd=None)
+
+        assert "CLAUDE_CODE_TMPDIR" not in result
+
+    def test_tmpdir_not_set_when_sdk_cwd_is_empty_string(self):
+        """CLAUDE_CODE_TMPDIR is NOT in the env when sdk_cwd is empty string."""
+        cfg = _make_config(use_openrouter=False)
+        with patch("backend.copilot.sdk.env.config", cfg):
+            from backend.copilot.sdk.env import build_sdk_env
+
+            result = build_sdk_env(sdk_cwd="")
+
+        assert "CLAUDE_CODE_TMPDIR" not in result
+
+    @patch("backend.copilot.sdk.env.validate_subscription")
+    def test_tmpdir_set_in_subscription_mode(self, mock_validate):
+        """CLAUDE_CODE_TMPDIR is set even in subscription mode."""
+        cfg = _make_config(use_claude_code_subscription=True)
+        with patch("backend.copilot.sdk.env.config", cfg):
+            from backend.copilot.sdk.env import build_sdk_env
+
+            result = build_sdk_env(sdk_cwd="/tmp/sub-workspace")
+
+        assert result["CLAUDE_CODE_TMPDIR"] == "/tmp/sub-workspace"
+        assert result["ANTHROPIC_API_KEY"] == ""

autogpt_platform/backend/backend/copilot/sdk/prompt_too_long_test.py

@@ -38,7 +38,7 @@ class TestFlattenAssistantContent:
 
     def test_tool_use_blocks(self):
         blocks = [{"type": "tool_use", "name": "read_file", "input": {}}]
-        assert _flatten_assistant_content(blocks) == "[tool_use: read_file]"
+        assert _flatten_assistant_content(blocks) == ""
 
     def test_mixed_blocks(self):
         blocks = [
@@ -47,19 +47,22 @@ class TestFlattenAssistantContent:
         ]
         result = _flatten_assistant_content(blocks)
         assert "Let me read that." in result
-        assert "[tool_use: Read]" in result
+        # tool_use blocks are dropped entirely to prevent model mimicry
+        assert "Read" not in result
 
     def test_raw_strings(self):
         assert _flatten_assistant_content(["hello", "world"]) == "hello\nworld"
 
-    def test_unknown_block_type_preserved_as_placeholder(self):
+    def test_unknown_block_type_dropped(self):
         blocks = [
             {"type": "text", "text": "See this image:"},
             {"type": "image", "source": {"type": "base64", "data": "..."}},
         ]
         result = _flatten_assistant_content(blocks)
         assert "See this image:" in result
-        assert "[__image__]" in result
+        # Unknown block types are dropped to prevent model mimicry
+        assert "[__image__]" not in result
+        assert "base64" not in result
 
     def test_empty(self):
         assert _flatten_assistant_content([]) == ""
@@ -279,7 +282,8 @@ class TestTranscriptToMessages:
         messages = _transcript_to_messages(content)
         assert len(messages) == 2
         assert "Let me check." in messages[0]["content"]
-        assert "[tool_use: read_file]" in messages[0]["content"]
+        # tool_use blocks are dropped entirely to prevent model mimicry
+        assert "read_file" not in messages[0]["content"]
         assert messages[1]["content"] == "file contents"
 
 
@@ -442,8 +446,11 @@ class TestCompactTranscript:
         assert result is not None
         assert validate_transcript(result)
         msgs = _transcript_to_messages(result)
-        assert len(msgs) == 2
+        # 3 messages: compressed prefix (2) + preserved last assistant (1)
+        assert len(msgs) == 3
         assert msgs[1]["content"] == "Summarized response"
+        # The last assistant entry is preserved verbatim from original
+        assert msgs[2]["content"] == "Details"
 
     @pytest.mark.asyncio
     async def test_returns_none_on_compression_failure(self, mock_chat_config):

autogpt_platform/backend/backend/copilot/sdk/query_builder_test.py

@@ -49,22 +49,22 @@ def test_format_assistant_tool_calls():
         )
     ]
     result = _format_conversation_context(msgs)
-    assert result is not None
-    assert 'You called tool: search({"q": "test"})' in result
+    # Assistant with no content and tool_calls omitted produces no lines
+    assert result is None
 
 
 def test_format_tool_result():
     msgs = [ChatMessage(role="tool", content='{"result": "ok"}')]
     result = _format_conversation_context(msgs)
     assert result is not None
-    assert 'Tool result: {"result": "ok"}' in result
+    assert 'Tool output: {"result": "ok"}' in result
 
 
 def test_format_tool_result_none_content():
     msgs = [ChatMessage(role="tool", content=None)]
     result = _format_conversation_context(msgs)
     assert result is not None
-    assert "Tool result: " in result
+    assert "Tool output: " in result
 
 
 def test_format_full_conversation():
@@ -84,8 +84,8 @@ def test_format_full_conversation():
     assert result is not None
     assert "User: find agents" in result
     assert "You responded: I'll search for agents." in result
-    assert "You called tool: find_agents" in result
-    assert "Tool result:" in result
+    # tool_calls are omitted to prevent model mimicry
+    assert "Tool output:" in result
     assert "You responded: Found Agent1." in result
 
 

autogpt_platform/backend/backend/copilot/sdk/response_adapter.py

@@ -15,6 +15,7 @@ from claude_agent_sdk import (
     ResultMessage,
     SystemMessage,
     TextBlock,
+    ThinkingBlock,
     ToolResultBlock,
     ToolUseBlock,
     UserMessage,
@@ -26,6 +27,7 @@ from backend.copilot.response_model import (
     StreamError,
     StreamFinish,
     StreamFinishStep,
+    StreamHeartbeat,
     StreamStart,
     StreamStartStep,
     StreamTextDelta,
@@ -75,6 +77,12 @@ class SDKResponseAdapter:
                 # Open the first step (matches non-SDK: StreamStart then StreamStartStep)
                 responses.append(StreamStartStep())
                 self.step_open = True
+            elif sdk_message.subtype == "task_progress":
+                # Emit a heartbeat so publish_chunk is called during long
+                # sub-agent runs. Without this, the Redis stream and meta
+                # key TTLs expire during gaps where no real chunks are
+                # produced (task_progress events were previously silent).
+                responses.append(StreamHeartbeat())
 
         elif isinstance(sdk_message, AssistantMessage):
             # Flush any SDK built-in tool calls that didn't get a UserMessage
@@ -100,6 +108,11 @@ class SDKResponseAdapter:
                             StreamTextDelta(id=self.text_block_id, delta=block.text)
                         )
 
+                elif isinstance(block, ThinkingBlock):
+                    # Thinking blocks are preserved in the transcript but
+                    # not streamed to the frontend — skip silently.
+                    pass
+
                 elif isinstance(block, ToolUseBlock):
                     self._end_text_if_open(responses)
 

autogpt_platform/backend/backend/copilot/sdk/response_adapter_test.py

@@ -18,6 +18,7 @@ from backend.copilot.response_model import (
     StreamError,
     StreamFinish,
     StreamFinishStep,
+    StreamHeartbeat,
     StreamStart,
     StreamStartStep,
     StreamTextDelta,
@@ -28,6 +29,7 @@ from backend.copilot.response_model import (
     StreamToolOutputAvailable,
 )
 
+from .compaction import compaction_events
 from .response_adapter import SDKResponseAdapter
 from .tool_adapter import MCP_TOOL_PREFIX
 from .tool_adapter import _pending_tool_outputs as _pto
@@ -59,6 +61,14 @@ def test_system_non_init_emits_nothing():
     assert results == []
 
 
+def test_task_progress_emits_heartbeat():
+    """task_progress events emit a StreamHeartbeat to keep Redis TTL alive."""
+    adapter = _adapter()
+    results = adapter.convert_message(SystemMessage(subtype="task_progress", data={}))
+    assert len(results) == 1
+    assert isinstance(results[0], StreamHeartbeat)
+
+
 # -- AssistantMessage with TextBlock -----------------------------------------
 
 
@@ -680,3 +690,102 @@ def test_already_resolved_tool_skipped_in_user_message():
     assert (
         len(output_events) == 0
     ), "Already-resolved tool should not emit duplicate output"
+
+
+# -- _end_text_if_open before compaction -------------------------------------
+
+
+def test_end_text_if_open_emits_text_end_before_finish_step():
+    """StreamTextEnd must be emitted before StreamFinishStep during compaction.
+
+    When ``emit_end_if_ready`` fires compaction events while a text block is
+    still open, ``_end_text_if_open`` must close it first.  If StreamFinishStep
+    arrives before StreamTextEnd, the Vercel AI SDK clears ``activeTextParts``
+    and raises "Received text-end for missing text part".
+    """
+    adapter = _adapter()
+
+    # Open a text block by processing an AssistantMessage with text
+    msg = AssistantMessage(content=[TextBlock(text="partial response")], model="test")
+    adapter.convert_message(msg)
+    assert adapter.has_started_text
+    assert not adapter.has_ended_text
+
+    # Simulate what service.py does before yielding compaction events
+    pre_close: list[StreamBaseResponse] = []
+    adapter._end_text_if_open(pre_close)
+    combined = pre_close + list(compaction_events("Compacted transcript"))
+
+    text_end_idx = next(
+        (i for i, e in enumerate(combined) if isinstance(e, StreamTextEnd)), None
+    )
+    finish_step_idx = next(
+        (i for i, e in enumerate(combined) if isinstance(e, StreamFinishStep)), None
+    )
+
+    assert text_end_idx is not None, "StreamTextEnd must be present"
+    assert finish_step_idx is not None, "StreamFinishStep must be present"
+    assert text_end_idx < finish_step_idx, (
+        f"StreamTextEnd (idx={text_end_idx}) must precede "
+        f"StreamFinishStep (idx={finish_step_idx}) — otherwise the Vercel AI SDK "
+        "clears activeTextParts before text-end arrives"
+    )
+
+
+def test_step_open_must_reset_after_compaction_finish_step():
+    """Adapter step_open must be reset when compaction emits StreamFinishStep.
+
+    Compaction events bypass the adapter, so service.py must explicitly clear
+    step_open after yielding a StreamFinishStep from compaction. Without this,
+    the next AssistantMessage skips StreamStartStep because the adapter still
+    thinks a step is open.
+    """
+    adapter = _adapter()
+
+    # Open a step + text block via an AssistantMessage
+    msg = AssistantMessage(content=[TextBlock(text="thinking...")], model="test")
+    adapter.convert_message(msg)
+    assert adapter.step_open is True
+
+    # Simulate what service.py does: close text, then check compaction events
+    pre_close: list[StreamBaseResponse] = []
+    adapter._end_text_if_open(pre_close)
+
+    events = list(compaction_events("Compacted transcript"))
+    if any(isinstance(ev, StreamFinishStep) for ev in events):
+        adapter.step_open = False
+
+    assert (
+        adapter.step_open is False
+    ), "step_open must be False after compaction emits StreamFinishStep"
+
+    # Next AssistantMessage must open a new step
+    msg2 = AssistantMessage(content=[TextBlock(text="continued")], model="test")
+    results = adapter.convert_message(msg2)
+    assert any(
+        isinstance(r, StreamStartStep) for r in results
+    ), "A new StreamStartStep must be emitted after compaction closed the step"
+
+
+def test_end_text_if_open_no_op_when_no_text_open():
+    """_end_text_if_open emits nothing when no text block is open."""
+    adapter = _adapter()
+    results: list[StreamBaseResponse] = []
+    adapter._end_text_if_open(results)
+    assert results == []
+
+
+def test_end_text_if_open_no_op_after_text_already_ended():
+    """_end_text_if_open emits nothing when the text block is already closed."""
+    adapter = _adapter()
+    msg = AssistantMessage(content=[TextBlock(text="hello")], model="test")
+    adapter.convert_message(msg)
+    # Close it once
+    first: list[StreamBaseResponse] = []
+    adapter._end_text_if_open(first)
+    assert len(first) == 1
+    assert isinstance(first[0], StreamTextEnd)
+    # Second call must be a no-op
+    second: list[StreamBaseResponse] = []
+    adapter._end_text_if_open(second)
+    assert second == []

autogpt_platform/backend/backend/copilot/sdk/retry_scenarios_test.py

@@ -124,8 +124,11 @@ class TestScenarioCompactAndRetry:
         assert result != original  # Must be different
         assert validate_transcript(result)
         msgs = _transcript_to_messages(result)
-        assert len(msgs) == 2
+        # 3 messages: compressed prefix (2) + preserved last assistant (1)
+        assert len(msgs) == 3
         assert msgs[0]["content"] == "[summary of conversation]"
+        # Last assistant preserved verbatim
+        assert msgs[2]["content"] == "Long answer 2"
 
     def test_compacted_transcript_loads_into_builder(self):
         """TranscriptBuilder can load a compacted transcript and continue."""
@@ -737,7 +740,10 @@ class TestRetryEdgeCases:
         assert result is not None
         assert result != transcript
         msgs = _transcript_to_messages(result)
-        assert len(msgs) == 2
+        # 3 messages: compressed prefix (2) + preserved last assistant (1)
+        assert len(msgs) == 3
+        # Last assistant preserved verbatim
+        assert msgs[2]["content"] == "Answer 19"
 
     def test_messages_to_transcript_roundtrip_preserves_content(self):
         """Verify messages → transcript → messages preserves all content."""
@@ -898,14 +904,14 @@ class TestTranscriptEdgeCases:
         assert restored[1]["content"] == "Second"
 
     def test_flatten_assistant_with_only_tool_use(self):
-        """Assistant message with only tool_use blocks (no text)."""
+        """Assistant message with only tool_use blocks (no text) flattens to empty."""
         blocks = [
             {"type": "tool_use", "name": "bash", "input": {"cmd": "ls"}},
             {"type": "tool_use", "name": "read", "input": {"path": "/f"}},
         ]
         result = _flatten_assistant_content(blocks)
-        assert "[tool_use: bash]" in result
-        assert "[tool_use: read]" in result
+        # tool_use blocks are dropped entirely to prevent model mimicry
+        assert result == ""
 
     def test_flatten_tool_result_nested_image(self):
         """Tool result containing image blocks uses placeholder."""
@@ -1004,7 +1010,7 @@ def _make_sdk_patches(
         (f"{_SVC}.create_security_hooks", dict(return_value=MagicMock())),
         (f"{_SVC}.get_copilot_tool_names", dict(return_value=[])),
         (f"{_SVC}.get_sdk_disallowed_tools", dict(return_value=[])),
-        (f"{_SVC}._build_sdk_env", dict(return_value=None)),
+        (f"{_SVC}.build_sdk_env", dict(return_value={})),
         (f"{_SVC}._resolve_sdk_model", dict(return_value=None)),
         (f"{_SVC}.set_execution_context", {}),
         (
@@ -1408,3 +1414,261 @@ class TestStreamChatCompletionRetryIntegration:
         # Verify user-friendly message (not raw SDK text)
         assert "Authentication" in errors[0].errorText
         assert any(isinstance(e, StreamStart) for e in events)
+
+    @pytest.mark.asyncio
+    async def test_result_message_prompt_too_long_triggers_compaction(self):
+        """CLI returns ResultMessage(subtype="error") with "Prompt is too long".
+
+        When the Claude CLI rejects the prompt pre-API (model=<synthetic>,
+        duration_api_ms=0), it sends a ResultMessage with is_error=True
+        instead of raising a Python exception.  The retry loop must still
+        detect this as a context-length error and trigger compaction.
+        """
+        import contextlib
+
+        from claude_agent_sdk import ResultMessage
+
+        from backend.copilot.response_model import StreamError, StreamStart
+        from backend.copilot.sdk.service import stream_chat_completion_sdk
+
+        session = self._make_session()
+        success_result = self._make_result_message()
+        attempt_count = [0]
+
+        error_result = ResultMessage(
+            subtype="error",
+            result="Prompt is too long",
+            duration_ms=100,
+            duration_api_ms=0,
+            is_error=True,
+            num_turns=0,
+            session_id="test-session-id",
+        )
+
+        def _client_factory(*args, **kwargs):
+            attempt_count[0] += 1
+            if attempt_count[0] == 1:
+                # First attempt: CLI returns error ResultMessage
+                return self._make_client_mock(result_message=error_result)
+            # Second attempt (after compaction): succeeds
+            return self._make_client_mock(result_message=success_result)
+
+        original_transcript = _build_transcript(
+            [("user", "prior question"), ("assistant", "prior answer")]
+        )
+        compacted_transcript = _build_transcript(
+            [("user", "[summary]"), ("assistant", "summary reply")]
+        )
+
+        patches = _make_sdk_patches(
+            session,
+            original_transcript=original_transcript,
+            compacted_transcript=compacted_transcript,
+            client_side_effect=_client_factory,
+        )
+
+        events = []
+        with contextlib.ExitStack() as stack:
+            for target, kwargs in patches:
+                stack.enter_context(patch(target, **kwargs))
+            async for event in stream_chat_completion_sdk(
+                session_id="test-session-id",
+                message="hello",
+                is_user_message=True,
+                user_id="test-user",
+                session=session,
+            ):
+                events.append(event)
+
+        assert attempt_count[0] == 2, (
+            f"Expected 2 SDK attempts (CLI error ResultMessage "
+            f"should trigger compaction retry), got {attempt_count[0]}"
+        )
+        errors = [e for e in events if isinstance(e, StreamError)]
+        assert not errors, f"Unexpected StreamError: {errors}"
+        assert any(isinstance(e, StreamStart) for e in events)
+
+    @pytest.mark.asyncio
+    async def test_result_message_success_subtype_prompt_too_long_triggers_compaction(
+        self,
+    ):
+        """CLI returns ResultMessage(subtype="success") with result="Prompt is too long".
+
+        The SDK internally compacts but the transcript is still too long.  It
+        returns subtype="success" (process completed) with result="Prompt is
+        too long" (the actual rejection message).  The retry loop must detect
+        this as a context-length error and trigger compaction — the subtype
+        "success" must not fool it into treating this as a real response.
+        """
+        import contextlib
+
+        from claude_agent_sdk import ResultMessage
+
+        from backend.copilot.response_model import StreamError, StreamStart
+        from backend.copilot.sdk.service import stream_chat_completion_sdk
+
+        session = self._make_session()
+        success_result = self._make_result_message()
+        attempt_count = [0]
+
+        error_result = ResultMessage(
+            subtype="success",
+            result="Prompt is too long",
+            duration_ms=100,
+            duration_api_ms=0,
+            is_error=False,
+            num_turns=1,
+            session_id="test-session-id",
+        )
+
+        def _client_factory(*args, **kwargs):
+            attempt_count[0] += 1
+
+            async def _receive_error():
+                yield error_result
+
+            async def _receive_success():
+                yield success_result
+
+            client = MagicMock()
+            client._transport = MagicMock()
+            client._transport.write = AsyncMock()
+            client.query = AsyncMock()
+            if attempt_count[0] == 1:
+                client.receive_response = _receive_error
+            else:
+                client.receive_response = _receive_success
+            cm = AsyncMock()
+            cm.__aenter__.return_value = client
+            cm.__aexit__.return_value = None
+            return cm
+
+        original_transcript = _build_transcript(
+            [("user", "prior question"), ("assistant", "prior answer")]
+        )
+        compacted_transcript = _build_transcript(
+            [("user", "[summary]"), ("assistant", "summary reply")]
+        )
+
+        patches = _make_sdk_patches(
+            session,
+            original_transcript=original_transcript,
+            compacted_transcript=compacted_transcript,
+            client_side_effect=_client_factory,
+        )
+
+        events = []
+        with contextlib.ExitStack() as stack:
+            for target, kwargs in patches:
+                stack.enter_context(patch(target, **kwargs))
+            async for event in stream_chat_completion_sdk(
+                session_id="test-session-id",
+                message="hello",
+                is_user_message=True,
+                user_id="test-user",
+                session=session,
+            ):
+                events.append(event)
+
+        assert attempt_count[0] == 2, (
+            f"Expected 2 SDK attempts (subtype='success' with 'Prompt is too long' "
+            f"result should trigger compaction retry), got {attempt_count[0]}"
+        )
+        errors = [e for e in events if isinstance(e, StreamError)]
+        assert not errors, f"Unexpected StreamError: {errors}"
+        assert any(isinstance(e, StreamStart) for e in events)
+
+    @pytest.mark.asyncio
+    async def test_assistant_message_error_content_prompt_too_long_triggers_compaction(
+        self,
+    ):
+        """AssistantMessage.error="invalid_request" with content "Prompt is too long".
+
+        The SDK returns error type "invalid_request" but puts the actual
+        rejection message ("Prompt is too long") in the content blocks.
+        The retry loop must detect this via content inspection (sdk_error
+        being set confirms it's an error message, not user content).
+        """
+        import contextlib
+
+        from claude_agent_sdk import AssistantMessage, ResultMessage, TextBlock
+
+        from backend.copilot.response_model import StreamError, StreamStart
+        from backend.copilot.sdk.service import stream_chat_completion_sdk
+
+        session = self._make_session()
+        success_result = self._make_result_message()
+        attempt_count = [0]
+
+        def _client_factory(*args, **kwargs):
+            attempt_count[0] += 1
+
+            async def _receive_error():
+                # SDK returns invalid_request with "Prompt is too long" in content.
+                # ResultMessage.result is a non-PTL value ("done") to isolate
+                # the AssistantMessage content detection path exclusively.
+                yield AssistantMessage(
+                    content=[TextBlock(text="Prompt is too long")],
+                    model="<synthetic>",
+                    error="invalid_request",
+                )
+                yield ResultMessage(
+                    subtype="success",
+                    result="done",
+                    duration_ms=100,
+                    duration_api_ms=0,
+                    is_error=False,
+                    num_turns=1,
+                    session_id="test-session-id",
+                )
+
+            async def _receive_success():
+                yield success_result
+
+            client = MagicMock()
+            client._transport = MagicMock()
+            client._transport.write = AsyncMock()
+            client.query = AsyncMock()
+            if attempt_count[0] == 1:
+                client.receive_response = _receive_error
+            else:
+                client.receive_response = _receive_success
+            cm = AsyncMock()
+            cm.__aenter__.return_value = client
+            cm.__aexit__.return_value = None
+            return cm
+
+        original_transcript = _build_transcript(
+            [("user", "prior question"), ("assistant", "prior answer")]
+        )
+        compacted_transcript = _build_transcript(
+            [("user", "[summary]"), ("assistant", "summary reply")]
+        )
+
+        patches = _make_sdk_patches(
+            session,
+            original_transcript=original_transcript,
+            compacted_transcript=compacted_transcript,
+            client_side_effect=_client_factory,
+        )
+
+        events = []
+        with contextlib.ExitStack() as stack:
+            for target, kwargs in patches:
+                stack.enter_context(patch(target, **kwargs))
+            async for event in stream_chat_completion_sdk(
+                session_id="test-session-id",
+                message="hello",
+                is_user_message=True,
+                user_id="test-user",
+                session=session,
+            ):
+                events.append(event)
+
+        assert attempt_count[0] == 2, (
+            f"Expected 2 SDK attempts (AssistantMessage error content 'Prompt is "
+            f"too long' should trigger compaction retry), got {attempt_count[0]}"
+        )
+        errors = [e for e in events if isinstance(e, StreamError)]
+        assert not errors, f"Unexpected StreamError: {errors}"
+        assert any(isinstance(e, StreamStart) for e in events)

autogpt_platform/backend/backend/copilot/sdk/security_hooks.py

@@ -22,6 +22,38 @@ from .tool_adapter import (
 
 logger = logging.getLogger(__name__)
 
+# The SDK CLI uses "Task" in older versions and "Agent" in v2.x+.
+# Shared across all sessions — used by security hooks for sub-agent detection.
+_SUBAGENT_TOOLS: frozenset[str] = frozenset({"Task", "Agent"})
+
+# Unicode ranges stripped by _sanitize():
+#   - BiDi overrides (U+202A-U+202E, U+2066-U+2069) can trick reviewers
+#     into misreading code/logs.
+#   - Zero-width characters (U+200B-U+200F, U+FEFF) can hide content.
+_BIDI_AND_ZW_CHARS = set(
+    chr(c)
+    for r in (range(0x202A, 0x202F), range(0x2066, 0x206A), range(0x200B, 0x2010))
+    for c in r
+) | {"\ufeff"}
+
+
+def _sanitize(value: str, max_len: int = 200) -> str:
+    """Strip control characters and truncate for safe logging.
+
+    Removes C0 (U+0000-U+001F), DEL (U+007F), C1 (U+0080-U+009F),
+    Unicode BiDi overrides, and zero-width characters to prevent
+    log injection and visual spoofing.
+    """
+    cleaned = "".join(
+        c
+        for c in value
+        if c >= " "
+        and c != "\x7f"
+        and not ("\x80" <= c <= "\x9f")
+        and c not in _BIDI_AND_ZW_CHARS
+    )
+    return cleaned[:max_len]
+
 
 def _deny(reason: str) -> dict[str, Any]:
     """Return a hook denial response."""
@@ -136,11 +168,13 @@ def create_security_hooks(
     - PostToolUse: Log successful tool executions
     - PostToolUseFailure: Log and handle failed tool executions
     - PreCompact: Log context compaction events (SDK handles compaction automatically)
+    - SubagentStart: Log sub-agent lifecycle start
+    - SubagentStop: Log sub-agent lifecycle end
 
     Args:
         user_id: Current user ID for isolation validation
         sdk_cwd: SDK working directory for workspace-scoped tool validation
-        max_subtasks: Maximum concurrent Task (sub-agent) spawns allowed per session
+        max_subtasks: Maximum concurrent sub-agent spawns allowed per session
         on_compact: Callback invoked when SDK starts compacting context.
             Receives the transcript_path from the hook input.
 
@@ -151,9 +185,19 @@ def create_security_hooks(
         from claude_agent_sdk import HookMatcher
         from claude_agent_sdk.types import HookContext, HookInput, SyncHookJSONOutput
 
-        # Per-session tracking for Task sub-agent concurrency.
+        # Per-session tracking for sub-agent concurrency.
         # Set of tool_use_ids that consumed a slot — len() is the active count.
-        task_tool_use_ids: set[str] = set()
+        #
+        # LIMITATION: For background (async) agents the SDK returns the
+        # Agent/Task tool immediately with {isAsync: true}, which triggers
+        # PostToolUse and releases the slot while the agent is still running.
+        # SubagentStop fires later when the background process finishes but
+        # does not currently hold a slot.  This means the concurrency limit
+        # only gates *launches*, not true concurrent execution.  To fix this
+        # we would need to track background agent_ids separately and release
+        # in SubagentStop, but the SDK does not guarantee SubagentStop fires
+        # for every background agent (e.g. on session abort).
+        subagent_tool_use_ids: set[str] = set()
 
         async def pre_tool_use_hook(
             input_data: HookInput,
@@ -165,29 +209,22 @@ def create_security_hooks(
             tool_name = cast(str, input_data.get("tool_name", ""))
             tool_input = cast(dict[str, Any], input_data.get("tool_input", {}))
 
-            # Rate-limit Task (sub-agent) spawns per session
-            if tool_name == "Task":
-                # Block background task execution first — denied calls
-                # should not consume a subtask slot.
-                if tool_input.get("run_in_background"):
-                    logger.info(f"[SDK] Blocked background Task, user={user_id}")
-                    return cast(
-                        SyncHookJSONOutput,
-                        _deny(
-                            "Background task execution is not supported. "
-                            "Run tasks in the foreground instead "
-                            "(remove the run_in_background parameter)."
-                        ),
-                    )
-                if len(task_tool_use_ids) >= max_subtasks:
+            # Rate-limit sub-agent spawns per session.
+            # The SDK CLI renamed "Task" → "Agent" in v2.x; handle both.
+            if tool_name in _SUBAGENT_TOOLS:
+                # Background agents are allowed — the SDK returns immediately
+                # with {isAsync: true} and the model polls via TaskOutput.
+                # Still count them against the concurrency limit.
+                if len(subagent_tool_use_ids) >= max_subtasks:
                     logger.warning(
-                        f"[SDK] Task limit reached ({max_subtasks}), user={user_id}"
+                        f"[SDK] Sub-agent limit reached ({max_subtasks}), "
+                        f"user={user_id}"
                     )
                     return cast(
                         SyncHookJSONOutput,
                         _deny(
-                            f"Maximum {max_subtasks} concurrent sub-tasks. "
-                            "Wait for running sub-tasks to finish, "
+                            f"Maximum {max_subtasks} concurrent sub-agents. "
+                            "Wait for running sub-agents to finish, "
                             "or continue in the main conversation."
                         ),
                     )
@@ -208,20 +245,20 @@ def create_security_hooks(
             if result:
                 return cast(SyncHookJSONOutput, result)
 
-            # Reserve the Task slot only after all validations pass
-            if tool_name == "Task" and tool_use_id is not None:
-                task_tool_use_ids.add(tool_use_id)
+            # Reserve the sub-agent slot only after all validations pass
+            if tool_name in _SUBAGENT_TOOLS and tool_use_id is not None:
+                subagent_tool_use_ids.add(tool_use_id)
 
             logger.debug(f"[SDK] Tool start: {tool_name}, user={user_id}")
             return cast(SyncHookJSONOutput, {})
 
-        def _release_task_slot(tool_name: str, tool_use_id: str | None) -> None:
-            """Release a Task concurrency slot if one was reserved."""
-            if tool_name == "Task" and tool_use_id in task_tool_use_ids:
-                task_tool_use_ids.discard(tool_use_id)
+        def _release_subagent_slot(tool_name: str, tool_use_id: str | None) -> None:
+            """Release a sub-agent concurrency slot if one was reserved."""
+            if tool_name in _SUBAGENT_TOOLS and tool_use_id in subagent_tool_use_ids:
+                subagent_tool_use_ids.discard(tool_use_id)
                 logger.info(
-                    "[SDK] Task slot released, active=%d/%d, user=%s",
-                    len(task_tool_use_ids),
+                    "[SDK] Sub-agent slot released, active=%d/%d, user=%s",
+                    len(subagent_tool_use_ids),
                     max_subtasks,
                     user_id,
                 )
@@ -241,13 +278,14 @@ def create_security_hooks(
             _ = context
             tool_name = cast(str, input_data.get("tool_name", ""))
 
-            _release_task_slot(tool_name, tool_use_id)
+            _release_subagent_slot(tool_name, tool_use_id)
             is_builtin = not tool_name.startswith(MCP_TOOL_PREFIX)
+            safe_tool_use_id = _sanitize(str(tool_use_id or ""), max_len=12)
             logger.info(
                 "[SDK] PostToolUse: %s (builtin=%s, tool_use_id=%s)",
                 tool_name,
                 is_builtin,
-                (tool_use_id or "")[:12],
+                safe_tool_use_id,
             )
 
             # Stash output for SDK built-in tools so the response adapter can
@@ -256,7 +294,7 @@ def create_security_hooks(
             if is_builtin:
                 tool_response = input_data.get("tool_response")
                 if tool_response is not None:
-                    resp_preview = str(tool_response)[:100]
+                    resp_preview = _sanitize(str(tool_response), max_len=100)
                     logger.info(
                         "[SDK] Stashing builtin output for %s (%d chars): %s...",
                         tool_name,
@@ -280,13 +318,17 @@ def create_security_hooks(
             """Log failed tool executions for debugging."""
             _ = context
             tool_name = cast(str, input_data.get("tool_name", ""))
-            error = input_data.get("error", "Unknown error")
+            error = _sanitize(str(input_data.get("error", "Unknown error")))
+            safe_tool_use_id = _sanitize(str(tool_use_id or ""))
             logger.warning(
-                f"[SDK] Tool failed: {tool_name}, error={error}, "
-                f"user={user_id}, tool_use_id={tool_use_id}"
+                "[SDK] Tool failed: %s, error=%s, user=%s, tool_use_id=%s",
+                tool_name,
+                error,
+                user_id,
+                safe_tool_use_id,
             )
 
-            _release_task_slot(tool_name, tool_use_id)
+            _release_subagent_slot(tool_name, tool_use_id)
 
             return cast(SyncHookJSONOutput, {})
 
@@ -301,20 +343,17 @@ def create_security_hooks(
             This hook provides visibility into when compaction happens.
             """
             _ = context, tool_use_id
-            trigger = input_data.get("trigger", "auto")
+            trigger = _sanitize(str(input_data.get("trigger", "auto")), max_len=50)
             # Sanitize untrusted input: strip control chars for logging AND
             # for the value passed downstream.  read_compacted_entries()
             # validates against _projects_base() as defence-in-depth, but
             # sanitizing here prevents log injection and rejects obviously
             # malformed paths early.
-            transcript_path = (
-                str(input_data.get("transcript_path", ""))
-                .replace("\n", "")
-                .replace("\r", "")
+            transcript_path = _sanitize(
+                str(input_data.get("transcript_path", "")), max_len=500
             )
             logger.info(
-                "[SDK] Context compaction triggered: %s, user=%s, "
-                "transcript_path=%s",
+                "[SDK] Context compaction triggered: %s, user=%s, transcript_path=%s",
                 trigger,
                 user_id,
                 transcript_path,
@@ -323,6 +362,44 @@ def create_security_hooks(
                 on_compact(transcript_path)
             return cast(SyncHookJSONOutput, {})
 
+        async def subagent_start_hook(
+            input_data: HookInput,
+            tool_use_id: str | None,
+            context: HookContext,
+        ) -> SyncHookJSONOutput:
+            """Log when a sub-agent starts execution."""
+            _ = context, tool_use_id
+            agent_id = _sanitize(str(input_data.get("agent_id", "?")))
+            agent_type = _sanitize(str(input_data.get("agent_type", "?")))
+            logger.info(
+                "[SDK] SubagentStart: agent_id=%s, type=%s, user=%s",
+                agent_id,
+                agent_type,
+                user_id,
+            )
+            return cast(SyncHookJSONOutput, {})
+
+        async def subagent_stop_hook(
+            input_data: HookInput,
+            tool_use_id: str | None,
+            context: HookContext,
+        ) -> SyncHookJSONOutput:
+            """Log when a sub-agent stops."""
+            _ = context, tool_use_id
+            agent_id = _sanitize(str(input_data.get("agent_id", "?")))
+            agent_type = _sanitize(str(input_data.get("agent_type", "?")))
+            transcript = _sanitize(
+                str(input_data.get("agent_transcript_path", "")), max_len=500
+            )
+            logger.info(
+                "[SDK] SubagentStop: agent_id=%s, type=%s, user=%s, transcript=%s",
+                agent_id,
+                agent_type,
+                user_id,
+                transcript,
+            )
+            return cast(SyncHookJSONOutput, {})
+
         hooks: dict[str, Any] = {
             "PreToolUse": [HookMatcher(matcher="*", hooks=[pre_tool_use_hook])],
             "PostToolUse": [HookMatcher(matcher="*", hooks=[post_tool_use_hook])],
@@ -330,6 +407,8 @@ def create_security_hooks(
                 HookMatcher(matcher="*", hooks=[post_tool_failure_hook])
             ],
             "PreCompact": [HookMatcher(matcher="*", hooks=[pre_compact_hook])],
+            "SubagentStart": [HookMatcher(matcher="*", hooks=[subagent_start_hook])],
+            "SubagentStop": [HookMatcher(matcher="*", hooks=[subagent_stop_hook])],
         }
 
         return hooks

autogpt_platform/backend/backend/copilot/sdk/security_hooks_test.py

@@ -5,13 +5,18 @@ They validate that the security hooks correctly block unauthorized paths,
 tool access, and dangerous input patterns.
 """
 
+import logging
 import os
 
 import pytest
 
 from backend.copilot.context import _current_project_dir
 
-from .security_hooks import _validate_tool_access, _validate_user_isolation
+from .security_hooks import (
+    _validate_tool_access,
+    _validate_user_isolation,
+    create_security_hooks,
+)
 
 SDK_CWD = "/tmp/copilot-abc123"
 
@@ -132,8 +137,20 @@ def test_read_tool_results_allowed():
         _current_project_dir.reset(token)
 
 
+def test_read_tool_outputs_allowed():
+    """tool-outputs/ paths should be allowed, same as tool-results/."""
+    home = os.path.expanduser("~")
+    path = f"{home}/.claude/projects/-tmp-copilot-abc123/a1b2c3d4-e5f6-7890-abcd-ef1234567890/tool-outputs/12345.txt"
+    token = _current_project_dir.set("-tmp-copilot-abc123")
+    try:
+        result = _validate_tool_access("Read", {"file_path": path}, sdk_cwd=SDK_CWD)
+        assert result == {}
+    finally:
+        _current_project_dir.reset(token)
+
+
 def test_read_claude_projects_settings_json_denied():
-    """SDK-internal artifacts like settings.json are NOT accessible — only tool-results/ is."""
+    """SDK-internal artifacts like settings.json are NOT accessible — only tool-results/tool-outputs is."""
     home = os.path.expanduser("~")
     path = f"{home}/.claude/projects/-tmp-copilot-abc123/settings.json"
     token = _current_project_dir.set("-tmp-copilot-abc123")
@@ -220,8 +237,6 @@ def test_bash_builtin_blocked_message_clarity():
 @pytest.fixture()
 def _hooks():
     """Create security hooks and return (pre, post, post_failure) handlers."""
-    from .security_hooks import create_security_hooks
-
     hooks = create_security_hooks(user_id="u1", sdk_cwd=SDK_CWD, max_subtasks=2)
     pre = hooks["PreToolUse"][0].hooks[0]
     post = hooks["PostToolUse"][0].hooks[0]
@@ -231,16 +246,15 @@ def _hooks():
 
 @pytest.mark.skipif(not _sdk_available(), reason="claude_agent_sdk not installed")
 @pytest.mark.asyncio
-async def test_task_background_blocked(_hooks):
-    """Task with run_in_background=true must be denied."""
+async def test_task_background_allowed(_hooks):
+    """Task with run_in_background=true is allowed (SDK handles async lifecycle)."""
     pre, _, _ = _hooks
     result = await pre(
         {"tool_name": "Task", "tool_input": {"run_in_background": True, "prompt": "x"}},
-        tool_use_id=None,
+        tool_use_id="tu-bg-1",
         context={},
     )
-    assert _is_denied(result)
-    assert "foreground" in _reason(result).lower()
+    assert not _is_denied(result)
 
 
 @pytest.mark.skipif(not _sdk_available(), reason="claude_agent_sdk not installed")
@@ -354,3 +368,303 @@ async def test_task_slot_released_on_failure(_hooks):
         context={},
     )
     assert not _is_denied(result)
+
+
+# ---------------------------------------------------------------------------
+# "Agent" tool name (SDK v2.x+ renamed "Task" → "Agent")
+# ---------------------------------------------------------------------------
+
+
+@pytest.mark.skipif(not _sdk_available(), reason="claude_agent_sdk not installed")
+@pytest.mark.asyncio
+async def test_agent_background_allowed(_hooks):
+    """Agent with run_in_background=true is allowed (SDK handles async lifecycle)."""
+    pre, _, _ = _hooks
+    result = await pre(
+        {
+            "tool_name": "Agent",
+            "tool_input": {"run_in_background": True, "prompt": "x"},
+        },
+        tool_use_id="tu-agent-bg-1",
+        context={},
+    )
+    assert not _is_denied(result)
+
+
+@pytest.mark.skipif(not _sdk_available(), reason="claude_agent_sdk not installed")
+@pytest.mark.asyncio
+async def test_agent_foreground_allowed(_hooks):
+    """Agent without run_in_background should be allowed."""
+    pre, _, _ = _hooks
+    result = await pre(
+        {"tool_name": "Agent", "tool_input": {"prompt": "do stuff"}},
+        tool_use_id="tu-agent-1",
+        context={},
+    )
+    assert not _is_denied(result)
+
+
+@pytest.mark.skipif(not _sdk_available(), reason="claude_agent_sdk not installed")
+@pytest.mark.asyncio
+async def test_background_agent_counts_against_limit(_hooks):
+    """Background agents still consume concurrency slots."""
+    pre, _, _ = _hooks
+    # Two background agents fill the limit
+    for i in range(2):
+        result = await pre(
+            {
+                "tool_name": "Agent",
+                "tool_input": {"run_in_background": True, "prompt": "bg"},
+            },
+            tool_use_id=f"tu-bglimit-{i}",
+            context={},
+        )
+        assert not _is_denied(result)
+    # Third (background or foreground) should be denied
+    result = await pre(
+        {
+            "tool_name": "Agent",
+            "tool_input": {"run_in_background": True, "prompt": "over"},
+        },
+        tool_use_id="tu-bglimit-2",
+        context={},
+    )
+    assert _is_denied(result)
+    assert "Maximum" in _reason(result)
+
+
+@pytest.mark.skipif(not _sdk_available(), reason="claude_agent_sdk not installed")
+@pytest.mark.asyncio
+async def test_agent_limit_enforced(_hooks):
+    """Agent spawns beyond max_subtasks should be denied."""
+    pre, _, _ = _hooks
+    # First two should pass
+    for i in range(2):
+        result = await pre(
+            {"tool_name": "Agent", "tool_input": {"prompt": "ok"}},
+            tool_use_id=f"tu-agent-limit-{i}",
+            context={},
+        )
+        assert not _is_denied(result)
+
+    # Third should be denied (limit=2)
+    result = await pre(
+        {"tool_name": "Agent", "tool_input": {"prompt": "over limit"}},
+        tool_use_id="tu-agent-limit-2",
+        context={},
+    )
+    assert _is_denied(result)
+    assert "Maximum" in _reason(result)
+
+
+@pytest.mark.skipif(not _sdk_available(), reason="claude_agent_sdk not installed")
+@pytest.mark.asyncio
+async def test_agent_slot_released_on_completion(_hooks):
+    """Completing an Agent should free a slot so new Agents can be spawned."""
+    pre, post, _ = _hooks
+    # Fill both slots
+    for i in range(2):
+        result = await pre(
+            {"tool_name": "Agent", "tool_input": {"prompt": "ok"}},
+            tool_use_id=f"tu-agent-comp-{i}",
+            context={},
+        )
+        assert not _is_denied(result)
+
+    # Third should be denied — at capacity
+    result = await pre(
+        {"tool_name": "Agent", "tool_input": {"prompt": "over"}},
+        tool_use_id="tu-agent-comp-2",
+        context={},
+    )
+    assert _is_denied(result)
+
+    # Complete first agent — frees a slot
+    await post(
+        {"tool_name": "Agent", "tool_input": {}},
+        tool_use_id="tu-agent-comp-0",
+        context={},
+    )
+
+    # Now a new Agent should be allowed
+    result = await pre(
+        {"tool_name": "Agent", "tool_input": {"prompt": "after release"}},
+        tool_use_id="tu-agent-comp-3",
+        context={},
+    )
+    assert not _is_denied(result)
+
+
+@pytest.mark.skipif(not _sdk_available(), reason="claude_agent_sdk not installed")
+@pytest.mark.asyncio
+async def test_agent_slot_released_on_failure(_hooks):
+    """A failed Agent should also free its concurrency slot."""
+    pre, _, post_failure = _hooks
+    # Fill both slots
+    for i in range(2):
+        result = await pre(
+            {"tool_name": "Agent", "tool_input": {"prompt": "ok"}},
+            tool_use_id=f"tu-agent-fail-{i}",
+            context={},
+        )
+        assert not _is_denied(result)
+
+    # At capacity
+    result = await pre(
+        {"tool_name": "Agent", "tool_input": {"prompt": "over"}},
+        tool_use_id="tu-agent-fail-2",
+        context={},
+    )
+    assert _is_denied(result)
+
+    # Fail first agent — should free a slot
+    await post_failure(
+        {"tool_name": "Agent", "tool_input": {}, "error": "something broke"},
+        tool_use_id="tu-agent-fail-0",
+        context={},
+    )
+
+    # New Agent should be allowed
+    result = await pre(
+        {"tool_name": "Agent", "tool_input": {"prompt": "after failure"}},
+        tool_use_id="tu-agent-fail-3",
+        context={},
+    )
+    assert not _is_denied(result)
+
+
+@pytest.mark.skipif(not _sdk_available(), reason="claude_agent_sdk not installed")
+@pytest.mark.asyncio
+async def test_mixed_task_agent_share_slots(_hooks):
+    """Task and Agent share the same concurrency pool."""
+    pre, post, _ = _hooks
+    # Fill one slot with Task, one with Agent
+    result = await pre(
+        {"tool_name": "Task", "tool_input": {"prompt": "ok"}},
+        tool_use_id="tu-mix-task",
+        context={},
+    )
+    assert not _is_denied(result)
+
+    result = await pre(
+        {"tool_name": "Agent", "tool_input": {"prompt": "ok"}},
+        tool_use_id="tu-mix-agent",
+        context={},
+    )
+    assert not _is_denied(result)
+
+    # Third (either name) should be denied
+    result = await pre(
+        {"tool_name": "Agent", "tool_input": {"prompt": "over"}},
+        tool_use_id="tu-mix-over",
+        context={},
+    )
+    assert _is_denied(result)
+
+    # Release the Task slot
+    await post(
+        {"tool_name": "Task", "tool_input": {}},
+        tool_use_id="tu-mix-task",
+        context={},
+    )
+
+    # Now an Agent should be allowed
+    result = await pre(
+        {"tool_name": "Agent", "tool_input": {"prompt": "after task release"}},
+        tool_use_id="tu-mix-new",
+        context={},
+    )
+    assert not _is_denied(result)
+
+
+# ---------------------------------------------------------------------------
+# SubagentStart / SubagentStop hooks
+# ---------------------------------------------------------------------------
+
+
+@pytest.fixture()
+def _subagent_hooks():
+    """Create hooks and return (subagent_start, subagent_stop) handlers."""
+    hooks = create_security_hooks(user_id="u1", sdk_cwd=SDK_CWD, max_subtasks=2)
+    start = hooks["SubagentStart"][0].hooks[0]
+    stop = hooks["SubagentStop"][0].hooks[0]
+    return start, stop
+
+
+@pytest.mark.skipif(not _sdk_available(), reason="claude_agent_sdk not installed")
+@pytest.mark.asyncio
+async def test_subagent_start_hook_returns_empty(_subagent_hooks):
+    """SubagentStart hook should return an empty dict (logging only)."""
+    start, _ = _subagent_hooks
+    result = await start(
+        {"agent_id": "sa-123", "agent_type": "research"},
+        tool_use_id=None,
+        context={},
+    )
+    assert result == {}
+
+
+@pytest.mark.skipif(not _sdk_available(), reason="claude_agent_sdk not installed")
+@pytest.mark.asyncio
+async def test_subagent_stop_hook_returns_empty(_subagent_hooks):
+    """SubagentStop hook should return an empty dict (logging only)."""
+    _, stop = _subagent_hooks
+    result = await stop(
+        {
+            "agent_id": "sa-123",
+            "agent_type": "research",
+            "agent_transcript_path": "/tmp/transcript.txt",
+        },
+        tool_use_id=None,
+        context={},
+    )
+    assert result == {}
+
+
+@pytest.mark.skipif(not _sdk_available(), reason="claude_agent_sdk not installed")
+@pytest.mark.asyncio
+async def test_subagent_hooks_sanitize_inputs(_subagent_hooks, caplog):
+    """SubagentStart/Stop should sanitize control chars from inputs."""
+    start, stop = _subagent_hooks
+    # Inject control characters (C0, DEL, C1, BiDi overrides, zero-width)
+    # — hook should not raise AND logs must be clean
+    with caplog.at_level(logging.DEBUG, logger="backend.copilot.sdk.security_hooks"):
+        result = await start(
+            {
+                "agent_id": "sa\n-injected\r\x00\x7f",
+                "agent_type": "safe\x80_type\x9f\ttab",
+            },
+            tool_use_id=None,
+            context={},
+        )
+    assert result == {}
+    # Control chars must be stripped from the logged values
+    for record in caplog.records:
+        assert "\x00" not in record.message
+        assert "\r" not in record.message
+        assert "\n" not in record.message
+        assert "\x7f" not in record.message
+        assert "\x80" not in record.message
+        assert "\x9f" not in record.message
+    assert "safe_type" in caplog.text
+
+    caplog.clear()
+    with caplog.at_level(logging.DEBUG, logger="backend.copilot.sdk.security_hooks"):
+        result = await stop(
+            {
+                "agent_id": "sa\n-injected\x7f",
+                "agent_type": "type\r\x80\x9f",
+                "agent_transcript_path": "/tmp/\x00malicious\npath\u202a\u200b",
+            },
+            tool_use_id=None,
+            context={},
+        )
+    assert result == {}
+    for record in caplog.records:
+        assert "\x00" not in record.message
+        assert "\r" not in record.message
+        assert "\n" not in record.message
+        assert "\x7f" not in record.message
+        assert "\u202a" not in record.message
+        assert "\u200b" not in record.message
+    assert "/tmp/maliciouspath" in caplog.text

autogpt_platform/backend/backend/copilot/sdk/service.py

@@ -59,11 +59,14 @@ from ..response_model import (
     StreamBaseResponse,
     StreamError,
     StreamFinish,
+    StreamFinishStep,
     StreamHeartbeat,
     StreamStart,
+    StreamStartStep,
     StreamStatus,
     StreamTextDelta,
     StreamToolInputAvailable,
+    StreamToolInputStart,
     StreamToolOutputAvailable,
     StreamUsage,
 )
@@ -77,15 +80,13 @@ from ..tools.e2b_sandbox import get_or_create_sandbox, pause_sandbox_direct
 from ..tools.sandbox import WORKSPACE_PREFIX, make_session_path
 from ..tracking import track_user_message
 from .compaction import CompactionTracker, filter_compaction_messages
+from .env import build_sdk_env  # noqa: F401 — re-export for backward compat
 from .response_adapter import SDKResponseAdapter
 from .security_hooks import create_security_hooks
-from .subscription import validate_subscription as _validate_claude_code_subscription
 from .tool_adapter import (
-    cancel_pending_tool_tasks,
     create_copilot_mcp_server,
     get_copilot_tool_names,
     get_sdk_disallowed_tools,
-    pre_launch_tool_call,
     reset_stash_event,
     reset_tool_failure_counters,
     set_execution_context,
@@ -115,9 +116,10 @@ _MAX_STREAM_ATTEMPTS = 3
 
 # Hard circuit breaker: abort the stream if the model sends this many
 # consecutive tool calls with empty parameters (a sign of context
-# saturation or serialization failure).  Empty input ({}) is never
-# legitimate — even one is suspicious, three is conclusive.
-_EMPTY_TOOL_CALL_LIMIT = 3
+# saturation or serialization failure).  The MCP wrapper now returns
+# guidance on the first empty call, giving the model a chance to
+# self-correct.  The limit is generous to allow recovery attempts.
+_EMPTY_TOOL_CALL_LIMIT = 5
 
 # User-facing error shown when the empty-tool-call circuit breaker trips.
 _CIRCUIT_BREAKER_ERROR_MSG = (
@@ -185,6 +187,24 @@ def _is_prompt_too_long(err: BaseException) -> bool:
     return False
 
 
+def _is_sdk_disconnect_error(exc: BaseException) -> bool:
+    """Return True if *exc* is an expected SDK cleanup error from client disconnect.
+
+    Two known patterns occur when ``GeneratorExit`` tears down the async
+    generator and the SDK's ``__aexit__`` runs in a different context/task:
+
+    * ``RuntimeError``: cancel scope exited in wrong task (anyio)
+    * ``ValueError``: ContextVar token created in a different Context (OTEL)
+
+    These are suppressed to avoid polluting Sentry with non-actionable noise.
+    """
+    if isinstance(exc, RuntimeError) and "cancel scope" in str(exc):
+        return True
+    if isinstance(exc, ValueError) and "was created in a different Context" in str(exc):
+        return True
+    return False
+
+
 def _is_tool_only_message(sdk_msg: object) -> bool:
     """Return True if *sdk_msg* is an AssistantMessage containing only ToolUseBlocks.
 
@@ -409,6 +429,63 @@ _HEARTBEAT_INTERVAL = 10.0  # seconds
 STREAM_LOCK_PREFIX = "copilot:stream:lock:"
 
 
+async def _safe_close_sdk_client(
+    sdk_client: ClaudeSDKClient,
+    log_prefix: str,
+) -> None:
+    """Close a ClaudeSDKClient, suppressing errors from client disconnect.
+
+    When the SSE client disconnects mid-stream, ``GeneratorExit`` propagates
+    through the async generator stack and causes ``ClaudeSDKClient.__aexit__``
+    to run in a different async context or task than where the client was
+    opened.  This triggers two known error classes:
+
+    * ``ValueError``: ``<Token var=<ContextVar name='current_context'>>
+      was created in a different Context`` — OpenTelemetry's
+      ``context.detach()`` fails because the OTEL context token was
+      created in the original generator coroutine but detach runs in
+      the GC / cleanup coroutine (Sentry: AUTOGPT-SERVER-8BT).
+
+    * ``RuntimeError``: ``Attempted to exit cancel scope in a different
+      task than it was entered in`` — anyio's ``TaskGroup.__aexit__``
+      detects that the cancel scope was entered in one task but is
+      being exited in another (Sentry: AUTOGPT-SERVER-8BW).
+
+    Both are harmless — the TCP connection is already dead and no
+    resources leak.  Logging them at ``debug`` level keeps observability
+    without polluting Sentry.
+    """
+    try:
+        await sdk_client.__aexit__(None, None, None)
+    except (ValueError, RuntimeError) as exc:
+        if _is_sdk_disconnect_error(exc):
+            # Expected during client disconnect — suppress to avoid Sentry noise.
+            logger.debug(
+                "%s SDK client cleanup error suppressed (client disconnect): %s: %s",
+                log_prefix,
+                type(exc).__name__,
+                exc,
+            )
+        else:
+            raise
+    except GeneratorExit:
+        # GeneratorExit can propagate through __aexit__ — suppress it here
+        # since the generator is already being torn down.
+        logger.debug(
+            "%s SDK client cleanup GeneratorExit suppressed (client disconnect)",
+            log_prefix,
+        )
+    except Exception:
+        # Unexpected cleanup error — log at error level so Sentry captures it
+        # (via its logging integration), but don't propagate since we're in
+        # teardown and the caller cannot meaningfully handle this.
+        logger.error(
+            "%s Unexpected SDK client cleanup error",
+            log_prefix,
+            exc_info=True,
+        )
+
+
 async def _iter_sdk_messages(
     client: ClaudeSDKClient,
 ) -> AsyncGenerator[Any, None]:
@@ -492,60 +569,6 @@ def _resolve_sdk_model() -> str | None:
     return model
 
 
-def _build_sdk_env(
-    session_id: str | None = None,
-    user_id: str | None = None,
-) -> dict[str, str]:
-    """Build env vars for the SDK CLI subprocess.
-
-    Three modes (checked in order):
-    1. **Subscription** — clears all keys; CLI uses `claude login` auth.
-    2. **Direct Anthropic** — returns `{}`; subprocess inherits
-       `ANTHROPIC_API_KEY` from the parent environment.
-    3. **OpenRouter** (default) — overrides base URL and auth token to
-       route through the proxy, with Langfuse trace headers.
-    """
-    # --- Mode 1: Claude Code subscription auth ---
-    if config.use_claude_code_subscription:
-        _validate_claude_code_subscription()
-        return {
-            "ANTHROPIC_API_KEY": "",
-            "ANTHROPIC_AUTH_TOKEN": "",
-            "ANTHROPIC_BASE_URL": "",
-        }
-
-    # --- Mode 2: Direct Anthropic (no proxy hop) ---
-    # `openrouter_active` checks the flag *and* credential presence.
-    if not config.openrouter_active:
-        return {}
-
-    # --- Mode 3: OpenRouter proxy ---
-    # Strip /v1 suffix — SDK expects the base URL without a version path.
-    base = (config.base_url or "").rstrip("/")
-    if base.endswith("/v1"):
-        base = base[:-3]
-    env: dict[str, str] = {
-        "ANTHROPIC_BASE_URL": base,
-        "ANTHROPIC_AUTH_TOKEN": config.api_key or "",
-        "ANTHROPIC_API_KEY": "",  # force CLI to use AUTH_TOKEN
-    }
-
-    # Inject broadcast headers so OpenRouter forwards traces to Langfuse.
-    def _safe(v: str) -> str:
-        """Sanitise a header value: strip newlines/whitespace and cap length."""
-        return v.replace("\r", "").replace("\n", "").strip()[:128]
-
-    parts = []
-    if session_id:
-        parts.append(f"x-session-id: {_safe(session_id)}")
-    if user_id:
-        parts.append(f"x-user-id: {_safe(user_id)}")
-    if parts:
-        env["ANTHROPIC_CUSTOM_HEADERS"] = "\n".join(parts)
-
-    return env
-
-
 def _make_sdk_cwd(session_id: str) -> str:
     """Create a safe, session-specific working directory path.
 
@@ -595,7 +618,9 @@ def _format_sdk_content_blocks(blocks: list) -> list[dict[str, Any]]:
     """Convert SDK content blocks to transcript format.
 
     Handles TextBlock, ToolUseBlock, ToolResultBlock, and ThinkingBlock.
-    Unknown block types are logged and skipped.
+    Raw dicts (e.g. ``redacted_thinking`` blocks that the SDK may not have
+    a typed class for) are passed through verbatim to preserve them in the
+    transcript.  Unknown typed block objects are logged and skipped.
     """
     result: list[dict[str, Any]] = []
     for block in blocks or []:
@@ -627,6 +652,9 @@ def _format_sdk_content_blocks(blocks: list) -> list[dict[str, Any]]:
                     "signature": block.signature,
                 }
             )
+        elif isinstance(block, dict) and "type" in block:
+            # Preserve raw dict blocks (e.g. redacted_thinking) verbatim.
+            result.append(block)
         else:
             logger.warning(
                 f"[SDK] Unknown content block type: {type(block).__name__}. "
@@ -720,15 +748,11 @@ def _format_conversation_context(messages: list[ChatMessage]) -> str | None:
         elif msg.role == "assistant":
             if msg.content:
                 lines.append(f"You responded: {msg.content}")
-            if msg.tool_calls:
-                for tc in msg.tool_calls:
-                    func = tc.get("function", {})
-                    tool_name = func.get("name", "unknown")
-                    tool_args = func.get("arguments", "")
-                    lines.append(f"You called tool: {tool_name}({tool_args})")
+            # Omit tool_calls — any text representation gets mimicked
+            # by the model. Tool results below provide the context.
         elif msg.role == "tool":
             content = msg.content or ""
-            lines.append(f"Tool result: {content}")
+            lines.append(f"Tool output: {content[:500]}")
 
     if not lines:
         return None
@@ -1188,7 +1212,25 @@ async def _run_stream_attempt(
 
     consecutive_empty_tool_calls = 0
 
-    async with ClaudeSDKClient(options=state.options) as client:
+    # --- Intermediate persistence tracking ---
+    # Flush session messages to DB periodically so page reloads show progress
+    # during long-running turns (see incident d2f7cba3: 82-min turn lost on refresh).
+    _last_flush_time = time.monotonic()
+    _msgs_since_flush = 0
+    _FLUSH_INTERVAL_SECONDS = 30.0
+    _FLUSH_MESSAGE_THRESHOLD = 10
+
+    # Use manual __aenter__/__aexit__ instead of ``async with`` so we can
+    # suppress SDK cleanup errors that occur when the SSE client disconnects
+    # mid-stream.  GeneratorExit causes the SDK's ``__aexit__`` to run in a
+    # different async context/task than where the client was opened, which
+    # triggers:
+    #   - ValueError: ContextVar token mismatch (AUTOGPT-SERVER-8BT)
+    #   - RuntimeError: cancel scope in wrong task  (AUTOGPT-SERVER-8BW)
+    # Both are harmless — the TCP connection is already dead.
+    sdk_client = ClaudeSDKClient(options=state.options)
+    client = await sdk_client.__aenter__()
+    try:
         logger.info(
             "%s Sending query — resume=%s, total_msgs=%d, "
             "query_len=%d, attached_files=%d, image_blocks=%d",
@@ -1264,6 +1306,27 @@ async def _run_stream_attempt(
                     error_preview,
                 )
 
+                # Intercept prompt-too-long errors surfaced as
+                # AssistantMessage.error (not as a Python exception).
+                # Re-raise so the outer retry loop can compact the
+                # transcript and retry with reduced context.
+                # Check both error_text and error_preview: sdk_error
+                # being set confirms this is an error message (not user
+                # content), so checking content is safe. The actual
+                # error description (e.g. "Prompt is too long") may be
+                # in the content, not the error type field
+                # (e.g. error="invalid_request", content="Prompt is
+                # too long").
+                if _is_prompt_too_long(Exception(error_text)) or _is_prompt_too_long(
+                    Exception(error_preview)
+                ):
+                    logger.warning(
+                        "%s Prompt-too-long detected via AssistantMessage "
+                        "error — raising for retry",
+                        ctx.log_prefix,
+                    )
+                    raise RuntimeError("Prompt is too long")
+
                 # Intercept transient API errors (socket closed,
                 # ECONNRESET) — replace the raw message with a
                 # user-friendly error text and use the retryable
@@ -1291,28 +1354,17 @@ async def _run_stream_attempt(
                     ended_with_stream_error = True
                     break
 
-            # Parallel tool execution: pre-launch every ToolUseBlock as an
-            # asyncio.Task the moment its AssistantMessage arrives.  The SDK
-            # sends one AssistantMessage per tool call when issuing parallel
-            # calls, so each message is pre-launched independently.  The MCP
-            # handlers will await the already-running task instead of executing
-            # fresh, making all concurrent tool calls run in parallel.
-            #
-            # Also determine if the message is a tool-only batch (all content
+            # Determine if the message is a tool-only batch (all content
             # items are ToolUseBlocks) — such messages have no text output yet,
             # so we skip the wait_for_stash flush below.
+            #
+            # Note: parallel execution of tools is handled natively by the
+            # SDK CLI via readOnlyHint annotations on tool definitions.
             is_tool_only = False
             if isinstance(sdk_msg, AssistantMessage) and sdk_msg.content:
-                is_tool_only = True
-                # NOTE: Pre-launches are sequential (each await completes
-                # file-ref expansion before the next starts).  This is fine
-                # since expansion is typically sub-ms; a future optimisation
-                # could gather all pre-launches concurrently.
-                for tool_use in sdk_msg.content:
-                    if isinstance(tool_use, ToolUseBlock):
-                        await pre_launch_tool_call(tool_use.name, tool_use.input)
-                    else:
-                        is_tool_only = False
+                is_tool_only = all(
+                    isinstance(item, ToolUseBlock) for item in sdk_msg.content
+                )
 
             # Race-condition fix: SDK hooks (PostToolUse) are
             # executed asynchronously via start_soon() — the next
@@ -1369,6 +1421,16 @@ async def _run_stream_attempt(
                         sdk_msg.result or "(no error message provided)",
                     )
 
+                # Check for prompt-too-long regardless of subtype — the
+                # SDK may return subtype="success" with result="Prompt is
+                # too long" when the CLI rejects the prompt before calling
+                # the API (cost_usd=0, no tokens consumed).  If we only
+                # check the "error" subtype path, the stream appears to
+                # complete normally, the synthetic error text is stored
+                # in the transcript, and the session grows without bound.
+                if _is_prompt_too_long(RuntimeError(sdk_msg.result or "")):
+                    raise RuntimeError("Prompt is too long")
+
                 # Capture token usage from ResultMessage.
                 # Anthropic reports cached tokens separately:
                 #   input_tokens = uncached only
@@ -1400,6 +1462,23 @@ async def _run_stream_attempt(
             # Emit compaction end if SDK finished compacting.
             # Sync TranscriptBuilder with the CLI's active context.
             compact_result = await ctx.compaction.emit_end_if_ready(ctx.session)
+            if compact_result.events:
+                # Compaction events end with StreamFinishStep, which maps to
+                # Vercel AI SDK's "finish-step" — that clears activeTextParts.
+                # Close any open text block BEFORE the compaction events so
+                # the text-end arrives before finish-step, preventing
+                # "text-end for missing text part" errors on the frontend.
+                pre_close: list[StreamBaseResponse] = []
+                state.adapter._end_text_if_open(pre_close)
+                # Compaction events bypass the adapter, so sync step state
+                # when a StreamFinishStep is present — otherwise the adapter
+                # will skip StreamStartStep on the next AssistantMessage.
+                if any(
+                    isinstance(ev, StreamFinishStep) for ev in compact_result.events
+                ):
+                    state.adapter.step_open = False
+                for r in pre_close:
+                    yield r
             for ev in compact_result.events:
                 yield ev
             entries_replaced = False
@@ -1446,8 +1525,38 @@ async def _run_stream_attempt(
                     model=sdk_msg.model,
                 )
 
+            # --- Intermediate persistence ---
+            # Flush session messages to DB periodically so page reloads
+            # show progress during long-running turns.
+            _msgs_since_flush += 1
+            now = time.monotonic()
+            if (
+                _msgs_since_flush >= _FLUSH_MESSAGE_THRESHOLD
+                or (now - _last_flush_time) >= _FLUSH_INTERVAL_SECONDS
+            ):
+                try:
+                    await asyncio.shield(upsert_chat_session(ctx.session))
+                    logger.debug(
+                        "%s Intermediate flush: %d messages "
+                        "(msgs_since=%d, elapsed=%.1fs)",
+                        ctx.log_prefix,
+                        len(ctx.session.messages),
+                        _msgs_since_flush,
+                        now - _last_flush_time,
+                    )
+                except Exception as flush_err:
+                    logger.warning(
+                        "%s Intermediate flush failed: %s",
+                        ctx.log_prefix,
+                        flush_err,
+                    )
+                _last_flush_time = now
+                _msgs_since_flush = 0
+
             if acc.stream_completed:
                 break
+    finally:
+        await _safe_close_sdk_client(sdk_client, ctx.log_prefix)
 
     # --- Post-stream processing (only on success) ---
     if state.adapter.has_unresolved_tool_calls:
@@ -1775,7 +1884,10 @@ async def stream_chat_completion_sdk(
         )
 
         # Fail fast when no API credentials are available at all.
-        sdk_env = _build_sdk_env(session_id=session_id, user_id=user_id)
+        # sdk_cwd routes the CLI's temp dir into the per-session workspace
+        # so sub-agent output files land inside sdk_cwd (see build_sdk_env).
+        sdk_env = build_sdk_env(session_id=session_id, user_id=user_id, sdk_cwd=sdk_cwd)
+
         if not config.api_key and not config.use_claude_code_subscription:
             raise RuntimeError(
                 "No API key configured. Set OPEN_ROUTER_API_KEY, "
@@ -1970,13 +2082,22 @@ async def stream_chat_completion_sdk(
 
             try:
                 async for event in _run_stream_attempt(stream_ctx, state):
-                    if not isinstance(event, StreamHeartbeat):
+                    if not isinstance(
+                        event,
+                        (
+                            StreamHeartbeat,
+                            # Compaction UI events are cosmetic and must not
+                            # block retry — they're emitted before the SDK
+                            # query on compacted attempts.
+                            StreamStartStep,
+                            StreamFinishStep,
+                            StreamToolInputStart,
+                            StreamToolInputAvailable,
+                            StreamToolOutputAvailable,
+                        ),
+                    ):
                         events_yielded += 1
                     yield event
-                # Cancel any pre-launched tasks that were never dispatched
-                # by the SDK (e.g. edge-case SDK behaviour changes). Symmetric
-                # with the three error-path await cancel_pending_tool_tasks() calls.
-                await cancel_pending_tool_tasks()
                 break  # Stream completed — exit retry loop
             except asyncio.CancelledError:
                 logger.warning(
@@ -1985,9 +2106,6 @@ async def stream_chat_completion_sdk(
                     attempt + 1,
                     _MAX_STREAM_ATTEMPTS,
                 )
-                # Cancel any pre-launched tasks so they don't continue executing
-                # against a rolled-back or abandoned session.
-                await cancel_pending_tool_tasks()
                 raise
             except _HandledStreamError as exc:
                 # _run_stream_attempt already yielded a StreamError and
@@ -2019,8 +2137,6 @@ async def stream_chat_completion_sdk(
                     retryable=True,
                 )
                 ended_with_stream_error = True
-                # Cancel any pre-launched tasks from the failed attempt.
-                await cancel_pending_tool_tasks()
                 break
             except Exception as e:
                 stream_err = e
@@ -2037,9 +2153,6 @@ async def stream_chat_completion_sdk(
                     exc_info=True,
                 )
                 session.messages = session.messages[:pre_attempt_msg_count]
-                # Cancel any pre-launched tasks from the failed attempt so they
-                # don't continue executing against the rolled-back session.
-                await cancel_pending_tool_tasks()
                 if events_yielded > 0:
                     # Events were already sent to the frontend and cannot be
                     # unsent.  Retrying would produce duplicate/inconsistent
@@ -2169,9 +2282,16 @@ async def stream_chat_completion_sdk(
             error_msg = "Operation cancelled"
         else:
             error_msg = str(e) or type(e).__name__
-            # SDK cleanup RuntimeError is expected during cancellation, log as warning
-            if isinstance(e, RuntimeError) and "cancel scope" in str(e):
-                logger.warning("%s SDK cleanup error: %s", log_prefix, error_msg)
+            # SDK cleanup errors are expected during client disconnect —
+            # log as warning rather than error to reduce Sentry noise.
+            # These are normally caught by _safe_close_sdk_client but
+            # can escape in edge cases (e.g. GeneratorExit timing).
+            if _is_sdk_disconnect_error(e):
+                logger.warning(
+                    "%s SDK cleanup error (client disconnect): %s",
+                    log_prefix,
+                    error_msg,
+                )
             else:
                 logger.error("%s Error: %s", log_prefix, error_msg, exc_info=True)
 
@@ -2193,10 +2313,11 @@ async def stream_chat_completion_sdk(
             )
 
         # Yield StreamError for immediate feedback (only for non-cancellation errors)
-        # Skip for CancelledError and RuntimeError cleanup issues (both are cancellations)
-        is_cancellation = isinstance(e, asyncio.CancelledError) or (
-            isinstance(e, RuntimeError) and "cancel scope" in str(e)
-        )
+        # Skip for CancelledError and SDK disconnect cleanup errors — these
+        # are not actionable by the user and the SSE connection is already dead.
+        is_cancellation = isinstance(
+            e, asyncio.CancelledError
+        ) or _is_sdk_disconnect_error(e)
         if not is_cancellation:
             yield StreamError(errorText=display_msg, code=code)
 

autogpt_platform/backend/backend/copilot/sdk/service_test.py

@@ -8,7 +8,12 @@ from unittest.mock import AsyncMock, MagicMock, patch
 
 import pytest
 
-from .service import _prepare_file_attachments, _resolve_sdk_model
+from .service import (
+    _is_sdk_disconnect_error,
+    _prepare_file_attachments,
+    _resolve_sdk_model,
+    _safe_close_sdk_client,
+)
 
 
 @dataclass
@@ -499,3 +504,111 @@ class TestResolveSdkModel:
         )
         monkeypatch.setattr("backend.copilot.sdk.service.config", cfg)
         assert _resolve_sdk_model() == "claude-opus-4-6"
+
+
+# ---------------------------------------------------------------------------
+# _is_sdk_disconnect_error — classify client disconnect cleanup errors
+# ---------------------------------------------------------------------------
+
+
+class TestIsSdkDisconnectError:
+    """Tests for _is_sdk_disconnect_error — identifies expected SDK cleanup errors."""
+
+    def test_cancel_scope_runtime_error(self):
+        """RuntimeError about cancel scope in wrong task is a disconnect error."""
+        exc = RuntimeError(
+            "Attempted to exit cancel scope in a different task than it was entered in"
+        )
+        assert _is_sdk_disconnect_error(exc) is True
+
+    def test_context_var_value_error(self):
+        """ValueError about ContextVar token mismatch is a disconnect error."""
+        exc = ValueError(
+            "<Token var=<ContextVar name='current_context'>> "
+            "was created in a different Context"
+        )
+        assert _is_sdk_disconnect_error(exc) is True
+
+    def test_unrelated_runtime_error(self):
+        """Unrelated RuntimeError should NOT be classified as disconnect error."""
+        exc = RuntimeError("something else went wrong")
+        assert _is_sdk_disconnect_error(exc) is False
+
+    def test_unrelated_value_error(self):
+        """Unrelated ValueError should NOT be classified as disconnect error."""
+        exc = ValueError("invalid argument")
+        assert _is_sdk_disconnect_error(exc) is False
+
+    def test_other_exception_types(self):
+        """Non-RuntimeError/ValueError should NOT be classified as disconnect error."""
+        assert _is_sdk_disconnect_error(TypeError("bad type")) is False
+        assert _is_sdk_disconnect_error(OSError("network down")) is False
+        assert _is_sdk_disconnect_error(asyncio.CancelledError()) is False
+
+
+# ---------------------------------------------------------------------------
+# _safe_close_sdk_client — suppress cleanup errors during disconnect
+# ---------------------------------------------------------------------------
+
+
+class TestSafeCloseSdkClient:
+    """Tests for _safe_close_sdk_client — suppresses expected SDK cleanup errors."""
+
+    @pytest.mark.asyncio
+    async def test_clean_exit(self):
+        """Normal __aexit__ (no error) should succeed silently."""
+        client = AsyncMock()
+        client.__aexit__ = AsyncMock(return_value=None)
+        await _safe_close_sdk_client(client, "[test]")
+        client.__aexit__.assert_awaited_once_with(None, None, None)
+
+    @pytest.mark.asyncio
+    async def test_cancel_scope_runtime_error_suppressed(self):
+        """RuntimeError from cancel scope mismatch should be suppressed."""
+        client = AsyncMock()
+        client.__aexit__ = AsyncMock(
+            side_effect=RuntimeError(
+                "Attempted to exit cancel scope in a different task"
+            )
+        )
+        # Should NOT raise
+        await _safe_close_sdk_client(client, "[test]")
+
+    @pytest.mark.asyncio
+    async def test_context_var_value_error_suppressed(self):
+        """ValueError from ContextVar token mismatch should be suppressed."""
+        client = AsyncMock()
+        client.__aexit__ = AsyncMock(
+            side_effect=ValueError(
+                "<Token var=<ContextVar name='current_context'>> "
+                "was created in a different Context"
+            )
+        )
+        # Should NOT raise
+        await _safe_close_sdk_client(client, "[test]")
+
+    @pytest.mark.asyncio
+    async def test_unexpected_exception_suppressed_with_error_log(self):
+        """Unexpected exceptions should be caught (not propagated) but logged at error."""
+        client = AsyncMock()
+        client.__aexit__ = AsyncMock(side_effect=OSError("unexpected"))
+        # Should NOT raise — unexpected errors are also suppressed to
+        # avoid crashing the generator during teardown.  Logged at error
+        # level so Sentry captures them via its logging integration.
+        await _safe_close_sdk_client(client, "[test]")
+
+    @pytest.mark.asyncio
+    async def test_unrelated_runtime_error_propagates(self):
+        """Non-cancel-scope RuntimeError should propagate (not suppressed)."""
+        client = AsyncMock()
+        client.__aexit__ = AsyncMock(side_effect=RuntimeError("something unrelated"))
+        with pytest.raises(RuntimeError, match="something unrelated"):
+            await _safe_close_sdk_client(client, "[test]")
+
+    @pytest.mark.asyncio
+    async def test_unrelated_value_error_propagates(self):
+        """Non-disconnect ValueError should propagate (not suppressed)."""
+        client = AsyncMock()
+        client.__aexit__ = AsyncMock(side_effect=ValueError("invalid argument"))
+        with pytest.raises(ValueError, match="invalid argument"):
+            await _safe_close_sdk_client(client, "[test]")

autogpt_platform/backend/backend/copilot/sdk/thinking_blocks_test.py

@@ -0,0 +1,823 @@
+"""Tests for thinking/redacted_thinking block preservation.
+
+Validates the fix for the Anthropic API error:
+    "thinking or redacted_thinking blocks in the latest assistant message
+    cannot be modified. These blocks must remain as they were in the
+    original response."
+
+The API requires that thinking blocks in the LAST assistant message are
+preserved value-identical. Older assistant messages may have thinking blocks
+stripped entirely. This test suite covers:
+
+  1. _flatten_assistant_content — strips thinking from older messages
+  2. compact_transcript — preserves last assistant's thinking blocks
+  3. response_adapter — handles ThinkingBlock without error
+  4. _format_sdk_content_blocks — preserves redacted_thinking blocks
+"""
+
+from __future__ import annotations
+
+from unittest.mock import AsyncMock, patch
+
+import pytest
+from claude_agent_sdk import AssistantMessage, TextBlock, ThinkingBlock
+
+from backend.copilot.response_model import (
+    StreamStartStep,
+    StreamTextDelta,
+    StreamTextStart,
+)
+from backend.util import json
+
+from .conftest import build_structured_transcript
+from .response_adapter import SDKResponseAdapter
+from .service import _format_sdk_content_blocks
+from .transcript import (
+    _find_last_assistant_entry,
+    _flatten_assistant_content,
+    _messages_to_transcript,
+    _rechain_tail,
+    _transcript_to_messages,
+    compact_transcript,
+    validate_transcript,
+)
+
+# ---------------------------------------------------------------------------
+# Fixtures: realistic thinking block content
+# ---------------------------------------------------------------------------
+
+THINKING_BLOCK = {
+    "type": "thinking",
+    "thinking": "Let me analyze the user's request carefully...",
+    "signature": "ErUBCkYIAxgCIkD0V2MsRXPkuGolGexaW9V1kluijxXGF",
+}
+
+REDACTED_THINKING_BLOCK = {
+    "type": "redacted_thinking",
+    "data": "EmwKAhgBEgy2VEE8PJaS2oLJCPkaT...",
+}
+
+
+def _make_thinking_transcript() -> str:
+    """Build a transcript with thinking blocks in multiple assistant turns.
+
+    Layout:
+      User 1 → Assistant 1 (thinking + text + tool_use)
+      User 2 (tool_result) → Assistant 2 (thinking + text)
+      User 3 → Assistant 3 (thinking + redacted_thinking + text) ← LAST
+    """
+    return build_structured_transcript(
+        [
+            ("user", "What files are in this project?"),
+            (
+                "assistant",
+                [
+                    {
+                        "type": "thinking",
+                        "thinking": "I should list the files.",
+                        "signature": "sig_old_1",
+                    },
+                    {"type": "text", "text": "Let me check the files."},
+                    {
+                        "type": "tool_use",
+                        "id": "tu1",
+                        "name": "list_files",
+                        "input": {"path": "/"},
+                    },
+                ],
+            ),
+            ("user", "Here are the files: a.py, b.py"),
+            (
+                "assistant",
+                [
+                    {
+                        "type": "thinking",
+                        "thinking": "Good, I see two Python files.",
+                        "signature": "sig_old_2",
+                    },
+                    {"type": "text", "text": "I found a.py and b.py."},
+                ],
+            ),
+            ("user", "Tell me about a.py"),
+            (
+                "assistant",
+                [
+                    THINKING_BLOCK,
+                    REDACTED_THINKING_BLOCK,
+                    {"type": "text", "text": "a.py contains the main entry point."},
+                ],
+            ),
+        ]
+    )
+
+
+def _last_assistant_content(transcript_jsonl: str) -> list[dict] | None:
+    """Extract the content blocks of the last assistant entry in a transcript."""
+    last_content = None
+    for line in transcript_jsonl.strip().split("\n"):
+        entry = json.loads(line)
+        msg = entry.get("message", {})
+        if msg.get("role") == "assistant":
+            last_content = msg.get("content")
+    return last_content
+
+
+# ---------------------------------------------------------------------------
+# _find_last_assistant_entry — unit tests
+# ---------------------------------------------------------------------------
+
+
+class TestFindLastAssistantEntry:
+    def test_splits_at_last_assistant(self):
+        """Prefix contains everything before last assistant; tail starts at it."""
+        transcript = build_structured_transcript(
+            [
+                ("user", "Hello"),
+                ("assistant", [{"type": "text", "text": "Hi"}]),
+                ("user", "More"),
+                ("assistant", [{"type": "text", "text": "Details"}]),
+            ]
+        )
+        prefix, tail = _find_last_assistant_entry(transcript)
+        # 3 entries in prefix (user, assistant, user), 1 in tail (last assistant)
+        assert len(prefix) == 3
+        assert len(tail) == 1
+
+    def test_no_assistant_returns_all_in_prefix(self):
+        """When there's no assistant, all lines are in prefix, tail is empty."""
+        transcript = build_structured_transcript(
+            [("user", "Hello"), ("user", "Another question")]
+        )
+        prefix, tail = _find_last_assistant_entry(transcript)
+        assert len(prefix) == 2
+        assert tail == []
+
+    def test_assistant_at_index_zero(self):
+        """When assistant is the first entry, prefix is empty."""
+        transcript = build_structured_transcript(
+            [("assistant", [{"type": "text", "text": "Start"}])]
+        )
+        prefix, tail = _find_last_assistant_entry(transcript)
+        assert prefix == []
+        assert len(tail) == 1
+
+    def test_trailing_user_included_in_tail(self):
+        """User message after last assistant is part of the tail."""
+        transcript = build_structured_transcript(
+            [
+                ("user", "Q1"),
+                ("assistant", [{"type": "text", "text": "A1"}]),
+                ("user", "Q2"),
+            ]
+        )
+        prefix, tail = _find_last_assistant_entry(transcript)
+        assert len(prefix) == 1  # first user
+        assert len(tail) == 2  # last assistant + trailing user
+
+    def test_multi_entry_turn_fully_preserved(self):
+        """An assistant turn spanning multiple JSONL entries (same message.id)
+        must be entirely in the tail, not split across prefix and tail."""
+        # Build manually because build_structured_transcript generates unique ids
+        lines = [
+            json.dumps(
+                {
+                    "type": "user",
+                    "uuid": "u1",
+                    "parentUuid": "",
+                    "message": {"role": "user", "content": "Hello"},
+                }
+            ),
+            json.dumps(
+                {
+                    "type": "assistant",
+                    "uuid": "a1-think",
+                    "parentUuid": "u1",
+                    "message": {
+                        "role": "assistant",
+                        "id": "msg_same_turn",
+                        "type": "message",
+                        "content": [THINKING_BLOCK],
+                        "stop_reason": None,
+                        "stop_sequence": None,
+                    },
+                }
+            ),
+            json.dumps(
+                {
+                    "type": "assistant",
+                    "uuid": "a1-tool",
+                    "parentUuid": "u1",
+                    "message": {
+                        "role": "assistant",
+                        "id": "msg_same_turn",
+                        "type": "message",
+                        "content": [
+                            {
+                                "type": "tool_use",
+                                "id": "tu1",
+                                "name": "Bash",
+                                "input": {},
+                            },
+                        ],
+                        "stop_reason": "tool_use",
+                        "stop_sequence": None,
+                    },
+                }
+            ),
+        ]
+        transcript = "\n".join(lines) + "\n"
+        prefix, tail = _find_last_assistant_entry(transcript)
+        # Both assistant entries share msg_same_turn → both in tail
+        assert len(prefix) == 1  # only the user entry
+        assert len(tail) == 2  # both assistant entries (thinking + tool_use)
+
+    def test_no_message_id_preserves_last_assistant(self):
+        """When the last assistant entry has no message.id, it should still
+        be preserved in the tail (fail closed) rather than being compressed."""
+        lines = [
+            json.dumps(
+                {
+                    "type": "user",
+                    "uuid": "u1",
+                    "parentUuid": "",
+                    "message": {"role": "user", "content": "Hello"},
+                }
+            ),
+            json.dumps(
+                {
+                    "type": "assistant",
+                    "uuid": "a1",
+                    "parentUuid": "u1",
+                    "message": {
+                        "role": "assistant",
+                        "content": [THINKING_BLOCK, {"type": "text", "text": "Hi"}],
+                    },
+                }
+            ),
+        ]
+        transcript = "\n".join(lines) + "\n"
+        prefix, tail = _find_last_assistant_entry(transcript)
+        assert len(prefix) == 1  # user entry
+        assert len(tail) == 1  # assistant entry preserved
+
+
+# ---------------------------------------------------------------------------
+# _rechain_tail — UUID chain patching
+# ---------------------------------------------------------------------------
+
+
+class TestRechainTail:
+    def test_patches_first_entry_parentuuid(self):
+        """First tail entry's parentUuid should point to last prefix uuid."""
+        prefix = _messages_to_transcript(
+            [
+                {"role": "user", "content": "Hello"},
+                {"role": "assistant", "content": "Hi"},
+            ]
+        )
+        # Get the last uuid from the prefix
+        last_prefix_uuid = None
+        for line in prefix.strip().split("\n"):
+            entry = json.loads(line)
+            last_prefix_uuid = entry.get("uuid")
+
+        tail_lines = [
+            json.dumps(
+                {
+                    "type": "assistant",
+                    "uuid": "tail-a1",
+                    "parentUuid": "old-parent",
+                    "message": {
+                        "role": "assistant",
+                        "content": [{"type": "text", "text": "Tail msg"}],
+                    },
+                }
+            )
+        ]
+        result = _rechain_tail(prefix, tail_lines)
+        entry = json.loads(result.strip())
+        assert entry["parentUuid"] == last_prefix_uuid
+        assert entry["uuid"] == "tail-a1"  # uuid preserved
+
+    def test_chains_multiple_tail_entries(self):
+        """Subsequent tail entries chain to each other."""
+        prefix = _messages_to_transcript([{"role": "user", "content": "Hi"}])
+        tail_lines = [
+            json.dumps(
+                {
+                    "type": "assistant",
+                    "uuid": "t1",
+                    "parentUuid": "old1",
+                    "message": {"role": "assistant", "content": []},
+                }
+            ),
+            json.dumps(
+                {
+                    "type": "user",
+                    "uuid": "t2",
+                    "parentUuid": "old2",
+                    "message": {"role": "user", "content": "Follow-up"},
+                }
+            ),
+        ]
+        result = _rechain_tail(prefix, tail_lines)
+        entries = [json.loads(ln) for ln in result.strip().split("\n")]
+        assert len(entries) == 2
+        # Second entry's parentUuid should be first entry's uuid
+        assert entries[1]["parentUuid"] == "t1"
+
+    def test_empty_tail_returns_empty(self):
+        """No tail entries → empty string."""
+        prefix = _messages_to_transcript([{"role": "user", "content": "Hi"}])
+        assert _rechain_tail(prefix, []) == ""
+
+    def test_preserves_message_content_verbatim(self):
+        """Tail message content (including thinking blocks) must not be modified."""
+        prefix = _messages_to_transcript([{"role": "user", "content": "Hi"}])
+        original_content = [
+            THINKING_BLOCK,
+            REDACTED_THINKING_BLOCK,
+            {"type": "text", "text": "Response"},
+        ]
+        tail_lines = [
+            json.dumps(
+                {
+                    "type": "assistant",
+                    "uuid": "t1",
+                    "parentUuid": "old",
+                    "message": {
+                        "role": "assistant",
+                        "content": original_content,
+                    },
+                }
+            )
+        ]
+        result = _rechain_tail(prefix, tail_lines)
+        entry = json.loads(result.strip())
+        assert entry["message"]["content"] == original_content
+
+
+# ---------------------------------------------------------------------------
+# _flatten_assistant_content — thinking blocks
+# ---------------------------------------------------------------------------
+
+
+class TestFlattenThinkingBlocks:
+    def test_thinking_blocks_are_stripped(self):
+        """Thinking blocks should not appear in flattened text for compression."""
+        blocks = [
+            {"type": "thinking", "thinking": "secret thoughts", "signature": "sig"},
+            {"type": "text", "text": "Hello user"},
+        ]
+        result = _flatten_assistant_content(blocks)
+        assert "secret thoughts" not in result
+        assert "Hello user" in result
+
+    def test_redacted_thinking_blocks_are_stripped(self):
+        """Redacted thinking blocks should not appear in flattened text."""
+        blocks = [
+            {"type": "redacted_thinking", "data": "encrypted_data"},
+            {"type": "text", "text": "Response text"},
+        ]
+        result = _flatten_assistant_content(blocks)
+        assert "encrypted_data" not in result
+        assert "Response text" in result
+
+    def test_thinking_only_message_flattens_to_empty(self):
+        """A message with only thinking blocks flattens to empty string."""
+        blocks = [
+            {"type": "thinking", "thinking": "just thinking...", "signature": "sig"},
+        ]
+        result = _flatten_assistant_content(blocks)
+        assert result == ""
+
+    def test_mixed_thinking_text_tool(self):
+        """Mixed blocks: only text survives flattening; thinking and tool_use dropped."""
+        blocks = [
+            {"type": "thinking", "thinking": "hmm", "signature": "sig"},
+            {"type": "redacted_thinking", "data": "xyz"},
+            {"type": "text", "text": "I'll read the file."},
+            {"type": "tool_use", "name": "Read", "input": {"path": "/x"}},
+        ]
+        result = _flatten_assistant_content(blocks)
+        assert "hmm" not in result
+        assert "xyz" not in result
+        assert "I'll read the file." in result
+        # tool_use blocks are dropped entirely to prevent model mimicry
+        assert "Read" not in result
+
+
+# ---------------------------------------------------------------------------
+# compact_transcript — thinking block preservation
+# ---------------------------------------------------------------------------
+
+
+class TestCompactTranscriptThinkingBlocks:
+    """Verify that compact_transcript preserves thinking blocks in the
+    last assistant message while stripping them from older messages."""
+
+    @pytest.mark.asyncio
+    async def test_last_assistant_thinking_blocks_preserved(self, mock_chat_config):
+        """After compaction, the last assistant entry must retain its
+        original thinking and redacted_thinking blocks verbatim."""
+        transcript = _make_thinking_transcript()
+
+        compacted_msgs = [
+            {"role": "user", "content": "[conversation summary]"},
+            {"role": "assistant", "content": "Summarized response"},
+        ]
+        mock_result = type(
+            "CompressResult",
+            (),
+            {
+                "was_compacted": True,
+                "messages": compacted_msgs,
+                "original_token_count": 800,
+                "token_count": 200,
+                "messages_summarized": 4,
+                "messages_dropped": 0,
+            },
+        )()
+        with patch(
+            "backend.copilot.sdk.transcript._run_compression",
+            new_callable=AsyncMock,
+            return_value=mock_result,
+        ):
+            result = await compact_transcript(transcript, model="test-model")
+
+        assert result is not None
+        assert validate_transcript(result)
+
+        last_content = _last_assistant_content(result)
+        assert last_content is not None, "No assistant entry found"
+        assert isinstance(last_content, list)
+
+        # The last assistant must have the thinking blocks preserved
+        block_types = [b["type"] for b in last_content]
+        assert (
+            "thinking" in block_types
+        ), "thinking block missing from last assistant message"
+        assert (
+            "redacted_thinking" in block_types
+        ), "redacted_thinking block missing from last assistant message"
+        assert "text" in block_types
+
+        # Verify the thinking block content is value-identical
+        thinking_blocks = [b for b in last_content if b["type"] == "thinking"]
+        assert len(thinking_blocks) == 1
+        assert thinking_blocks[0]["thinking"] == THINKING_BLOCK["thinking"]
+        assert thinking_blocks[0]["signature"] == THINKING_BLOCK["signature"]
+
+        redacted_blocks = [b for b in last_content if b["type"] == "redacted_thinking"]
+        assert len(redacted_blocks) == 1
+        assert redacted_blocks[0]["data"] == REDACTED_THINKING_BLOCK["data"]
+
+    @pytest.mark.asyncio
+    async def test_older_assistant_thinking_blocks_stripped(self, mock_chat_config):
+        """Older assistant messages should NOT retain thinking blocks
+        after compaction (they're compressed into summaries)."""
+        transcript = _make_thinking_transcript()
+
+        # The compressor will receive messages where older assistant
+        # entries have already had thinking blocks stripped.
+        captured_messages: list[dict] = []
+
+        async def mock_compression(messages, model, log_prefix):
+            captured_messages.extend(messages)
+            return type(
+                "CompressResult",
+                (),
+                {
+                    "was_compacted": True,
+                    "messages": messages,
+                    "original_token_count": 800,
+                    "token_count": 400,
+                    "messages_summarized": 2,
+                    "messages_dropped": 0,
+                },
+            )()
+
+        with patch(
+            "backend.copilot.sdk.transcript._run_compression",
+            side_effect=mock_compression,
+        ):
+            await compact_transcript(transcript, model="test-model")
+
+        # Check that the messages sent to compression don't contain
+        # thinking content from older assistant messages
+        for msg in captured_messages:
+            if msg["role"] == "assistant":
+                content = msg.get("content", "")
+                assert (
+                    "I should list the files." not in content
+                ), "Old thinking block content leaked into compression input"
+                assert (
+                    "Good, I see two Python files." not in content
+                ), "Old thinking block content leaked into compression input"
+
+    @pytest.mark.asyncio
+    async def test_trailing_user_message_after_last_assistant(self, mock_chat_config):
+        """When the last entry is a user message, the last *assistant*
+        message's thinking blocks should still be preserved."""
+        transcript = build_structured_transcript(
+            [
+                ("user", "Hello"),
+                (
+                    "assistant",
+                    [
+                        THINKING_BLOCK,
+                        {"type": "text", "text": "Hi there"},
+                    ],
+                ),
+                ("user", "Follow-up question"),
+            ]
+        )
+
+        # The compressor only receives the prefix (1 user message); the
+        # tail (assistant + trailing user) is preserved verbatim.
+        compacted_msgs = [
+            {"role": "user", "content": "Hello"},
+        ]
+        mock_result = type(
+            "CompressResult",
+            (),
+            {
+                "was_compacted": True,
+                "messages": compacted_msgs,
+                "original_token_count": 400,
+                "token_count": 100,
+                "messages_summarized": 0,
+                "messages_dropped": 0,
+            },
+        )()
+        with patch(
+            "backend.copilot.sdk.transcript._run_compression",
+            new_callable=AsyncMock,
+            return_value=mock_result,
+        ):
+            result = await compact_transcript(transcript, model="test-model")
+
+        assert result is not None
+
+        last_content = _last_assistant_content(result)
+        assert last_content is not None
+        assert isinstance(last_content, list)
+        block_types = [b["type"] for b in last_content]
+        assert (
+            "thinking" in block_types
+        ), "thinking block lost from last assistant despite trailing user msg"
+
+    @pytest.mark.asyncio
+    async def test_single_assistant_with_thinking_preserved(self, mock_chat_config):
+        """When there's only one assistant message (which is also the last),
+        its thinking blocks must be preserved."""
+        transcript = build_structured_transcript(
+            [
+                ("user", "Hello"),
+                (
+                    "assistant",
+                    [
+                        THINKING_BLOCK,
+                        {"type": "text", "text": "World"},
+                    ],
+                ),
+            ]
+        )
+
+        compacted_msgs = [
+            {"role": "user", "content": "Hello"},
+            {"role": "assistant", "content": "World"},
+        ]
+        mock_result = type(
+            "CompressResult",
+            (),
+            {
+                "was_compacted": True,
+                "messages": compacted_msgs,
+                "original_token_count": 200,
+                "token_count": 100,
+                "messages_summarized": 0,
+                "messages_dropped": 0,
+            },
+        )()
+        with patch(
+            "backend.copilot.sdk.transcript._run_compression",
+            new_callable=AsyncMock,
+            return_value=mock_result,
+        ):
+            result = await compact_transcript(transcript, model="test-model")
+
+        assert result is not None
+
+        last_content = _last_assistant_content(result)
+        assert last_content is not None
+        assert isinstance(last_content, list)
+        block_types = [b["type"] for b in last_content]
+        assert "thinking" in block_types
+
+    @pytest.mark.asyncio
+    async def test_tail_parentuuid_rewired_to_prefix(self, mock_chat_config):
+        """After compaction, the first tail entry's parentUuid must point to
+        the last entry in the compressed prefix — not its original parent."""
+        transcript = _make_thinking_transcript()
+
+        compacted_msgs = [
+            {"role": "user", "content": "[conversation summary]"},
+            {"role": "assistant", "content": "Summarized response"},
+        ]
+        mock_result = type(
+            "CompressResult",
+            (),
+            {
+                "was_compacted": True,
+                "messages": compacted_msgs,
+                "original_token_count": 800,
+                "token_count": 200,
+                "messages_summarized": 4,
+                "messages_dropped": 0,
+            },
+        )()
+        with patch(
+            "backend.copilot.sdk.transcript._run_compression",
+            new_callable=AsyncMock,
+            return_value=mock_result,
+        ):
+            result = await compact_transcript(transcript, model="test-model")
+
+        assert result is not None
+        lines = [ln for ln in result.strip().split("\n") if ln.strip()]
+        entries = [json.loads(ln) for ln in lines]
+
+        # Find the boundary: the compressed prefix ends just before the
+        # first tail entry (last assistant in original transcript).
+        tail_start = None
+        for i, entry in enumerate(entries):
+            msg = entry.get("message", {})
+            if isinstance(msg.get("content"), list):
+                # Structured content = preserved tail entry
+                tail_start = i
+                break
+
+        assert tail_start is not None, "Could not find preserved tail entry"
+        assert tail_start > 0, "Tail should not be the first entry"
+
+        # The tail entry's parentUuid must be the uuid of the preceding entry
+        prefix_last_uuid = entries[tail_start - 1]["uuid"]
+        tail_first_parent = entries[tail_start]["parentUuid"]
+        assert tail_first_parent == prefix_last_uuid, (
+            f"Tail parentUuid {tail_first_parent!r} != "
+            f"last prefix uuid {prefix_last_uuid!r}"
+        )
+
+    @pytest.mark.asyncio
+    async def test_no_thinking_blocks_still_works(self, mock_chat_config):
+        """Compaction should still work normally when there are no thinking
+        blocks in the transcript."""
+        transcript = build_structured_transcript(
+            [
+                ("user", "Hello"),
+                ("assistant", [{"type": "text", "text": "Hi"}]),
+                ("user", "More"),
+                ("assistant", [{"type": "text", "text": "Details"}]),
+            ]
+        )
+
+        compacted_msgs = [
+            {"role": "user", "content": "[summary]"},
+            {"role": "assistant", "content": "Summary"},
+        ]
+        mock_result = type(
+            "CompressResult",
+            (),
+            {
+                "was_compacted": True,
+                "messages": compacted_msgs,
+                "original_token_count": 200,
+                "token_count": 50,
+                "messages_summarized": 2,
+                "messages_dropped": 0,
+            },
+        )()
+        with patch(
+            "backend.copilot.sdk.transcript._run_compression",
+            new_callable=AsyncMock,
+            return_value=mock_result,
+        ):
+            result = await compact_transcript(transcript, model="test-model")
+
+        assert result is not None
+        assert validate_transcript(result)
+        # Verify last assistant content is preserved even without thinking blocks
+        last_content = _last_assistant_content(result)
+        assert last_content is not None
+        assert last_content == [{"type": "text", "text": "Details"}]
+
+
+# ---------------------------------------------------------------------------
+# _transcript_to_messages — thinking block handling
+# ---------------------------------------------------------------------------
+
+
+class TestTranscriptToMessagesThinking:
+    def test_thinking_blocks_excluded_from_flattened_content(self):
+        """When _transcript_to_messages flattens content, thinking block
+        text should not leak into the message content string."""
+        transcript = build_structured_transcript(
+            [
+                ("user", "Hello"),
+                (
+                    "assistant",
+                    [
+                        {
+                            "type": "thinking",
+                            "thinking": "SECRET_THOUGHT",
+                            "signature": "sig",
+                        },
+                        {"type": "text", "text": "Visible response"},
+                    ],
+                ),
+            ]
+        )
+        messages = _transcript_to_messages(transcript)
+        assistant_msg = [m for m in messages if m["role"] == "assistant"][0]
+        assert "SECRET_THOUGHT" not in assistant_msg["content"]
+        assert "Visible response" in assistant_msg["content"]
+
+
+# ---------------------------------------------------------------------------
+# response_adapter — ThinkingBlock handling
+# ---------------------------------------------------------------------------
+
+
+class TestResponseAdapterThinkingBlock:
+    def test_thinking_block_does_not_crash(self):
+        """ThinkingBlock in AssistantMessage should not cause an error."""
+        adapter = SDKResponseAdapter(message_id="msg-1", session_id="sess-1")
+        msg = AssistantMessage(
+            content=[
+                ThinkingBlock(
+                    thinking="Let me think about this...",
+                    signature="sig_test_123",
+                ),
+                TextBlock(text="Here is my response."),
+            ],
+            model="claude-test",
+        )
+        results = adapter.convert_message(msg)
+        # Should produce stream events for text only, no crash
+        types = [type(r) for r in results]
+        assert StreamStartStep in types
+        assert StreamTextStart in types or StreamTextDelta in types
+
+    def test_thinking_block_does_not_emit_stream_events(self):
+        """ThinkingBlock should NOT produce any StreamTextDelta events
+        containing thinking content."""
+        adapter = SDKResponseAdapter(message_id="msg-1", session_id="sess-1")
+        msg = AssistantMessage(
+            content=[
+                ThinkingBlock(
+                    thinking="My secret thoughts",
+                    signature="sig_test_456",
+                ),
+                TextBlock(text="Public response"),
+            ],
+            model="claude-test",
+        )
+        results = adapter.convert_message(msg)
+        text_deltas = [r for r in results if isinstance(r, StreamTextDelta)]
+        for delta in text_deltas:
+            assert "secret thoughts" not in (delta.delta or "")
+
+
+# ---------------------------------------------------------------------------
+# _format_sdk_content_blocks — redacted_thinking handling
+# ---------------------------------------------------------------------------
+
+
+class TestFormatSdkContentBlocks:
+    def test_thinking_block_preserved(self):
+        """ThinkingBlock should be serialized with type, thinking, and signature."""
+        blocks = [
+            ThinkingBlock(thinking="My thoughts", signature="sig123"),
+            TextBlock(text="Response"),
+        ]
+        result = _format_sdk_content_blocks(blocks)
+        assert len(result) == 2
+        assert result[0] == {
+            "type": "thinking",
+            "thinking": "My thoughts",
+            "signature": "sig123",
+        }
+        assert result[1] == {"type": "text", "text": "Response"}
+
+    def test_raw_dict_redacted_thinking_preserved(self):
+        """Raw dict blocks (e.g. redacted_thinking) pass through unchanged."""
+        raw_block = {"type": "redacted_thinking", "data": "EmwKAh...encrypted"}
+        blocks = [
+            raw_block,
+            TextBlock(text="Response"),
+        ]
+        result = _format_sdk_content_blocks(blocks)
+        assert len(result) == 2
+        assert result[0] == raw_block
+        assert result[1] == {"type": "text", "text": "Response"}

autogpt_platform/backend/backend/copilot/sdk/tool_adapter.py

@@ -14,6 +14,7 @@ from contextvars import ContextVar
 from typing import TYPE_CHECKING, Any
 
 from claude_agent_sdk import create_sdk_mcp_server, tool
+from mcp.types import ToolAnnotations
 
 from backend.copilot.context import (
     _current_permissions,
@@ -37,7 +38,7 @@ from backend.copilot.tools import TOOL_REGISTRY
 from backend.copilot.tools.base import BaseTool
 from backend.util.truncate import truncate
 
-from .e2b_file_tools import E2B_FILE_TOOL_NAMES, E2B_FILE_TOOLS
+from .e2b_file_tools import E2B_FILE_TOOL_NAMES, E2B_FILE_TOOLS, bridge_and_annotate
 
 if TYPE_CHECKING:
     from e2b import AsyncSandbox
@@ -53,14 +54,6 @@ _MCP_MAX_CHARS = 500_000
 MCP_SERVER_NAME = "copilot"
 MCP_TOOL_PREFIX = f"mcp__{MCP_SERVER_NAME}__"
 
-# Map from tool_name -> Queue of pre-launched (task, args) pairs.
-# Initialised per-session in set_execution_context() so concurrent sessions
-# never share the same dict.
-_TaskQueueItem = tuple[asyncio.Task[dict[str, Any]], dict[str, Any]]
-_tool_task_queues: ContextVar[dict[str, asyncio.Queue[_TaskQueueItem]] | None] = (
-    ContextVar("_tool_task_queues", default=None)
-)
-
 # Stash for MCP tool outputs before the SDK potentially truncates them.
 # Keyed by tool_name → full output string. Consumed (popped) by the
 # response adapter when it builds StreamToolOutputAvailable.
@@ -115,7 +108,6 @@ def set_execution_context(
     _current_permissions.set(permissions)
     _pending_tool_outputs.set({})
     _stash_event.set(asyncio.Event())
-    _tool_task_queues.set({})
     _consecutive_tool_failures.set({})
 
 
@@ -132,48 +124,6 @@ def reset_stash_event() -> None:
         event.clear()
 
 
-async def cancel_pending_tool_tasks() -> None:
-    """Cancel all queued pre-launched tasks for the current execution context.
-
-    Call this when a stream attempt aborts (error, cancellation) to prevent
-    pre-launched tasks from continuing to execute against a rolled-back session.
-    Tasks that are already done are skipped; in-flight tasks are cancelled and
-    awaited so that any cleanup (``finally`` blocks, DB rollbacks) completes
-    before the next retry starts.
-    """
-    queues = _tool_task_queues.get()
-    if not queues:
-        return
-    cancelled_tasks: list[asyncio.Task] = []
-    for tool_name, queue in list(queues.items()):
-        cancelled = 0
-        while not queue.empty():
-            task, _args = queue.get_nowait()
-            if not task.done():
-                task.cancel()
-                cancelled_tasks.append(task)
-                cancelled += 1
-        if cancelled:
-            logger.debug(
-                "Cancelled %d pre-launched task(s) for tool '%s'", cancelled, tool_name
-            )
-    queues.clear()
-    # Await all cancelled tasks so their cleanup (finally blocks, DB rollbacks)
-    # completes before the next retry attempt starts new pre-launches.
-    # Use a timeout to prevent hanging indefinitely if a task's cleanup is stuck.
-    if cancelled_tasks:
-        try:
-            await asyncio.wait_for(
-                asyncio.gather(*cancelled_tasks, return_exceptions=True),
-                timeout=5.0,
-            )
-        except TimeoutError:
-            logger.warning(
-                "Timed out waiting for %d cancelled task(s) to clean up",
-                len(cancelled_tasks),
-            )
-
-
 def reset_tool_failure_counters() -> None:
     """Reset all tool-level circuit breaker counters.
 
@@ -249,10 +199,6 @@ async def wait_for_stash(timeout: float = 2.0) -> bool:
     Uses ``asyncio.Event.wait()`` so it returns the instant the hook signals —
     the timeout is purely a safety net for the case where the hook never fires.
     Returns ``True`` if the stash signal was received, ``False`` on timeout.
-
-    The 2.0 s default was chosen to accommodate slower tool startup in cloud
-    sandboxes while still failing fast when the hook genuinely will not fire.
-    With the parallel pre-launch path, hooks typically fire well under 1 ms.
     """
     event = _stash_event.get(None)
     if event is None:
@@ -271,95 +217,13 @@ async def wait_for_stash(timeout: float = 2.0) -> bool:
         return False
 
 
-async def pre_launch_tool_call(tool_name: str, args: dict[str, Any]) -> None:
-    """Pre-launch a tool as a background task so parallel calls run concurrently.
-
-    Called when an AssistantMessage with ToolUseBlocks is received, before the
-    SDK dispatches the MCP tool/call requests. The tool_handler will await the
-    pre-launched task instead of executing fresh.
-
-    The tool_name may include an MCP prefix (e.g. ``mcp__copilot__run_block``);
-    the prefix is stripped automatically before looking up the tool.
-
-    Ordering guarantee: the Claude Agent SDK dispatches MCP ``tools/call`` requests
-    in the same order as the ToolUseBlocks appear in the AssistantMessage.
-    Pre-launched tasks are queued FIFO per tool name, so the N-th handler for a
-    given tool name dequeues the N-th pre-launched task — result and args always
-    correspond when the SDK preserves order (which it does in the current SDK).
-    """
-    queues = _tool_task_queues.get()
-    if queues is None:
-        return
-
-    # Strip the MCP server prefix (e.g. "mcp__copilot__") to get the bare tool name.
-    # Use removeprefix so tool names that themselves contain "__" are handled correctly.
-    bare_name = tool_name.removeprefix(MCP_TOOL_PREFIX)
-
-    base_tool = TOOL_REGISTRY.get(bare_name)
-    if base_tool is None:
-        return
-
-    user_id, session = get_execution_context()
-    if session is None:
-        return
-
-    # Expand @@agptfile: references before launching the task.
-    # The _truncating wrapper (which normally handles expansion) runs AFTER
-    # pre_launch_tool_call — the pre-launched task would otherwise receive raw
-    # @@agptfile: tokens and fail to resolve them inside _execute_tool_sync.
-    # Use _build_input_schema (same path as _truncating) for schema-aware expansion.
-    input_schema: dict[str, Any] | None
-    try:
-        input_schema = _build_input_schema(base_tool)
-    except Exception:
-        input_schema = None  # schema unavailable — skip schema-aware expansion
-    try:
-        args = await expand_file_refs_in_args(
-            args, user_id, session, input_schema=input_schema
-        )
-    except FileRefExpansionError as exc:
-        logger.warning(
-            "pre_launch_tool_call: @@agptfile expansion failed for %s: %s — skipping pre-launch",
-            bare_name,
-            exc,
-        )
-        return
-
-    task = asyncio.create_task(_execute_tool_sync(base_tool, user_id, session, args))
-    # Log unhandled exceptions so "Task exception was never retrieved" warnings
-    # do not pollute stderr when a task is pre-launched but never dequeued.
-    task.add_done_callback(
-        lambda t, name=bare_name: (
-            logger.warning(
-                "Pre-launched task for %s raised unhandled: %s",
-                name,
-                t.exception(),
-            )
-            if not t.cancelled() and t.exception()
-            else None
-        )
-    )
-
-    if bare_name not in queues:
-        queues[bare_name] = asyncio.Queue[_TaskQueueItem]()
-    # Store (task, args) so the handler can log a warning if the SDK dispatches
-    # calls in a different order than the ToolUseBlocks appeared in the message.
-    queues[bare_name].put_nowait((task, args))
-
-
 async def _execute_tool_sync(
     base_tool: BaseTool,
     user_id: str | None,
     session: ChatSession,
     args: dict[str, Any],
 ) -> dict[str, Any]:
-    """Execute a tool synchronously and return MCP-formatted response.
-
-    Note: ``@@agptfile:`` expansion should be performed by the caller before
-    invoking this function.  For the normal (non-parallel) path it is handled
-    by the ``_truncating`` wrapper; for the pre-launched parallel path it is
-    handled in :func:`pre_launch_tool_call` before the task is created.
-    """
+    """Execute a tool synchronously and return MCP-formatted response."""
     effective_id = f"sdk-{uuid.uuid4().hex[:12]}"
     result = await base_tool.execute(
         user_id=user_id,
@@ -455,83 +319,7 @@ def create_tool_handler(base_tool: BaseTool):
     """
 
     async def tool_handler(args: dict[str, Any]) -> dict[str, Any]:
-        """Execute the wrapped tool and return MCP-formatted response.
-
-        If a pre-launched task exists (from parallel tool pre-launch in the
-        message loop), await it instead of executing fresh.
-        """
-        queues = _tool_task_queues.get()
-        if queues and base_tool.name in queues:
-            queue = queues[base_tool.name]
-            if not queue.empty():
-                task, launch_args = queue.get_nowait()
-                # Sanity-check: warn if the args don't match — this can happen
-                # if the SDK dispatches tool calls in a different order than the
-                # ToolUseBlocks appeared in the AssistantMessage (unlikely but
-                # could occur in future SDK versions or with SDK bugs).
-                # We compare full values (not just keys) so that two run_block
-                # calls with different block_id values are caught even though
-                # both have the same key set.
-                if launch_args != args:
-                    logger.warning(
-                        "Pre-launched task for %s: arg mismatch "
-                        "(launch_keys=%s, call_keys=%s) — cancelling "
-                        "pre-launched task and falling back to direct execution",
-                        base_tool.name,
-                        (
-                            sorted(launch_args.keys())
-                            if isinstance(launch_args, dict)
-                            else type(launch_args).__name__
-                        ),
-                        (
-                            sorted(args.keys())
-                            if isinstance(args, dict)
-                            else type(args).__name__
-                        ),
-                    )
-                    if not task.done():
-                        task.cancel()
-                        # Await cancellation to prevent duplicate concurrent
-                        # execution for blocks with side effects.
-                        try:
-                            await task
-                        except (asyncio.CancelledError, Exception):
-                            pass
-                    # Fall through to the direct-execution path below.
-                else:
-                    # Args match — await the pre-launched task.
-                    try:
-                        result = await task
-                    except asyncio.CancelledError:
-                        # Re-raise: CancelledError may be propagating from the
-                        # outer streaming loop being cancelled — swallowing it
-                        # would mask the cancellation and prevent proper cleanup.
-                        logger.warning(
-                            "Pre-launched tool %s was cancelled — re-raising",
-                            base_tool.name,
-                        )
-                        raise
-                    except Exception as e:
-                        logger.error(
-                            "Pre-launched tool %s failed: %s",
-                            base_tool.name,
-                            e,
-                            exc_info=True,
-                        )
-                        return _mcp_error(
-                            f"Failed to execute {base_tool.name}. "
-                            "Check server logs for details."
-                        )
-
-                    # Pre-truncate the result so the _truncating wrapper (which
-                    # wraps this handler) receives an already-within-budget
-                    # value. _truncating handles stashing — we must NOT stash
-                    # here or the output will be appended twice to the FIFO
-                    # queue and pop_pending_tool_output would return a duplicate
-                    # entry on the second call for the same tool.
-                    return truncate(result, _MCP_MAX_CHARS)
-
-        # No pre-launched task — execute directly (fallback for non-parallel calls).
+        """Execute the wrapped tool and return MCP-formatted response."""
         user_id, session = get_execution_context()
 
         if session is None:
@@ -599,7 +387,16 @@ async def _read_file_handler(args: dict[str, Any]) -> dict[str, Any]:
             selected = list(itertools.islice(f, offset, offset + limit))
         # Cleanup happens in _cleanup_sdk_tool_results after session ends;
         # don't delete here — the SDK may read in multiple chunks.
-        return _mcp_ok("".join(selected))
+        #
+        # When E2B is active, also copy the file into the sandbox so
+        # bash_exec can process it (the model often uses Read then bash).
+        text = "".join(selected)
+        sandbox = _current_sandbox.get(None)
+        if sandbox is not None:
+            annotation = await bridge_and_annotate(sandbox, resolved, offset, limit)
+            if annotation:
+                text += annotation
+        return _mcp_ok(text)
     except FileNotFoundError:
         return _mcp_err(f"File not found: {file_path}")
     except Exception as e:
@@ -648,9 +445,19 @@ def _text_from_mcp_result(result: dict[str, Any]) -> str:
     )
 
 
+_PARALLEL_ANNOTATION = ToolAnnotations(readOnlyHint=True)
+
+
 def create_copilot_mcp_server(*, use_e2b: bool = False):
     """Create an in-process MCP server configuration for CoPilot tools.
 
+    All tools are annotated with ``readOnlyHint=True`` so the SDK CLI
+    dispatches concurrent tool calls in parallel rather than sequentially.
+    This is a deliberate override: even side-effect tools use the hint
+    because the MCP tools are already individually sandboxed and the
+    pre-launch duplicate-execution bug (SECRT-2204) is worse than
+    sequential dispatch.
+
     When *use_e2b* is True, five additional MCP file tools are registered
     that route directly to the E2B sandbox filesystem, and the caller should
     disable the corresponding SDK built-in tools via
@@ -668,6 +475,28 @@ def create_copilot_mcp_server(*, use_e2b: bool = False):
         Applied once to every registered tool."""
 
         async def wrapper(args: dict[str, Any]) -> dict[str, Any]:
+            # Empty tool args = model's output was truncated by the API's
+            # max_tokens limit.  Instead of letting the tool fail with a
+            # confusing error (and eventually tripping the circuit breaker),
+            # return clear guidance so the model can self-correct.
+            if not args and input_schema and input_schema.get("required"):
+                logger.warning(
+                    "[MCP] %s called with empty args (likely output "
+                    "token truncation) — returning guidance",
+                    tool_name,
+                )
+                return _mcp_error(
+                    f"Your call to {tool_name} had empty arguments — "
+                    f"this means your previous response was too long and "
+                    f"the tool call input was truncated by the API. "
+                    f"To fix this: break your work into smaller steps. "
+                    f"For large content, first write it to a file using "
+                    f"bash_exec with cat >> (append section by section), "
+                    f"then pass it via @@agptfile:filename reference. "
+                    f"Do NOT retry with the same approach — it will "
+                    f"be truncated again."
+                )
+
             # Circuit breaker: stop infinite retry loops with identical args.
             # Use the original (pre-expansion) args for fingerprinting so
             # check and record always use the same key — @@agptfile:
@@ -718,24 +547,35 @@ def create_copilot_mcp_server(*, use_e2b: bool = False):
     for tool_name, base_tool in TOOL_REGISTRY.items():
         handler = create_tool_handler(base_tool)
         schema = _build_input_schema(base_tool)
+        # All tools annotated readOnlyHint=True to enable parallel dispatch.
+        # The SDK CLI uses this hint to dispatch concurrent tool calls in
+        # parallel rather than sequentially.  Side-effect safety is ensured
+        # by the tool implementations themselves (idempotency, credentials).
         decorated = tool(
             tool_name,
             base_tool.description,
             schema,
+            annotations=_PARALLEL_ANNOTATION,
         )(_truncating(handler, tool_name, input_schema=schema))
         sdk_tools.append(decorated)
 
     # E2B file tools replace SDK built-in Read/Write/Edit/Glob/Grep.
     if use_e2b:
         for name, desc, schema, handler in E2B_FILE_TOOLS:
-            decorated = tool(name, desc, schema)(_truncating(handler, name))
+            decorated = tool(
+                name,
+                desc,
+                schema,
+                annotations=_PARALLEL_ANNOTATION,
+            )(_truncating(handler, name))
             sdk_tools.append(decorated)
 
-    # Read tool for SDK-truncated tool results (always needed).
+    # Read tool for SDK-truncated tool results (always needed, read-only).
     read_tool = tool(
         _READ_TOOL_NAME,
         _READ_TOOL_DESCRIPTION,
         _READ_TOOL_SCHEMA,
+        annotations=_PARALLEL_ANNOTATION,
     )(_truncating(_read_file_handler, _READ_TOOL_NAME))
     sdk_tools.append(read_tool)
 
@@ -750,13 +590,14 @@ def create_copilot_mcp_server(*, use_e2b: bool = False):
 # Security hooks validate that file paths stay within sdk_cwd.
 # Bash is NOT included — use the sandboxed MCP bash_exec tool instead,
 # which provides kernel-level network isolation via unshare --net.
-# Task allows spawning sub-agents (rate-limited by security hooks).
+# Task/Agent allows spawning sub-agents (rate-limited by security hooks).
+#   The CLI renamed "Task" → "Agent" in v2.x; both are listed for compat.
 # WebSearch uses Brave Search via Anthropic's API — safe, no SSRF risk.
 # TodoWrite manages the task checklist shown in the UI — no security concern.
 # In E2B mode, all five are disabled — MCP equivalents provide direct sandbox
 # access.  read_file also handles local tool-results and ephemeral reads.
 _SDK_BUILTIN_FILE_TOOLS = ["Read", "Write", "Edit", "Glob", "Grep"]
-_SDK_BUILTIN_ALWAYS = ["Task", "WebSearch", "TodoWrite"]
+_SDK_BUILTIN_ALWAYS = ["Task", "Agent", "WebSearch", "TodoWrite"]
 _SDK_BUILTIN_TOOLS = [*_SDK_BUILTIN_FILE_TOOLS, *_SDK_BUILTIN_ALWAYS]
 
 # SDK built-in tools that must be explicitly blocked.

autogpt_platform/backend/backend/copilot/sdk/tool_adapter_test.py

@@ -1,22 +1,21 @@
-"""Tests for tool_adapter helpers: truncation, stash, context vars, parallel pre-launch."""
+"""Tests for tool_adapter: truncation, stash, context vars, readOnlyHint annotations."""
 
 import asyncio
-from unittest.mock import AsyncMock, MagicMock, patch
+from unittest.mock import AsyncMock, MagicMock
 
 import pytest
+from mcp.types import ToolAnnotations
 
 from backend.copilot.context import get_sdk_cwd
 from backend.copilot.response_model import StreamToolOutputAvailable
-from backend.copilot.sdk.file_ref import FileRefExpansionError
 from backend.util.truncate import truncate
 
 from .tool_adapter import (
     _MCP_MAX_CHARS,
+    SDK_DISALLOWED_TOOLS,
     _text_from_mcp_result,
-    cancel_pending_tool_tasks,
     create_tool_handler,
     pop_pending_tool_output,
-    pre_launch_tool_call,
     reset_stash_event,
     set_execution_context,
     stash_pending_tool_output,
@@ -244,7 +243,7 @@ class TestTruncationAndStashIntegration:
 
 
 # ---------------------------------------------------------------------------
-# Parallel pre-launch infrastructure
+# create_tool_handler (direct execution, no pre-launch)
 # ---------------------------------------------------------------------------
 
 
@@ -277,169 +276,18 @@ def _init_ctx(session=None):
     )
 
 
-class TestPreLaunchToolCall:
-    """Tests for pre_launch_tool_call and the queue-based parallel dispatch."""
+class TestCreateToolHandler:
+    """Tests for create_tool_handler — direct tool execution."""
 
     @pytest.fixture(autouse=True)
     def _init(self):
         _init_ctx(session=_make_mock_session())
 
     @pytest.mark.asyncio
-    async def test_unknown_tool_is_silently_ignored(self):
-        """pre_launch_tool_call does nothing for tools not in TOOL_REGISTRY."""
-        # Should not raise even if the tool name is completely unknown
-        await pre_launch_tool_call("nonexistent_tool", {})
-
-    @pytest.mark.asyncio
-    async def test_mcp_prefix_stripped_before_registry_lookup(self):
-        """mcp__copilot__run_block is looked up as 'run_block'."""
-        mock_tool = _make_mock_tool("run_block")
-        with patch(
-            "backend.copilot.sdk.tool_adapter.TOOL_REGISTRY",
-            {"run_block": mock_tool},
-        ):
-            await pre_launch_tool_call("mcp__copilot__run_block", {"block_id": "b1"})
-
-        # The task was enqueued — mock_tool.execute should be called once
-        # (may not complete immediately but should start)
-        await asyncio.sleep(0)  # yield to event loop
-        mock_tool.execute.assert_awaited_once()
-
-    @pytest.mark.asyncio
-    async def test_bare_tool_name_without_prefix(self):
-        """Tool names without __ separator are looked up as-is."""
-        mock_tool = _make_mock_tool("run_block")
-        with patch(
-            "backend.copilot.sdk.tool_adapter.TOOL_REGISTRY",
-            {"run_block": mock_tool},
-        ):
-            await pre_launch_tool_call("run_block", {"block_id": "b1"})
-
-        await asyncio.sleep(0)
-        mock_tool.execute.assert_awaited_once()
-
-    @pytest.mark.asyncio
-    async def test_task_enqueued_fifo_for_same_tool(self):
-        """Two pre-launched calls for the same tool name are enqueued FIFO."""
-        results = []
-
-        async def slow_execute(*args, **kwargs):
-            results.append(len(results))
-            return StreamToolOutputAvailable(
-                toolCallId="id",
-                output=str(len(results) - 1),
-                toolName="t",
-                success=True,
-            )
-
-        mock_tool = _make_mock_tool("t")
-        mock_tool.execute = AsyncMock(side_effect=slow_execute)
-
-        with patch(
-            "backend.copilot.sdk.tool_adapter.TOOL_REGISTRY",
-            {"t": mock_tool},
-        ):
-            await pre_launch_tool_call("t", {"n": 1})
-            await pre_launch_tool_call("t", {"n": 2})
-            await asyncio.sleep(0)
-
-        assert mock_tool.execute.await_count == 2
-
-    @pytest.mark.asyncio
-    async def test_file_ref_expansion_failure_skips_pre_launch(self):
-        """When @@agptfile: expansion fails, pre_launch_tool_call skips the task.
-
-        The handler should then fall back to direct execution (which will also
-        fail with a proper MCP error via _truncating's own expansion).
-        """
-        mock_tool = _make_mock_tool("run_block", output="should-not-execute")
-
-        with (
-            patch(
-                "backend.copilot.sdk.tool_adapter.TOOL_REGISTRY",
-                {"run_block": mock_tool},
-            ),
-            patch(
-                "backend.copilot.sdk.tool_adapter.expand_file_refs_in_args",
-                AsyncMock(side_effect=FileRefExpansionError("@@agptfile:missing.txt")),
-            ),
-        ):
-            # Should not raise — expansion failure is handled gracefully
-            await pre_launch_tool_call("run_block", {"text": "@@agptfile:missing.txt"})
-            await asyncio.sleep(0)
-
-        # No task was pre-launched — execute was not called
-        mock_tool.execute.assert_not_awaited()
-
-
-class TestCreateToolHandlerParallel:
-    """Tests for create_tool_handler using pre-launched tasks."""
-
-    @pytest.fixture(autouse=True)
-    def _init(self):
-        _init_ctx(session=_make_mock_session())
-
-    @pytest.mark.asyncio
-    async def test_handler_uses_prelaunched_task(self):
-        """Handler pops and awaits the pre-launched task rather than re-executing."""
-        mock_tool = _make_mock_tool("run_block", output="pre-launched result")
-
-        with patch(
-            "backend.copilot.sdk.tool_adapter.TOOL_REGISTRY",
-            {"run_block": mock_tool},
-        ):
-            await pre_launch_tool_call("run_block", {"block_id": "b1"})
-            await asyncio.sleep(0)  # let task start
-
-            handler = create_tool_handler(mock_tool)
-            result = await handler({"block_id": "b1"})
-
-        assert result["isError"] is False
-        text = result["content"][0]["text"]
-        assert "pre-launched result" in text
-        # Should only have been called once (the pre-launched task), not twice
-        mock_tool.execute.assert_awaited_once()
-
-    @pytest.mark.asyncio
-    async def test_handler_does_not_double_stash_for_prelaunched_task(self):
-        """Pre-launched task result must NOT be stashed by tool_handler directly.
-
-        The _truncating wrapper wraps tool_handler and handles stashing after
-        tool_handler returns.  If tool_handler also stashed, the output would be
-        appended twice to the FIFO queue and pop_pending_tool_output would return
-        a duplicate on the second call.
-
-        This test calls tool_handler directly (without _truncating) and asserts
-        that nothing was stashed — confirming stashing is deferred to _truncating.
-        """
-        mock_tool = _make_mock_tool("run_block", output="stash-me")
-
-        with patch(
-            "backend.copilot.sdk.tool_adapter.TOOL_REGISTRY",
-            {"run_block": mock_tool},
-        ):
-            await pre_launch_tool_call("run_block", {"block_id": "b1"})
-            await asyncio.sleep(0)
-
-            handler = create_tool_handler(mock_tool)
-            result = await handler({"block_id": "b1"})
-
-        assert result["isError"] is False
-        assert "stash-me" in result["content"][0]["text"]
-        # tool_handler must NOT stash — _truncating (which wraps handler) does it.
-        # Calling pop here (without going through _truncating) should return None.
-        not_stashed = pop_pending_tool_output("run_block")
-        assert not_stashed is None, (
-            "tool_handler must not stash directly — _truncating handles stashing "
-            "to prevent double-stash in the FIFO queue"
-        )
-
-    @pytest.mark.asyncio
-    async def test_handler_falls_back_when_queue_empty(self):
-        """When no pre-launched task exists, handler executes directly."""
+    async def test_handler_executes_tool_directly(self):
+        """Handler executes the tool and returns MCP-formatted result."""
         mock_tool = _make_mock_tool("run_block", output="direct result")
 
-        # Don't call pre_launch_tool_call — queue is empty
         handler = create_tool_handler(mock_tool)
         result = await handler({"block_id": "b1"})
 
@@ -449,104 +297,9 @@ class TestCreateToolHandlerParallel:
         mock_tool.execute.assert_awaited_once()
 
     @pytest.mark.asyncio
-    async def test_handler_cancelled_error_propagates(self):
-        """CancelledError from a pre-launched task is re-raised to preserve cancellation semantics."""
+    async def test_handler_returns_error_on_no_session(self):
+        """When session is None, handler returns MCP error."""
         mock_tool = _make_mock_tool("run_block")
-        mock_tool.execute = AsyncMock(side_effect=asyncio.CancelledError())
-
-        with patch(
-            "backend.copilot.sdk.tool_adapter.TOOL_REGISTRY",
-            {"run_block": mock_tool},
-        ):
-            await pre_launch_tool_call("run_block", {"block_id": "b1"})
-            await asyncio.sleep(0)
-
-            handler = create_tool_handler(mock_tool)
-            with pytest.raises(asyncio.CancelledError):
-                await handler({"block_id": "b1"})
-
-    @pytest.mark.asyncio
-    async def test_handler_exception_returns_mcp_error(self):
-        """Exception from a pre-launched task is caught and returned as MCP error."""
-        mock_tool = _make_mock_tool("run_block")
-        mock_tool.execute = AsyncMock(side_effect=RuntimeError("block exploded"))
-
-        with patch(
-            "backend.copilot.sdk.tool_adapter.TOOL_REGISTRY",
-            {"run_block": mock_tool},
-        ):
-            await pre_launch_tool_call("run_block", {"block_id": "b1"})
-            await asyncio.sleep(0)
-
-            handler = create_tool_handler(mock_tool)
-            result = await handler({"block_id": "b1"})
-
-        assert result["isError"] is True
-        assert "Failed to execute run_block" in result["content"][0]["text"]
-
-    @pytest.mark.asyncio
-    async def test_two_same_tool_calls_dispatched_in_order(self):
-        """Two pre-launched tasks for the same tool are consumed in FIFO order."""
-        call_order = []
-
-        async def execute_with_tag(*args, **kwargs):
-            tag = kwargs.get("block_id", "?")
-            call_order.append(tag)
-            return StreamToolOutputAvailable(
-                toolCallId="id", output=f"out-{tag}", toolName="run_block", success=True
-            )
-
-        mock_tool = _make_mock_tool("run_block")
-        mock_tool.execute = AsyncMock(side_effect=execute_with_tag)
-
-        with patch(
-            "backend.copilot.sdk.tool_adapter.TOOL_REGISTRY",
-            {"run_block": mock_tool},
-        ):
-            await pre_launch_tool_call("run_block", {"block_id": "first"})
-            await pre_launch_tool_call("run_block", {"block_id": "second"})
-            await asyncio.sleep(0)
-
-            handler = create_tool_handler(mock_tool)
-            r1 = await handler({"block_id": "first"})
-            r2 = await handler({"block_id": "second"})
-
-        assert "out-first" in r1["content"][0]["text"]
-        assert "out-second" in r2["content"][0]["text"]
-        assert call_order == [
-            "first",
-            "second",
-        ], f"Expected FIFO dispatch order but got {call_order}"
-
-    @pytest.mark.asyncio
-    async def test_arg_mismatch_falls_back_to_direct_execution(self):
-        """When pre-launched args differ from SDK args, handler cancels pre-launched
-        task and falls back to direct execution with the correct args."""
-        mock_tool = _make_mock_tool("run_block", output="direct-result")
-
-        with patch(
-            "backend.copilot.sdk.tool_adapter.TOOL_REGISTRY",
-            {"run_block": mock_tool},
-        ):
-            # Pre-launch with args {"block_id": "wrong"}
-            await pre_launch_tool_call("run_block", {"block_id": "wrong"})
-            await asyncio.sleep(0)
-
-            # SDK dispatches with different args
-            handler = create_tool_handler(mock_tool)
-            result = await handler({"block_id": "correct"})
-
-        assert result["isError"] is False
-        # The tool was called twice: once by pre-launch (wrong args), once by
-        # direct fallback (correct args). The result should come from the
-        # direct execution path.
-        assert mock_tool.execute.await_count == 2
-
-    @pytest.mark.asyncio
-    async def test_no_session_falls_back_gracefully(self):
-        """When session is None and no pre-launched task, handler returns MCP error."""
-        mock_tool = _make_mock_tool("run_block")
-        # session=None means get_execution_context returns (user_id, None)
         set_execution_context(user_id="u", session=None, sandbox=None)  # type: ignore[arg-type]
 
         handler = create_tool_handler(mock_tool)
@@ -555,220 +308,406 @@ class TestCreateToolHandlerParallel:
         assert result["isError"] is True
         assert "session" in result["content"][0]["text"].lower()
 
-
-# ---------------------------------------------------------------------------
-# cancel_pending_tool_tasks
-# ---------------------------------------------------------------------------
-
-
-class TestCancelPendingToolTasks:
-    """Tests for cancel_pending_tool_tasks — the stream-abort cleanup helper."""
-
-    @pytest.fixture(autouse=True)
-    def _init(self):
-        _init_ctx(session=_make_mock_session())
-
     @pytest.mark.asyncio
-    async def test_cancels_queued_tasks(self):
-        """Queued tasks are cancelled and the queue is cleared."""
-        ran = False
-
-        async def never_run(*_args, **_kwargs):
-            nonlocal ran
-            await asyncio.sleep(10)  # long enough to still be pending
-            ran = True
-
+    async def test_handler_returns_error_on_exception(self):
+        """Exception from tool execution is caught and returned as MCP error."""
         mock_tool = _make_mock_tool("run_block")
-        mock_tool.execute = AsyncMock(side_effect=never_run)
+        mock_tool.execute = AsyncMock(side_effect=RuntimeError("block exploded"))
 
-        with patch(
-            "backend.copilot.sdk.tool_adapter.TOOL_REGISTRY",
-            {"run_block": mock_tool},
-        ):
-            await pre_launch_tool_call("run_block", {"block_id": "b1"})
-            await asyncio.sleep(0)  # let task start
-            await cancel_pending_tool_tasks()
-            await asyncio.sleep(0)  # let cancellation propagate
+        handler = create_tool_handler(mock_tool)
+        result = await handler({"block_id": "b1"})
 
-        assert not ran, "Task should have been cancelled before completing"
+        assert result["isError"] is True
+        assert "Failed to execute run_block" in result["content"][0]["text"]
 
     @pytest.mark.asyncio
-    async def test_noop_when_no_tasks_queued(self):
-        """cancel_pending_tool_tasks does not raise when queues are empty."""
-        await cancel_pending_tool_tasks()  # should not raise
+    async def test_handler_executes_once_per_call(self):
+        """Each handler call executes the tool exactly once — no duplicate execution."""
+        mock_tool = _make_mock_tool("run_block", output="single-execution")
 
-    @pytest.mark.asyncio
-    async def test_handler_does_not_find_cancelled_task(self):
-        """After cancel, tool_handler falls back to direct execution."""
-        mock_tool = _make_mock_tool("run_block", output="direct-fallback")
+        handler = create_tool_handler(mock_tool)
+        await handler({"block_id": "b1"})
+        await handler({"block_id": "b2"})
 
-        with patch(
-            "backend.copilot.sdk.tool_adapter.TOOL_REGISTRY",
-            {"run_block": mock_tool},
-        ):
-            await pre_launch_tool_call("run_block", {"block_id": "b1"})
-            await asyncio.sleep(0)
-            await cancel_pending_tool_tasks()
-
-            # Queue is now empty — handler should execute directly
-            handler = create_tool_handler(mock_tool)
-            result = await handler({"block_id": "b1"})
-
-        assert result["isError"] is False
-        assert "direct-fallback" in result["content"][0]["text"]
-
-
-# ---------------------------------------------------------------------------
-# Concurrent / parallel pre-launch scenarios
-# ---------------------------------------------------------------------------
-
-
-class TestAllParallelToolsPrelaunchedIndependently:
-    """Simulate SDK sending N separate AssistantMessages for the same tool concurrently."""
-
-    @pytest.fixture(autouse=True)
-    def _init(self):
-        _init_ctx(session=_make_mock_session())
-
-    @pytest.mark.asyncio
-    async def test_all_parallel_tools_prelaunched_independently(self):
-        """5 pre-launches for the same tool all enqueue independently and run concurrently.
-
-        Each task sleeps for PER_TASK_S seconds. If they ran sequentially the total
-        wall time would be ~5*PER_TASK_S. Running concurrently it should finish in
-        roughly PER_TASK_S (plus scheduling overhead).
-        """
-        PER_TASK_S = 0.05
-        N = 5
-        started: list[int] = []
-        finished: list[int] = []
-
-        async def slow_execute(*args, **kwargs):
-            idx = len(started)
-            started.append(idx)
-            await asyncio.sleep(PER_TASK_S)
-            finished.append(idx)
-            return StreamToolOutputAvailable(
-                toolCallId=f"id-{idx}",
-                output=f"result-{idx}",
-                toolName="bash_exec",
-                success=True,
-            )
-
-        mock_tool = _make_mock_tool("bash_exec")
-        mock_tool.execute = AsyncMock(side_effect=slow_execute)
-
-        with patch(
-            "backend.copilot.sdk.tool_adapter.TOOL_REGISTRY",
-            {"bash_exec": mock_tool},
-        ):
-            for i in range(N):
-                await pre_launch_tool_call("bash_exec", {"cmd": f"echo {i}"})
-
-            # Measure only the concurrent execution window, not pre-launch overhead.
-            # Starting the timer here avoids false failures on slow CI runners where
-            # the pre_launch_tool_call setup takes longer than the concurrent sleep.
-            t0 = asyncio.get_running_loop().time()
-            await asyncio.sleep(PER_TASK_S * 2)
-            elapsed = asyncio.get_running_loop().time() - t0
-
-        assert mock_tool.execute.await_count == N
-        assert len(finished) == N
-        # Wall time of the sleep window should be well under N * PER_TASK_S
-        # (sequential would be ~0.25s; concurrent finishes in ~PER_TASK_S = 0.05s)
-        assert elapsed < N * PER_TASK_S, (
-            f"Expected concurrent execution (<{N * PER_TASK_S:.2f}s) "
-            f"but sleep window took {elapsed:.2f}s"
-        )
-
-
-class TestHandlerReturnsResultFromCorrectPrelaunchedTask:
-    """Pop pre-launched tasks in order and verify each returns its own result."""
-
-    @pytest.fixture(autouse=True)
-    def _init(self):
-        _init_ctx(session=_make_mock_session())
-
-    @pytest.mark.asyncio
-    async def test_handler_returns_result_from_correct_prelaunched_task(self):
-        """Two pre-launches for the same tool: first handler gets first result, second gets second."""
-
-        async def execute_with_cmd(*args, **kwargs):
-            cmd = kwargs.get("cmd", "?")
-            return StreamToolOutputAvailable(
-                toolCallId="id",
-                output=f"output-for-{cmd}",
-                toolName="bash_exec",
-                success=True,
-            )
-
-        mock_tool = _make_mock_tool("bash_exec")
-        mock_tool.execute = AsyncMock(side_effect=execute_with_cmd)
-
-        with patch(
-            "backend.copilot.sdk.tool_adapter.TOOL_REGISTRY",
-            {"bash_exec": mock_tool},
-        ):
-            await pre_launch_tool_call("bash_exec", {"cmd": "alpha"})
-            await pre_launch_tool_call("bash_exec", {"cmd": "beta"})
-            await asyncio.sleep(0)  # let both tasks start
-
-            handler = create_tool_handler(mock_tool)
-            r1 = await handler({"cmd": "alpha"})
-            r2 = await handler({"cmd": "beta"})
-
-        text1 = r1["content"][0]["text"]
-        text2 = r2["content"][0]["text"]
-        assert "output-for-alpha" in text1, f"Expected alpha result, got: {text1}"
-        assert "output-for-beta" in text2, f"Expected beta result, got: {text2}"
         assert mock_tool.execute.await_count == 2
 
 
-class TestFiveConcurrentPrelaunchAllComplete:
-    """Pre-launch 5 tasks; consume all 5 via handlers; assert all succeed."""
+# ---------------------------------------------------------------------------
+# Regression tests: bugs fixed by removing pre-launch mechanism
+#
+# Each test class includes a _buggy_handler fixture that reproduces the old
+# pre-launch implementation inline.  Tests run against BOTH the buggy handler
+# (xfail — proves the bug exists) and the current clean handler (must pass).
+# ---------------------------------------------------------------------------
+
+
+def _make_execute_fn(tool_name: str = "run_block"):
+    """Return (execute_fn, call_log) — execute_fn records every call."""
+    call_log: list[dict] = []
+
+    async def execute_fn(*args, **kwargs):
+        call_log.append(kwargs)
+        return StreamToolOutputAvailable(
+            toolCallId=f"id-{len(call_log)}",
+            output=f"result-{len(call_log)}",
+            toolName=tool_name,
+            success=True,
+        )
+
+    return execute_fn, call_log
+
+
+async def _buggy_prelaunch_handler(mock_tool, pre_launch_args, dispatch_args):
+    """Simulate the OLD buggy pre-launch flow.
+
+    1. pre_launch_tool_call fires _execute_tool_sync with pre_launch_args
+    2. SDK dispatches handler with dispatch_args
+    3. Handler compares args — on mismatch, cancels + re-executes (BUG)
+
+    Returns the handler result.
+    """
+    from backend.copilot.sdk.tool_adapter import _execute_tool_sync
+
+    user_id, session = "user-1", _make_mock_session()
+
+    # Step 1: pre-launch fires immediately (speculative)
+    task = asyncio.create_task(
+        _execute_tool_sync(mock_tool, user_id, session, pre_launch_args)
+    )
+    await asyncio.sleep(0)  # let task start
+
+    # Step 2: SDK dispatches with (potentially different) args
+    if pre_launch_args != dispatch_args:
+        # Arg mismatch path: cancel pre-launched task + re-execute
+        if not task.done():
+            task.cancel()
+            try:
+                await task
+            except (asyncio.CancelledError, Exception):
+                pass
+        # Fall through to direct execution (duplicate!)
+        return await _execute_tool_sync(mock_tool, user_id, session, dispatch_args)
+    else:
+        return await task
+
+
+class TestBug1DuplicateExecution:
+    """Bug 1 (SECRT-2204): arg mismatch causes duplicate execution.
+
+    Pre-launch fires with raw args, SDK dispatches with normalised args.
+    Mismatch → cancel (too late) + re-execute → 2 API calls.
+    """
 
     @pytest.fixture(autouse=True)
     def _init(self):
         _init_ctx(session=_make_mock_session())
 
+    @pytest.mark.xfail(reason="Old pre-launch code causes duplicate execution")
     @pytest.mark.asyncio
-    async def test_five_concurrent_prelaunch_all_complete(self):
-        """All 5 pre-launched tasks complete and return successful results."""
-        N = 5
-        call_count = 0
+    async def test_old_code_duplicates_on_arg_mismatch(self):
+        """OLD CODE: pre-launch with args A, dispatch with args B → 2 calls."""
+        execute_fn, call_log = _make_execute_fn()
+        mock_tool = _make_mock_tool("run_block")
+        mock_tool.execute = AsyncMock(side_effect=execute_fn)
 
-        async def counting_execute(*args, **kwargs):
-            nonlocal call_count
-            call_count += 1
-            n = call_count
+        pre_launch_args = {"block_id": "b1", "input_data": {"title": "Test"}}
+        dispatch_args = {
+            "block_id": "b1",
+            "input_data": {"title": "Test", "priority": None},
+        }
+
+        await _buggy_prelaunch_handler(mock_tool, pre_launch_args, dispatch_args)
+
+        # BUG: pre-launch executed once + fallback executed again = 2
+        assert len(call_log) == 1, (
+            f"Expected 1 execution but got {len(call_log)} — "
+            f"duplicate execution bug!"
+        )
+
+    @pytest.mark.asyncio
+    async def test_current_code_no_duplicate(self):
+        """FIXED: handler executes exactly once regardless of arg shape."""
+        execute_fn, call_log = _make_execute_fn()
+        mock_tool = _make_mock_tool("run_block")
+        mock_tool.execute = AsyncMock(side_effect=execute_fn)
+
+        handler = create_tool_handler(mock_tool)
+        await handler({"block_id": "b1", "input_data": {"title": "Test"}})
+
+        assert len(call_log) == 1, f"Expected 1 execution but got {len(call_log)}"
+
+
+class TestBug2FIFODesync:
+    """Bug 2: FIFO desync when security hook denies a tool.
+
+    Pre-launch queues [task_A, task_B]. Tool A denied (no MCP dispatch).
+    Tool B's handler dequeues task_A → returns wrong result.
+    """
+
+    @pytest.fixture(autouse=True)
+    def _init(self):
+        _init_ctx(session=_make_mock_session())
+
+    @pytest.mark.xfail(reason="Old FIFO queue returns wrong result on denial")
+    @pytest.mark.asyncio
+    async def test_old_code_fifo_desync_on_denial(self):
+        """OLD CODE: denied tool's task stays in queue, next tool gets wrong result."""
+        from backend.copilot.sdk.tool_adapter import _execute_tool_sync
+
+        call_log: list[str] = []
+
+        async def tagged_execute(*args, **kwargs):
+            tag = kwargs.get("block_id", "?")
+            call_log.append(tag)
             return StreamToolOutputAvailable(
-                toolCallId=f"id-{n}",
-                output=f"done-{n}",
-                toolName="bash_exec",
+                toolCallId="id",
+                output=f"result-for-{tag}",
+                toolName="run_block",
                 success=True,
             )
 
-        mock_tool = _make_mock_tool("bash_exec")
-        mock_tool.execute = AsyncMock(side_effect=counting_execute)
+        mock_tool = _make_mock_tool("run_block")
+        mock_tool.execute = AsyncMock(side_effect=tagged_execute)
+        user_id, session = "user-1", _make_mock_session()
 
-        with patch(
-            "backend.copilot.sdk.tool_adapter.TOOL_REGISTRY",
-            {"bash_exec": mock_tool},
-        ):
-            for i in range(N):
-                await pre_launch_tool_call("bash_exec", {"cmd": f"task-{i}"})
+        # Simulate old FIFO queue
+        queue: asyncio.Queue = asyncio.Queue()
 
-            await asyncio.sleep(0)  # let all tasks start
+        # Pre-launch for tool A and tool B
+        task_a = asyncio.create_task(
+            _execute_tool_sync(mock_tool, user_id, session, {"block_id": "A"})
+        )
+        task_b = asyncio.create_task(
+            _execute_tool_sync(mock_tool, user_id, session, {"block_id": "B"})
+        )
+        queue.put_nowait(task_a)
+        queue.put_nowait(task_b)
+        await asyncio.sleep(0)  # let both tasks run
 
-            handler = create_tool_handler(mock_tool)
-            results = []
-            for i in range(N):
-                results.append(await handler({"cmd": f"task-{i}"}))
+        # Tool A is DENIED by security hook — no MCP dispatch, no dequeue
+        # Tool B's handler dequeues from FIFO → gets task_A!
+        dequeued_task = queue.get_nowait()
+        result = await dequeued_task
+        result_text = result["content"][0]["text"]
 
-        assert (
-            mock_tool.execute.await_count == N
-        ), f"Expected {N} execute calls, got {mock_tool.execute.await_count}"
-        for i, result in enumerate(results):
-            assert result["isError"] is False, f"Result {i} should not be an error"
-            text = result["content"][0]["text"]
-            assert "done-" in text, f"Result {i} missing expected output: {text}"
+        # BUG: handler for B got task_A's result
+        assert "result-for-B" in result_text, (
+            f"Expected result for B but got: {result_text} — "
+            f"FIFO desync: B got A's result!"
+        )
+
+    @pytest.mark.asyncio
+    async def test_current_code_no_fifo_desync(self):
+        """FIXED: each handler call executes independently, no shared queue."""
+        call_log: list[str] = []
+
+        async def tagged_execute(*args, **kwargs):
+            tag = kwargs.get("block_id", "?")
+            call_log.append(tag)
+            return StreamToolOutputAvailable(
+                toolCallId="id",
+                output=f"result-for-{tag}",
+                toolName="run_block",
+                success=True,
+            )
+
+        mock_tool = _make_mock_tool("run_block")
+        mock_tool.execute = AsyncMock(side_effect=tagged_execute)
+
+        handler = create_tool_handler(mock_tool)
+
+        # Tool A denied (never called). Tool B dispatched normally.
+        result_b = await handler({"block_id": "B"})
+
+        assert "result-for-B" in result_b["content"][0]["text"]
+        assert call_log == ["B"]
+
+
+class TestBug3CancelRace:
+    """Bug 3: cancel race — task completes before cancel arrives.
+
+    Pre-launch fires fast HTTP call (< 1s). By the time handler detects
+    mismatch and calls task.cancel(), the API call already completed.
+    Side effect (Linear issue created) is irreversible.
+    """
+
+    @pytest.fixture(autouse=True)
+    def _init(self):
+        _init_ctx(session=_make_mock_session())
+
+    @pytest.mark.xfail(reason="Old code: cancel arrives after task completes")
+    @pytest.mark.asyncio
+    async def test_old_code_cancel_arrives_too_late(self):
+        """OLD CODE: fast task completes before cancel, side effect persists."""
+        side_effects: list[str] = []
+
+        async def fast_execute_with_side_effect(*args, **kwargs):
+            # Side effect happens immediately (like an HTTP POST to Linear)
+            side_effects.append("created-issue")
+            return StreamToolOutputAvailable(
+                toolCallId="id",
+                output="issue-created",
+                toolName="run_block",
+                success=True,
+            )
+
+        mock_tool = _make_mock_tool("run_block")
+        mock_tool.execute = AsyncMock(side_effect=fast_execute_with_side_effect)
+
+        # Pre-launch fires immediately
+        pre_launch_args = {"block_id": "b1"}
+        dispatch_args = {"block_id": "b1", "extra": "normalised"}
+
+        await _buggy_prelaunch_handler(mock_tool, pre_launch_args, dispatch_args)
+
+        # BUG: side effect happened TWICE (pre-launch + fallback)
+        assert len(side_effects) == 1, (
+            f"Expected 1 side effect but got {len(side_effects)} — "
+            f"cancel race: pre-launch completed before cancel!"
+        )
+
+    @pytest.mark.asyncio
+    async def test_current_code_single_side_effect(self):
+        """FIXED: no speculative execution, exactly 1 side effect per call."""
+        side_effects: list[str] = []
+
+        async def execute_with_side_effect(*args, **kwargs):
+            side_effects.append("created-issue")
+            return StreamToolOutputAvailable(
+                toolCallId="id",
+                output="issue-created",
+                toolName="run_block",
+                success=True,
+            )
+
+        mock_tool = _make_mock_tool("run_block")
+        mock_tool.execute = AsyncMock(side_effect=execute_with_side_effect)
+
+        handler = create_tool_handler(mock_tool)
+        await handler({"block_id": "b1"})
+
+        assert len(side_effects) == 1
+
+
+# ---------------------------------------------------------------------------
+# readOnlyHint annotations
+# ---------------------------------------------------------------------------
+
+
+class TestReadOnlyAnnotations:
+    """Tests that all tools get readOnlyHint=True for parallel dispatch."""
+
+    def test_parallel_annotation_constant(self):
+        """_PARALLEL_ANNOTATION is a ToolAnnotations with readOnlyHint=True."""
+        from .tool_adapter import _PARALLEL_ANNOTATION
+
+        assert isinstance(_PARALLEL_ANNOTATION, ToolAnnotations)
+        assert _PARALLEL_ANNOTATION.readOnlyHint is True
+
+
+# ---------------------------------------------------------------------------
+# SDK_DISALLOWED_TOOLS
+# ---------------------------------------------------------------------------
+
+
+class TestSDKDisallowedTools:
+    """Verify that dangerous SDK built-in tools are in the disallowed list."""
+
+    def test_bash_tool_is_disallowed(self):
+        assert "Bash" in SDK_DISALLOWED_TOOLS
+
+    def test_webfetch_tool_is_disallowed(self):
+        """WebFetch is disallowed due to SSRF risk."""
+        assert "WebFetch" in SDK_DISALLOWED_TOOLS
+
+
+# ---------------------------------------------------------------------------
+# _read_file_handler — bridge_and_annotate integration
+# ---------------------------------------------------------------------------
+
+
+class TestReadFileHandlerBridge:
+    """Verify that _read_file_handler calls bridge_and_annotate when a sandbox is active."""
+
+    @pytest.fixture(autouse=True)
+    def _init_context(self):
+        set_execution_context(
+            user_id="test",
+            session=None,  # type: ignore[arg-type]
+            sandbox=None,
+            sdk_cwd="/tmp/copilot-bridge-test",
+        )
+
+    @pytest.mark.asyncio
+    async def test_bridge_called_when_sandbox_active(self, tmp_path, monkeypatch):
+        """When a sandbox is set, bridge_and_annotate is called and its annotation appended."""
+        from backend.copilot.context import _current_sandbox
+
+        from .tool_adapter import _read_file_handler
+
+        test_file = tmp_path / "tool-results" / "data.json"
+        test_file.parent.mkdir(parents=True, exist_ok=True)
+        test_file.write_text('{"ok": true}\n')
+
+        monkeypatch.setattr(
+            "backend.copilot.sdk.tool_adapter.is_allowed_local_path",
+            lambda path, cwd: True,
+        )
+
+        fake_sandbox = object()
+        token = _current_sandbox.set(fake_sandbox)  # type: ignore[arg-type]
+        try:
+            bridge_calls: list[tuple] = []
+
+            async def fake_bridge_and_annotate(sandbox, file_path, offset, limit):
+                bridge_calls.append((sandbox, file_path, offset, limit))
+                return "\n[Sandbox copy available at /tmp/abc-data.json]"
+
+            monkeypatch.setattr(
+                "backend.copilot.sdk.tool_adapter.bridge_and_annotate",
+                fake_bridge_and_annotate,
+            )
+
+            result = await _read_file_handler(
+                {"file_path": str(test_file), "offset": 0, "limit": 2000}
+            )
+
+            assert result["isError"] is False
+            assert len(bridge_calls) == 1
+            assert bridge_calls[0][0] is fake_sandbox
+            assert "/tmp/abc-data.json" in result["content"][0]["text"]
+        finally:
+            _current_sandbox.reset(token)
+
+    @pytest.mark.asyncio
+    async def test_bridge_not_called_without_sandbox(self, tmp_path, monkeypatch):
+        """When no sandbox is set, bridge_and_annotate is not called."""
+        from .tool_adapter import _read_file_handler
+
+        test_file = tmp_path / "tool-results" / "data.json"
+        test_file.parent.mkdir(parents=True, exist_ok=True)
+        test_file.write_text('{"ok": true}\n')
+
+        monkeypatch.setattr(
+            "backend.copilot.sdk.tool_adapter.is_allowed_local_path",
+            lambda path, cwd: True,
+        )
+
+        bridge_calls: list[tuple] = []
+
+        async def fake_bridge_and_annotate(sandbox, file_path, offset, limit):
+            bridge_calls.append((sandbox, file_path, offset, limit))
+            return "\n[Sandbox copy available at /tmp/abc-data.json]"
+
+        monkeypatch.setattr(
+            "backend.copilot.sdk.tool_adapter.bridge_and_annotate",
+            fake_bridge_and_annotate,
+        )
+
+        result = await _read_file_handler(
+            {"file_path": str(test_file), "offset": 0, "limit": 2000}
+        )
+
+        assert result["isError"] is False
+        assert len(bridge_calls) == 0
+        assert "Sandbox copy" not in result["content"][0]["text"]

autogpt_platform/backend/backend/copilot/sdk/transcript.py

@@ -43,6 +43,10 @@ STRIPPABLE_TYPES = frozenset(
     {"progress", "file-history-snapshot", "queue-operation", "summary", "pr-link"}
 )
 
+# Thinking block types that can be stripped from non-last assistant entries.
+# The Anthropic API only requires these in the *last* assistant message.
+_THINKING_BLOCK_TYPES = frozenset({"thinking", "redacted_thinking"})
+
 
 @dataclass
 class TranscriptDownload:
@@ -450,6 +454,83 @@ def _build_meta_storage_path(user_id: str, session_id: str, backend: object) ->
     )
 
 
+def strip_stale_thinking_blocks(content: str) -> str:
+    """Remove thinking/redacted_thinking blocks from non-last assistant entries.
+
+    The Anthropic API only requires thinking blocks in the **last** assistant
+    message to be value-identical to the original response.  Older assistant
+    entries carry stale thinking blocks that consume significant tokens
+    (often 10-50K each) without providing useful context for ``--resume``.
+
+    Stripping them before upload prevents the CLI from triggering compaction
+    every turn just to compress away the stale thinking bloat.
+    """
+    lines = content.strip().split("\n")
+    if not lines:
+        return content
+
+    parsed: list[tuple[str, dict | None]] = []
+    for line in lines:
+        parsed.append((line, json.loads(line, fallback=None)))
+
+    # Reverse scan to find the last assistant message ID and index.
+    last_asst_msg_id: str | None = None
+    last_asst_idx: int | None = None
+    for i in range(len(parsed) - 1, -1, -1):
+        _line, entry = parsed[i]
+        if not isinstance(entry, dict):
+            continue
+        msg = entry.get("message", {})
+        if msg.get("role") == "assistant":
+            last_asst_msg_id = msg.get("id")
+            last_asst_idx = i
+            break
+
+    if last_asst_idx is None:
+        return content
+
+    result_lines: list[str] = []
+    stripped_count = 0
+    for i, (line, entry) in enumerate(parsed):
+        if not isinstance(entry, dict):
+            result_lines.append(line)
+            continue
+
+        msg = entry.get("message", {})
+        # Only strip from assistant entries that are NOT the last turn.
+        # Use msg_id matching when available; fall back to index for entries
+        # without an id field.
+        is_last_turn = (
+            last_asst_msg_id is not None and msg.get("id") == last_asst_msg_id
+        ) or (last_asst_msg_id is None and i == last_asst_idx)
+        if (
+            msg.get("role") == "assistant"
+            and not is_last_turn
+            and isinstance(msg.get("content"), list)
+        ):
+            content_blocks = msg["content"]
+            filtered = [
+                b
+                for b in content_blocks
+                if not (isinstance(b, dict) and b.get("type") in _THINKING_BLOCK_TYPES)
+            ]
+            if len(filtered) < len(content_blocks):
+                stripped_count += len(content_blocks) - len(filtered)
+                entry = {**entry, "message": {**msg, "content": filtered}}
+                result_lines.append(json.dumps(entry, separators=(",", ":")))
+                continue
+
+        result_lines.append(line)
+
+    if stripped_count:
+        logger.info(
+            "[Transcript] Stripped %d stale thinking block(s) from non-last entries",
+            stripped_count,
+        )
+
+    return "\n".join(result_lines) + "\n"
+
+
 async def upload_transcript(
     user_id: str,
     session_id: str,
@@ -472,6 +553,9 @@ async def upload_transcript(
     # Strip metadata entries (progress, file-history-snapshot, etc.)
     # Note: SDK-built transcripts shouldn't have these, but strip for safety
     stripped = strip_progress_entries(content)
+    # Strip stale thinking blocks from older assistant entries — these consume
+    # significant tokens and trigger unnecessary CLI compaction every turn.
+    stripped = strip_stale_thinking_blocks(stripped)
     if not validate_transcript(stripped):
         # Log entry types for debugging — helps identify why validation failed
         entry_types = [
@@ -609,24 +693,35 @@ def _flatten_assistant_content(blocks: list) -> str:
     """Flatten assistant content blocks into a single plain-text string.
 
     Structured ``tool_use`` blocks are converted to ``[tool_use: name]``
-    placeholders.  This is intentional: ``compress_context`` requires plain
-    text for token counting and LLM summarization.  The structural loss is
-    acceptable because compaction only runs when the original transcript was
-    already too large for the model — a summarized plain-text version is
-    better than no context at all.
+    placeholders.  ``thinking`` and ``redacted_thinking`` blocks are
+    silently dropped — they carry no useful context for compression
+    summaries and must not leak into compacted transcripts (the Anthropic
+    API requires thinking blocks in the last assistant message to be
+    value-identical to the original response; including stale thinking
+    text would violate that constraint).
+
+    This is intentional: ``compress_context`` requires plain text for
+    token counting and LLM summarization.  The structural loss is
+    acceptable because compaction only runs when the original transcript
+    was already too large for the model.
     """
     parts: list[str] = []
     for block in blocks:
         if isinstance(block, dict):
             btype = block.get("type", "")
+            if btype in _THINKING_BLOCK_TYPES:
+                continue
             if btype == "text":
                 parts.append(block.get("text", ""))
             elif btype == "tool_use":
-                parts.append(f"[tool_use: {block.get('name', '?')}]")
+                # Drop tool_use entirely — any text representation gets
+                # mimicked by the model as plain text instead of actual
+                # structured tool calls. The tool results (in the
+                # following user/tool_result entry) provide sufficient
+                # context about what happened.
+                continue
             else:
-                # Preserve non-text blocks (e.g. image) as placeholders.
-                # Use __prefix__ to distinguish from literal user text.
-                parts.append(f"[__{btype}__]")
+                continue
         elif isinstance(block, str):
             parts.append(block)
     return "\n".join(parts) if parts else ""
@@ -805,6 +900,68 @@ async def _run_compression(
         )
 
 
+def _find_last_assistant_entry(
+    content: str,
+) -> tuple[list[str], list[str]]:
+    """Split JSONL lines into (compressible_prefix, preserved_tail).
+
+    The tail starts at the **first** entry of the last assistant turn and
+    includes everything after it (typically trailing user messages).  An
+    assistant turn can span multiple consecutive JSONL entries sharing the
+    same ``message.id`` (e.g., a thinking entry followed by a tool_use
+    entry).  All entries of the turn are preserved verbatim.
+
+    The Anthropic API requires that ``thinking`` and ``redacted_thinking``
+    blocks in the **last** assistant message remain value-identical to the
+    original response (the API validates parsed signature values, not raw
+    JSON bytes).  By excluding the entire turn from compression we
+    guarantee those blocks are never altered.
+
+    Returns ``(all_lines, [])`` when no assistant entry is found.
+    """
+    lines = [ln for ln in content.strip().split("\n") if ln.strip()]
+
+    # Parse all lines once to avoid double JSON deserialization.
+    # json.loads with fallback=None returns Any; non-dict entries are
+    # safely skipped by the isinstance(entry, dict) guards below.
+    parsed: list = [json.loads(ln, fallback=None) for ln in lines]
+
+    # Reverse scan: find the message.id and index of the last assistant entry.
+    last_asst_msg_id: str | None = None
+    last_asst_idx: int | None = None
+    for i in range(len(parsed) - 1, -1, -1):
+        entry = parsed[i]
+        if not isinstance(entry, dict):
+            continue
+        msg = entry.get("message", {})
+        if msg.get("role") == "assistant":
+            last_asst_idx = i
+            last_asst_msg_id = msg.get("id")
+            break
+
+    if last_asst_idx is None:
+        return lines, []
+
+    # If the assistant entry has no message.id, fall back to preserving
+    # from that single entry onward — safer than compressing everything.
+    if last_asst_msg_id is None:
+        return lines[:last_asst_idx], lines[last_asst_idx:]
+
+    # Forward scan: find the first entry of this turn (same message.id).
+    first_turn_idx: int | None = None
+    for i, entry in enumerate(parsed):
+        if not isinstance(entry, dict):
+            continue
+        msg = entry.get("message", {})
+        if msg.get("role") == "assistant" and msg.get("id") == last_asst_msg_id:
+            first_turn_idx = i
+            break
+
+    if first_turn_idx is None:
+        return lines, []
+    return lines[:first_turn_idx], lines[first_turn_idx:]
+
+
 async def compact_transcript(
     content: str,
     *,
@@ -816,42 +973,50 @@ async def compact_transcript(
     Converts transcript entries to plain messages, runs ``compress_context``
     (the same compressor used for pre-query history), and rebuilds JSONL.
 
-    Structured content (``tool_use`` blocks, ``tool_result`` nesting, images)
-    is flattened to plain text for compression.  This matches the fidelity of
-    the Plan C (DB compression) fallback path, where
-    ``_format_conversation_context`` similarly renders tool calls as
-    ``You called tool: name(args)`` and results as ``Tool result: ...``.
-    Neither path preserves structured API content blocks — the compacted
-    context serves as text history for the LLM, which creates proper
-    structured tool calls going forward.
+    The **last assistant entry** (and any entries after it) are preserved
+    verbatim — never flattened or compressed.  The Anthropic API requires
+    ``thinking`` and ``redacted_thinking`` blocks in the latest assistant
+    message to be value-identical to the original response (the API
+    validates parsed signature values, not raw JSON bytes); compressing
+    them would destroy the cryptographic signatures and cause
+    ``invalid_request_error``.
 
-    Images are per-turn attachments loaded from workspace storage by file ID
-    (via ``_prepare_file_attachments``), not part of the conversation history.
-    They are re-attached each turn and are unaffected by compaction.
+    Structured content in *older* assistant entries (``tool_use`` blocks,
+    ``thinking`` blocks, ``tool_result`` nesting, images) is flattened to
+    plain text for compression.  This matches the fidelity of the Plan C
+    (DB compression) fallback path.
 
     Returns the compacted JSONL string, or ``None`` on failure.
 
     See also:
         ``_compress_messages`` in ``service.py`` — compresses ``ChatMessage``
-        lists for pre-query DB history.  Both share ``compress_context()``
-        but operate on different input formats (JSONL transcript entries
-        here vs. ChatMessage dicts there).
+        lists for pre-query DB history.
     """
-    messages = _transcript_to_messages(content)
-    if len(messages) < 2:
-        logger.warning("%s Too few messages to compact (%d)", log_prefix, len(messages))
+    prefix_lines, tail_lines = _find_last_assistant_entry(content)
+
+    # Build the JSONL string for the compressible prefix
+    prefix_content = "\n".join(prefix_lines) + "\n" if prefix_lines else ""
+    messages = _transcript_to_messages(prefix_content) if prefix_content else []
+
+    if len(messages) + len(tail_lines) < 2:
+        total = len(messages) + len(tail_lines)
+        logger.warning("%s Too few messages to compact (%d)", log_prefix, total)
+        return None
+    if not messages:
+        logger.warning("%s Nothing to compress (only tail entries remain)", log_prefix)
         return None
     try:
         result = await _run_compression(messages, model, log_prefix)
         if not result.was_compacted:
-            # Compressor says it's within budget, but the SDK rejected it.
-            # Return None so the caller falls through to DB fallback.
             logger.warning(
                 "%s Compressor reports within budget but SDK rejected — "
                 "signalling failure",
                 log_prefix,
             )
             return None
+        if not result.messages:
+            logger.warning("%s Compressor returned empty messages", log_prefix)
+            return None
         logger.info(
             "%s Compacted transcript: %d->%d tokens (%d summarized, %d dropped)",
             log_prefix,
@@ -860,7 +1025,29 @@ async def compact_transcript(
             result.messages_summarized,
             result.messages_dropped,
         )
-        compacted = _messages_to_transcript(result.messages)
+        compressed_part = _messages_to_transcript(result.messages)
+
+        # Re-append the preserved tail (last assistant + trailing entries)
+        # with parentUuid patched to chain onto the compressed prefix.
+        tail_part = _rechain_tail(compressed_part, tail_lines)
+        compacted = compressed_part + tail_part
+
+        if len(compacted) >= len(content):
+            # Byte count can increase due to preserved tail entries
+            # (thinking blocks, JSON overhead) even when token count
+            # decreased.  Log a warning but still return — the API
+            # validates tokens not bytes, and the caller falls through
+            # to DB fallback if the transcript is still too large.
+            logger.warning(
+                "%s Compacted transcript (%d bytes) is not smaller than "
+                "original (%d bytes) — may still reduce token count",
+                log_prefix,
+                len(compacted),
+                len(content),
+            )
+        # Authoritative validation — the caller (_reduce_context) also
+        # validates, but this is the canonical check that guarantees we
+        # never return a malformed transcript from this function.
         if not validate_transcript(compacted):
             logger.warning("%s Compacted transcript failed validation", log_prefix)
             return None
@@ -870,3 +1057,43 @@ async def compact_transcript(
             "%s Transcript compaction failed: %s", log_prefix, e, exc_info=True
         )
         return None
+
+
+def _rechain_tail(compressed_prefix: str, tail_lines: list[str]) -> str:
+    """Patch tail entries so their parentUuid chain links to the compressed prefix.
+
+    The first tail entry's ``parentUuid`` is set to the ``uuid`` of the
+    last entry in the compressed prefix.  Subsequent tail entries are
+    rechained to point to their predecessor in the tail — their original
+    ``parentUuid`` values may reference entries that were compressed away.
+    """
+    if not tail_lines:
+        return ""
+    # Find the last uuid in the compressed prefix
+    last_prefix_uuid = ""
+    for line in reversed(compressed_prefix.strip().split("\n")):
+        if not line.strip():
+            continue
+        entry = json.loads(line, fallback=None)
+        if isinstance(entry, dict) and "uuid" in entry:
+            last_prefix_uuid = entry["uuid"]
+            break
+
+    result_lines: list[str] = []
+    prev_uuid: str | None = None
+    for i, line in enumerate(tail_lines):
+        entry = json.loads(line, fallback=None)
+        if not isinstance(entry, dict):
+            # Safety guard: _find_last_assistant_entry already filters empty
+            # lines, and well-formed JSONL always parses to dicts.  Non-dict
+            # lines are passed through unchanged; prev_uuid is intentionally
+            # NOT updated so the next dict entry chains to the last known uuid.
+            result_lines.append(line)
+            continue
+        if i == 0:
+            entry["parentUuid"] = last_prefix_uuid
+        elif prev_uuid is not None:
+            entry["parentUuid"] = prev_uuid
+        prev_uuid = entry.get("uuid")
+        result_lines.append(json.dumps(entry, separators=(",", ":")))
+    return "\n".join(result_lines) + "\n"

autogpt_platform/backend/backend/copilot/sdk/transcript_test.py

@@ -13,6 +13,7 @@ from .transcript import (
     delete_transcript,
     read_compacted_entries,
     strip_progress_entries,
+    strip_stale_thinking_blocks,
     validate_transcript,
     write_transcript_to_tempfile,
 )
@@ -1200,3 +1201,170 @@ class TestCleanupStaleProjectDirs:
         removed = cleanup_stale_project_dirs(encoded_cwd="some-other-project")
         assert removed == 0
         assert non_copilot.exists()
+
+
+# ---------------------------------------------------------------------------
+# strip_stale_thinking_blocks
+# ---------------------------------------------------------------------------
+
+
+class TestStripStaleThinkingBlocks:
+    """Tests for strip_stale_thinking_blocks — removes thinking/redacted_thinking
+    blocks from non-last assistant entries to reduce transcript bloat."""
+
+    def _asst_entry(
+        self, msg_id: str, content: list, uuid: str = "u1", parent: str = ""
+    ) -> dict:
+        return {
+            "type": "assistant",
+            "uuid": uuid,
+            "parentUuid": parent,
+            "message": {
+                "role": "assistant",
+                "id": msg_id,
+                "type": "message",
+                "content": content,
+            },
+        }
+
+    def _user_entry(self, text: str, uuid: str = "u0", parent: str = "") -> dict:
+        return {
+            "type": "user",
+            "uuid": uuid,
+            "parentUuid": parent,
+            "message": {"role": "user", "content": text},
+        }
+
+    def test_strips_thinking_from_older_assistant(self) -> None:
+        """Thinking blocks in non-last assistant entries should be removed."""
+        old_asst = self._asst_entry(
+            "msg_old",
+            [
+                {"type": "thinking", "thinking": "deep thoughts..."},
+                {"type": "text", "text": "hello"},
+                {"type": "redacted_thinking", "data": "secret"},
+            ],
+            uuid="a1",
+        )
+        new_asst = self._asst_entry(
+            "msg_new",
+            [
+                {"type": "thinking", "thinking": "latest thoughts"},
+                {"type": "text", "text": "world"},
+            ],
+            uuid="a2",
+            parent="a1",
+        )
+        content = _make_jsonl(old_asst, new_asst)
+        result = strip_stale_thinking_blocks(content)
+        lines = [json.loads(ln) for ln in result.strip().split("\n")]
+
+        # Old assistant should have thinking blocks stripped
+        old_content = lines[0]["message"]["content"]
+        assert len(old_content) == 1
+        assert old_content[0]["type"] == "text"
+
+        # New (last) assistant should be untouched
+        new_content = lines[1]["message"]["content"]
+        assert len(new_content) == 2
+        assert new_content[0]["type"] == "thinking"
+        assert new_content[1]["type"] == "text"
+
+    def test_preserves_last_assistant_thinking(self) -> None:
+        """The last assistant entry's thinking blocks must be preserved."""
+        entry = self._asst_entry(
+            "msg_only",
+            [
+                {"type": "thinking", "thinking": "must keep"},
+                {"type": "text", "text": "response"},
+            ],
+        )
+        content = _make_jsonl(entry)
+        result = strip_stale_thinking_blocks(content)
+        lines = [json.loads(ln) for ln in result.strip().split("\n")]
+        assert len(lines[0]["message"]["content"]) == 2
+
+    def test_no_assistant_entries_returns_unchanged(self) -> None:
+        """Transcripts with only user entries should pass through unchanged."""
+        user = self._user_entry("hello")
+        content = _make_jsonl(user)
+        assert strip_stale_thinking_blocks(content) == content
+
+    def test_empty_content_returns_unchanged(self) -> None:
+        assert strip_stale_thinking_blocks("") == ""
+
+    def test_multiple_turns_strips_all_but_last(self) -> None:
+        """With 3 assistant turns, only the last keeps thinking blocks."""
+        entries = [
+            self._asst_entry(
+                "msg_1",
+                [
+                    {"type": "thinking", "thinking": "t1"},
+                    {"type": "text", "text": "a1"},
+                ],
+                uuid="a1",
+            ),
+            self._user_entry("q2", uuid="u2", parent="a1"),
+            self._asst_entry(
+                "msg_2",
+                [
+                    {"type": "thinking", "thinking": "t2"},
+                    {"type": "text", "text": "a2"},
+                ],
+                uuid="a2",
+                parent="u2",
+            ),
+            self._user_entry("q3", uuid="u3", parent="a2"),
+            self._asst_entry(
+                "msg_3",
+                [
+                    {"type": "thinking", "thinking": "t3"},
+                    {"type": "text", "text": "a3"},
+                ],
+                uuid="a3",
+                parent="u3",
+            ),
+        ]
+        content = _make_jsonl(*entries)
+        result = strip_stale_thinking_blocks(content)
+        lines = [json.loads(ln) for ln in result.strip().split("\n")]
+
+        # msg_1: thinking stripped
+        assert len(lines[0]["message"]["content"]) == 1
+        assert lines[0]["message"]["content"][0]["type"] == "text"
+        # msg_2: thinking stripped
+        assert len(lines[2]["message"]["content"]) == 1
+        # msg_3 (last): thinking preserved
+        assert len(lines[4]["message"]["content"]) == 2
+        assert lines[4]["message"]["content"][0]["type"] == "thinking"
+
+    def test_same_msg_id_multi_entry_turn(self) -> None:
+        """Multiple entries sharing the same message.id (same turn) are preserved."""
+        entries = [
+            self._asst_entry(
+                "msg_old",
+                [{"type": "thinking", "thinking": "old"}],
+                uuid="a1",
+            ),
+            self._asst_entry(
+                "msg_last",
+                [{"type": "thinking", "thinking": "t_part1"}],
+                uuid="a2",
+                parent="a1",
+            ),
+            self._asst_entry(
+                "msg_last",
+                [{"type": "text", "text": "response"}],
+                uuid="a3",
+                parent="a2",
+            ),
+        ]
+        content = _make_jsonl(*entries)
+        result = strip_stale_thinking_blocks(content)
+        lines = [json.loads(ln) for ln in result.strip().split("\n")]
+
+        # Old entry stripped
+        assert lines[0]["message"]["content"] == []
+        # Both entries of last turn (msg_last) preserved
+        assert lines[1]["message"]["content"][0]["type"] == "thinking"
+        assert lines[2]["message"]["content"][0]["type"] == "text"

autogpt_platform/backend/backend/copilot/service_test.py

@@ -30,7 +30,7 @@ async def test_sdk_resume_multi_turn(setup_test_user, test_user_id):
     if not cfg.claude_agent_use_resume:
         return pytest.skip("CLAUDE_AGENT_USE_RESUME is not enabled, skipping test")
 
-    session = await create_chat_session(test_user_id)
+    session = await create_chat_session(test_user_id, dry_run=False)
     session = await upsert_chat_session(session)
 
     # --- Turn 1: send a message with a unique keyword ---