fix(chat/tools): remove hard output truncation from sandbox and web_fetch

The SDK already handles oversized tool output by writing to tool-results
files and reading back via MCP. Our 50K char truncation was cutting off
output before the agent could see it — the SDK's mechanism is the proper
way to handle large results.

.dockerignore

@@ -5,42 +5,13 @@
 !docs/
 
 # Platform - Libs
-!autogpt_platform/autogpt_libs/autogpt_libs/
+!autogpt_platform/autogpt_libs/
-!autogpt_platform/autogpt_libs/pyproject.toml
-!autogpt_platform/autogpt_libs/poetry.lock
-!autogpt_platform/autogpt_libs/README.md
 
 # Platform - Backend
-!autogpt_platform/backend/backend/
+!autogpt_platform/backend/
-!autogpt_platform/backend/test/e2e_test_data.py
-!autogpt_platform/backend/migrations/
-!autogpt_platform/backend/schema.prisma
-!autogpt_platform/backend/pyproject.toml
-!autogpt_platform/backend/poetry.lock
-!autogpt_platform/backend/README.md
-!autogpt_platform/backend/.env
-!autogpt_platform/backend/gen_prisma_types_stub.py
-
-# Platform - Market
-!autogpt_platform/market/market/
-!autogpt_platform/market/scripts.py
-!autogpt_platform/market/schema.prisma
-!autogpt_platform/market/pyproject.toml
-!autogpt_platform/market/poetry.lock
-!autogpt_platform/market/README.md
 
 # Platform - Frontend
-!autogpt_platform/frontend/src/
+!autogpt_platform/frontend/
-!autogpt_platform/frontend/public/
-!autogpt_platform/frontend/scripts/
-!autogpt_platform/frontend/package.json
-!autogpt_platform/frontend/pnpm-lock.yaml
-!autogpt_platform/frontend/tsconfig.json
-!autogpt_platform/frontend/README.md
-## config
-!autogpt_platform/frontend/*.config.*
-!autogpt_platform/frontend/.env.*
-!autogpt_platform/frontend/.env
 
 # Classic - AutoGPT
 !classic/original_autogpt/autogpt/
@@ -64,6 +35,38 @@
 # Classic - Frontend
 !classic/frontend/build/web/
 
-# Explicitly re-ignore some folders
+# Explicitly re-ignore unwanted files from whitelisted directories
-.*
+# Note: These patterns MUST come after the whitelist rules to take effect
-**/__pycache__
+
+# Hidden files and directories (but keep frontend .env files needed for build)
+**/.*
+!autogpt_platform/frontend/.env
+!autogpt_platform/frontend/.env.default
+!autogpt_platform/frontend/.env.production
+
+# Python artifacts
+**/__pycache__/
+**/*.pyc
+**/*.pyo
+**/.venv/
+**/.ruff_cache/
+**/.pytest_cache/
+**/.coverage
+**/htmlcov/
+
+# Node artifacts
+**/node_modules/
+**/.next/
+**/storybook-static/
+**/playwright-report/
+**/test-results/
+
+# Build artifacts
+**/dist/
+**/build/
+!autogpt_platform/frontend/src/**/build/
+**/target/
+
+# Logs and temp files
+**/*.log
+**/*.tmp

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

@@ -26,7 +26,6 @@ jobs:
   setup:
     runs-on: ubuntu-latest
     outputs:
-      cache-key: ${{ steps.cache-key.outputs.key }}
       components-changed: ${{ steps.filter.outputs.components }}
 
     steps:
@@ -41,28 +40,17 @@ jobs:
             components:
               - 'autogpt_platform/frontend/src/components/**'
 
-      - name: Set up Node.js
-        uses: actions/setup-node@v6
-        with:
-          node-version: "22.18.0"
-
       - name: Enable corepack
         run: corepack enable
 
-      - name: Generate cache key
+      - name: Set up Node
-        id: cache-key
+        uses: actions/setup-node@v6
-        run: echo "key=${{ runner.os }}-pnpm-${{ hashFiles('autogpt_platform/frontend/pnpm-lock.yaml', 'autogpt_platform/frontend/package.json') }}" >> $GITHUB_OUTPUT
-
-      - name: Cache dependencies
-        uses: actions/cache@v5
         with:
-          path: ~/.pnpm-store
+          node-version: "22.18.0"
-          key: ${{ steps.cache-key.outputs.key }}
+          cache: "pnpm"
-          restore-keys: |
+          cache-dependency-path: autogpt_platform/frontend/pnpm-lock.yaml
-            ${{ runner.os }}-pnpm-${{ hashFiles('autogpt_platform/frontend/pnpm-lock.yaml') }}
-            ${{ runner.os }}-pnpm-
 
-      - name: Install dependencies
+      - name: Install dependencies to populate cache
         run: pnpm install --frozen-lockfile
 
   lint:
@@ -73,22 +61,15 @@ jobs:
       - name: Checkout repository
         uses: actions/checkout@v6
 
-      - name: Set up Node.js
-        uses: actions/setup-node@v6
-        with:
-          node-version: "22.18.0"
-
       - name: Enable corepack
         run: corepack enable
 
-      - name: Restore dependencies cache
+      - name: Set up Node
-        uses: actions/cache@v5
+        uses: actions/setup-node@v6
         with:
-          path: ~/.pnpm-store
+          node-version: "22.18.0"
-          key: ${{ needs.setup.outputs.cache-key }}
+          cache: "pnpm"
-          restore-keys: |
+          cache-dependency-path: autogpt_platform/frontend/pnpm-lock.yaml
-            ${{ runner.os }}-pnpm-${{ hashFiles('autogpt_platform/frontend/pnpm-lock.yaml') }}
-            ${{ runner.os }}-pnpm-
 
       - name: Install dependencies
         run: pnpm install --frozen-lockfile
@@ -111,22 +92,15 @@ jobs:
         with:
           fetch-depth: 0
 
-      - name: Set up Node.js
-        uses: actions/setup-node@v6
-        with:
-          node-version: "22.18.0"
-
       - name: Enable corepack
         run: corepack enable
 
-      - name: Restore dependencies cache
+      - name: Set up Node
-        uses: actions/cache@v5
+        uses: actions/setup-node@v6
         with:
-          path: ~/.pnpm-store
+          node-version: "22.18.0"
-          key: ${{ needs.setup.outputs.cache-key }}
+          cache: "pnpm"
-          restore-keys: |
+          cache-dependency-path: autogpt_platform/frontend/pnpm-lock.yaml
-            ${{ runner.os }}-pnpm-${{ hashFiles('autogpt_platform/frontend/pnpm-lock.yaml') }}
-            ${{ runner.os }}-pnpm-
 
       - name: Install dependencies
         run: pnpm install --frozen-lockfile
@@ -141,10 +115,8 @@ jobs:
           exitOnceUploaded: true
 
   e2e_test:
+    name: end-to-end tests
     runs-on: big-boi
-    needs: setup
-    strategy:
-      fail-fast: false
 
     steps:
       - name: Checkout repository
@@ -152,19 +124,11 @@ jobs:
         with:
           submodules: recursive
 
-      - name: Set up Node.js
+      - name: Set up Platform - Copy default supabase .env
-        uses: actions/setup-node@v6
-        with:
-          node-version: "22.18.0"
-
-      - name: Enable corepack
-        run: corepack enable
-
-      - name: Copy default supabase .env
         run: |
           cp ../.env.default ../.env
 
-      - name: Copy backend .env and set OpenAI API key
+      - name: Set up Platform - Copy backend .env and set OpenAI API key
         run: |
           cp ../backend/.env.default ../backend/.env
           echo "OPENAI_INTERNAL_API_KEY=${{ secrets.OPENAI_API_KEY }}" >> ../backend/.env
@@ -172,77 +136,125 @@ jobs:
           # Used by E2E test data script to generate embeddings for approved store agents
           OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
 
-      - name: Set up Docker Buildx
+      - name: Set up Platform - Set up Docker Buildx
         uses: docker/setup-buildx-action@v3
+        with:
+          driver: docker-container
+          driver-opts: network=host
 
-      - name: Cache Docker layers
+      - name: Set up Platform - Expose GHA cache to docker buildx CLI
+        uses: crazy-max/ghaction-github-runtime@v3
+
+      - name: Set up Platform - Build Docker images (with cache)
+        working-directory: autogpt_platform
+        run: |
+          pip install pyyaml
+
+          # Resolve extends and generate a flat compose file that bake can understand
+          docker compose -f docker-compose.yml config > docker-compose.resolved.yml
+
+          # Add cache configuration to the resolved compose file
+          python ../.github/workflows/scripts/docker-ci-fix-compose-build-cache.py \
+            --source docker-compose.resolved.yml \
+            --cache-from "type=gha" \
+            --cache-to "type=gha,mode=max" \
+            --backend-hash "${{ hashFiles('autogpt_platform/backend/Dockerfile', 'autogpt_platform/backend/poetry.lock', 'autogpt_platform/backend/backend') }}" \
+            --frontend-hash "${{ hashFiles('autogpt_platform/frontend/Dockerfile', 'autogpt_platform/frontend/pnpm-lock.yaml', 'autogpt_platform/frontend/src') }}" \
+            --git-ref "${{ github.ref }}"
+
+          # Build with bake using the resolved compose file (now includes cache config)
+          docker buildx bake --allow=fs.read=.. -f docker-compose.resolved.yml --load
+        env:
+          NEXT_PUBLIC_PW_TEST: true
+
+      - name: Set up tests - Cache E2E test data
+        id: e2e-data-cache
         uses: actions/cache@v5
         with:
-          path: /tmp/.buildx-cache
+          path: /tmp/e2e_test_data.sql
-          key: ${{ runner.os }}-buildx-frontend-test-${{ hashFiles('autogpt_platform/docker-compose.yml', 'autogpt_platform/backend/Dockerfile', 'autogpt_platform/backend/pyproject.toml', 'autogpt_platform/backend/poetry.lock') }}
+          key: e2e-test-data-${{ hashFiles('autogpt_platform/backend/test/e2e_test_data.py', 'autogpt_platform/backend/migrations/**', '.github/workflows/platform-frontend-ci.yml') }}
-          restore-keys: |
-            ${{ runner.os }}-buildx-frontend-test-
 
-      - name: Run docker compose
+      - name: Set up Platform - Start Supabase DB + Auth
         run: |
-          NEXT_PUBLIC_PW_TEST=true docker compose -f ../docker-compose.yml up -d
+          docker compose -f ../docker-compose.resolved.yml up -d db auth --no-build
+          echo "Waiting for database to be ready..."
+          timeout 60 sh -c 'until docker compose -f ../docker-compose.resolved.yml exec -T db pg_isready -U postgres 2>/dev/null; do sleep 2; done'
+          echo "Waiting for auth service to be ready..."
+          timeout 60 sh -c 'until docker compose -f ../docker-compose.resolved.yml exec -T db psql -U postgres -d postgres -c "SELECT 1 FROM auth.users LIMIT 1" 2>/dev/null; do sleep 2; done' || echo "Auth schema check timeout, continuing..."
+
+      - name: Set up Platform - Run migrations
+        run: |
+          echo "Running migrations..."
+          docker compose -f ../docker-compose.resolved.yml run --rm migrate
+          echo "✅ Migrations completed"
         env:
-          DOCKER_BUILDKIT: 1
+          NEXT_PUBLIC_PW_TEST: true
-          BUILDX_CACHE_FROM: type=local,src=/tmp/.buildx-cache
-          BUILDX_CACHE_TO: type=local,dest=/tmp/.buildx-cache-new,mode=max
 
-      - name: Move cache
+      - name: Set up tests - Load cached E2E test data
+        if: steps.e2e-data-cache.outputs.cache-hit == 'true'
         run: |
-          rm -rf /tmp/.buildx-cache
+          echo "✅ Found cached E2E test data, restoring..."
-          if [ -d "/tmp/.buildx-cache-new" ]; then
+          {
-            mv /tmp/.buildx-cache-new /tmp/.buildx-cache
+            echo "SET session_replication_role = 'replica';"
-          fi
+            cat /tmp/e2e_test_data.sql
+            echo "SET session_replication_role = 'origin';"
+          } | docker compose -f ../docker-compose.resolved.yml exec -T db psql -U postgres -d postgres -b
+          # Refresh materialized views after restore
+          docker compose -f ../docker-compose.resolved.yml exec -T db \
+            psql -U postgres -d postgres -b -c "SET search_path TO platform; SELECT refresh_store_materialized_views();" || true
 
-      - name: Wait for services to be ready
+          echo "✅ E2E test data restored from cache"
+
+      - name: Set up Platform - Start (all other services)
         run: |
+          docker compose -f ../docker-compose.resolved.yml up -d --no-build
           echo "Waiting for rest_server to be ready..."
           timeout 60 sh -c 'until curl -f http://localhost:8006/health 2>/dev/null; do sleep 2; done' || echo "Rest server health check timeout, continuing..."
-          echo "Waiting for database to be ready..."
+        env:
-          timeout 60 sh -c 'until docker compose -f ../docker-compose.yml exec -T db pg_isready -U postgres 2>/dev/null; do sleep 2; done' || echo "Database ready check timeout, continuing..."
+          NEXT_PUBLIC_PW_TEST: true
 
-      - name: Create E2E test data
+      - name: Set up tests - Create E2E test data
+        if: steps.e2e-data-cache.outputs.cache-hit != 'true'
         run: |
           echo "Creating E2E test data..."
-          # First try to run the script from inside the container
+          docker cp ../backend/test/e2e_test_data.py $(docker compose -f ../docker-compose.resolved.yml ps -q rest_server):/tmp/e2e_test_data.py
-          if docker compose -f ../docker-compose.yml exec -T rest_server test -f /app/autogpt_platform/backend/test/e2e_test_data.py; then
+          docker compose -f ../docker-compose.resolved.yml exec -T rest_server sh -c "cd /app/autogpt_platform && python /tmp/e2e_test_data.py" || {
-            echo "✅ Found e2e_test_data.py in container, running it..."
+            echo "❌ E2E test data creation failed!"
-            docker compose -f ../docker-compose.yml exec -T rest_server sh -c "cd /app/autogpt_platform && python backend/test/e2e_test_data.py" || {
+            docker compose -f ../docker-compose.resolved.yml logs --tail=50 rest_server
-              echo "❌ E2E test data creation failed!"
+            exit 1
-              docker compose -f ../docker-compose.yml logs --tail=50 rest_server
+          }
-              exit 1
-            }
-          else
-            echo "⚠️ e2e_test_data.py not found in container, copying and running..."
-            # Copy the script into the container and run it
-            docker cp ../backend/test/e2e_test_data.py $(docker compose -f ../docker-compose.yml ps -q rest_server):/tmp/e2e_test_data.py || {
-              echo "❌ Failed to copy script to container"
-              exit 1
-            }
-            docker compose -f ../docker-compose.yml exec -T rest_server sh -c "cd /app/autogpt_platform && python /tmp/e2e_test_data.py" || {
-              echo "❌ E2E test data creation failed!"
-              docker compose -f ../docker-compose.yml logs --tail=50 rest_server
-              exit 1
-            }
-          fi
 
-      - name: Restore dependencies cache
+          # Dump auth.users + platform schema for cache (two separate dumps)
-        uses: actions/cache@v5
+          echo "Dumping database for cache..."
+          {
+            docker compose -f ../docker-compose.resolved.yml exec -T db \
+              pg_dump -U postgres --data-only --column-inserts \
+              --table='auth.users' postgres
+            docker compose -f ../docker-compose.resolved.yml exec -T db \
+              pg_dump -U postgres --data-only --column-inserts \
+              --schema=platform \
+              --exclude-table='platform._prisma_migrations' \
+              --exclude-table='platform.apscheduler_jobs' \
+              --exclude-table='platform.apscheduler_jobs_batched_notifications' \
+              postgres
+          } > /tmp/e2e_test_data.sql
+
+          echo "✅ Database dump created for caching ($(wc -l < /tmp/e2e_test_data.sql) lines)"
+
+      - name: Set up tests - Enable corepack
+        run: corepack enable
+
+      - name: Set up tests - Set up Node
+        uses: actions/setup-node@v6
         with:
-          path: ~/.pnpm-store
+          node-version: "22.18.0"
-          key: ${{ needs.setup.outputs.cache-key }}
+          cache: "pnpm"
-          restore-keys: |
+          cache-dependency-path: autogpt_platform/frontend/pnpm-lock.yaml
-            ${{ runner.os }}-pnpm-${{ hashFiles('autogpt_platform/frontend/pnpm-lock.yaml') }}
-            ${{ runner.os }}-pnpm-
 
-      - name: Install dependencies
+      - name: Set up tests - Install dependencies
         run: pnpm install --frozen-lockfile
 
-      - name: Install Browser 'chromium'
+      - name: Set up tests - Install browser 'chromium'
         run: pnpm playwright install --with-deps chromium
 
       - name: Run Playwright tests
@@ -269,7 +281,7 @@ jobs:
 
       - name: Print Final Docker Compose logs
         if: always()
-        run: docker compose -f ../docker-compose.yml logs
+        run: docker compose -f ../docker-compose.resolved.yml logs
 
   integration_test:
     runs-on: ubuntu-latest
@@ -281,22 +293,15 @@ jobs:
         with:
           submodules: recursive
 
-      - name: Set up Node.js
-        uses: actions/setup-node@v6
-        with:
-          node-version: "22.18.0"
-
       - name: Enable corepack
         run: corepack enable
 
-      - name: Restore dependencies cache
+      - name: Set up Node
-        uses: actions/cache@v5
+        uses: actions/setup-node@v6
         with:
-          path: ~/.pnpm-store
+          node-version: "22.18.0"
-          key: ${{ needs.setup.outputs.cache-key }}
+          cache: "pnpm"
-          restore-keys: |
+          cache-dependency-path: autogpt_platform/frontend/pnpm-lock.yaml
-            ${{ runner.os }}-pnpm-${{ hashFiles('autogpt_platform/frontend/pnpm-lock.yaml') }}
-            ${{ runner.os }}-pnpm-
 
       - name: Install dependencies
         run: pnpm install --frozen-lockfile

.github/workflows/scripts/docker-ci-fix-compose-build-cache.py

@@ -0,0 +1,195 @@
+#!/usr/bin/env python3
+"""
+Add cache configuration to a resolved docker-compose file for all services
+that have a build key, and ensure image names match what docker compose expects.
+"""
+
+import argparse
+
+import yaml
+
+
+DEFAULT_BRANCH = "dev"
+CACHE_BUILDS_FOR_COMPONENTS = ["backend", "frontend"]
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Add cache config to a resolved compose file"
+    )
+    parser.add_argument(
+        "--source",
+        required=True,
+        help="Source compose file to read (should be output of `docker compose config`)",
+    )
+    parser.add_argument(
+        "--cache-from",
+        default="type=gha",
+        help="Cache source configuration",
+    )
+    parser.add_argument(
+        "--cache-to",
+        default="type=gha,mode=max",
+        help="Cache destination configuration",
+    )
+    for component in CACHE_BUILDS_FOR_COMPONENTS:
+        parser.add_argument(
+            f"--{component}-hash",
+            default="",
+            help=f"Hash for {component} cache scope (e.g., from hashFiles())",
+        )
+    parser.add_argument(
+        "--git-ref",
+        default="",
+        help="Git ref for branch-based cache scope (e.g., refs/heads/master)",
+    )
+    args = parser.parse_args()
+
+    # Normalize git ref to a safe scope name (e.g., refs/heads/master -> master)
+    git_ref_scope = ""
+    if args.git_ref:
+        git_ref_scope = args.git_ref.replace("refs/heads/", "").replace("/", "-")
+
+    with open(args.source, "r") as f:
+        compose = yaml.safe_load(f)
+
+    # Get project name from compose file or default
+    project_name = compose.get("name", "autogpt_platform")
+
+    def get_image_name(dockerfile: str, target: str) -> str:
+        """Generate image name based on Dockerfile folder and build target."""
+        dockerfile_parts = dockerfile.replace("\\", "/").split("/")
+        if len(dockerfile_parts) >= 2:
+            folder_name = dockerfile_parts[-2]  # e.g., "backend" or "frontend"
+        else:
+            folder_name = "app"
+        return f"{project_name}-{folder_name}:{target}"
+
+    def get_build_key(dockerfile: str, target: str) -> str:
+        """Generate a unique key for a Dockerfile+target combination."""
+        return f"{dockerfile}:{target}"
+
+    def get_component(dockerfile: str) -> str | None:
+        """Get component name (frontend/backend) from dockerfile path."""
+        for component in CACHE_BUILDS_FOR_COMPONENTS:
+            if component in dockerfile:
+                return component
+        return None
+
+    # First pass: collect all services with build configs and identify duplicates
+    # Track which (dockerfile, target) combinations we've seen
+    build_key_to_first_service: dict[str, str] = {}
+    services_to_build: list[str] = []
+    services_to_dedupe: list[str] = []
+
+    for service_name, service_config in compose.get("services", {}).items():
+        if "build" not in service_config:
+            continue
+
+        build_config = service_config["build"]
+        dockerfile = build_config.get("dockerfile", "Dockerfile")
+        target = build_config.get("target", "default")
+        build_key = get_build_key(dockerfile, target)
+
+        if build_key not in build_key_to_first_service:
+            # First service with this build config - it will do the actual build
+            build_key_to_first_service[build_key] = service_name
+            services_to_build.append(service_name)
+        else:
+            # Duplicate - will just use the image from the first service
+            services_to_dedupe.append(service_name)
+
+    # Second pass: configure builds and deduplicate
+    modified_services = []
+    for service_name, service_config in compose.get("services", {}).items():
+        if "build" not in service_config:
+            continue
+
+        build_config = service_config["build"]
+        dockerfile = build_config.get("dockerfile", "Dockerfile")
+        target = build_config.get("target", "latest")
+        image_name = get_image_name(dockerfile, target)
+
+        # Set image name for all services (needed for both builders and deduped)
+        service_config["image"] = image_name
+
+        if service_name in services_to_dedupe:
+            # Remove build config - this service will use the pre-built image
+            del service_config["build"]
+            continue
+
+        # This service will do the actual build - add cache config
+        cache_from_list = []
+        cache_to_list = []
+
+        component = get_component(dockerfile)
+        if not component:
+            # Skip services that don't clearly match frontend/backend
+            continue
+
+        # Get the hash for this component
+        component_hash = getattr(args, f"{component}_hash")
+
+        # Scope format: platform-{component}-{target}-{hash|ref}
+        # Example: platform-backend-server-abc123
+
+        if "type=gha" in args.cache_from:
+            # 1. Primary: exact hash match (most specific)
+            if component_hash:
+                hash_scope = f"platform-{component}-{target}-{component_hash}"
+                cache_from_list.append(f"{args.cache_from},scope={hash_scope}")
+
+            # 2. Fallback: branch-based cache
+            if git_ref_scope:
+                ref_scope = f"platform-{component}-{target}-{git_ref_scope}"
+                cache_from_list.append(f"{args.cache_from},scope={ref_scope}")
+
+            # 3. Fallback: dev branch cache (for PRs/feature branches)
+            if git_ref_scope and git_ref_scope != DEFAULT_BRANCH:
+                master_scope = f"platform-{component}-{target}-{DEFAULT_BRANCH}"
+                cache_from_list.append(f"{args.cache_from},scope={master_scope}")
+
+        if "type=gha" in args.cache_to:
+            # Write to both hash-based and branch-based scopes
+            if component_hash:
+                hash_scope = f"platform-{component}-{target}-{component_hash}"
+                cache_to_list.append(f"{args.cache_to},scope={hash_scope}")
+
+            if git_ref_scope:
+                ref_scope = f"platform-{component}-{target}-{git_ref_scope}"
+                cache_to_list.append(f"{args.cache_to},scope={ref_scope}")
+
+        # Ensure we have at least one cache source/target
+        if not cache_from_list:
+            cache_from_list.append(args.cache_from)
+        if not cache_to_list:
+            cache_to_list.append(args.cache_to)
+
+        build_config["cache_from"] = cache_from_list
+        build_config["cache_to"] = cache_to_list
+        modified_services.append(service_name)
+
+    # Write back to the same file
+    with open(args.source, "w") as f:
+        yaml.dump(compose, f, default_flow_style=False, sort_keys=False)
+
+    print(f"Added cache config to {len(modified_services)} services in {args.source}:")
+    for svc in modified_services:
+        svc_config = compose["services"][svc]
+        build_cfg = svc_config.get("build", {})
+        cache_from_list = build_cfg.get("cache_from", ["none"])
+        cache_to_list = build_cfg.get("cache_to", ["none"])
+        print(f"  - {svc}")
+        print(f"      image: {svc_config.get('image', 'N/A')}")
+        print(f"      cache_from: {cache_from_list}")
+        print(f"      cache_to: {cache_to_list}")
+    if services_to_dedupe:
+        print(
+            f"Deduplicated {len(services_to_dedupe)} services (will use pre-built images):"
+        )
+        for svc in services_to_dedupe:
+            print(f"  - {svc} -> {compose['services'][svc].get('image', 'N/A')}")
+
+
+if __name__ == "__main__":
+    main()

autogpt_platform/CLAUDE.md

@@ -45,6 +45,11 @@ AutoGPT Platform is a monorepo containing:
 - 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.

autogpt_platform/autogpt_libs/poetry.lock

@@ -448,61 +448,61 @@ toml = ["tomli ; python_full_version <= \"3.11.0a6\""]
 
 [[package]]
 name = "cryptography"
-version = "46.0.4"
+version = "46.0.5"
 description = "cryptography is a package which provides cryptographic recipes and primitives to Python developers."
 optional = false
 python-versions = "!=3.9.0,!=3.9.1,>=3.8"
 groups = ["main"]
 files = [
-    {file = "cryptography-46.0.4-cp311-abi3-macosx_10_9_universal2.whl", hash = "sha256:281526e865ed4166009e235afadf3a4c4cba6056f99336a99efba65336fd5485"},
+    {file = "cryptography-46.0.5-cp311-abi3-macosx_10_9_universal2.whl", hash = "sha256:351695ada9ea9618b3500b490ad54c739860883df6c1f555e088eaf25b1bbaad"},
-    {file = "cryptography-46.0.4-cp311-abi3-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:5f14fba5bf6f4390d7ff8f086c566454bff0411f6d8aa7af79c88b6f9267aecc"},
+    {file = "cryptography-46.0.5-cp311-abi3-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:c18ff11e86df2e28854939acde2d003f7984f721eba450b56a200ad90eeb0e6b"},
-    {file = "cryptography-46.0.4-cp311-abi3-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:47bcd19517e6389132f76e2d5303ded6cf3f78903da2158a671be8de024f4cd0"},
+    {file = "cryptography-46.0.5-cp311-abi3-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:4d7e3d356b8cd4ea5aff04f129d5f66ebdc7b6f8eae802b93739ed520c47c79b"},
-    {file = "cryptography-46.0.4-cp311-abi3-manylinux_2_28_aarch64.whl", hash = "sha256:01df4f50f314fbe7009f54046e908d1754f19d0c6d3070df1e6268c5a4af09fa"},
+    {file = "cryptography-46.0.5-cp311-abi3-manylinux_2_28_aarch64.whl", hash = "sha256:50bfb6925eff619c9c023b967d5b77a54e04256c4281b0e21336a130cd7fc263"},
-    {file = "cryptography-46.0.4-cp311-abi3-manylinux_2_28_ppc64le.whl", hash = "sha256:5aa3e463596b0087b3da0dbe2b2487e9fc261d25da85754e30e3b40637d61f81"},
+    {file = "cryptography-46.0.5-cp311-abi3-manylinux_2_28_ppc64le.whl", hash = "sha256:803812e111e75d1aa73690d2facc295eaefd4439be1023fefc4995eaea2af90d"},
-    {file = "cryptography-46.0.4-cp311-abi3-manylinux_2_28_x86_64.whl", hash = "sha256:0a9ad24359fee86f131836a9ac3bffc9329e956624a2d379b613f8f8abaf5255"},
+    {file = "cryptography-46.0.5-cp311-abi3-manylinux_2_28_x86_64.whl", hash = "sha256:3ee190460e2fbe447175cda91b88b84ae8322a104fc27766ad09428754a618ed"},
-    {file = "cryptography-46.0.4-cp311-abi3-manylinux_2_31_armv7l.whl", hash = "sha256:dc1272e25ef673efe72f2096e92ae39dea1a1a450dd44918b15351f72c5a168e"},
+    {file = "cryptography-46.0.5-cp311-abi3-manylinux_2_31_armv7l.whl", hash = "sha256:f145bba11b878005c496e93e257c1e88f154d278d2638e6450d17e0f31e558d2"},
-    {file = "cryptography-46.0.4-cp311-abi3-manylinux_2_34_aarch64.whl", hash = "sha256:de0f5f4ec8711ebc555f54735d4c673fc34b65c44283895f1a08c2b49d2fd99c"},
+    {file = "cryptography-46.0.5-cp311-abi3-manylinux_2_34_aarch64.whl", hash = "sha256:e9251e3be159d1020c4030bd2e5f84d6a43fe54b6c19c12f51cde9542a2817b2"},
-    {file = "cryptography-46.0.4-cp311-abi3-manylinux_2_34_ppc64le.whl", hash = "sha256:eeeb2e33d8dbcccc34d64651f00a98cb41b2dc69cef866771a5717e6734dfa32"},
+    {file = "cryptography-46.0.5-cp311-abi3-manylinux_2_34_ppc64le.whl", hash = "sha256:47fb8a66058b80e509c47118ef8a75d14c455e81ac369050f20ba0d23e77fee0"},
-    {file = "cryptography-46.0.4-cp311-abi3-manylinux_2_34_x86_64.whl", hash = "sha256:3d425eacbc9aceafd2cb429e42f4e5d5633c6f873f5e567077043ef1b9bbf616"},
+    {file = "cryptography-46.0.5-cp311-abi3-manylinux_2_34_x86_64.whl", hash = "sha256:4c3341037c136030cb46e4b1e17b7418ea4cbd9dd207e4a6f3b2b24e0d4ac731"},
-    {file = "cryptography-46.0.4-cp311-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:91627ebf691d1ea3976a031b61fb7bac1ccd745afa03602275dda443e11c8de0"},
+    {file = "cryptography-46.0.5-cp311-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:890bcb4abd5a2d3f852196437129eb3667d62630333aacc13dfd470fad3aaa82"},
-    {file = "cryptography-46.0.4-cp311-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:2d08bc22efd73e8854b0b7caff402d735b354862f1145d7be3b9c0f740fef6a0"},
+    {file = "cryptography-46.0.5-cp311-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:80a8d7bfdf38f87ca30a5391c0c9ce4ed2926918e017c29ddf643d0ed2778ea1"},
-    {file = "cryptography-46.0.4-cp311-abi3-win32.whl", hash = "sha256:82a62483daf20b8134f6e92898da70d04d0ef9a75829d732ea1018678185f4f5"},
+    {file = "cryptography-46.0.5-cp311-abi3-win32.whl", hash = "sha256:60ee7e19e95104d4c03871d7d7dfb3d22ef8a9b9c6778c94e1c8fcc8365afd48"},
-    {file = "cryptography-46.0.4-cp311-abi3-win_amd64.whl", hash = "sha256:6225d3ebe26a55dbc8ead5ad1265c0403552a63336499564675b29eb3184c09b"},
+    {file = "cryptography-46.0.5-cp311-abi3-win_amd64.whl", hash = "sha256:38946c54b16c885c72c4f59846be9743d699eee2b69b6988e0a00a01f46a61a4"},
-    {file = "cryptography-46.0.4-cp314-cp314t-macosx_10_9_universal2.whl", hash = "sha256:485e2b65d25ec0d901bca7bcae0f53b00133bf3173916d8e421f6fddde103908"},
+    {file = "cryptography-46.0.5-cp314-cp314t-macosx_10_9_universal2.whl", hash = "sha256:94a76daa32eb78d61339aff7952ea819b1734b46f73646a07decb40e5b3448e2"},
-    {file = "cryptography-46.0.4-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:078e5f06bd2fa5aea5a324f2a09f914b1484f1d0c2a4d6a8a28c74e72f65f2da"},
+    {file = "cryptography-46.0.5-cp314-cp314t-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:5be7bf2fb40769e05739dd0046e7b26f9d4670badc7b032d6ce4db64dddc0678"},
-    {file = "cryptography-46.0.4-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:dce1e4f068f03008da7fa51cc7abc6ddc5e5de3e3d1550334eaf8393982a5829"},
+    {file = "cryptography-46.0.5-cp314-cp314t-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:fe346b143ff9685e40192a4960938545c699054ba11d4f9029f94751e3f71d87"},
-    {file = "cryptography-46.0.4-cp314-cp314t-manylinux_2_28_aarch64.whl", hash = "sha256:2067461c80271f422ee7bdbe79b9b4be54a5162e90345f86a23445a0cf3fd8a2"},
+    {file = "cryptography-46.0.5-cp314-cp314t-manylinux_2_28_aarch64.whl", hash = "sha256:c69fd885df7d089548a42d5ec05be26050ebcd2283d89b3d30676eb32ff87dee"},
-    {file = "cryptography-46.0.4-cp314-cp314t-manylinux_2_28_ppc64le.whl", hash = "sha256:c92010b58a51196a5f41c3795190203ac52edfd5dc3ff99149b4659eba9d2085"},
+    {file = "cryptography-46.0.5-cp314-cp314t-manylinux_2_28_ppc64le.whl", hash = "sha256:8293f3dea7fc929ef7240796ba231413afa7b68ce38fd21da2995549f5961981"},
-    {file = "cryptography-46.0.4-cp314-cp314t-manylinux_2_28_x86_64.whl", hash = "sha256:829c2b12bbc5428ab02d6b7f7e9bbfd53e33efd6672d21341f2177470171ad8b"},
+    {file = "cryptography-46.0.5-cp314-cp314t-manylinux_2_28_x86_64.whl", hash = "sha256:1abfdb89b41c3be0365328a410baa9df3ff8a9110fb75e7b52e66803ddabc9a9"},
-    {file = "cryptography-46.0.4-cp314-cp314t-manylinux_2_31_armv7l.whl", hash = "sha256:62217ba44bf81b30abaeda1488686a04a702a261e26f87db51ff61d9d3510abd"},
+    {file = "cryptography-46.0.5-cp314-cp314t-manylinux_2_31_armv7l.whl", hash = "sha256:d66e421495fdb797610a08f43b05269e0a5ea7f5e652a89bfd5a7d3c1dee3648"},
-    {file = "cryptography-46.0.4-cp314-cp314t-manylinux_2_34_aarch64.whl", hash = "sha256:9c2da296c8d3415b93e6053f5a728649a87a48ce084a9aaf51d6e46c87c7f2d2"},
+    {file = "cryptography-46.0.5-cp314-cp314t-manylinux_2_34_aarch64.whl", hash = "sha256:4e817a8920bfbcff8940ecfd60f23d01836408242b30f1a708d93198393a80b4"},
-    {file = "cryptography-46.0.4-cp314-cp314t-manylinux_2_34_ppc64le.whl", hash = "sha256:9b34d8ba84454641a6bf4d6762d15847ecbd85c1316c0a7984e6e4e9f748ec2e"},
+    {file = "cryptography-46.0.5-cp314-cp314t-manylinux_2_34_ppc64le.whl", hash = "sha256:68f68d13f2e1cb95163fa3b4db4bf9a159a418f5f6e7242564fc75fcae667fd0"},
-    {file = "cryptography-46.0.4-cp314-cp314t-manylinux_2_34_x86_64.whl", hash = "sha256:df4a817fa7138dd0c96c8c8c20f04b8aaa1fac3bbf610913dcad8ea82e1bfd3f"},
+    {file = "cryptography-46.0.5-cp314-cp314t-manylinux_2_34_x86_64.whl", hash = "sha256:a3d1fae9863299076f05cb8a778c467578262fae09f9dc0ee9b12eb4268ce663"},
-    {file = "cryptography-46.0.4-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:b1de0ebf7587f28f9190b9cb526e901bf448c9e6a99655d2b07fff60e8212a82"},
+    {file = "cryptography-46.0.5-cp314-cp314t-musllinux_1_2_aarch64.whl", hash = "sha256:c4143987a42a2397f2fc3b4d7e3a7d313fbe684f67ff443999e803dd75a76826"},
-    {file = "cryptography-46.0.4-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:9b4d17bc7bd7cdd98e3af40b441feaea4c68225e2eb2341026c84511ad246c0c"},
+    {file = "cryptography-46.0.5-cp314-cp314t-musllinux_1_2_x86_64.whl", hash = "sha256:7d731d4b107030987fd61a7f8ab512b25b53cef8f233a97379ede116f30eb67d"},
-    {file = "cryptography-46.0.4-cp314-cp314t-win32.whl", hash = "sha256:c411f16275b0dea722d76544a61d6421e2cc829ad76eec79280dbdc9ddf50061"},
+    {file = "cryptography-46.0.5-cp314-cp314t-win32.whl", hash = "sha256:c3bcce8521d785d510b2aad26ae2c966092b7daa8f45dd8f44734a104dc0bc1a"},
-    {file = "cryptography-46.0.4-cp314-cp314t-win_amd64.whl", hash = "sha256:728fedc529efc1439eb6107b677f7f7558adab4553ef8669f0d02d42d7b959a7"},
+    {file = "cryptography-46.0.5-cp314-cp314t-win_amd64.whl", hash = "sha256:4d8ae8659ab18c65ced284993c2265910f6c9e650189d4e3f68445ef82a810e4"},
-    {file = "cryptography-46.0.4-cp38-abi3-macosx_10_9_universal2.whl", hash = "sha256:a9556ba711f7c23f77b151d5798f3ac44a13455cc68db7697a1096e6d0563cab"},
+    {file = "cryptography-46.0.5-cp38-abi3-macosx_10_9_universal2.whl", hash = "sha256:4108d4c09fbbf2789d0c926eb4152ae1760d5a2d97612b92d508d96c861e4d31"},
-    {file = "cryptography-46.0.4-cp38-abi3-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:8bf75b0259e87fa70bddc0b8b4078b76e7fd512fd9afae6c1193bcf440a4dbef"},
+    {file = "cryptography-46.0.5-cp38-abi3-manylinux2014_aarch64.manylinux_2_17_aarch64.whl", hash = "sha256:7d1f30a86d2757199cb2d56e48cce14deddf1f9c95f1ef1b64ee91ea43fe2e18"},
-    {file = "cryptography-46.0.4-cp38-abi3-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:3c268a3490df22270955966ba236d6bc4a8f9b6e4ffddb78aac535f1a5ea471d"},
+    {file = "cryptography-46.0.5-cp38-abi3-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:039917b0dc418bb9f6edce8a906572d69e74bd330b0b3fea4f79dab7f8ddd235"},
-    {file = "cryptography-46.0.4-cp38-abi3-manylinux_2_28_aarch64.whl", hash = "sha256:812815182f6a0c1d49a37893a303b44eaac827d7f0d582cecfc81b6427f22973"},
+    {file = "cryptography-46.0.5-cp38-abi3-manylinux_2_28_aarch64.whl", hash = "sha256:ba2a27ff02f48193fc4daeadf8ad2590516fa3d0adeeb34336b96f7fa64c1e3a"},
-    {file = "cryptography-46.0.4-cp38-abi3-manylinux_2_28_ppc64le.whl", hash = "sha256:a90e43e3ef65e6dcf969dfe3bb40cbf5aef0d523dff95bfa24256be172a845f4"},
+    {file = "cryptography-46.0.5-cp38-abi3-manylinux_2_28_ppc64le.whl", hash = "sha256:61aa400dce22cb001a98014f647dc21cda08f7915ceb95df0c9eaf84b4b6af76"},
-    {file = "cryptography-46.0.4-cp38-abi3-manylinux_2_28_x86_64.whl", hash = "sha256:a05177ff6296644ef2876fce50518dffb5bcdf903c85250974fc8bc85d54c0af"},
+    {file = "cryptography-46.0.5-cp38-abi3-manylinux_2_28_x86_64.whl", hash = "sha256:3ce58ba46e1bc2aac4f7d9290223cead56743fa6ab94a5d53292ffaac6a91614"},
-    {file = "cryptography-46.0.4-cp38-abi3-manylinux_2_31_armv7l.whl", hash = "sha256:daa392191f626d50f1b136c9b4cf08af69ca8279d110ea24f5c2700054d2e263"},
+    {file = "cryptography-46.0.5-cp38-abi3-manylinux_2_31_armv7l.whl", hash = "sha256:420d0e909050490d04359e7fdb5ed7e667ca5c3c402b809ae2563d7e66a92229"},
-    {file = "cryptography-46.0.4-cp38-abi3-manylinux_2_34_aarch64.whl", hash = "sha256:e07ea39c5b048e085f15923511d8121e4a9dc45cee4e3b970ca4f0d338f23095"},
+    {file = "cryptography-46.0.5-cp38-abi3-manylinux_2_34_aarch64.whl", hash = "sha256:582f5fcd2afa31622f317f80426a027f30dc792e9c80ffee87b993200ea115f1"},
-    {file = "cryptography-46.0.4-cp38-abi3-manylinux_2_34_ppc64le.whl", hash = "sha256:d5a45ddc256f492ce42a4e35879c5e5528c09cd9ad12420828c972951d8e016b"},
+    {file = "cryptography-46.0.5-cp38-abi3-manylinux_2_34_ppc64le.whl", hash = "sha256:bfd56bb4b37ed4f330b82402f6f435845a5f5648edf1ad497da51a8452d5d62d"},
-    {file = "cryptography-46.0.4-cp38-abi3-manylinux_2_34_x86_64.whl", hash = "sha256:6bb5157bf6a350e5b28aee23beb2d84ae6f5be390b2f8ee7ea179cda077e1019"},
+    {file = "cryptography-46.0.5-cp38-abi3-manylinux_2_34_x86_64.whl", hash = "sha256:a3d507bb6a513ca96ba84443226af944b0f7f47dcc9a399d110cd6146481d24c"},
-    {file = "cryptography-46.0.4-cp38-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:dd5aba870a2c40f87a3af043e0dee7d9eb02d4aff88a797b48f2b43eff8c3ab4"},
+    {file = "cryptography-46.0.5-cp38-abi3-musllinux_1_2_aarch64.whl", hash = "sha256:9f16fbdf4da055efb21c22d81b89f155f02ba420558db21288b3d0035bafd5f4"},
-    {file = "cryptography-46.0.4-cp38-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:93d8291da8d71024379ab2cb0b5c57915300155ad42e07f76bea6ad838d7e59b"},
+    {file = "cryptography-46.0.5-cp38-abi3-musllinux_1_2_x86_64.whl", hash = "sha256:ced80795227d70549a411a4ab66e8ce307899fad2220ce5ab2f296e687eacde9"},
-    {file = "cryptography-46.0.4-cp38-abi3-win32.whl", hash = "sha256:0563655cb3c6d05fb2afe693340bc050c30f9f34e15763361cf08e94749401fc"},
+    {file = "cryptography-46.0.5-cp38-abi3-win32.whl", hash = "sha256:02f547fce831f5096c9a567fd41bc12ca8f11df260959ecc7c3202555cc47a72"},
-    {file = "cryptography-46.0.4-cp38-abi3-win_amd64.whl", hash = "sha256:fa0900b9ef9c49728887d1576fd8d9e7e3ea872fa9b25ef9b64888adc434e976"},
+    {file = "cryptography-46.0.5-cp38-abi3-win_amd64.whl", hash = "sha256:556e106ee01aa13484ce9b0239bca667be5004efb0aabbed28d353df86445595"},
-    {file = "cryptography-46.0.4-pp311-pypy311_pp73-macosx_11_0_arm64.whl", hash = "sha256:766330cce7416c92b5e90c3bb71b1b79521760cdcfc3a6a1a182d4c9fab23d2b"},
+    {file = "cryptography-46.0.5-pp311-pypy311_pp73-macosx_11_0_arm64.whl", hash = "sha256:3b4995dc971c9fb83c25aa44cf45f02ba86f71ee600d81091c2f0cbae116b06c"},
-    {file = "cryptography-46.0.4-pp311-pypy311_pp73-manylinux_2_28_aarch64.whl", hash = "sha256:c236a44acfb610e70f6b3e1c3ca20ff24459659231ef2f8c48e879e2d32b73da"},
+    {file = "cryptography-46.0.5-pp311-pypy311_pp73-manylinux_2_28_aarch64.whl", hash = "sha256:bc84e875994c3b445871ea7181d424588171efec3e185dced958dad9e001950a"},
-    {file = "cryptography-46.0.4-pp311-pypy311_pp73-manylinux_2_28_x86_64.whl", hash = "sha256:8a15fb869670efa8f83cbffbc8753c1abf236883225aed74cd179b720ac9ec80"},
+    {file = "cryptography-46.0.5-pp311-pypy311_pp73-manylinux_2_28_x86_64.whl", hash = "sha256:2ae6971afd6246710480e3f15824ed3029a60fc16991db250034efd0b9fb4356"},
-    {file = "cryptography-46.0.4-pp311-pypy311_pp73-manylinux_2_34_aarch64.whl", hash = "sha256:fdc3daab53b212472f1524d070735b2f0c214239df131903bae1d598016fa822"},
+    {file = "cryptography-46.0.5-pp311-pypy311_pp73-manylinux_2_34_aarch64.whl", hash = "sha256:d861ee9e76ace6cf36a6a89b959ec08e7bc2493ee39d07ffe5acb23ef46d27da"},
-    {file = "cryptography-46.0.4-pp311-pypy311_pp73-manylinux_2_34_x86_64.whl", hash = "sha256:44cc0675b27cadb71bdbb96099cca1fa051cd11d2ade09e5cd3a2edb929ed947"},
+    {file = "cryptography-46.0.5-pp311-pypy311_pp73-manylinux_2_34_x86_64.whl", hash = "sha256:2b7a67c9cd56372f3249b39699f2ad479f6991e62ea15800973b956f4b73e257"},
-    {file = "cryptography-46.0.4-pp311-pypy311_pp73-win_amd64.whl", hash = "sha256:be8c01a7d5a55f9a47d1888162b76c8f49d62b234d88f0ff91a9fbebe32ffbc3"},
+    {file = "cryptography-46.0.5-pp311-pypy311_pp73-win_amd64.whl", hash = "sha256:8456928655f856c6e1533ff59d5be76578a7157224dbd9ce6872f25055ab9ab7"},
-    {file = "cryptography-46.0.4.tar.gz", hash = "sha256:bfd019f60f8abc2ed1b9be4ddc21cfef059c841d86d710bb69909a688cbb8f59"},
+    {file = "cryptography-46.0.5.tar.gz", hash = "sha256:abace499247268e3757271b2f1e244b36b06f8515cf27c4d49468fc9eb16e93d"},
 ]
 
 [package.dependencies]
@@ -516,7 +516,7 @@ nox = ["nox[uv] (>=2024.4.15)"]
 pep8test = ["check-sdist", "click (>=8.0.1)", "mypy (>=1.14)", "ruff (>=0.11.11)"]
 sdist = ["build (>=1.0.0)"]
 ssh = ["bcrypt (>=3.1.5)"]
-test = ["certifi (>=2024)", "cryptography-vectors (==46.0.4)", "pretend (>=0.7)", "pytest (>=7.4.0)", "pytest-benchmark (>=4.0)", "pytest-cov (>=2.10.1)", "pytest-xdist (>=3.5.0)"]
+test = ["certifi (>=2024)", "cryptography-vectors (==46.0.5)", "pretend (>=0.7)", "pytest (>=7.4.0)", "pytest-benchmark (>=4.0)", "pytest-cov (>=2.10.1)", "pytest-xdist (>=3.5.0)"]
 test-randomorder = ["pytest-randomly"]
 
 [[package]]
@@ -570,24 +570,25 @@ tests = ["coverage", "coveralls", "dill", "mock", "nose"]
 
 [[package]]
 name = "fastapi"
-version = "0.128.0"
+version = "0.128.7"
 description = "FastAPI framework, high performance, easy to learn, fast to code, ready for production"
 optional = false
 python-versions = ">=3.9"
 groups = ["main"]
 files = [
-    {file = "fastapi-0.128.0-py3-none-any.whl", hash = "sha256:aebd93f9716ee3b4f4fcfe13ffb7cf308d99c9f3ab5622d8877441072561582d"},
+    {file = "fastapi-0.128.7-py3-none-any.whl", hash = "sha256:6bd9bd31cb7047465f2d3fa3ba3f33b0870b17d4eaf7cdb36d1576ab060ad662"},
-    {file = "fastapi-0.128.0.tar.gz", hash = "sha256:1cc179e1cef10a6be60ffe429f79b829dce99d8de32d7acb7e6c8dfdf7f2645a"},
+    {file = "fastapi-0.128.7.tar.gz", hash = "sha256:783c273416995486c155ad2c0e2b45905dedfaf20b9ef8d9f6a9124670639a24"},
 ]
 
 [package.dependencies]
 annotated-doc = ">=0.0.2"
 pydantic = ">=2.7.0"
-starlette = ">=0.40.0,<0.51.0"
+starlette = ">=0.40.0,<1.0.0"
 typing-extensions = ">=4.8.0"
+typing-inspection = ">=0.4.2"
 
 [package.extras]
-all = ["email-validator (>=2.0.0)", "fastapi-cli[standard] (>=0.0.8)", "httpx (>=0.23.0,<1.0.0)", "itsdangerous (>=1.1.0)", "jinja2 (>=3.1.5)", "orjson (>=3.2.1)", "pydantic-extra-types (>=2.0.0)", "pydantic-settings (>=2.0.0)", "python-multipart (>=0.0.18)", "pyyaml (>=5.3.1)", "ujson (>=4.0.1,!=4.0.2,!=4.1.0,!=4.2.0,!=4.3.0,!=5.0.0,!=5.1.0)", "uvicorn[standard] (>=0.12.0)"]
+all = ["email-validator (>=2.0.0)", "fastapi-cli[standard] (>=0.0.8)", "httpx (>=0.23.0,<1.0.0)", "itsdangerous (>=1.1.0)", "jinja2 (>=3.1.5)", "orjson (>=3.9.3)", "pydantic-extra-types (>=2.0.0)", "pydantic-settings (>=2.0.0)", "python-multipart (>=0.0.18)", "pyyaml (>=5.3.1)", "ujson (>=5.8.0)", "uvicorn[standard] (>=0.12.0)"]
 standard = ["email-validator (>=2.0.0)", "fastapi-cli[standard] (>=0.0.8)", "httpx (>=0.23.0,<1.0.0)", "jinja2 (>=3.1.5)", "pydantic-extra-types (>=2.0.0)", "pydantic-settings (>=2.0.0)", "python-multipart (>=0.0.18)", "uvicorn[standard] (>=0.12.0)"]
 standard-no-fastapi-cloud-cli = ["email-validator (>=2.0.0)", "fastapi-cli[standard-no-fastapi-cloud-cli] (>=0.0.8)", "httpx (>=0.23.0,<1.0.0)", "jinja2 (>=3.1.5)", "pydantic-extra-types (>=2.0.0)", "pydantic-settings (>=2.0.0)", "python-multipart (>=0.0.18)", "uvicorn[standard] (>=0.12.0)"]
 
@@ -1062,14 +1063,14 @@ urllib3 = ">=1.26.0,<3"
 
 [[package]]
 name = "launchdarkly-server-sdk"
-version = "9.14.1"
+version = "9.15.0"
 description = "LaunchDarkly SDK for Python"
 optional = false
-python-versions = ">=3.9"
+python-versions = ">=3.10"
 groups = ["main"]
 files = [
-    {file = "launchdarkly_server_sdk-9.14.1-py3-none-any.whl", hash = "sha256:a9e2bd9ecdef845cd631ae0d4334a1115e5b44257c42eb2349492be4bac7815c"},
+    {file = "launchdarkly_server_sdk-9.15.0-py3-none-any.whl", hash = "sha256:c267e29bfa3fb5e2a06a208448ada6ed5557a2924979b8d79c970b45d227c668"},
-    {file = "launchdarkly_server_sdk-9.14.1.tar.gz", hash = "sha256:1df44baf0a0efa74d8c1dad7a00592b98bce7d19edded7f770da8dbc49922213"},
+    {file = "launchdarkly_server_sdk-9.15.0.tar.gz", hash = "sha256:f31441b74bc1a69c381db57c33116509e407a2612628ad6dff0a7dbb39d5020b"},
 ]
 
 [package.dependencies]
@@ -1478,14 +1479,14 @@ testing = ["coverage", "pytest", "pytest-benchmark"]
 
 [[package]]
 name = "postgrest"
-version = "2.27.2"
+version = "2.28.0"
 description = "PostgREST client for Python. This library provides an ORM interface to PostgREST."
 optional = false
 python-versions = ">=3.9"
 groups = ["main"]
 files = [
-    {file = "postgrest-2.27.2-py3-none-any.whl", hash = "sha256:1666fef3de05ca097a314433dd5ae2f2d71c613cb7b233d0f468c4ffe37277da"},
+    {file = "postgrest-2.28.0-py3-none-any.whl", hash = "sha256:7bca2f24dd1a1bf8a3d586c7482aba6cd41662da6733045fad585b63b7f7df75"},
-    {file = "postgrest-2.27.2.tar.gz", hash = "sha256:55407d530b5af3d64e883a71fec1f345d369958f723ce4a8ab0b7d169e313242"},
+    {file = "postgrest-2.28.0.tar.gz", hash = "sha256:c36b38646d25ea4255321d3d924ce70f8d20ec7799cb42c1221d6a818d4f6515"},
 ]
 
 [package.dependencies]
@@ -2248,14 +2249,14 @@ cli = ["click (>=5.0)"]
 
 [[package]]
 name = "realtime"
-version = "2.27.2"
+version = "2.28.0"
 description = ""
 optional = false
 python-versions = ">=3.9"
 groups = ["main"]
 files = [
-    {file = "realtime-2.27.2-py3-none-any.whl", hash = "sha256:34a9cbb26a274e707e8fc9e3ee0a66de944beac0fe604dc336d1e985db2c830f"},
+    {file = "realtime-2.28.0-py3-none-any.whl", hash = "sha256:db1bd59bab9b1fcc9f9d3b1a073bed35bf4994d720e6751f10031a58d57a3836"},
-    {file = "realtime-2.27.2.tar.gz", hash = "sha256:b960a90294d2cea1b3f1275ecb89204304728e08fff1c393cc1b3150739556b3"},
+    {file = "realtime-2.28.0.tar.gz", hash = "sha256:d18cedcebd6a8f22fcd509bc767f639761eb218b7b2b6f14fc4205b6259b50fc"},
 ]
 
 [package.dependencies]
@@ -2436,14 +2437,14 @@ full = ["httpx (>=0.27.0,<0.29.0)", "itsdangerous", "jinja2", "python-multipart
 
 [[package]]
 name = "storage3"
-version = "2.27.2"
+version = "2.28.0"
 description = "Supabase Storage client for Python."
 optional = false
 python-versions = ">=3.9"
 groups = ["main"]
 files = [
-    {file = "storage3-2.27.2-py3-none-any.whl", hash = "sha256:e6f16e7a260729e7b1f46e9bf61746805a02e30f5e419ee1291007c432e3ec63"},
+    {file = "storage3-2.28.0-py3-none-any.whl", hash = "sha256:ecb50efd2ac71dabbdf97e99ad346eafa630c4c627a8e5a138ceb5fbbadae716"},
-    {file = "storage3-2.27.2.tar.gz", hash = "sha256:cb4807b7f86b4bb1272ac6fdd2f3cfd8ba577297046fa5f88557425200275af5"},
+    {file = "storage3-2.28.0.tar.gz", hash = "sha256:bc1d008aff67de7a0f2bd867baee7aadbcdb6f78f5a310b4f7a38e8c13c19865"},
 ]
 
 [package.dependencies]
@@ -2487,35 +2488,35 @@ python-dateutil = ">=2.6.0"
 
 [[package]]
 name = "supabase"
-version = "2.27.2"
+version = "2.28.0"
 description = "Supabase client for Python."
 optional = false
 python-versions = ">=3.9"
 groups = ["main"]
 files = [
-    {file = "supabase-2.27.2-py3-none-any.whl", hash = "sha256:d4dce00b3a418ee578017ec577c0e5be47a9a636355009c76f20ed2faa15bc54"},
+    {file = "supabase-2.28.0-py3-none-any.whl", hash = "sha256:42776971c7d0ccca16034df1ab96a31c50228eb1eb19da4249ad2f756fc20272"},
-    {file = "supabase-2.27.2.tar.gz", hash = "sha256:2aed40e4f3454438822442a1e94a47be6694c2c70392e7ae99b51a226d4293f7"},
+    {file = "supabase-2.28.0.tar.gz", hash = "sha256:aea299aaab2a2eed3c57e0be7fc035c6807214194cce795a3575add20268ece1"},
 ]
 
 [package.dependencies]
 httpx = ">=0.26,<0.29"
-postgrest = "2.27.2"
+postgrest = "2.28.0"
-realtime = "2.27.2"
+realtime = "2.28.0"
-storage3 = "2.27.2"
+storage3 = "2.28.0"
-supabase-auth = "2.27.2"
+supabase-auth = "2.28.0"
-supabase-functions = "2.27.2"
+supabase-functions = "2.28.0"
 yarl = ">=1.22.0"
 
 [[package]]
 name = "supabase-auth"
-version = "2.27.2"
+version = "2.28.0"
 description = "Python Client Library for Supabase Auth"
 optional = false
 python-versions = ">=3.9"
 groups = ["main"]
 files = [
-    {file = "supabase_auth-2.27.2-py3-none-any.whl", hash = "sha256:78ec25b11314d0a9527a7205f3b1c72560dccdc11b38392f80297ef98664ee91"},
+    {file = "supabase_auth-2.28.0-py3-none-any.whl", hash = "sha256:2ac85026cc285054c7fa6d41924f3a333e9ec298c013e5b5e1754039ba7caec9"},
-    {file = "supabase_auth-2.27.2.tar.gz", hash = "sha256:0f5bcc79b3677cb42e9d321f3c559070cfa40d6a29a67672cc8382fb7dc2fe97"},
+    {file = "supabase_auth-2.28.0.tar.gz", hash = "sha256:2bb8f18ff39934e44b28f10918db965659f3735cd6fbfcc022fe0b82dbf8233e"},
 ]
 
 [package.dependencies]
@@ -2525,14 +2526,14 @@ pyjwt = {version = ">=2.10.1", extras = ["crypto"]}
 
 [[package]]
 name = "supabase-functions"
-version = "2.27.2"
+version = "2.28.0"
 description = "Library for Supabase Functions"
 optional = false
 python-versions = ">=3.9"
 groups = ["main"]
 files = [
-    {file = "supabase_functions-2.27.2-py3-none-any.whl", hash = "sha256:db480efc669d0bca07605b9b6f167312af43121adcc842a111f79bea416ef754"},
+    {file = "supabase_functions-2.28.0-py3-none-any.whl", hash = "sha256:30bf2d586f8df285faf0621bb5d5bb3ec3157234fc820553ca156f009475e4ae"},
-    {file = "supabase_functions-2.27.2.tar.gz", hash = "sha256:d0c8266207a94371cb3fd35ad3c7f025b78a97cf026861e04ccd35ac1775f80b"},
+    {file = "supabase_functions-2.28.0.tar.gz", hash = "sha256:db3dddfc37aca5858819eb461130968473bd8c75bd284581013958526dac718b"},
 ]
 
 [package.dependencies]
@@ -2911,4 +2912,4 @@ type = ["pytest-mypy"]
 [metadata]
 lock-version = "2.1"
 python-versions = ">=3.10,<4.0"
-content-hash = "40eae94995dc0a388fa832ed4af9b6137f28d5b5ced3aaea70d5f91d4d9a179d"
+content-hash = "9619cae908ad38fa2c48016a58bcf4241f6f5793aa0e6cc140276e91c433cbbb"

autogpt_platform/autogpt_libs/pyproject.toml

@@ -11,14 +11,14 @@ python = ">=3.10,<4.0"
 colorama = "^0.4.6"
 cryptography = "^46.0"
 expiringdict = "^1.2.2"
-fastapi = "^0.128.0"
+fastapi = "^0.128.7"
 google-cloud-logging = "^3.13.0"
-launchdarkly-server-sdk = "^9.14.1"
+launchdarkly-server-sdk = "^9.15.0"
 pydantic = "^2.12.5"
 pydantic-settings = "^2.12.0"
 pyjwt = { version = "^2.11.0", extras = ["crypto"] }
 redis = "^6.2.0"
-supabase = "^2.27.2"
+supabase = "^2.28.0"
 uvicorn = "^0.40.0"
 
 [tool.poetry.group.dev.dependencies]

autogpt_platform/backend/Dockerfile

@@ -1,3 +1,5 @@
+# ============================ DEPENDENCY BUILDER ============================ #
+
 FROM debian:13-slim AS builder
 
 # Set environment variables
@@ -51,7 +53,9 @@ COPY autogpt_platform/backend/backend/data/partial_types.py ./backend/data/parti
 COPY autogpt_platform/backend/gen_prisma_types_stub.py ./
 RUN poetry run prisma generate && poetry run gen-prisma-stub
 
-FROM debian:13-slim AS server_dependencies
+# ============================== BACKEND SERVER ============================== #
+
+FROM debian:13-slim AS server
 
 WORKDIR /app
 
@@ -62,16 +66,21 @@ ENV POETRY_HOME=/opt/poetry \
     DEBIAN_FRONTEND=noninteractive
 ENV PATH=/opt/poetry/bin:$PATH
 
-# Install Python, FFmpeg, and ImageMagick (required for video processing blocks)
+# Install Python, FFmpeg, ImageMagick, and CLI tools for agent use.
-RUN apt-get update && apt-get install -y \
+# bubblewrap provides OS-level sandbox (whitelist-only FS + no network)
+# for the bash_exec MCP tool.
+# Using --no-install-recommends saves ~650MB by skipping unnecessary deps like llvm, mesa, etc.
+RUN apt-get update && apt-get install -y --no-install-recommends \
     python3.13 \
     python3-pip \
     ffmpeg \
     imagemagick \
+    jq \
+    ripgrep \
+    tree \
+    bubblewrap \
     && rm -rf /var/lib/apt/lists/*
 
-# Copy only necessary files from builder
-COPY --from=builder /app /app
 COPY --from=builder /usr/local/lib/python3* /usr/local/lib/python3*
 COPY --from=builder /usr/local/bin/poetry /usr/local/bin/poetry
 # Copy Node.js installation for Prisma
@@ -81,30 +90,54 @@ COPY --from=builder /usr/bin/npm /usr/bin/npm
 COPY --from=builder /usr/bin/npx /usr/bin/npx
 COPY --from=builder /root/.cache/prisma-python/binaries /root/.cache/prisma-python/binaries
 
-ENV PATH="/app/autogpt_platform/backend/.venv/bin:$PATH"
-
-RUN mkdir -p /app/autogpt_platform/autogpt_libs
-RUN mkdir -p /app/autogpt_platform/backend
-
-COPY autogpt_platform/autogpt_libs /app/autogpt_platform/autogpt_libs
-
-COPY autogpt_platform/backend/poetry.lock autogpt_platform/backend/pyproject.toml /app/autogpt_platform/backend/
-
 WORKDIR /app/autogpt_platform/backend
 
-FROM server_dependencies AS migrate
+# Copy only the .venv from builder (not the entire /app directory)
+# The .venv includes the generated Prisma client
+COPY --from=builder /app/autogpt_platform/backend/.venv ./.venv
+ENV PATH="/app/autogpt_platform/backend/.venv/bin:$PATH"
 
-# Migration stage only needs schema and migrations - much lighter than full backend
+# Copy dependency files + autogpt_libs (path dependency)
-COPY autogpt_platform/backend/schema.prisma /app/autogpt_platform/backend/
+COPY autogpt_platform/autogpt_libs /app/autogpt_platform/autogpt_libs
-COPY autogpt_platform/backend/backend/data/partial_types.py /app/autogpt_platform/backend/backend/data/partial_types.py
+COPY autogpt_platform/backend/poetry.lock autogpt_platform/backend/pyproject.toml ./
-COPY autogpt_platform/backend/migrations /app/autogpt_platform/backend/migrations
 
-FROM server_dependencies AS server
+# Copy backend code + docs (for Copilot docs search)
-
+COPY autogpt_platform/backend ./
-COPY autogpt_platform/backend /app/autogpt_platform/backend
 COPY docs /app/docs
 RUN poetry install --no-ansi --only-root
 
 ENV PORT=8000
 
 CMD ["poetry", "run", "rest"]
+
+# =============================== DB MIGRATOR =============================== #
+
+# Lightweight migrate stage - only needs Prisma CLI, not full Python environment
+FROM debian:13-slim AS migrate
+
+WORKDIR /app/autogpt_platform/backend
+
+ENV DEBIAN_FRONTEND=noninteractive
+
+# Install only what's needed for prisma migrate: Node.js and minimal Python for prisma-python
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    python3.13 \
+    python3-pip \
+    ca-certificates \
+    && rm -rf /var/lib/apt/lists/*
+
+# Copy Node.js from builder (needed for Prisma CLI)
+COPY --from=builder /usr/bin/node /usr/bin/node
+COPY --from=builder /usr/lib/node_modules /usr/lib/node_modules
+COPY --from=builder /usr/bin/npm /usr/bin/npm
+
+# Copy Prisma binaries
+COPY --from=builder /root/.cache/prisma-python/binaries /root/.cache/prisma-python/binaries
+
+# Install prisma-client-py directly (much smaller than copying full venv)
+RUN pip3 install prisma>=0.15.0 --break-system-packages
+
+COPY autogpt_platform/backend/schema.prisma ./
+COPY autogpt_platform/backend/backend/data/partial_types.py ./backend/data/partial_types.py
+COPY autogpt_platform/backend/gen_prisma_types_stub.py ./
+COPY autogpt_platform/backend/migrations ./migrations

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

@@ -27,12 +27,11 @@ class ChatConfig(BaseSettings):
     session_ttl: int = Field(default=43200, description="Session TTL in seconds")
 
     # Streaming Configuration
-    max_context_messages: int = Field(
-        default=50, ge=1, le=200, description="Maximum context messages"
-    )
-
     stream_timeout: int = Field(default=300, description="Stream timeout in seconds")
-    max_retries: int = Field(default=3, description="Maximum number of retries")
+    max_retries: int = Field(
+        default=3,
+        description="Max retries for fallback path (SDK handles retries internally)",
+    )
     max_agent_runs: int = Field(default=30, description="Maximum number of agent runs")
     max_agent_schedules: int = Field(
         default=30, description="Maximum number of agent schedules"
@@ -93,6 +92,31 @@ class ChatConfig(BaseSettings):
         description="Name of the prompt in Langfuse to fetch",
     )
 
+    # Claude Agent SDK Configuration
+    use_claude_agent_sdk: bool = Field(
+        default=True,
+        description="Use Claude Agent SDK for chat completions",
+    )
+    claude_agent_model: str | None = Field(
+        default=None,
+        description="Model for the Claude Agent SDK path. If None, derives from "
+        "the `model` field by stripping the OpenRouter provider prefix.",
+    )
+    claude_agent_max_buffer_size: int = Field(
+        default=10 * 1024 * 1024,  # 10MB (default SDK is 1MB)
+        description="Max buffer size in bytes for Claude Agent SDK JSON message parsing. "
+        "Increase if tool outputs exceed the limit.",
+    )
+    claude_agent_max_subtasks: int = Field(
+        default=10,
+        description="Max number of sub-agent Tasks the SDK can spawn per session.",
+    )
+    claude_agent_use_resume: bool = Field(
+        default=True,
+        description="Use --resume for multi-turn conversations instead of "
+        "history compression. Falls back to compression when unavailable.",
+    )
+
     # Extended thinking configuration for Claude models
     thinking_enabled: bool = Field(
         default=True,
@@ -138,6 +162,17 @@ class ChatConfig(BaseSettings):
             v = os.getenv("CHAT_INTERNAL_API_KEY")
         return v
 
+    @field_validator("use_claude_agent_sdk", mode="before")
+    @classmethod
+    def get_use_claude_agent_sdk(cls, v):
+        """Get use_claude_agent_sdk from environment if not provided."""
+        # Check environment variable - default to True if not set
+        env_val = os.getenv("CHAT_USE_CLAUDE_AGENT_SDK", "").lower()
+        if env_val:
+            return env_val in ("true", "1", "yes", "on")
+        # Default to True (SDK enabled by default)
+        return True if v is None else v
+
     # Prompt paths for different contexts
     PROMPT_PATHS: dict[str, str] = {
         "default": "prompts/chat_system.md",

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

@@ -334,9 +334,8 @@ async def _get_session_from_cache(session_id: str) -> ChatSession | None:
     try:
         session = ChatSession.model_validate_json(raw_session)
         logger.info(
-            f"Loading session {session_id} from cache: "
+            f"[CACHE] Loaded session {session_id}: {len(session.messages)} messages, "
-            f"message_count={len(session.messages)}, "
+            f"last_roles={[m.role for m in session.messages[-3:]]}"  # Last 3 roles
-            f"roles={[m.role for m in session.messages]}"
         )
         return session
     except Exception as e:
@@ -378,11 +377,9 @@ async def _get_session_from_db(session_id: str) -> ChatSession | None:
         return None
 
     messages = prisma_session.Messages
-    logger.info(
+    logger.debug(
-        f"Loading session {session_id} from DB: "
+        f"[DB] Loaded session {session_id}: {len(messages) if messages else 0} messages, "
-        f"has_messages={messages is not None}, "
+        f"roles={[m.role for m in messages[-3:]] if messages else []}"  # Last 3 roles
-        f"message_count={len(messages) if messages else 0}, "
-        f"roles={[m.role for m in messages] if messages else []}"
     )
 
     return ChatSession.from_db(prisma_session, messages)
@@ -433,10 +430,9 @@ async def _save_session_to_db(
                     "function_call": msg.function_call,
                 }
             )
-        logger.info(
+        logger.debug(
-            f"Saving {len(new_messages)} new messages to DB for session {session.session_id}: "
+            f"[DB] Saving {len(new_messages)} messages to session {session.session_id}, "
-            f"roles={[m['role'] for m in messages_data]}, "
+            f"roles={[m['role'] for m in messages_data]}"
-            f"start_sequence={existing_message_count}"
         )
         await chat_db.add_chat_messages_batch(
             session_id=session.session_id,
@@ -476,7 +472,7 @@ async def get_chat_session(
         logger.warning(f"Unexpected cache error for session {session_id}: {e}")
 
     # Fall back to database
-    logger.info(f"Session {session_id} not in cache, checking database")
+    logger.debug(f"Session {session_id} not in cache, checking database")
     session = await _get_session_from_db(session_id)
 
     if session is None:
@@ -493,7 +489,6 @@ async def get_chat_session(
     # Cache the session from DB
     try:
         await _cache_session(session)
-        logger.info(f"Cached session {session_id} from database")
     except Exception as e:
         logger.warning(f"Failed to cache session {session_id}: {e}")
 
@@ -558,6 +553,40 @@ async def upsert_chat_session(
         return session
 
 
+async def append_and_save_message(session_id: str, message: ChatMessage) -> ChatSession:
+    """Atomically append a message to a session and persist it.
+
+    Acquires the session lock, re-fetches the latest session state,
+    appends the message, and saves — preventing message loss when
+    concurrent requests modify the same session.
+    """
+    lock = await _get_session_lock(session_id)
+
+    async with lock:
+        session = await get_chat_session(session_id)
+        if session is None:
+            raise ValueError(f"Session {session_id} not found")
+
+        session.messages.append(message)
+        existing_message_count = await chat_db.get_chat_session_message_count(
+            session_id
+        )
+
+        try:
+            await _save_session_to_db(session, existing_message_count)
+        except Exception as e:
+            raise DatabaseError(
+                f"Failed to persist message to session {session_id}"
+            ) from e
+
+        try:
+            await _cache_session(session)
+        except Exception as e:
+            logger.warning(f"Cache write failed for session {session_id}: {e}")
+
+        return session
+
+
 async def create_chat_session(user_id: str) -> ChatSession:
     """Create a new chat session and persist it.
 
@@ -664,13 +693,19 @@ async def update_session_title(session_id: str, title: str) -> bool:
             logger.warning(f"Session {session_id} not found for title update")
             return False
 
-        # Invalidate cache so next fetch gets updated title
+        # Update title in cache if it exists (instead of invalidating).
+        # This prevents race conditions where cache invalidation causes
+        # the frontend to see stale DB data while streaming is still in progress.
         try:
-            redis_key = _get_session_cache_key(session_id)
+            cached = await _get_session_from_cache(session_id)
-            async_redis = await get_redis_async()
+            if cached:
-            await async_redis.delete(redis_key)
+                cached.title = title
+                await _cache_session(cached)
         except Exception as e:
-            logger.warning(f"Failed to invalidate cache for session {session_id}: {e}")
+            # Not critical - title will be correct on next full cache refresh
+            logger.warning(
+                f"Failed to update title in cache for session {session_id}: {e}"
+            )
 
         return True
     except Exception as e:

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

@@ -1,5 +1,6 @@
 """Chat API routes for chat session management and streaming via SSE."""
 
+import asyncio
 import logging
 import uuid as uuid_module
 from collections.abc import AsyncGenerator
@@ -11,19 +12,29 @@ from fastapi.responses import StreamingResponse
 from pydantic import BaseModel
 
 from backend.util.exceptions import NotFoundError
+from backend.util.feature_flag import Flag, is_feature_enabled
 
 from . import service as chat_service
 from . import stream_registry
 from .completion_handler import process_operation_failure, process_operation_success
 from .config import ChatConfig
-from .model import ChatSession, create_chat_session, get_chat_session, get_user_sessions
+from .model import (
-from .response_model import StreamFinish, StreamHeartbeat
+    ChatMessage,
+    ChatSession,
+    append_and_save_message,
+    create_chat_session,
+    get_chat_session,
+    get_user_sessions,
+)
+from .response_model import StreamError, StreamFinish, StreamHeartbeat, StreamStart
+from .sdk import service as sdk_service
 from .tools.models import (
     AgentDetailsResponse,
     AgentOutputResponse,
     AgentPreviewResponse,
     AgentSavedResponse,
     AgentsFoundResponse,
+    BlockDetailsResponse,
     BlockListResponse,
     BlockOutputResponse,
     ClarificationNeededResponse,
@@ -40,6 +51,7 @@ from .tools.models import (
     SetupRequirementsResponse,
     UnderstandingUpdatedResponse,
 )
+from .tracking import track_user_message
 
 config = ChatConfig()
 
@@ -231,6 +243,10 @@ async def get_session(
     active_task, last_message_id = await stream_registry.get_active_task_for_session(
         session_id, user_id
     )
+    logger.info(
+        f"[GET_SESSION] session={session_id}, active_task={active_task is not None}, "
+        f"msg_count={len(messages)}, last_role={messages[-1].get('role') if messages else 'none'}"
+    )
     if active_task:
         # Filter out the in-progress assistant message from the session response.
         # The client will receive the complete assistant response through the SSE
@@ -300,10 +316,9 @@ async def stream_chat_post(
         f"user={user_id}, message_len={len(request.message)}",
         extra={"json_fields": log_meta},
     )
-
     session = await _validate_and_get_session(session_id, user_id)
     logger.info(
-        f"[TIMING] session validated in {(time.perf_counter() - stream_start_time)*1000:.1f}ms",
+        f"[TIMING] session validated in {(time.perf_counter() - stream_start_time) * 1000:.1f}ms",
         extra={
             "json_fields": {
                 **log_meta,
@@ -312,6 +327,25 @@ async def stream_chat_post(
         },
     )
 
+    # Atomically append user message to session BEFORE creating task to avoid
+    # race condition where GET_SESSION sees task as "running" but message isn't
+    # saved yet.  append_and_save_message re-fetches inside a lock to prevent
+    # message loss from concurrent requests.
+    if request.message:
+        message = ChatMessage(
+            role="user" if request.is_user_message else "assistant",
+            content=request.message,
+        )
+        if request.is_user_message:
+            track_user_message(
+                user_id=user_id,
+                session_id=session_id,
+                message_length=len(request.message),
+            )
+        logger.info(f"[STREAM] Saving user message to session {session_id}")
+        session = await append_and_save_message(session_id, message)
+        logger.info(f"[STREAM] User message saved for session {session_id}")
+
     # Create a task in the stream registry for reconnection support
     task_id = str(uuid_module.uuid4())
     operation_id = str(uuid_module.uuid4())
@@ -327,7 +361,7 @@ async def stream_chat_post(
         operation_id=operation_id,
     )
     logger.info(
-        f"[TIMING] create_task completed in {(time.perf_counter() - task_create_start)*1000:.1f}ms",
+        f"[TIMING] create_task completed in {(time.perf_counter() - task_create_start) * 1000:.1f}ms",
         extra={
             "json_fields": {
                 **log_meta,
@@ -348,15 +382,47 @@ async def stream_chat_post(
         first_chunk_time, ttfc = None, None
         chunk_count = 0
         try:
-            async for chunk in chat_service.stream_chat_completion(
+            # Emit a start event with task_id for reconnection
+            start_chunk = StreamStart(messageId=task_id, taskId=task_id)
+            await stream_registry.publish_chunk(task_id, start_chunk)
+            logger.info(
+                f"[TIMING] StreamStart published at {(time_module.perf_counter() - gen_start_time) * 1000:.1f}ms",
+                extra={
+                    "json_fields": {
+                        **log_meta,
+                        "elapsed_ms": (time_module.perf_counter() - gen_start_time)
+                        * 1000,
+                    }
+                },
+            )
+
+            # Choose service based on LaunchDarkly flag (falls back to config default)
+            use_sdk = await is_feature_enabled(
+                Flag.COPILOT_SDK,
+                user_id or "anonymous",
+                default=config.use_claude_agent_sdk,
+            )
+            stream_fn = (
+                sdk_service.stream_chat_completion_sdk
+                if use_sdk
+                else chat_service.stream_chat_completion
+            )
+            logger.info(
+                f"[TIMING] Calling {'sdk' if use_sdk else 'standard'} stream_chat_completion",
+                extra={"json_fields": log_meta},
+            )
+            # Pass message=None since we already added it to the session above
+            async for chunk in stream_fn(
                 session_id,
-                request.message,
+                None,  # Message already in session
                 is_user_message=request.is_user_message,
                 user_id=user_id,
-                session=session,  # Pass pre-fetched session to avoid double-fetch
+                session=session,  # Pass session with message already added
                 context=request.context,
-                _task_id=task_id,  # Pass task_id so service emits start with taskId for reconnection
             ):
+                # Skip duplicate StreamStart — we already published one above
+                if isinstance(chunk, StreamStart):
+                    continue
                 chunk_count += 1
                 if first_chunk_time is None:
                     first_chunk_time = time_module.perf_counter()
@@ -377,7 +443,7 @@ async def stream_chat_post(
             gen_end_time = time_module.perf_counter()
             total_time = (gen_end_time - gen_start_time) * 1000
             logger.info(
-                f"[TIMING] run_ai_generation FINISHED in {total_time/1000:.1f}s; "
+                f"[TIMING] run_ai_generation FINISHED in {total_time / 1000:.1f}s; "
                 f"task={task_id}, session={session_id}, "
                 f"ttfc={ttfc or -1:.2f}s, n_chunks={chunk_count}",
                 extra={
@@ -404,6 +470,17 @@ async def stream_chat_post(
                     }
                 },
             )
+            # Publish a StreamError so the frontend can display an error message
+            try:
+                await stream_registry.publish_chunk(
+                    task_id,
+                    StreamError(
+                        errorText="An error occurred. Please try again.",
+                        code="stream_error",
+                    ),
+                )
+            except Exception:
+                pass  # Best-effort; mark_task_completed will publish StreamFinish
             await stream_registry.mark_task_completed(task_id, "failed")
 
     # Start the AI generation in a background task
@@ -506,8 +583,14 @@ async def stream_chat_post(
                     "json_fields": {**log_meta, "elapsed_ms": elapsed, "error": str(e)}
                 },
             )
+            # Surface error to frontend so it doesn't appear stuck
+            yield StreamError(
+                errorText="An error occurred. Please try again.",
+                code="stream_error",
+            ).to_sse()
+            yield StreamFinish().to_sse()
         finally:
-            # Unsubscribe when client disconnects or stream ends to prevent resource leak
+            # Unsubscribe when client disconnects or stream ends
             if subscriber_queue is not None:
                 try:
                     await stream_registry.unsubscribe_from_task(
@@ -751,8 +834,6 @@ async def stream_task(
         )
 
     async def event_generator() -> AsyncGenerator[str, None]:
-        import asyncio
-
         heartbeat_interval = 15.0  # Send heartbeat every 15 seconds
         try:
             while True:
@@ -971,6 +1052,7 @@ ToolResponseUnion = (
     | AgentSavedResponse
     | ClarificationNeededResponse
     | BlockListResponse
+    | BlockDetailsResponse
     | BlockOutputResponse
     | DocSearchResultsResponse
     | DocPageResponse

autogpt_platform/backend/backend/api/features/chat/sdk/__init__.py

@@ -0,0 +1,14 @@
+"""Claude Agent SDK integration for CoPilot.
+
+This module provides the integration layer between the Claude Agent SDK
+and the existing CoPilot tool system, enabling drop-in replacement of
+the current LLM orchestration with the battle-tested Claude Agent SDK.
+"""
+
+from .service import stream_chat_completion_sdk
+from .tool_adapter import create_copilot_mcp_server
+
+__all__ = [
+    "stream_chat_completion_sdk",
+    "create_copilot_mcp_server",
+]

autogpt_platform/backend/backend/api/features/chat/sdk/response_adapter.py

@@ -0,0 +1,203 @@
+"""Response adapter for converting Claude Agent SDK messages to Vercel AI SDK format.
+
+This module provides the adapter layer that converts streaming messages from
+the Claude Agent SDK into the Vercel AI SDK UI Stream Protocol format that
+the frontend expects.
+"""
+
+import json
+import logging
+import uuid
+
+from claude_agent_sdk import (
+    AssistantMessage,
+    Message,
+    ResultMessage,
+    SystemMessage,
+    TextBlock,
+    ToolResultBlock,
+    ToolUseBlock,
+    UserMessage,
+)
+
+from backend.api.features.chat.response_model import (
+    StreamBaseResponse,
+    StreamError,
+    StreamFinish,
+    StreamFinishStep,
+    StreamStart,
+    StreamStartStep,
+    StreamTextDelta,
+    StreamTextEnd,
+    StreamTextStart,
+    StreamToolInputAvailable,
+    StreamToolInputStart,
+    StreamToolOutputAvailable,
+)
+from backend.api.features.chat.sdk.tool_adapter import (
+    MCP_TOOL_PREFIX,
+    pop_pending_tool_output,
+)
+
+logger = logging.getLogger(__name__)
+
+
+class SDKResponseAdapter:
+    """Adapter for converting Claude Agent SDK messages to Vercel AI SDK format.
+
+    This class maintains state during a streaming session to properly track
+    text blocks, tool calls, and message lifecycle.
+    """
+
+    def __init__(self, message_id: str | None = None):
+        self.message_id = message_id or str(uuid.uuid4())
+        self.text_block_id = str(uuid.uuid4())
+        self.has_started_text = False
+        self.has_ended_text = False
+        self.current_tool_calls: dict[str, dict[str, str]] = {}
+        self.task_id: str | None = None
+        self.step_open = False
+
+    def set_task_id(self, task_id: str) -> None:
+        """Set the task ID for reconnection support."""
+        self.task_id = task_id
+
+    def convert_message(self, sdk_message: Message) -> list[StreamBaseResponse]:
+        """Convert a single SDK message to Vercel AI SDK format."""
+        responses: list[StreamBaseResponse] = []
+
+        if isinstance(sdk_message, SystemMessage):
+            if sdk_message.subtype == "init":
+                responses.append(
+                    StreamStart(messageId=self.message_id, taskId=self.task_id)
+                )
+                # Open the first step (matches non-SDK: StreamStart then StreamStartStep)
+                responses.append(StreamStartStep())
+                self.step_open = True
+
+        elif isinstance(sdk_message, AssistantMessage):
+            # After tool results, the SDK sends a new AssistantMessage for the
+            # next LLM turn. Open a new step if the previous one was closed.
+            if not self.step_open:
+                responses.append(StreamStartStep())
+                self.step_open = True
+
+            for block in sdk_message.content:
+                if isinstance(block, TextBlock):
+                    if block.text:
+                        self._ensure_text_started(responses)
+                        responses.append(
+                            StreamTextDelta(id=self.text_block_id, delta=block.text)
+                        )
+
+                elif isinstance(block, ToolUseBlock):
+                    self._end_text_if_open(responses)
+
+                    # Strip MCP prefix so frontend sees "find_block"
+                    # instead of "mcp__copilot__find_block".
+                    tool_name = block.name.removeprefix(MCP_TOOL_PREFIX)
+
+                    responses.append(
+                        StreamToolInputStart(toolCallId=block.id, toolName=tool_name)
+                    )
+                    responses.append(
+                        StreamToolInputAvailable(
+                            toolCallId=block.id,
+                            toolName=tool_name,
+                            input=block.input,
+                        )
+                    )
+                    self.current_tool_calls[block.id] = {"name": tool_name}
+
+        elif isinstance(sdk_message, UserMessage):
+            # UserMessage carries tool results back from tool execution.
+            content = sdk_message.content
+            blocks = content if isinstance(content, list) else []
+            for block in blocks:
+                if isinstance(block, ToolResultBlock) and block.tool_use_id:
+                    tool_info = self.current_tool_calls.get(block.tool_use_id, {})
+                    tool_name = tool_info.get("name", "unknown")
+
+                    # Prefer the stashed full output over the SDK's
+                    # (potentially truncated) ToolResultBlock content.
+                    # The SDK truncates large results, writing them to disk,
+                    # which breaks frontend widget parsing.
+                    output = pop_pending_tool_output(tool_name) or (
+                        _extract_tool_output(block.content)
+                    )
+
+                    responses.append(
+                        StreamToolOutputAvailable(
+                            toolCallId=block.tool_use_id,
+                            toolName=tool_name,
+                            output=output,
+                            success=not (block.is_error or False),
+                        )
+                    )
+
+            # Close the current step after tool results — the next
+            # AssistantMessage will open a new step for the continuation.
+            if self.step_open:
+                responses.append(StreamFinishStep())
+                self.step_open = False
+
+        elif isinstance(sdk_message, ResultMessage):
+            self._end_text_if_open(responses)
+            # Close the step before finishing.
+            if self.step_open:
+                responses.append(StreamFinishStep())
+                self.step_open = False
+
+            if sdk_message.subtype == "success":
+                responses.append(StreamFinish())
+            elif sdk_message.subtype in ("error", "error_during_execution"):
+                error_msg = getattr(sdk_message, "result", None) or "Unknown error"
+                responses.append(
+                    StreamError(errorText=str(error_msg), code="sdk_error")
+                )
+                responses.append(StreamFinish())
+            else:
+                logger.warning(
+                    f"Unexpected ResultMessage subtype: {sdk_message.subtype}"
+                )
+                responses.append(StreamFinish())
+
+        else:
+            logger.debug(f"Unhandled SDK message type: {type(sdk_message).__name__}")
+
+        return responses
+
+    def _ensure_text_started(self, responses: list[StreamBaseResponse]) -> None:
+        """Start (or restart) a text block if needed."""
+        if not self.has_started_text or self.has_ended_text:
+            if self.has_ended_text:
+                self.text_block_id = str(uuid.uuid4())
+                self.has_ended_text = False
+            responses.append(StreamTextStart(id=self.text_block_id))
+            self.has_started_text = True
+
+    def _end_text_if_open(self, responses: list[StreamBaseResponse]) -> None:
+        """End the current text block if one is open."""
+        if self.has_started_text and not self.has_ended_text:
+            responses.append(StreamTextEnd(id=self.text_block_id))
+            self.has_ended_text = True
+
+
+def _extract_tool_output(content: str | list[dict[str, str]] | None) -> str:
+    """Extract a string output from a ToolResultBlock's content field."""
+    if isinstance(content, str):
+        return content
+    if isinstance(content, list):
+        parts = [item.get("text", "") for item in content if item.get("type") == "text"]
+        if parts:
+            return "".join(parts)
+        try:
+            return json.dumps(content)
+        except (TypeError, ValueError):
+            return str(content)
+    if content is None:
+        return ""
+    try:
+        return json.dumps(content)
+    except (TypeError, ValueError):
+        return str(content)

autogpt_platform/backend/backend/api/features/chat/sdk/response_adapter_test.py

@@ -0,0 +1,366 @@
+"""Unit tests for the SDK response adapter."""
+
+from claude_agent_sdk import (
+    AssistantMessage,
+    ResultMessage,
+    SystemMessage,
+    TextBlock,
+    ToolResultBlock,
+    ToolUseBlock,
+    UserMessage,
+)
+
+from backend.api.features.chat.response_model import (
+    StreamBaseResponse,
+    StreamError,
+    StreamFinish,
+    StreamFinishStep,
+    StreamStart,
+    StreamStartStep,
+    StreamTextDelta,
+    StreamTextEnd,
+    StreamTextStart,
+    StreamToolInputAvailable,
+    StreamToolInputStart,
+    StreamToolOutputAvailable,
+)
+
+from .response_adapter import SDKResponseAdapter
+from .tool_adapter import MCP_TOOL_PREFIX
+
+
+def _adapter() -> SDKResponseAdapter:
+    a = SDKResponseAdapter(message_id="msg-1")
+    a.set_task_id("task-1")
+    return a
+
+
+# -- SystemMessage -----------------------------------------------------------
+
+
+def test_system_init_emits_start_and_step():
+    adapter = _adapter()
+    results = adapter.convert_message(SystemMessage(subtype="init", data={}))
+    assert len(results) == 2
+    assert isinstance(results[0], StreamStart)
+    assert results[0].messageId == "msg-1"
+    assert results[0].taskId == "task-1"
+    assert isinstance(results[1], StreamStartStep)
+
+
+def test_system_non_init_emits_nothing():
+    adapter = _adapter()
+    results = adapter.convert_message(SystemMessage(subtype="other", data={}))
+    assert results == []
+
+
+# -- AssistantMessage with TextBlock -----------------------------------------
+
+
+def test_text_block_emits_step_start_and_delta():
+    adapter = _adapter()
+    msg = AssistantMessage(content=[TextBlock(text="hello")], model="test")
+    results = adapter.convert_message(msg)
+    assert len(results) == 3
+    assert isinstance(results[0], StreamStartStep)
+    assert isinstance(results[1], StreamTextStart)
+    assert isinstance(results[2], StreamTextDelta)
+    assert results[2].delta == "hello"
+
+
+def test_empty_text_block_emits_only_step():
+    adapter = _adapter()
+    msg = AssistantMessage(content=[TextBlock(text="")], model="test")
+    results = adapter.convert_message(msg)
+    # Empty text skipped, but step still opens
+    assert len(results) == 1
+    assert isinstance(results[0], StreamStartStep)
+
+
+def test_multiple_text_deltas_reuse_block_id():
+    adapter = _adapter()
+    msg1 = AssistantMessage(content=[TextBlock(text="a")], model="test")
+    msg2 = AssistantMessage(content=[TextBlock(text="b")], model="test")
+    r1 = adapter.convert_message(msg1)
+    r2 = adapter.convert_message(msg2)
+    # First gets step+start+delta, second only delta (block & step already started)
+    assert len(r1) == 3
+    assert isinstance(r1[0], StreamStartStep)
+    assert isinstance(r1[1], StreamTextStart)
+    assert len(r2) == 1
+    assert isinstance(r2[0], StreamTextDelta)
+    assert r1[1].id == r2[0].id  # same block ID
+
+
+# -- AssistantMessage with ToolUseBlock --------------------------------------
+
+
+def test_tool_use_emits_input_start_and_available():
+    """Tool names arrive with MCP prefix and should be stripped for the frontend."""
+    adapter = _adapter()
+    msg = AssistantMessage(
+        content=[
+            ToolUseBlock(
+                id="tool-1",
+                name=f"{MCP_TOOL_PREFIX}find_agent",
+                input={"q": "x"},
+            )
+        ],
+        model="test",
+    )
+    results = adapter.convert_message(msg)
+    assert len(results) == 3
+    assert isinstance(results[0], StreamStartStep)
+    assert isinstance(results[1], StreamToolInputStart)
+    assert results[1].toolCallId == "tool-1"
+    assert results[1].toolName == "find_agent"  # prefix stripped
+    assert isinstance(results[2], StreamToolInputAvailable)
+    assert results[2].toolName == "find_agent"  # prefix stripped
+    assert results[2].input == {"q": "x"}
+
+
+def test_text_then_tool_ends_text_block():
+    adapter = _adapter()
+    text_msg = AssistantMessage(content=[TextBlock(text="thinking...")], model="test")
+    tool_msg = AssistantMessage(
+        content=[ToolUseBlock(id="t1", name=f"{MCP_TOOL_PREFIX}tool", input={})],
+        model="test",
+    )
+    adapter.convert_message(text_msg)  # opens step + text
+    results = adapter.convert_message(tool_msg)
+    # Step already open, so: TextEnd, ToolInputStart, ToolInputAvailable
+    assert len(results) == 3
+    assert isinstance(results[0], StreamTextEnd)
+    assert isinstance(results[1], StreamToolInputStart)
+
+
+# -- UserMessage with ToolResultBlock ----------------------------------------
+
+
+def test_tool_result_emits_output_and_finish_step():
+    adapter = _adapter()
+    # First register the tool call (opens step) — SDK sends prefixed name
+    tool_msg = AssistantMessage(
+        content=[ToolUseBlock(id="t1", name=f"{MCP_TOOL_PREFIX}find_agent", input={})],
+        model="test",
+    )
+    adapter.convert_message(tool_msg)
+
+    # Now send tool result
+    result_msg = UserMessage(
+        content=[ToolResultBlock(tool_use_id="t1", content="found 3 agents")]
+    )
+    results = adapter.convert_message(result_msg)
+    assert len(results) == 2
+    assert isinstance(results[0], StreamToolOutputAvailable)
+    assert results[0].toolCallId == "t1"
+    assert results[0].toolName == "find_agent"  # prefix stripped
+    assert results[0].output == "found 3 agents"
+    assert results[0].success is True
+    assert isinstance(results[1], StreamFinishStep)
+
+
+def test_tool_result_error():
+    adapter = _adapter()
+    adapter.convert_message(
+        AssistantMessage(
+            content=[
+                ToolUseBlock(id="t1", name=f"{MCP_TOOL_PREFIX}run_agent", input={})
+            ],
+            model="test",
+        )
+    )
+    result_msg = UserMessage(
+        content=[ToolResultBlock(tool_use_id="t1", content="timeout", is_error=True)]
+    )
+    results = adapter.convert_message(result_msg)
+    assert isinstance(results[0], StreamToolOutputAvailable)
+    assert results[0].success is False
+    assert isinstance(results[1], StreamFinishStep)
+
+
+def test_tool_result_list_content():
+    adapter = _adapter()
+    adapter.convert_message(
+        AssistantMessage(
+            content=[ToolUseBlock(id="t1", name=f"{MCP_TOOL_PREFIX}tool", input={})],
+            model="test",
+        )
+    )
+    result_msg = UserMessage(
+        content=[
+            ToolResultBlock(
+                tool_use_id="t1",
+                content=[
+                    {"type": "text", "text": "line1"},
+                    {"type": "text", "text": "line2"},
+                ],
+            )
+        ]
+    )
+    results = adapter.convert_message(result_msg)
+    assert isinstance(results[0], StreamToolOutputAvailable)
+    assert results[0].output == "line1line2"
+    assert isinstance(results[1], StreamFinishStep)
+
+
+def test_string_user_message_ignored():
+    """A plain string UserMessage (not tool results) produces no output."""
+    adapter = _adapter()
+    results = adapter.convert_message(UserMessage(content="hello"))
+    assert results == []
+
+
+# -- ResultMessage -----------------------------------------------------------
+
+
+def test_result_success_emits_finish_step_and_finish():
+    adapter = _adapter()
+    # Start some text first (opens step)
+    adapter.convert_message(
+        AssistantMessage(content=[TextBlock(text="done")], model="test")
+    )
+    msg = ResultMessage(
+        subtype="success",
+        duration_ms=100,
+        duration_api_ms=50,
+        is_error=False,
+        num_turns=1,
+        session_id="s1",
+    )
+    results = adapter.convert_message(msg)
+    # TextEnd + FinishStep + StreamFinish
+    assert len(results) == 3
+    assert isinstance(results[0], StreamTextEnd)
+    assert isinstance(results[1], StreamFinishStep)
+    assert isinstance(results[2], StreamFinish)
+
+
+def test_result_error_emits_error_and_finish():
+    adapter = _adapter()
+    msg = ResultMessage(
+        subtype="error",
+        duration_ms=100,
+        duration_api_ms=50,
+        is_error=True,
+        num_turns=0,
+        session_id="s1",
+        result="API rate limited",
+    )
+    results = adapter.convert_message(msg)
+    # No step was open, so no FinishStep — just Error + Finish
+    assert len(results) == 2
+    assert isinstance(results[0], StreamError)
+    assert "API rate limited" in results[0].errorText
+    assert isinstance(results[1], StreamFinish)
+
+
+# -- Text after tools (new block ID) ----------------------------------------
+
+
+def test_text_after_tool_gets_new_block_id():
+    adapter = _adapter()
+    # Text -> Tool -> ToolResult -> Text should get a new text block ID and step
+    adapter.convert_message(
+        AssistantMessage(content=[TextBlock(text="before")], model="test")
+    )
+    adapter.convert_message(
+        AssistantMessage(
+            content=[ToolUseBlock(id="t1", name=f"{MCP_TOOL_PREFIX}tool", input={})],
+            model="test",
+        )
+    )
+    # Send tool result (closes step)
+    adapter.convert_message(
+        UserMessage(content=[ToolResultBlock(tool_use_id="t1", content="ok")])
+    )
+    results = adapter.convert_message(
+        AssistantMessage(content=[TextBlock(text="after")], model="test")
+    )
+    # Should get StreamStartStep (new step) + StreamTextStart (new block) + StreamTextDelta
+    assert len(results) == 3
+    assert isinstance(results[0], StreamStartStep)
+    assert isinstance(results[1], StreamTextStart)
+    assert isinstance(results[2], StreamTextDelta)
+    assert results[2].delta == "after"
+
+
+# -- Full conversation flow --------------------------------------------------
+
+
+def test_full_conversation_flow():
+    """Simulate a complete conversation: init -> text -> tool -> result -> text -> finish."""
+    adapter = _adapter()
+    all_responses: list[StreamBaseResponse] = []
+
+    # 1. Init
+    all_responses.extend(
+        adapter.convert_message(SystemMessage(subtype="init", data={}))
+    )
+    # 2. Assistant text
+    all_responses.extend(
+        adapter.convert_message(
+            AssistantMessage(content=[TextBlock(text="Let me search")], model="test")
+        )
+    )
+    # 3. Tool use
+    all_responses.extend(
+        adapter.convert_message(
+            AssistantMessage(
+                content=[
+                    ToolUseBlock(
+                        id="t1",
+                        name=f"{MCP_TOOL_PREFIX}find_agent",
+                        input={"query": "email"},
+                    )
+                ],
+                model="test",
+            )
+        )
+    )
+    # 4. Tool result
+    all_responses.extend(
+        adapter.convert_message(
+            UserMessage(
+                content=[ToolResultBlock(tool_use_id="t1", content="Found 2 agents")]
+            )
+        )
+    )
+    # 5. More text
+    all_responses.extend(
+        adapter.convert_message(
+            AssistantMessage(content=[TextBlock(text="I found 2")], model="test")
+        )
+    )
+    # 6. Result
+    all_responses.extend(
+        adapter.convert_message(
+            ResultMessage(
+                subtype="success",
+                duration_ms=500,
+                duration_api_ms=400,
+                is_error=False,
+                num_turns=2,
+                session_id="s1",
+            )
+        )
+    )
+
+    types = [type(r).__name__ for r in all_responses]
+    assert types == [
+        "StreamStart",
+        "StreamStartStep",  # step 1: text + tool call
+        "StreamTextStart",
+        "StreamTextDelta",  # "Let me search"
+        "StreamTextEnd",  # closed before tool
+        "StreamToolInputStart",
+        "StreamToolInputAvailable",
+        "StreamToolOutputAvailable",  # tool result
+        "StreamFinishStep",  # step 1 closed after tool result
+        "StreamStartStep",  # step 2: continuation text
+        "StreamTextStart",  # new block after tool
+        "StreamTextDelta",  # "I found 2"
+        "StreamTextEnd",  # closed by result
+        "StreamFinishStep",  # step 2 closed
+        "StreamFinish",
+    ]

autogpt_platform/backend/backend/api/features/chat/sdk/security_hooks.py

@@ -0,0 +1,335 @@
+"""Security hooks for Claude Agent SDK integration.
+
+This module provides security hooks that validate tool calls before execution,
+ensuring multi-user isolation and preventing unauthorized operations.
+"""
+
+import json
+import logging
+import os
+import re
+from collections.abc import Callable
+from typing import Any, cast
+
+from backend.api.features.chat.sdk.tool_adapter import MCP_TOOL_PREFIX
+
+logger = logging.getLogger(__name__)
+
+# Tools that are blocked entirely (CLI/system access).
+# "Bash" (capital) is the SDK built-in — it's NOT in allowed_tools but blocked
+# here as defence-in-depth.  The agent uses mcp__copilot__bash_exec instead,
+# which has kernel-level network isolation (unshare --net).
+BLOCKED_TOOLS = {
+    "Bash",
+    "bash",
+    "shell",
+    "exec",
+    "terminal",
+    "command",
+}
+
+# Tools allowed only when their path argument stays within the SDK workspace.
+# The SDK uses these to handle oversized tool results (writes to tool-results/
+# files, then reads them back) and for workspace file operations.
+WORKSPACE_SCOPED_TOOLS = {"Read", "Write", "Edit", "Glob", "Grep"}
+
+# Dangerous patterns in tool inputs
+DANGEROUS_PATTERNS = [
+    r"sudo",
+    r"rm\s+-rf",
+    r"dd\s+if=",
+    r"/etc/passwd",
+    r"/etc/shadow",
+    r"chmod\s+777",
+    r"curl\s+.*\|.*sh",
+    r"wget\s+.*\|.*sh",
+    r"eval\s*\(",
+    r"exec\s*\(",
+    r"__import__",
+    r"os\.system",
+    r"subprocess",
+]
+
+
+def _deny(reason: str) -> dict[str, Any]:
+    """Return a hook denial response."""
+    return {
+        "hookSpecificOutput": {
+            "hookEventName": "PreToolUse",
+            "permissionDecision": "deny",
+            "permissionDecisionReason": reason,
+        }
+    }
+
+
+def _validate_workspace_path(
+    tool_name: str, tool_input: dict[str, Any], sdk_cwd: str | None
+) -> dict[str, Any]:
+    """Validate that a workspace-scoped tool only accesses allowed paths.
+
+    Allowed directories:
+    - The SDK working directory (``/tmp/copilot-<session>/``)
+    - The SDK tool-results directory (``~/.claude/projects/…/tool-results/``)
+    """
+    path = tool_input.get("file_path") or tool_input.get("path") or ""
+    if not path:
+        # Glob/Grep without a path default to cwd which is already sandboxed
+        return {}
+
+    # Resolve relative paths against sdk_cwd (the SDK sets cwd so the LLM
+    # naturally uses relative paths like "test.txt" instead of absolute ones).
+    # Tilde paths (~/) are home-dir references, not relative — expand first.
+    if path.startswith("~"):
+        resolved = os.path.realpath(os.path.expanduser(path))
+    elif not os.path.isabs(path) and sdk_cwd:
+        resolved = os.path.realpath(os.path.join(sdk_cwd, path))
+    else:
+        resolved = os.path.realpath(path)
+
+    # Allow access within the SDK working directory
+    if sdk_cwd:
+        norm_cwd = os.path.realpath(sdk_cwd)
+        if resolved.startswith(norm_cwd + os.sep) or resolved == norm_cwd:
+            return {}
+
+    # Allow access to ~/.claude/projects/*/tool-results/ (big tool results)
+    claude_dir = os.path.realpath(os.path.expanduser("~/.claude/projects"))
+    tool_results_seg = os.sep + "tool-results" + os.sep
+    if resolved.startswith(claude_dir + os.sep) and tool_results_seg in resolved:
+        return {}
+
+    logger.warning(
+        f"Blocked {tool_name} outside workspace: {path} (resolved={resolved})"
+    )
+    workspace_hint = f" Allowed workspace: {sdk_cwd}" if sdk_cwd else ""
+    return _deny(
+        f"[SECURITY] Tool '{tool_name}' can only access files within the workspace "
+        f"directory.{workspace_hint} "
+        "This is enforced by the platform and cannot be bypassed."
+    )
+
+
+def _validate_tool_access(
+    tool_name: str, tool_input: dict[str, Any], sdk_cwd: str | None = None
+) -> dict[str, Any]:
+    """Validate that a tool call is allowed.
+
+    Returns:
+        Empty dict to allow, or dict with hookSpecificOutput to deny
+    """
+    # Block forbidden tools
+    if tool_name in BLOCKED_TOOLS:
+        logger.warning(f"Blocked tool access attempt: {tool_name}")
+        return _deny(
+            f"[SECURITY] Tool '{tool_name}' is blocked for security. "
+            "This is enforced by the platform and cannot be bypassed. "
+            "Use the CoPilot-specific MCP tools instead."
+        )
+
+    # Workspace-scoped tools: allowed only within the SDK workspace directory
+    if tool_name in WORKSPACE_SCOPED_TOOLS:
+        return _validate_workspace_path(tool_name, tool_input, sdk_cwd)
+
+    # Check for dangerous patterns in tool input
+    # Use json.dumps for predictable format (str() produces Python repr)
+    input_str = json.dumps(tool_input) if tool_input else ""
+
+    for pattern in DANGEROUS_PATTERNS:
+        if re.search(pattern, input_str, re.IGNORECASE):
+            logger.warning(
+                f"Blocked dangerous pattern in tool input: {pattern} in {tool_name}"
+            )
+            return _deny(
+                "[SECURITY] Input contains a blocked pattern. "
+                "This is enforced by the platform and cannot be bypassed."
+            )
+
+    return {}
+
+
+def _validate_user_isolation(
+    tool_name: str, tool_input: dict[str, Any], user_id: str | None
+) -> dict[str, Any]:
+    """Validate that tool calls respect user isolation."""
+    # For workspace file tools, ensure path doesn't escape
+    if "workspace" in tool_name.lower():
+        path = tool_input.get("path", "") or tool_input.get("file_path", "")
+        if path:
+            # Check for path traversal
+            if ".." in path or path.startswith("/"):
+                logger.warning(
+                    f"Blocked path traversal attempt: {path} by user {user_id}"
+                )
+                return {
+                    "hookSpecificOutput": {
+                        "hookEventName": "PreToolUse",
+                        "permissionDecision": "deny",
+                        "permissionDecisionReason": "Path traversal not allowed",
+                    }
+                }
+
+    return {}
+
+
+def create_security_hooks(
+    user_id: str | None,
+    sdk_cwd: str | None = None,
+    max_subtasks: int = 3,
+    on_stop: Callable[[str, str], None] | None = None,
+) -> dict[str, Any]:
+    """Create the security hooks configuration for Claude Agent SDK.
+
+    Includes security validation and observability hooks:
+    - PreToolUse: Security validation before tool execution
+    - PostToolUse: Log successful tool executions
+    - PostToolUseFailure: Log and handle failed tool executions
+    - PreCompact: Log context compaction events (SDK handles compaction automatically)
+    - Stop: Capture transcript path for stateless resume (when *on_stop* is provided)
+
+    Args:
+        user_id: Current user ID for isolation validation
+        sdk_cwd: SDK working directory for workspace-scoped tool validation
+        max_subtasks: Maximum Task (sub-agent) spawns allowed per session
+        on_stop: Callback ``(transcript_path, sdk_session_id)`` invoked when
+            the SDK finishes processing — used to read the JSONL transcript
+            before the CLI process exits.
+
+    Returns:
+        Hooks configuration dict for ClaudeAgentOptions
+    """
+    try:
+        from claude_agent_sdk import HookMatcher
+        from claude_agent_sdk.types import HookContext, HookInput, SyncHookJSONOutput
+
+        # Per-session counter for Task sub-agent spawns
+        task_spawn_count = 0
+
+        async def pre_tool_use_hook(
+            input_data: HookInput,
+            tool_use_id: str | None,
+            context: HookContext,
+        ) -> SyncHookJSONOutput:
+            """Combined pre-tool-use validation hook."""
+            nonlocal task_spawn_count
+            _ = context  # unused but required by signature
+            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":
+                task_spawn_count += 1
+                if task_spawn_count > max_subtasks:
+                    logger.warning(
+                        f"[SDK] Task limit reached ({max_subtasks}), user={user_id}"
+                    )
+                    return cast(
+                        SyncHookJSONOutput,
+                        _deny(
+                            f"Maximum {max_subtasks} sub-tasks per session. "
+                            "Please continue in the main conversation."
+                        ),
+                    )
+
+            # Strip MCP prefix for consistent validation
+            is_copilot_tool = tool_name.startswith(MCP_TOOL_PREFIX)
+            clean_name = tool_name.removeprefix(MCP_TOOL_PREFIX)
+
+            # Only block non-CoPilot tools; our MCP-registered tools
+            # (including Read for oversized results) are already sandboxed.
+            if not is_copilot_tool:
+                result = _validate_tool_access(clean_name, tool_input, sdk_cwd)
+                if result:
+                    return cast(SyncHookJSONOutput, result)
+
+            # Validate user isolation
+            result = _validate_user_isolation(clean_name, tool_input, user_id)
+            if result:
+                return cast(SyncHookJSONOutput, result)
+
+            logger.debug(f"[SDK] Tool start: {tool_name}, user={user_id}")
+            return cast(SyncHookJSONOutput, {})
+
+        async def post_tool_use_hook(
+            input_data: HookInput,
+            tool_use_id: str | None,
+            context: HookContext,
+        ) -> SyncHookJSONOutput:
+            """Log successful tool executions for observability."""
+            _ = context
+            tool_name = cast(str, input_data.get("tool_name", ""))
+            logger.debug(f"[SDK] Tool success: {tool_name}, tool_use_id={tool_use_id}")
+            return cast(SyncHookJSONOutput, {})
+
+        async def post_tool_failure_hook(
+            input_data: HookInput,
+            tool_use_id: str | None,
+            context: HookContext,
+        ) -> SyncHookJSONOutput:
+            """Log failed tool executions for debugging."""
+            _ = context
+            tool_name = cast(str, input_data.get("tool_name", ""))
+            error = input_data.get("error", "Unknown error")
+            logger.warning(
+                f"[SDK] Tool failed: {tool_name}, error={error}, "
+                f"user={user_id}, tool_use_id={tool_use_id}"
+            )
+            return cast(SyncHookJSONOutput, {})
+
+        async def pre_compact_hook(
+            input_data: HookInput,
+            tool_use_id: str | None,
+            context: HookContext,
+        ) -> SyncHookJSONOutput:
+            """Log when SDK triggers context compaction.
+
+            The SDK automatically compacts conversation history when it grows too large.
+            This hook provides visibility into when compaction happens.
+            """
+            _ = context, tool_use_id
+            trigger = input_data.get("trigger", "auto")
+            logger.info(
+                f"[SDK] Context compaction triggered: {trigger}, user={user_id}"
+            )
+            return cast(SyncHookJSONOutput, {})
+
+        # --- Stop hook: capture transcript path for stateless resume ---
+        async def stop_hook(
+            input_data: HookInput,
+            tool_use_id: str | None,
+            context: HookContext,
+        ) -> SyncHookJSONOutput:
+            """Capture transcript path when SDK finishes processing.
+
+            The Stop hook fires while the CLI process is still alive, giving us
+            a reliable window to read the JSONL transcript before SIGTERM.
+            """
+            _ = context, tool_use_id
+            transcript_path = cast(str, input_data.get("transcript_path", ""))
+            sdk_session_id = cast(str, input_data.get("session_id", ""))
+
+            if transcript_path and on_stop:
+                logger.info(
+                    f"[SDK] Stop hook: transcript_path={transcript_path}, "
+                    f"sdk_session_id={sdk_session_id[:12]}..."
+                )
+                on_stop(transcript_path, sdk_session_id)
+
+            return cast(SyncHookJSONOutput, {})
+
+        hooks: dict[str, Any] = {
+            "PreToolUse": [HookMatcher(matcher="*", hooks=[pre_tool_use_hook])],
+            "PostToolUse": [HookMatcher(matcher="*", hooks=[post_tool_use_hook])],
+            "PostToolUseFailure": [
+                HookMatcher(matcher="*", hooks=[post_tool_failure_hook])
+            ],
+            "PreCompact": [HookMatcher(matcher="*", hooks=[pre_compact_hook])],
+        }
+
+        if on_stop is not None:
+            hooks["Stop"] = [HookMatcher(matcher=None, hooks=[stop_hook])]
+
+        return hooks
+    except ImportError:
+        # Fallback for when SDK isn't available - return empty hooks
+        logger.warning("claude-agent-sdk not available, security hooks disabled")
+        return {}

autogpt_platform/backend/backend/api/features/chat/sdk/security_hooks_test.py

@@ -0,0 +1,165 @@
+"""Unit tests for SDK security hooks."""
+
+import os
+
+from .security_hooks import _validate_tool_access, _validate_user_isolation
+
+SDK_CWD = "/tmp/copilot-abc123"
+
+
+def _is_denied(result: dict) -> bool:
+    hook = result.get("hookSpecificOutput", {})
+    return hook.get("permissionDecision") == "deny"
+
+
+# -- Blocked tools -----------------------------------------------------------
+
+
+def test_blocked_tools_denied():
+    for tool in ("bash", "shell", "exec", "terminal", "command"):
+        result = _validate_tool_access(tool, {})
+        assert _is_denied(result), f"{tool} should be blocked"
+
+
+def test_unknown_tool_allowed():
+    result = _validate_tool_access("SomeCustomTool", {})
+    assert result == {}
+
+
+# -- Workspace-scoped tools --------------------------------------------------
+
+
+def test_read_within_workspace_allowed():
+    result = _validate_tool_access(
+        "Read", {"file_path": f"{SDK_CWD}/file.txt"}, sdk_cwd=SDK_CWD
+    )
+    assert result == {}
+
+
+def test_write_within_workspace_allowed():
+    result = _validate_tool_access(
+        "Write", {"file_path": f"{SDK_CWD}/output.json"}, sdk_cwd=SDK_CWD
+    )
+    assert result == {}
+
+
+def test_edit_within_workspace_allowed():
+    result = _validate_tool_access(
+        "Edit", {"file_path": f"{SDK_CWD}/src/main.py"}, sdk_cwd=SDK_CWD
+    )
+    assert result == {}
+
+
+def test_glob_within_workspace_allowed():
+    result = _validate_tool_access("Glob", {"path": f"{SDK_CWD}/src"}, sdk_cwd=SDK_CWD)
+    assert result == {}
+
+
+def test_grep_within_workspace_allowed():
+    result = _validate_tool_access("Grep", {"path": f"{SDK_CWD}/src"}, sdk_cwd=SDK_CWD)
+    assert result == {}
+
+
+def test_read_outside_workspace_denied():
+    result = _validate_tool_access(
+        "Read", {"file_path": "/etc/passwd"}, sdk_cwd=SDK_CWD
+    )
+    assert _is_denied(result)
+
+
+def test_write_outside_workspace_denied():
+    result = _validate_tool_access(
+        "Write", {"file_path": "/home/user/secrets.txt"}, sdk_cwd=SDK_CWD
+    )
+    assert _is_denied(result)
+
+
+def test_traversal_attack_denied():
+    result = _validate_tool_access(
+        "Read",
+        {"file_path": f"{SDK_CWD}/../../etc/passwd"},
+        sdk_cwd=SDK_CWD,
+    )
+    assert _is_denied(result)
+
+
+def test_no_path_allowed():
+    """Glob/Grep without a path argument defaults to cwd — should pass."""
+    result = _validate_tool_access("Glob", {}, sdk_cwd=SDK_CWD)
+    assert result == {}
+
+
+def test_read_no_cwd_denies_absolute():
+    """If no sdk_cwd is set, absolute paths are denied."""
+    result = _validate_tool_access("Read", {"file_path": "/tmp/anything"})
+    assert _is_denied(result)
+
+
+# -- Tool-results directory --------------------------------------------------
+
+
+def test_read_tool_results_allowed():
+    home = os.path.expanduser("~")
+    path = f"{home}/.claude/projects/-tmp-copilot-abc123/tool-results/12345.txt"
+    result = _validate_tool_access("Read", {"file_path": path}, sdk_cwd=SDK_CWD)
+    assert result == {}
+
+
+def test_read_claude_projects_without_tool_results_denied():
+    home = os.path.expanduser("~")
+    path = f"{home}/.claude/projects/-tmp-copilot-abc123/settings.json"
+    result = _validate_tool_access("Read", {"file_path": path}, sdk_cwd=SDK_CWD)
+    assert _is_denied(result)
+
+
+# -- Built-in Bash is blocked (use bash_exec MCP tool instead) ---------------
+
+
+def test_bash_builtin_always_blocked():
+    """SDK built-in Bash is blocked — bash_exec MCP tool with bubblewrap is used instead."""
+    result = _validate_tool_access("Bash", {"command": "echo hello"}, sdk_cwd=SDK_CWD)
+    assert _is_denied(result)
+
+
+# -- Dangerous patterns ------------------------------------------------------
+
+
+def test_dangerous_pattern_blocked():
+    result = _validate_tool_access("SomeTool", {"cmd": "sudo rm -rf /"})
+    assert _is_denied(result)
+
+
+def test_subprocess_pattern_blocked():
+    result = _validate_tool_access("SomeTool", {"code": "subprocess.run(...)"})
+    assert _is_denied(result)
+
+
+# -- User isolation ----------------------------------------------------------
+
+
+def test_workspace_path_traversal_blocked():
+    result = _validate_user_isolation(
+        "workspace_read", {"path": "../../../etc/shadow"}, user_id="user-1"
+    )
+    assert _is_denied(result)
+
+
+def test_workspace_absolute_path_blocked():
+    result = _validate_user_isolation(
+        "workspace_read", {"path": "/etc/passwd"}, user_id="user-1"
+    )
+    assert _is_denied(result)
+
+
+def test_workspace_normal_path_allowed():
+    result = _validate_user_isolation(
+        "workspace_read", {"path": "src/main.py"}, user_id="user-1"
+    )
+    assert result == {}
+
+
+def test_non_workspace_tool_passes_isolation():
+    result = _validate_user_isolation(
+        "find_agent", {"query": "email"}, user_id="user-1"
+    )
+    assert result == {}

autogpt_platform/backend/backend/api/features/chat/sdk/service.py

@@ -0,0 +1,751 @@
+"""Claude Agent SDK service layer for CoPilot chat completions."""
+
+import asyncio
+import json
+import logging
+import os
+import uuid
+from collections.abc import AsyncGenerator
+from dataclasses import dataclass
+from typing import Any
+
+from backend.util.exceptions import NotFoundError
+
+from .. import stream_registry
+from ..config import ChatConfig
+from ..model import (
+    ChatMessage,
+    ChatSession,
+    get_chat_session,
+    update_session_title,
+    upsert_chat_session,
+)
+from ..response_model import (
+    StreamBaseResponse,
+    StreamError,
+    StreamFinish,
+    StreamStart,
+    StreamTextDelta,
+    StreamToolInputAvailable,
+    StreamToolOutputAvailable,
+)
+from ..service import (
+    _build_system_prompt,
+    _execute_long_running_tool_with_streaming,
+    _generate_session_title,
+)
+from ..tools.models import OperationPendingResponse, OperationStartedResponse
+from ..tools.sandbox import WORKSPACE_PREFIX, make_session_path
+from ..tracking import track_user_message
+from .response_adapter import SDKResponseAdapter
+from .security_hooks import create_security_hooks
+from .tool_adapter import (
+    COPILOT_TOOL_NAMES,
+    LongRunningCallback,
+    create_copilot_mcp_server,
+    set_execution_context,
+)
+from .transcript import (
+    download_transcript,
+    read_transcript_file,
+    upload_transcript,
+    validate_transcript,
+    write_transcript_to_tempfile,
+)
+
+logger = logging.getLogger(__name__)
+config = ChatConfig()
+
+# Set to hold background tasks to prevent garbage collection
+_background_tasks: set[asyncio.Task[Any]] = set()
+
+
+@dataclass
+class CapturedTranscript:
+    """Info captured by the SDK Stop hook for stateless --resume."""
+
+    path: str = ""
+    sdk_session_id: str = ""
+
+    @property
+    def available(self) -> bool:
+        return bool(self.path)
+
+
+_SDK_CWD_PREFIX = WORKSPACE_PREFIX
+
+# Appended to the system prompt to inform the agent about available tools.
+# The SDK built-in Bash is NOT available — use mcp__copilot__bash_exec instead,
+# which has kernel-level network isolation (unshare --net).
+_SDK_TOOL_SUPPLEMENT = """
+
+## Tool notes
+
+- The SDK built-in Bash tool is NOT available.  Use the `bash_exec` MCP tool
+  for shell commands — it runs in a network-isolated sandbox.
+- **Shared workspace**: The SDK Read/Write tools and `bash_exec` share the
+  same working directory. Files created by one are readable by the other.
+  These files are **ephemeral** — they exist only for the current session.
+- **Persistent storage**: Use `write_workspace_file` / `read_workspace_file`
+  for files that should persist across sessions (stored in cloud storage).
+- Long-running tools (create_agent, edit_agent, etc.) are handled
+  asynchronously.  You will receive an immediate response; the actual result
+  is delivered to the user via a background stream.
+"""
+
+
+def _build_long_running_callback(user_id: str | None) -> LongRunningCallback:
+    """Build a callback that delegates long-running tools to the non-SDK infrastructure.
+
+    Long-running tools (create_agent, edit_agent, etc.) are delegated to the
+    existing background infrastructure: stream_registry (Redis Streams),
+    database persistence, and SSE reconnection.  This means results survive
+    page refreshes / pod restarts, and the frontend shows the proper loading
+    widget with progress updates.
+
+    The returned callback matches the ``LongRunningCallback`` signature:
+    ``(tool_name, args, session) -> MCP response dict``.
+    """
+
+    async def _callback(
+        tool_name: str, args: dict[str, Any], session: ChatSession
+    ) -> dict[str, Any]:
+        operation_id = str(uuid.uuid4())
+        task_id = str(uuid.uuid4())
+        tool_call_id = f"sdk-{uuid.uuid4().hex[:12]}"
+        session_id = session.session_id
+
+        # --- Build user-friendly messages (matches non-SDK service) ---
+        if tool_name == "create_agent":
+            desc = args.get("description", "")
+            desc_preview = (desc[:100] + "...") if len(desc) > 100 else desc
+            pending_msg = (
+                f"Creating your agent: {desc_preview}"
+                if desc_preview
+                else "Creating agent... This may take a few minutes."
+            )
+            started_msg = (
+                "Agent creation started. You can close this tab - "
+                "check your library in a few minutes."
+            )
+        elif tool_name == "edit_agent":
+            changes = args.get("changes", "")
+            changes_preview = (changes[:100] + "...") if len(changes) > 100 else changes
+            pending_msg = (
+                f"Editing agent: {changes_preview}"
+                if changes_preview
+                else "Editing agent... This may take a few minutes."
+            )
+            started_msg = (
+                "Agent edit started. You can close this tab - "
+                "check your library in a few minutes."
+            )
+        else:
+            pending_msg = f"Running {tool_name}... This may take a few minutes."
+            started_msg = (
+                f"{tool_name} started. You can close this tab - "
+                "check back in a few minutes."
+            )
+
+        # --- Register task in Redis for SSE reconnection ---
+        await stream_registry.create_task(
+            task_id=task_id,
+            session_id=session_id,
+            user_id=user_id,
+            tool_call_id=tool_call_id,
+            tool_name=tool_name,
+            operation_id=operation_id,
+        )
+
+        # --- Save OperationPendingResponse to chat history ---
+        pending_message = ChatMessage(
+            role="tool",
+            content=OperationPendingResponse(
+                message=pending_msg,
+                operation_id=operation_id,
+                tool_name=tool_name,
+            ).model_dump_json(),
+            tool_call_id=tool_call_id,
+        )
+        session.messages.append(pending_message)
+        await upsert_chat_session(session)
+
+        # --- Spawn background task (reuses non-SDK infrastructure) ---
+        bg_task = asyncio.create_task(
+            _execute_long_running_tool_with_streaming(
+                tool_name=tool_name,
+                parameters=args,
+                tool_call_id=tool_call_id,
+                operation_id=operation_id,
+                task_id=task_id,
+                session_id=session_id,
+                user_id=user_id,
+            )
+        )
+        _background_tasks.add(bg_task)
+        bg_task.add_done_callback(_background_tasks.discard)
+        await stream_registry.set_task_asyncio_task(task_id, bg_task)
+
+        logger.info(
+            f"[SDK] Long-running tool {tool_name} delegated to background "
+            f"(operation_id={operation_id}, task_id={task_id})"
+        )
+
+        # --- Return OperationStartedResponse as MCP tool result ---
+        # This flows through SDK → response adapter → frontend, triggering
+        # the loading widget with SSE reconnection support.
+        started_json = OperationStartedResponse(
+            message=started_msg,
+            operation_id=operation_id,
+            tool_name=tool_name,
+            task_id=task_id,
+        ).model_dump_json()
+
+        return {
+            "content": [{"type": "text", "text": started_json}],
+            "isError": False,
+        }
+
+    return _callback
+
+
+def _resolve_sdk_model() -> str | None:
+    """Resolve the model name for the Claude Agent SDK CLI.
+
+    Uses ``config.claude_agent_model`` if set, otherwise derives from
+    ``config.model`` by stripping the OpenRouter provider prefix (e.g.,
+    ``"anthropic/claude-opus-4.6"`` → ``"claude-opus-4.6"``).
+    """
+    if config.claude_agent_model:
+        return config.claude_agent_model
+    model = config.model
+    if "/" in model:
+        return model.split("/", 1)[1]
+    return model
+
+
+def _build_sdk_env() -> dict[str, str]:
+    """Build env vars for the SDK CLI process.
+
+    Routes API calls through OpenRouter (or a custom base_url) using
+    the same ``config.api_key`` / ``config.base_url`` as the non-SDK path.
+    This gives per-call token and cost tracking on the OpenRouter dashboard.
+
+    Only overrides ``ANTHROPIC_API_KEY`` when a valid proxy URL and auth
+    token are both present — otherwise returns an empty dict so the SDK
+    falls back to its default credentials.
+    """
+    env: dict[str, str] = {}
+    if config.api_key and config.base_url:
+        # Strip /v1 suffix — SDK expects the base URL without a version path
+        base = config.base_url.rstrip("/")
+        if base.endswith("/v1"):
+            base = base[:-3]
+        if not base or not base.startswith("http"):
+            # Invalid base_url — don't override SDK defaults
+            return env
+        env["ANTHROPIC_BASE_URL"] = base
+        env["ANTHROPIC_AUTH_TOKEN"] = config.api_key
+        # Must be explicitly empty so the CLI uses AUTH_TOKEN instead
+        env["ANTHROPIC_API_KEY"] = ""
+    return env
+
+
+def _make_sdk_cwd(session_id: str) -> str:
+    """Create a safe, session-specific working directory path.
+
+    Delegates to :func:`~backend.api.features.chat.tools.sandbox.make_session_path`
+    (single source of truth for path sanitization) and adds a defence-in-depth
+    assertion.
+    """
+    cwd = make_session_path(session_id)
+    # Defence-in-depth: normpath + startswith is a CodeQL-recognised sanitizer
+    cwd = os.path.normpath(cwd)
+    if not cwd.startswith(_SDK_CWD_PREFIX):
+        raise ValueError(f"SDK cwd escaped prefix: {cwd}")
+    return cwd
+
+
+def _cleanup_sdk_tool_results(cwd: str) -> None:
+    """Remove SDK tool-result files for a specific session working directory.
+
+    The SDK creates tool-result files under ~/.claude/projects/<encoded-cwd>/tool-results/.
+    We clean only the specific cwd's results to avoid race conditions between
+    concurrent sessions.
+
+    Security: cwd MUST be created by _make_sdk_cwd() which sanitizes session_id.
+    """
+    import shutil
+
+    # Validate cwd is under the expected prefix
+    normalized = os.path.normpath(cwd)
+    if not normalized.startswith(_SDK_CWD_PREFIX):
+        logger.warning(f"[SDK] Rejecting cleanup for path outside workspace: {cwd}")
+        return
+
+    # SDK encodes the cwd path by replacing '/' with '-'
+    encoded_cwd = normalized.replace("/", "-")
+
+    # Construct the project directory path (known-safe home expansion)
+    claude_projects = os.path.expanduser("~/.claude/projects")
+    project_dir = os.path.join(claude_projects, encoded_cwd)
+
+    # Security check 3: Validate project_dir is under ~/.claude/projects
+    project_dir = os.path.normpath(project_dir)
+    if not project_dir.startswith(claude_projects):
+        logger.warning(
+            f"[SDK] Rejecting cleanup for escaped project path: {project_dir}"
+        )
+        return
+
+    results_dir = os.path.join(project_dir, "tool-results")
+    if os.path.isdir(results_dir):
+        for filename in os.listdir(results_dir):
+            file_path = os.path.join(results_dir, filename)
+            try:
+                if os.path.isfile(file_path):
+                    os.remove(file_path)
+            except OSError:
+                pass
+
+    # Also clean up the temp cwd directory itself
+    try:
+        shutil.rmtree(normalized, ignore_errors=True)
+    except OSError:
+        pass
+
+
+async def _compress_conversation_history(
+    session: ChatSession,
+) -> list[ChatMessage]:
+    """Compress prior conversation messages if they exceed the token threshold.
+
+    Uses the shared compress_context() from prompt.py which supports:
+    - LLM summarization of old messages (keeps recent ones intact)
+    - Progressive content truncation as fallback
+    - Middle-out deletion as last resort
+
+    Returns the compressed prior messages (everything except the current message).
+    """
+    prior = session.messages[:-1]
+    if len(prior) < 2:
+        return prior
+
+    from backend.util.prompt import compress_context
+
+    # Convert ChatMessages to dicts for compress_context
+    messages_dict = []
+    for msg in prior:
+        msg_dict: dict[str, Any] = {"role": msg.role}
+        if msg.content:
+            msg_dict["content"] = msg.content
+        if msg.tool_calls:
+            msg_dict["tool_calls"] = msg.tool_calls
+        if msg.tool_call_id:
+            msg_dict["tool_call_id"] = msg.tool_call_id
+        messages_dict.append(msg_dict)
+
+    try:
+        import openai
+
+        async with openai.AsyncOpenAI(
+            api_key=config.api_key, base_url=config.base_url, timeout=30.0
+        ) as client:
+            result = await compress_context(
+                messages=messages_dict,
+                model=config.model,
+                client=client,
+            )
+    except Exception as e:
+        logger.warning(f"[SDK] Context compression with LLM failed: {e}")
+        # Fall back to truncation-only (no LLM summarization)
+        result = await compress_context(
+            messages=messages_dict,
+            model=config.model,
+            client=None,
+        )
+
+    if result.was_compacted:
+        logger.info(
+            f"[SDK] Context compacted: {result.original_token_count} -> "
+            f"{result.token_count} tokens "
+            f"({result.messages_summarized} summarized, "
+            f"{result.messages_dropped} dropped)"
+        )
+        # Convert compressed dicts back to ChatMessages
+        return [
+            ChatMessage(
+                role=m["role"],
+                content=m.get("content"),
+                tool_calls=m.get("tool_calls"),
+                tool_call_id=m.get("tool_call_id"),
+            )
+            for m in result.messages
+        ]
+
+    return prior
+
+
+def _format_conversation_context(messages: list[ChatMessage]) -> str | None:
+    """Format conversation messages into a context prefix for the user message.
+
+    Returns a string like:
+        <conversation_history>
+        User: hello
+        You responded: Hi! How can I help?
+        </conversation_history>
+
+    Returns None if there are no messages to format.
+    """
+    if not messages:
+        return None
+
+    lines: list[str] = []
+    for msg in messages:
+        if not msg.content:
+            continue
+        if msg.role == "user":
+            lines.append(f"User: {msg.content}")
+        elif msg.role == "assistant":
+            lines.append(f"You responded: {msg.content}")
+        # Skip tool messages — they're internal details
+
+    if not lines:
+        return None
+
+    return "<conversation_history>\n" + "\n".join(lines) + "\n</conversation_history>"
+
+
+async def stream_chat_completion_sdk(
+    session_id: str,
+    message: str | None = None,
+    tool_call_response: str | None = None,  # noqa: ARG001
+    is_user_message: bool = True,
+    user_id: str | None = None,
+    retry_count: int = 0,  # noqa: ARG001
+    session: ChatSession | None = None,
+    context: dict[str, str] | None = None,  # noqa: ARG001
+) -> AsyncGenerator[StreamBaseResponse, None]:
+    """Stream chat completion using Claude Agent SDK.
+
+    Drop-in replacement for stream_chat_completion with improved reliability.
+    """
+
+    if session is None:
+        session = await get_chat_session(session_id, user_id)
+
+    if not session:
+        raise NotFoundError(
+            f"Session {session_id} not found. Please create a new session first."
+        )
+
+    if message:
+        session.messages.append(
+            ChatMessage(
+                role="user" if is_user_message else "assistant", content=message
+            )
+        )
+        if is_user_message:
+            track_user_message(
+                user_id=user_id, session_id=session_id, message_length=len(message)
+            )
+
+    session = await upsert_chat_session(session)
+
+    # Generate title for new sessions (first user message)
+    if is_user_message and not session.title:
+        user_messages = [m for m in session.messages if m.role == "user"]
+        if len(user_messages) == 1:
+            first_message = user_messages[0].content or message or ""
+            if first_message:
+                task = asyncio.create_task(
+                    _update_title_async(session_id, first_message, user_id)
+                )
+                _background_tasks.add(task)
+                task.add_done_callback(_background_tasks.discard)
+
+    # Build system prompt (reuses non-SDK path with Langfuse support)
+    has_history = len(session.messages) > 1
+    system_prompt, _ = await _build_system_prompt(
+        user_id, has_conversation_history=has_history
+    )
+    system_prompt += _SDK_TOOL_SUPPLEMENT
+    message_id = str(uuid.uuid4())
+    task_id = str(uuid.uuid4())
+
+    yield StreamStart(messageId=message_id, taskId=task_id)
+
+    stream_completed = False
+    # Initialise sdk_cwd before the try so the finally can reference it
+    # even if _make_sdk_cwd raises (in that case it stays as "").
+    sdk_cwd = ""
+    use_resume = False
+
+    try:
+        # Use a session-specific temp dir to avoid cleanup race conditions
+        # between concurrent sessions.
+        sdk_cwd = _make_sdk_cwd(session_id)
+        os.makedirs(sdk_cwd, exist_ok=True)
+
+        set_execution_context(
+            user_id,
+            session,
+            long_running_callback=_build_long_running_callback(user_id),
+        )
+        try:
+            from claude_agent_sdk import ClaudeAgentOptions, ClaudeSDKClient
+
+            # Fail fast when no API credentials are available at all
+            sdk_env = _build_sdk_env()
+            if not sdk_env and not os.environ.get("ANTHROPIC_API_KEY"):
+                raise RuntimeError(
+                    "No API key configured. Set OPEN_ROUTER_API_KEY "
+                    "(or CHAT_API_KEY) for OpenRouter routing, "
+                    "or ANTHROPIC_API_KEY for direct Anthropic access."
+                )
+
+            mcp_server = create_copilot_mcp_server()
+
+            sdk_model = _resolve_sdk_model()
+
+            # --- Transcript capture via Stop hook ---
+            captured_transcript = CapturedTranscript()
+
+            def _on_stop(transcript_path: str, sdk_session_id: str) -> None:
+                captured_transcript.path = transcript_path
+                captured_transcript.sdk_session_id = sdk_session_id
+
+            security_hooks = create_security_hooks(
+                user_id,
+                sdk_cwd=sdk_cwd,
+                max_subtasks=config.claude_agent_max_subtasks,
+                on_stop=_on_stop if config.claude_agent_use_resume else None,
+            )
+
+            # --- Resume strategy: download transcript from bucket ---
+            resume_file: str | None = None
+            use_resume = False
+
+            if config.claude_agent_use_resume and user_id and len(session.messages) > 1:
+                transcript_content = await download_transcript(user_id, session_id)
+                if transcript_content and validate_transcript(transcript_content):
+                    resume_file = write_transcript_to_tempfile(
+                        transcript_content, session_id, sdk_cwd
+                    )
+                    if resume_file:
+                        use_resume = True
+                        logger.info(
+                            f"[SDK] Using --resume with transcript "
+                            f"({len(transcript_content)} bytes)"
+                        )
+
+            sdk_options_kwargs: dict[str, Any] = {
+                "system_prompt": system_prompt,
+                "mcp_servers": {"copilot": mcp_server},
+                "allowed_tools": COPILOT_TOOL_NAMES,
+                "disallowed_tools": ["Bash"],
+                "hooks": security_hooks,
+                "cwd": sdk_cwd,
+                "max_buffer_size": config.claude_agent_max_buffer_size,
+            }
+            if sdk_env:
+                sdk_options_kwargs["model"] = sdk_model
+                sdk_options_kwargs["env"] = sdk_env
+            if use_resume and resume_file:
+                sdk_options_kwargs["resume"] = resume_file
+
+            options = ClaudeAgentOptions(**sdk_options_kwargs)  # type: ignore[arg-type]
+
+            adapter = SDKResponseAdapter(message_id=message_id)
+            adapter.set_task_id(task_id)
+
+            async with ClaudeSDKClient(options=options) as client:
+                current_message = message or ""
+                if not current_message and session.messages:
+                    last_user = [m for m in session.messages if m.role == "user"]
+                    if last_user:
+                        current_message = last_user[-1].content or ""
+
+                if not current_message.strip():
+                    yield StreamError(
+                        errorText="Message cannot be empty.",
+                        code="empty_prompt",
+                    )
+                    yield StreamFinish()
+                    return
+
+                # Build query: with --resume the CLI already has full
+                # context, so we only send the new message.  Without
+                # resume, compress history into a context prefix.
+                query_message = current_message
+                if not use_resume and len(session.messages) > 1:
+                    logger.warning(
+                        f"[SDK] Using compression fallback for session "
+                        f"{session_id} ({len(session.messages)} messages) — "
+                        f"no transcript available for --resume"
+                    )
+                    compressed = await _compress_conversation_history(session)
+                    history_context = _format_conversation_context(compressed)
+                    if history_context:
+                        query_message = (
+                            f"{history_context}\n\n"
+                            f"Now, the user says:\n{current_message}"
+                        )
+
+                logger.info(
+                    f"[SDK] Sending query ({len(session.messages)} msgs in session)"
+                )
+                logger.debug(f"[SDK] Query preview: {current_message[:80]!r}")
+                await client.query(query_message, session_id=session_id)
+
+                assistant_response = ChatMessage(role="assistant", content="")
+                accumulated_tool_calls: list[dict[str, Any]] = []
+                has_appended_assistant = False
+                has_tool_results = False
+
+                async for sdk_msg in client.receive_messages():
+                    logger.debug(
+                        f"[SDK] Received: {type(sdk_msg).__name__} "
+                        f"{getattr(sdk_msg, 'subtype', '')}"
+                    )
+                    for response in adapter.convert_message(sdk_msg):
+                        if isinstance(response, StreamStart):
+                            continue
+
+                        yield response
+
+                        if isinstance(response, StreamTextDelta):
+                            delta = response.delta or ""
+                            # After tool results, start a new assistant
+                            # message for the post-tool text.
+                            if has_tool_results and has_appended_assistant:
+                                assistant_response = ChatMessage(
+                                    role="assistant", content=delta
+                                )
+                                accumulated_tool_calls = []
+                                has_appended_assistant = False
+                                has_tool_results = False
+                                session.messages.append(assistant_response)
+                                has_appended_assistant = True
+                            else:
+                                assistant_response.content = (
+                                    assistant_response.content or ""
+                                ) + delta
+                                if not has_appended_assistant:
+                                    session.messages.append(assistant_response)
+                                    has_appended_assistant = True
+
+                        elif isinstance(response, StreamToolInputAvailable):
+                            accumulated_tool_calls.append(
+                                {
+                                    "id": response.toolCallId,
+                                    "type": "function",
+                                    "function": {
+                                        "name": response.toolName,
+                                        "arguments": json.dumps(response.input or {}),
+                                    },
+                                }
+                            )
+                            assistant_response.tool_calls = accumulated_tool_calls
+                            if not has_appended_assistant:
+                                session.messages.append(assistant_response)
+                                has_appended_assistant = True
+
+                        elif isinstance(response, StreamToolOutputAvailable):
+                            session.messages.append(
+                                ChatMessage(
+                                    role="tool",
+                                    content=(
+                                        response.output
+                                        if isinstance(response.output, str)
+                                        else str(response.output)
+                                    ),
+                                    tool_call_id=response.toolCallId,
+                                )
+                            )
+                            has_tool_results = True
+
+                        elif isinstance(response, StreamFinish):
+                            stream_completed = True
+
+                    if stream_completed:
+                        break
+
+                if (
+                    assistant_response.content or assistant_response.tool_calls
+                ) and not has_appended_assistant:
+                    session.messages.append(assistant_response)
+
+                # --- Capture transcript while CLI is still alive ---
+                # Must happen INSIDE async with: close() sends SIGTERM
+                # which kills the CLI before it can flush the JSONL.
+                if (
+                    config.claude_agent_use_resume
+                    and user_id
+                    and captured_transcript.available
+                ):
+                    # Give CLI time to flush JSONL writes before we read
+                    await asyncio.sleep(0.5)
+                    raw_transcript = read_transcript_file(captured_transcript.path)
+                    if raw_transcript:
+                        task = asyncio.create_task(
+                            _upload_transcript_bg(user_id, session_id, raw_transcript)
+                        )
+                        _background_tasks.add(task)
+                        task.add_done_callback(_background_tasks.discard)
+                    else:
+                        logger.debug("[SDK] Stop hook fired but transcript not usable")
+
+        except ImportError:
+            raise RuntimeError(
+                "claude-agent-sdk is not installed. "
+                "Disable SDK mode (CHAT_USE_CLAUDE_AGENT_SDK=false) "
+                "to use the OpenAI-compatible fallback."
+            )
+
+        await upsert_chat_session(session)
+        logger.debug(
+            f"[SDK] Session {session_id} saved with {len(session.messages)} messages"
+        )
+        if not stream_completed:
+            yield StreamFinish()
+
+    except Exception as e:
+        logger.error(f"[SDK] Error: {e}", exc_info=True)
+        try:
+            await upsert_chat_session(session)
+        except Exception as save_err:
+            logger.error(f"[SDK] Failed to save session on error: {save_err}")
+        yield StreamError(
+            errorText="An error occurred. Please try again.",
+            code="sdk_error",
+        )
+        yield StreamFinish()
+    finally:
+        if sdk_cwd:
+            _cleanup_sdk_tool_results(sdk_cwd)
+
+
+async def _upload_transcript_bg(
+    user_id: str, session_id: str, raw_content: str
+) -> None:
+    """Background task to strip progress entries and upload transcript."""
+    try:
+        await upload_transcript(user_id, session_id, raw_content)
+    except Exception as e:
+        logger.error(f"[SDK] Failed to upload transcript for {session_id}: {e}")
+
+
+async def _update_title_async(
+    session_id: str, message: str, user_id: str | None = None
+) -> None:
+    """Background task to update session title."""
+    try:
+        title = await _generate_session_title(
+            message, user_id=user_id, session_id=session_id
+        )
+        if title:
+            await update_session_title(session_id, title)
+            logger.debug(f"[SDK] Generated title for {session_id}: {title}")
+    except Exception as e:
+        logger.warning(f"[SDK] Failed to update session title: {e}")

autogpt_platform/backend/backend/api/features/chat/sdk/tool_adapter.py

@@ -0,0 +1,325 @@
+"""Tool adapter for wrapping existing CoPilot tools as Claude Agent SDK MCP tools.
+
+This module provides the adapter layer that converts existing BaseTool implementations
+into in-process MCP tools that can be used with the Claude Agent SDK.
+
+Long-running tools (``is_long_running=True``) are delegated to the non-SDK
+background infrastructure (stream_registry, Redis persistence, SSE reconnection)
+via a callback provided by the service layer.  This avoids wasteful SDK polling
+and makes results survive page refreshes.
+"""
+
+import json
+import logging
+import os
+import uuid
+from collections.abc import Awaitable, Callable
+from contextvars import ContextVar
+from typing import Any
+
+from backend.api.features.chat.model import ChatSession
+from backend.api.features.chat.tools import TOOL_REGISTRY
+from backend.api.features.chat.tools.base import BaseTool
+
+logger = logging.getLogger(__name__)
+
+# Allowed base directory for the Read tool (SDK saves oversized tool results here).
+# Restricted to ~/.claude/projects/ and further validated to require "tool-results"
+# in the path — prevents reading settings, credentials, or other sensitive files.
+_SDK_PROJECTS_DIR = os.path.expanduser("~/.claude/projects/")
+
+# MCP server naming - the SDK prefixes tool names as "mcp__{server_name}__{tool}"
+MCP_SERVER_NAME = "copilot"
+MCP_TOOL_PREFIX = f"mcp__{MCP_SERVER_NAME}__"
+
+# Context variables to pass user/session info to tool execution
+_current_user_id: ContextVar[str | None] = ContextVar("current_user_id", default=None)
+_current_session: ContextVar[ChatSession | None] = ContextVar(
+    "current_session", 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.
+_pending_tool_outputs: ContextVar[dict[str, str]] = ContextVar(
+    "pending_tool_outputs", default=None  # type: ignore[arg-type]
+)
+
+# Callback type for delegating long-running tools to the non-SDK infrastructure.
+# Args: (tool_name, arguments, session) → MCP-formatted response dict.
+LongRunningCallback = Callable[
+    [str, dict[str, Any], ChatSession], Awaitable[dict[str, Any]]
+]
+
+# ContextVar so the service layer can inject the callback per-request.
+_long_running_callback: ContextVar[LongRunningCallback | None] = ContextVar(
+    "long_running_callback", default=None
+)
+
+
+def set_execution_context(
+    user_id: str | None,
+    session: ChatSession,
+    long_running_callback: LongRunningCallback | None = None,
+) -> None:
+    """Set the execution context for tool calls.
+
+    This must be called before streaming begins to ensure tools have access
+    to user_id and session information.
+
+    Args:
+        user_id: Current user's ID.
+        session: Current chat session.
+        long_running_callback: Optional callback to delegate long-running tools
+            to the non-SDK background infrastructure (stream_registry + Redis).
+    """
+    _current_user_id.set(user_id)
+    _current_session.set(session)
+    _pending_tool_outputs.set({})
+    _long_running_callback.set(long_running_callback)
+
+
+def get_execution_context() -> tuple[str | None, ChatSession | None]:
+    """Get the current execution context."""
+    return (
+        _current_user_id.get(),
+        _current_session.get(),
+    )
+
+
+def pop_pending_tool_output(tool_name: str) -> str | None:
+    """Pop and return the stashed full output for *tool_name*.
+
+    The SDK CLI may truncate large tool results (writing them to disk and
+    replacing the content with a file reference). This stash keeps the
+    original MCP output so the response adapter can forward it to the
+    frontend for proper widget rendering.
+
+    Returns ``None`` if nothing was stashed for *tool_name*.
+    """
+    pending = _pending_tool_outputs.get(None)
+    if pending is None:
+        return None
+    return pending.pop(tool_name, None)
+
+
+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."""
+    effective_id = f"sdk-{uuid.uuid4().hex[:12]}"
+    result = await base_tool.execute(
+        user_id=user_id,
+        session=session,
+        tool_call_id=effective_id,
+        **args,
+    )
+
+    text = (
+        result.output if isinstance(result.output, str) else json.dumps(result.output)
+    )
+
+    # Stash the full output before the SDK potentially truncates it.
+    pending = _pending_tool_outputs.get(None)
+    if pending is not None:
+        pending[base_tool.name] = text
+
+    return {
+        "content": [{"type": "text", "text": text}],
+        "isError": not result.success,
+    }
+
+
+def _mcp_error(message: str) -> dict[str, Any]:
+    return {
+        "content": [
+            {"type": "text", "text": json.dumps({"error": message, "type": "error"})}
+        ],
+        "isError": True,
+    }
+
+
+def create_tool_handler(base_tool: BaseTool):
+    """Create an async handler function for a BaseTool.
+
+    This wraps the existing BaseTool._execute method to be compatible
+    with the Claude Agent SDK MCP tool format.
+
+    Long-running tools (``is_long_running=True``) are delegated to the
+    non-SDK background infrastructure via a callback set in the execution
+    context.  The callback persists the operation in Redis (stream_registry)
+    so results survive page refreshes and pod restarts.
+    """
+
+    async def tool_handler(args: dict[str, Any]) -> dict[str, Any]:
+        """Execute the wrapped tool and return MCP-formatted response."""
+        user_id, session = get_execution_context()
+
+        if session is None:
+            return _mcp_error("No session context available")
+
+        # --- Long-running: delegate to non-SDK background infrastructure ---
+        if base_tool.is_long_running:
+            callback = _long_running_callback.get(None)
+            if callback:
+                try:
+                    return await callback(base_tool.name, args, session)
+                except Exception as e:
+                    logger.error(
+                        f"Long-running callback failed for {base_tool.name}: {e}",
+                        exc_info=True,
+                    )
+                    return _mcp_error(f"Failed to start {base_tool.name}: {e}")
+            # No callback — fall through to synchronous execution
+            logger.warning(
+                f"[SDK] No long-running callback for {base_tool.name}, "
+                f"executing synchronously (may block)"
+            )
+
+        # --- Normal (fast) tool: execute synchronously ---
+        try:
+            return await _execute_tool_sync(base_tool, user_id, session, args)
+        except Exception as e:
+            logger.error(f"Error executing tool {base_tool.name}: {e}", exc_info=True)
+            return _mcp_error(f"Failed to execute {base_tool.name}: {e}")
+
+    return tool_handler
+
+
+def _build_input_schema(base_tool: BaseTool) -> dict[str, Any]:
+    """Build a JSON Schema input schema for a tool."""
+    return {
+        "type": "object",
+        "properties": base_tool.parameters.get("properties", {}),
+        "required": base_tool.parameters.get("required", []),
+    }
+
+
+async def _read_file_handler(args: dict[str, Any]) -> dict[str, Any]:
+    """Read a file with optional offset/limit. Restricted to SDK working directory.
+
+    After reading, the file is deleted to prevent accumulation in long-running pods.
+    """
+    file_path = args.get("file_path", "")
+    offset = args.get("offset", 0)
+    limit = args.get("limit", 2000)
+
+    # Security: only allow reads under ~/.claude/projects/**/tool-results/
+    real_path = os.path.realpath(file_path)
+    if not real_path.startswith(_SDK_PROJECTS_DIR) or "tool-results" not in real_path:
+        return {
+            "content": [{"type": "text", "text": f"Access denied: {file_path}"}],
+            "isError": True,
+        }
+
+    try:
+        with open(real_path) as f:
+            lines = f.readlines()
+        selected = lines[offset : offset + limit]
+        content = "".join(selected)
+        # Clean up to prevent accumulation in long-running pods
+        try:
+            os.remove(real_path)
+        except OSError:
+            pass
+        return {"content": [{"type": "text", "text": content}], "isError": False}
+    except FileNotFoundError:
+        return {
+            "content": [{"type": "text", "text": f"File not found: {file_path}"}],
+            "isError": True,
+        }
+    except Exception as e:
+        return {
+            "content": [{"type": "text", "text": f"Error reading file: {e}"}],
+            "isError": True,
+        }
+
+
+_READ_TOOL_NAME = "Read"
+_READ_TOOL_DESCRIPTION = (
+    "Read a file from the local filesystem. "
+    "Use offset and limit to read specific line ranges for large files."
+)
+_READ_TOOL_SCHEMA = {
+    "type": "object",
+    "properties": {
+        "file_path": {
+            "type": "string",
+            "description": "The absolute path to the file to read",
+        },
+        "offset": {
+            "type": "integer",
+            "description": "Line number to start reading from (0-indexed). Default: 0",
+        },
+        "limit": {
+            "type": "integer",
+            "description": "Number of lines to read. Default: 2000",
+        },
+    },
+    "required": ["file_path"],
+}
+
+
+# Create the MCP server configuration
+def create_copilot_mcp_server():
+    """Create an in-process MCP server configuration for CoPilot tools.
+
+    This can be passed to ClaudeAgentOptions.mcp_servers.
+
+    Note: The actual SDK MCP server creation depends on the claude-agent-sdk
+    package being available. This function returns the configuration that
+    can be used with the SDK.
+    """
+    try:
+        from claude_agent_sdk import create_sdk_mcp_server, tool
+
+        # Create decorated tool functions
+        sdk_tools = []
+
+        for tool_name, base_tool in TOOL_REGISTRY.items():
+            handler = create_tool_handler(base_tool)
+            decorated = tool(
+                tool_name,
+                base_tool.description,
+                _build_input_schema(base_tool),
+            )(handler)
+            sdk_tools.append(decorated)
+
+        # Add the Read tool so the SDK can read back oversized tool results
+        read_tool = tool(
+            _READ_TOOL_NAME,
+            _READ_TOOL_DESCRIPTION,
+            _READ_TOOL_SCHEMA,
+        )(_read_file_handler)
+        sdk_tools.append(read_tool)
+
+        server = create_sdk_mcp_server(
+            name=MCP_SERVER_NAME,
+            version="1.0.0",
+            tools=sdk_tools,
+        )
+
+        return server
+
+    except ImportError:
+        # Let ImportError propagate so service.py handles the fallback
+        raise
+
+
+# SDK built-in tools allowed within the workspace directory.
+# 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).
+_SDK_BUILTIN_TOOLS = ["Read", "Write", "Edit", "Glob", "Grep", "Task"]
+
+# List of tool names for allowed_tools configuration
+# Include MCP tools, the MCP Read tool for oversized results,
+# and SDK built-in file tools for workspace operations.
+COPILOT_TOOL_NAMES = [
+    *[f"{MCP_TOOL_PREFIX}{name}" for name in TOOL_REGISTRY.keys()],
+    f"{MCP_TOOL_PREFIX}{_READ_TOOL_NAME}",
+    *_SDK_BUILTIN_TOOLS,
+]

autogpt_platform/backend/backend/api/features/chat/sdk/transcript.py

@@ -0,0 +1,355 @@
+"""JSONL transcript management for stateless multi-turn resume.
+
+The Claude Code CLI persists conversations as JSONL files (one JSON object per
+line).  When the SDK's ``Stop`` hook fires we read this file, strip bloat
+(progress entries, metadata), and upload the result to bucket storage.  On the
+next turn we download the transcript, write it to a temp file, and pass
+``--resume`` so the CLI can reconstruct the full conversation.
+
+Storage is handled via ``WorkspaceStorageBackend`` (GCS in prod, local
+filesystem for self-hosted) — no DB column needed.
+"""
+
+import json
+import logging
+import os
+import re
+
+logger = logging.getLogger(__name__)
+
+# UUIDs are hex + hyphens; strip everything else to prevent path injection.
+_SAFE_ID_RE = re.compile(r"[^0-9a-fA-F-]")
+
+# Entry types that can be safely removed from the transcript without breaking
+# the parentUuid conversation tree that ``--resume`` relies on.
+# - progress: UI progress ticks, no message content (avg 97KB for agent_progress)
+# - file-history-snapshot: undo tracking metadata
+# - queue-operation: internal queue bookkeeping
+# - summary: session summaries
+# - pr-link: PR link metadata
+STRIPPABLE_TYPES = frozenset(
+    {"progress", "file-history-snapshot", "queue-operation", "summary", "pr-link"}
+)
+
+# Workspace storage constants — deterministic path from session_id.
+TRANSCRIPT_STORAGE_PREFIX = "chat-transcripts"
+
+
+# ---------------------------------------------------------------------------
+# Progress stripping
+# ---------------------------------------------------------------------------
+
+
+def strip_progress_entries(content: str) -> str:
+    """Remove progress/metadata entries from a JSONL transcript.
+
+    Removes entries whose ``type`` is in ``STRIPPABLE_TYPES`` and reparents
+    any remaining child entries so the ``parentUuid`` chain stays intact.
+    Typically reduces transcript size by ~30%.
+    """
+    lines = content.strip().split("\n")
+
+    entries: list[dict] = []
+    for line in lines:
+        try:
+            entries.append(json.loads(line))
+        except json.JSONDecodeError:
+            # Keep unparseable lines as-is (safety)
+            entries.append({"_raw": line})
+
+    stripped_uuids: set[str] = set()
+    uuid_to_parent: dict[str, str] = {}
+    kept: list[dict] = []
+
+    for entry in entries:
+        if "_raw" in entry:
+            kept.append(entry)
+            continue
+        uid = entry.get("uuid", "")
+        parent = entry.get("parentUuid", "")
+        entry_type = entry.get("type", "")
+
+        if uid:
+            uuid_to_parent[uid] = parent
+
+        if entry_type in STRIPPABLE_TYPES:
+            if uid:
+                stripped_uuids.add(uid)
+        else:
+            kept.append(entry)
+
+    # Reparent: walk up chain through stripped entries to find surviving ancestor
+    for entry in kept:
+        if "_raw" in entry:
+            continue
+        parent = entry.get("parentUuid", "")
+        original_parent = parent
+        while parent in stripped_uuids:
+            parent = uuid_to_parent.get(parent, "")
+        if parent != original_parent:
+            entry["parentUuid"] = parent
+
+    result_lines: list[str] = []
+    for entry in kept:
+        if "_raw" in entry:
+            result_lines.append(entry["_raw"])
+        else:
+            result_lines.append(json.dumps(entry, separators=(",", ":")))
+
+    return "\n".join(result_lines) + "\n"
+
+
+# ---------------------------------------------------------------------------
+# Local file I/O (read from CLI's JSONL, write temp file for --resume)
+# ---------------------------------------------------------------------------
+
+
+def read_transcript_file(transcript_path: str) -> str | None:
+    """Read a JSONL transcript file from disk.
+
+    Returns the raw JSONL content, or ``None`` if the file is missing, empty,
+    or only contains metadata (≤2 lines with no conversation messages).
+    """
+    if not transcript_path or not os.path.isfile(transcript_path):
+        logger.debug(f"[Transcript] File not found: {transcript_path}")
+        return None
+
+    try:
+        with open(transcript_path) as f:
+            content = f.read()
+
+        if not content.strip():
+            logger.debug(f"[Transcript] Empty file: {transcript_path}")
+            return None
+
+        lines = content.strip().split("\n")
+        if len(lines) < 2:
+            # Metadata-only files have 1 line (single queue-operation or snapshot).
+            logger.debug(
+                f"[Transcript] Too few lines ({len(lines)}): {transcript_path}"
+            )
+            return None
+
+        # Quick structural validation — parse first and last lines.
+        json.loads(lines[0])
+        json.loads(lines[-1])
+
+        logger.info(
+            f"[Transcript] Read {len(lines)} lines, "
+            f"{len(content)} bytes from {transcript_path}"
+        )
+        return content
+
+    except (json.JSONDecodeError, OSError) as e:
+        logger.warning(f"[Transcript] Failed to read {transcript_path}: {e}")
+        return None
+
+
+def _sanitize_id(raw_id: str, max_len: int = 36) -> str:
+    """Sanitize an ID for safe use in file paths.
+
+    Session/user IDs are expected to be UUIDs (hex + hyphens).  Strip
+    everything else and truncate to *max_len* so the result cannot introduce
+    path separators or other special characters.
+    """
+    cleaned = _SAFE_ID_RE.sub("", raw_id or "")[:max_len]
+    return cleaned or "unknown"
+
+
+_SAFE_CWD_PREFIX = os.path.realpath("/tmp/copilot-")
+
+
+def write_transcript_to_tempfile(
+    transcript_content: str,
+    session_id: str,
+    cwd: str,
+) -> str | None:
+    """Write JSONL transcript to a temp file inside *cwd* for ``--resume``.
+
+    The file lives in the session working directory so it is cleaned up
+    automatically when the session ends.
+
+    Returns the absolute path to the file, or ``None`` on failure.
+    """
+    # Validate cwd is under the expected sandbox prefix (CodeQL sanitizer).
+    real_cwd = os.path.realpath(cwd)
+    if not real_cwd.startswith(_SAFE_CWD_PREFIX):
+        logger.warning(f"[Transcript] cwd outside sandbox: {cwd}")
+        return None
+
+    try:
+        os.makedirs(real_cwd, exist_ok=True)
+        safe_id = _sanitize_id(session_id, max_len=8)
+        jsonl_path = os.path.realpath(
+            os.path.join(real_cwd, f"transcript-{safe_id}.jsonl")
+        )
+        if not jsonl_path.startswith(real_cwd):
+            logger.warning(f"[Transcript] Path escaped cwd: {jsonl_path}")
+            return None
+
+        with open(jsonl_path, "w") as f:
+            f.write(transcript_content)
+
+        logger.info(f"[Transcript] Wrote resume file: {jsonl_path}")
+        return jsonl_path
+
+    except OSError as e:
+        logger.warning(f"[Transcript] Failed to write resume file: {e}")
+        return None
+
+
+def validate_transcript(content: str | None) -> bool:
+    """Check that a transcript has actual conversation messages.
+
+    A valid transcript for resume needs at least one user message and one
+    assistant message (not just queue-operation / file-history-snapshot
+    metadata).
+    """
+    if not content or not content.strip():
+        return False
+
+    lines = content.strip().split("\n")
+    if len(lines) < 2:
+        return False
+
+    has_user = False
+    has_assistant = False
+
+    for line in lines:
+        try:
+            entry = json.loads(line)
+            msg_type = entry.get("type")
+            if msg_type == "user":
+                has_user = True
+            elif msg_type == "assistant":
+                has_assistant = True
+        except json.JSONDecodeError:
+            return False
+
+    return has_user and has_assistant
+
+
+# ---------------------------------------------------------------------------
+# Bucket storage (GCS / local via WorkspaceStorageBackend)
+# ---------------------------------------------------------------------------
+
+
+def _storage_path_parts(user_id: str, session_id: str) -> tuple[str, str, str]:
+    """Return (workspace_id, file_id, filename) for a session's transcript.
+
+    Path structure: ``chat-transcripts/{user_id}/{session_id}.jsonl``
+    IDs are sanitized to hex+hyphen to prevent path traversal.
+    """
+    return (
+        TRANSCRIPT_STORAGE_PREFIX,
+        _sanitize_id(user_id),
+        f"{_sanitize_id(session_id)}.jsonl",
+    )
+
+
+def _build_storage_path(user_id: str, session_id: str, backend: object) -> str:
+    """Build the full storage path string that ``retrieve()`` expects.
+
+    ``store()`` returns a path like ``gcs://bucket/workspaces/...`` or
+    ``local://workspace_id/file_id/filename``.  Since we use deterministic
+    arguments we can reconstruct the same path for download/delete without
+    having stored the return value.
+    """
+    from backend.util.workspace_storage import GCSWorkspaceStorage
+
+    wid, fid, fname = _storage_path_parts(user_id, session_id)
+
+    if isinstance(backend, GCSWorkspaceStorage):
+        blob = f"workspaces/{wid}/{fid}/{fname}"
+        return f"gcs://{backend.bucket_name}/{blob}"
+    else:
+        # LocalWorkspaceStorage returns local://{relative_path}
+        return f"local://{wid}/{fid}/{fname}"
+
+
+async def upload_transcript(user_id: str, session_id: str, content: str) -> None:
+    """Strip progress entries and upload transcript to bucket storage.
+
+    Safety: only overwrites when the new (stripped) transcript is larger than
+    what is already stored.  Since JSONL is append-only, the latest transcript
+    is always the longest.  This prevents a slow/stale background task from
+    clobbering a newer upload from a concurrent turn.
+    """
+    from backend.util.workspace_storage import get_workspace_storage
+
+    stripped = strip_progress_entries(content)
+    if not validate_transcript(stripped):
+        logger.warning(
+            f"[Transcript] Skipping upload — stripped content is not a valid "
+            f"transcript for session {session_id}"
+        )
+        return
+
+    storage = await get_workspace_storage()
+    wid, fid, fname = _storage_path_parts(user_id, session_id)
+    encoded = stripped.encode("utf-8")
+    new_size = len(encoded)
+
+    # Check existing transcript size to avoid overwriting newer with older
+    path = _build_storage_path(user_id, session_id, storage)
+    try:
+        existing = await storage.retrieve(path)
+        if len(existing) >= new_size:
+            logger.info(
+                f"[Transcript] Skipping upload — existing transcript "
+                f"({len(existing)}B) >= new ({new_size}B) for session "
+                f"{session_id}"
+            )
+            return
+    except (FileNotFoundError, Exception):
+        pass  # No existing transcript or retrieval error — proceed with upload
+
+    await storage.store(
+        workspace_id=wid,
+        file_id=fid,
+        filename=fname,
+        content=encoded,
+    )
+    logger.info(
+        f"[Transcript] Uploaded {new_size} bytes "
+        f"(stripped from {len(content)}) for session {session_id}"
+    )
+
+
+async def download_transcript(user_id: str, session_id: str) -> str | None:
+    """Download transcript from bucket storage.
+
+    Returns the JSONL content string, or ``None`` if not found.
+    """
+    from backend.util.workspace_storage import get_workspace_storage
+
+    storage = await get_workspace_storage()
+    path = _build_storage_path(user_id, session_id, storage)
+
+    try:
+        data = await storage.retrieve(path)
+        content = data.decode("utf-8")
+        logger.info(
+            f"[Transcript] Downloaded {len(content)} bytes for session {session_id}"
+        )
+        return content
+    except FileNotFoundError:
+        logger.debug(f"[Transcript] No transcript in storage for {session_id}")
+        return None
+    except Exception as e:
+        logger.warning(f"[Transcript] Failed to download transcript: {e}")
+        return None
+
+
+async def delete_transcript(user_id: str, session_id: str) -> None:
+    """Delete transcript from bucket storage (e.g. after resume failure)."""
+    from backend.util.workspace_storage import get_workspace_storage
+
+    storage = await get_workspace_storage()
+    path = _build_storage_path(user_id, session_id, storage)
+
+    try:
+        await storage.delete(path)
+        logger.info(f"[Transcript] Deleted transcript for session {session_id}")
+    except Exception as e:
+        logger.warning(f"[Transcript] Failed to delete transcript: {e}")

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

@@ -245,12 +245,16 @@ async def _get_system_prompt_template(context: str) -> str:
     return DEFAULT_SYSTEM_PROMPT.format(users_information=context)
 
 
-async def _build_system_prompt(user_id: str | None) -> tuple[str, Any]:
+async def _build_system_prompt(
+    user_id: str | None, has_conversation_history: bool = False
+) -> tuple[str, Any]:
     """Build the full system prompt including business understanding if available.
 
     Args:
-        user_id: The user ID for fetching business understanding
+        user_id: The user ID for fetching business understanding.
-                     If "default" and this is the user's first session, will use "onboarding" instead.
+        has_conversation_history: Whether there's existing conversation history.
+            If True, we don't tell the model to greet/introduce (since they're
+            already in a conversation).
 
     Returns:
         Tuple of (compiled prompt string, business understanding object)
@@ -266,6 +270,8 @@ async def _build_system_prompt(user_id: str | None) -> tuple[str, Any]:
 
     if understanding:
         context = format_understanding_for_prompt(understanding)
+    elif has_conversation_history:
+        context = "No prior understanding saved yet. Continue the existing conversation naturally."
     else:
         context = "This is the first time you are meeting the user. Greet them and introduce them to the platform"
 
@@ -374,7 +380,6 @@ async def stream_chat_completion(
 
     Raises:
         NotFoundError: If session_id is invalid
-        ValueError: If max_context_messages is exceeded
 
     """
     completion_start = time.monotonic()
@@ -459,8 +464,9 @@ async def stream_chat_completion(
 
     # Generate title for new sessions on first user message (non-blocking)
     # Check: is_user_message, no title yet, and this is the first user message
-    if is_user_message and message and not session.title:
+    user_messages = [m for m in session.messages if m.role == "user"]
-        user_messages = [m for m in session.messages if m.role == "user"]
+    first_user_msg = message or (user_messages[0].content if user_messages else None)
+    if is_user_message and first_user_msg and not session.title:
         if len(user_messages) == 1:
             # First user message - generate title in background
             import asyncio
@@ -468,7 +474,7 @@ async def stream_chat_completion(
             # Capture only the values we need (not the session object) to avoid
             # stale data issues when the main flow modifies the session
             captured_session_id = session_id
-            captured_message = message
+            captured_message = first_user_msg
             captured_user_id = user_id
 
             async def _update_title():
@@ -1237,7 +1243,7 @@ async def _stream_chat_chunks(
 
                 total_time = (time_module.perf_counter() - stream_chunks_start) * 1000
                 logger.info(
-                    f"[TIMING] _stream_chat_chunks COMPLETED in {total_time/1000:.1f}s; "
+                    f"[TIMING] _stream_chat_chunks COMPLETED in {total_time / 1000:.1f}s; "
                     f"session={session.session_id}, user={session.user_id}",
                     extra={"json_fields": {**log_meta, "total_time_ms": total_time}},
                 )

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

@@ -1,3 +1,4 @@
+import asyncio
 import logging
 from os import getenv
 
@@ -11,6 +12,8 @@ from .response_model import (
     StreamTextDelta,
     StreamToolOutputAvailable,
 )
+from .sdk import service as sdk_service
+from .sdk.transcript import download_transcript
 
 logger = logging.getLogger(__name__)
 
@@ -80,3 +83,96 @@ async def test_stream_chat_completion_with_tool_calls(setup_test_user, test_user
     session = await get_chat_session(session.session_id)
     assert session, "Session not found"
     assert session.usage, "Usage is empty"
+
+
+@pytest.mark.asyncio(loop_scope="session")
+async def test_sdk_resume_multi_turn(setup_test_user, test_user_id):
+    """Test that the SDK --resume path captures and uses transcripts across turns.
+
+    Turn 1: Send a message containing a unique keyword.
+    Turn 2: Ask the model to recall that keyword — proving the transcript was
+    persisted and restored via --resume.
+    """
+    api_key: str | None = getenv("OPEN_ROUTER_API_KEY")
+    if not api_key:
+        return pytest.skip("OPEN_ROUTER_API_KEY is not set, skipping test")
+
+    from .config import ChatConfig
+
+    cfg = ChatConfig()
+    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 upsert_chat_session(session)
+
+    # --- Turn 1: send a message with a unique keyword ---
+    keyword = "ZEPHYR42"
+    turn1_msg = (
+        f"Please remember this special keyword: {keyword}. "
+        "Just confirm you've noted it, keep your response brief."
+    )
+    turn1_text = ""
+    turn1_errors: list[str] = []
+    turn1_ended = False
+
+    async for chunk in sdk_service.stream_chat_completion_sdk(
+        session.session_id,
+        turn1_msg,
+        user_id=test_user_id,
+    ):
+        if isinstance(chunk, StreamTextDelta):
+            turn1_text += chunk.delta
+        elif isinstance(chunk, StreamError):
+            turn1_errors.append(chunk.errorText)
+        elif isinstance(chunk, StreamFinish):
+            turn1_ended = True
+
+    assert turn1_ended, "Turn 1 did not finish"
+    assert not turn1_errors, f"Turn 1 errors: {turn1_errors}"
+    assert turn1_text, "Turn 1 produced no text"
+
+    # Wait for background upload task to complete (retry up to 5s)
+    transcript = None
+    for _ in range(10):
+        await asyncio.sleep(0.5)
+        transcript = await download_transcript(test_user_id, session.session_id)
+        if transcript:
+            break
+    assert transcript, (
+        "Transcript was not uploaded to bucket after turn 1 — "
+        "Stop hook may not have fired or transcript was too small"
+    )
+    logger.info(f"Turn 1 transcript uploaded: {len(transcript)} bytes")
+
+    # Reload session for turn 2
+    session = await get_chat_session(session.session_id, test_user_id)
+    assert session, "Session not found after turn 1"
+
+    # --- Turn 2: ask model to recall the keyword ---
+    turn2_msg = "What was the special keyword I asked you to remember?"
+    turn2_text = ""
+    turn2_errors: list[str] = []
+    turn2_ended = False
+
+    async for chunk in sdk_service.stream_chat_completion_sdk(
+        session.session_id,
+        turn2_msg,
+        user_id=test_user_id,
+        session=session,
+    ):
+        if isinstance(chunk, StreamTextDelta):
+            turn2_text += chunk.delta
+        elif isinstance(chunk, StreamError):
+            turn2_errors.append(chunk.errorText)
+        elif isinstance(chunk, StreamFinish):
+            turn2_ended = True
+
+    assert turn2_ended, "Turn 2 did not finish"
+    assert not turn2_errors, f"Turn 2 errors: {turn2_errors}"
+    assert turn2_text, "Turn 2 produced no text"
+    assert keyword in turn2_text, (
+        f"Model did not recall keyword '{keyword}' in turn 2. "
+        f"Response: {turn2_text[:200]}"
+    )
+    logger.info(f"Turn 2 recalled keyword successfully: {turn2_text[:100]}")

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

@@ -814,6 +814,28 @@ async def get_active_task_for_session(
                 if task_user_id and user_id != task_user_id:
                     continue
 
+                # Auto-expire stale tasks that exceeded stream_timeout
+                created_at_str = meta.get("created_at", "")
+                if created_at_str:
+                    try:
+                        created_at = datetime.fromisoformat(created_at_str)
+                        age_seconds = (
+                            datetime.now(timezone.utc) - created_at
+                        ).total_seconds()
+                        if age_seconds > config.stream_timeout:
+                            logger.warning(
+                                f"[TASK_LOOKUP] Auto-expiring stale task {task_id[:8]}... "
+                                f"(age={age_seconds:.0f}s > timeout={config.stream_timeout}s)"
+                            )
+                            await mark_task_completed(task_id, "failed")
+                            continue
+                    except (ValueError, TypeError):
+                        pass
+
+                logger.info(
+                    f"[TASK_LOOKUP] Found running task {task_id[:8]}... for session {session_id[:8]}..."
+                )
+
                 # Get the last message ID from Redis Stream
                 stream_key = _get_task_stream_key(task_id)
                 last_id = "0-0"

autogpt_platform/backend/backend/api/features/chat/tools/__init__.py

@@ -9,6 +9,8 @@ from backend.api.features.chat.tracking import track_tool_called
 from .add_understanding import AddUnderstandingTool
 from .agent_output import AgentOutputTool
 from .base import BaseTool
+from .bash_exec import BashExecTool
+from .check_operation_status import CheckOperationStatusTool
 from .create_agent import CreateAgentTool
 from .customize_agent import CustomizeAgentTool
 from .edit_agent import EditAgentTool
@@ -19,6 +21,7 @@ from .get_doc_page import GetDocPageTool
 from .run_agent import RunAgentTool
 from .run_block import RunBlockTool
 from .search_docs import SearchDocsTool
+from .web_fetch import WebFetchTool
 from .workspace_files import (
     DeleteWorkspaceFileTool,
     ListWorkspaceFilesTool,
@@ -43,9 +46,14 @@ TOOL_REGISTRY: dict[str, BaseTool] = {
     "run_agent": RunAgentTool(),
     "run_block": RunBlockTool(),
     "view_agent_output": AgentOutputTool(),
+    "check_operation_status": CheckOperationStatusTool(),
     "search_docs": SearchDocsTool(),
     "get_doc_page": GetDocPageTool(),
-    # Workspace tools for CoPilot file operations
+    # Web fetch for safe URL retrieval
+    "web_fetch": WebFetchTool(),
+    # Sandboxed code execution (bubblewrap)
+    "bash_exec": BashExecTool(),
+    # Persistent workspace tools (cloud storage, survives across sessions)
     "list_workspace_files": ListWorkspaceFilesTool(),
     "read_workspace_file": ReadWorkspaceFileTool(),
     "write_workspace_file": WriteWorkspaceFileTool(),

autogpt_platform/backend/backend/api/features/chat/tools/agent_generator/dummy.py

@@ -0,0 +1,154 @@
+"""Dummy Agent Generator for testing.
+
+Returns mock responses matching the format expected from the external service.
+Enable via AGENTGENERATOR_USE_DUMMY=true in settings.
+
+WARNING: This is for testing only. Do not use in production.
+"""
+
+import asyncio
+import logging
+import uuid
+from typing import Any
+
+logger = logging.getLogger(__name__)
+
+# Dummy decomposition result (instructions type)
+DUMMY_DECOMPOSITION_RESULT: dict[str, Any] = {
+    "type": "instructions",
+    "steps": [
+        {
+            "description": "Get input from user",
+            "action": "input",
+            "block_name": "AgentInputBlock",
+        },
+        {
+            "description": "Process the input",
+            "action": "process",
+            "block_name": "TextFormatterBlock",
+        },
+        {
+            "description": "Return output to user",
+            "action": "output",
+            "block_name": "AgentOutputBlock",
+        },
+    ],
+}
+
+# Block IDs from backend/blocks/io.py
+AGENT_INPUT_BLOCK_ID = "c0a8e994-ebf1-4a9c-a4d8-89d09c86741b"
+AGENT_OUTPUT_BLOCK_ID = "363ae599-353e-4804-937e-b2ee3cef3da4"
+
+
+def _generate_dummy_agent_json() -> dict[str, Any]:
+    """Generate a minimal valid agent JSON for testing."""
+    input_node_id = str(uuid.uuid4())
+    output_node_id = str(uuid.uuid4())
+
+    return {
+        "id": str(uuid.uuid4()),
+        "version": 1,
+        "is_active": True,
+        "name": "Dummy Test Agent",
+        "description": "A dummy agent generated for testing purposes",
+        "nodes": [
+            {
+                "id": input_node_id,
+                "block_id": AGENT_INPUT_BLOCK_ID,
+                "input_default": {
+                    "name": "input",
+                    "title": "Input",
+                    "description": "Enter your input",
+                    "placeholder_values": [],
+                },
+                "metadata": {"position": {"x": 0, "y": 0}},
+            },
+            {
+                "id": output_node_id,
+                "block_id": AGENT_OUTPUT_BLOCK_ID,
+                "input_default": {
+                    "name": "output",
+                    "title": "Output",
+                    "description": "Agent output",
+                    "format": "{output}",
+                },
+                "metadata": {"position": {"x": 400, "y": 0}},
+            },
+        ],
+        "links": [
+            {
+                "id": str(uuid.uuid4()),
+                "source_id": input_node_id,
+                "sink_id": output_node_id,
+                "source_name": "result",
+                "sink_name": "value",
+                "is_static": False,
+            },
+        ],
+    }
+
+
+async def decompose_goal_dummy(
+    description: str,
+    context: str = "",
+    library_agents: list[dict[str, Any]] | None = None,
+) -> dict[str, Any]:
+    """Return dummy decomposition result."""
+    logger.info("Using dummy agent generator for decompose_goal")
+    return DUMMY_DECOMPOSITION_RESULT.copy()
+
+
+async def generate_agent_dummy(
+    instructions: dict[str, Any],
+    library_agents: list[dict[str, Any]] | None = None,
+    operation_id: str | None = None,
+    task_id: str | None = None,
+) -> dict[str, Any]:
+    """Return dummy agent JSON after a simulated delay."""
+    logger.info("Using dummy agent generator for generate_agent (30s delay)")
+    await asyncio.sleep(30)
+    return _generate_dummy_agent_json()
+
+
+async def generate_agent_patch_dummy(
+    update_request: str,
+    current_agent: dict[str, Any],
+    library_agents: list[dict[str, Any]] | None = None,
+    operation_id: str | None = None,
+    task_id: str | None = None,
+) -> dict[str, Any]:
+    """Return dummy patched agent (returns the current agent with updated description)."""
+    logger.info("Using dummy agent generator for generate_agent_patch")
+    patched = current_agent.copy()
+    patched["description"] = (
+        f"{current_agent.get('description', '')} (updated: {update_request})"
+    )
+    return patched
+
+
+async def customize_template_dummy(
+    template_agent: dict[str, Any],
+    modification_request: str,
+    context: str = "",
+) -> dict[str, Any]:
+    """Return dummy customized template (returns template with updated description)."""
+    logger.info("Using dummy agent generator for customize_template")
+    customized = template_agent.copy()
+    customized["description"] = (
+        f"{template_agent.get('description', '')} (customized: {modification_request})"
+    )
+    return customized
+
+
+async def get_blocks_dummy() -> list[dict[str, Any]]:
+    """Return dummy blocks list."""
+    logger.info("Using dummy agent generator for get_blocks")
+    return [
+        {"id": AGENT_INPUT_BLOCK_ID, "name": "AgentInputBlock"},
+        {"id": AGENT_OUTPUT_BLOCK_ID, "name": "AgentOutputBlock"},
+    ]
+
+
+async def health_check_dummy() -> bool:
+    """Always returns healthy for dummy service."""
+    return True

autogpt_platform/backend/backend/api/features/chat/tools/agent_generator/service.py

@@ -12,8 +12,19 @@ import httpx
 
 from backend.util.settings import Settings
 
+from .dummy import (
+    customize_template_dummy,
+    decompose_goal_dummy,
+    generate_agent_dummy,
+    generate_agent_patch_dummy,
+    get_blocks_dummy,
+    health_check_dummy,
+)
+
 logger = logging.getLogger(__name__)
 
+_dummy_mode_warned = False
+
 
 def _create_error_response(
     error_message: str,
@@ -90,10 +101,26 @@ def _get_settings() -> Settings:
     return _settings
 
 
-def is_external_service_configured() -> bool:
+def _is_dummy_mode() -> bool:
-    """Check if external Agent Generator service is configured."""
+    """Check if dummy mode is enabled for testing."""
+    global _dummy_mode_warned
     settings = _get_settings()
-    return bool(settings.config.agentgenerator_host)
+    is_dummy = bool(settings.config.agentgenerator_use_dummy)
+    if is_dummy and not _dummy_mode_warned:
+        logger.warning(
+            "Agent Generator running in DUMMY MODE - returning mock responses. "
+            "Do not use in production!"
+        )
+        _dummy_mode_warned = True
+    return is_dummy
+
+
+def is_external_service_configured() -> bool:
+    """Check if external Agent Generator service is configured (or dummy mode)."""
+    settings = _get_settings()
+    return bool(settings.config.agentgenerator_host) or bool(
+        settings.config.agentgenerator_use_dummy
+    )
 
 
 def _get_base_url() -> str:
@@ -137,6 +164,9 @@ async def decompose_goal_external(
         - {"type": "error", "error": "...", "error_type": "..."} on error
         Or None on unexpected error
     """
+    if _is_dummy_mode():
+        return await decompose_goal_dummy(description, context, library_agents)
+
     client = _get_client()
 
     if context:
@@ -226,6 +256,11 @@ async def generate_agent_external(
     Returns:
         Agent JSON dict, {"status": "accepted"} for async, or error dict {"type": "error", ...} on error
     """
+    if _is_dummy_mode():
+        return await generate_agent_dummy(
+            instructions, library_agents, operation_id, task_id
+        )
+
     client = _get_client()
 
     # Build request payload
@@ -297,6 +332,11 @@ async def generate_agent_patch_external(
     Returns:
         Updated agent JSON, clarifying questions dict, {"status": "accepted"} for async, or error dict on error
     """
+    if _is_dummy_mode():
+        return await generate_agent_patch_dummy(
+            update_request, current_agent, library_agents, operation_id, task_id
+        )
+
     client = _get_client()
 
     # Build request payload
@@ -383,6 +423,11 @@ async def customize_template_external(
     Returns:
         Customized agent JSON, clarifying questions dict, or error dict on error
     """
+    if _is_dummy_mode():
+        return await customize_template_dummy(
+            template_agent, modification_request, context
+        )
+
     client = _get_client()
 
     request = modification_request
@@ -445,6 +490,9 @@ async def get_blocks_external() -> list[dict[str, Any]] | None:
     Returns:
         List of block info dicts or None on error
     """
+    if _is_dummy_mode():
+        return await get_blocks_dummy()
+
     client = _get_client()
 
     try:
@@ -478,6 +526,9 @@ async def health_check() -> bool:
     if not is_external_service_configured():
         return False
 
+    if _is_dummy_mode():
+        return await health_check_dummy()
+
     client = _get_client()
 
     try:

autogpt_platform/backend/backend/api/features/chat/tools/bash_exec.py

@@ -0,0 +1,131 @@
+"""Bash execution tool — run shell commands in a bubblewrap sandbox.
+
+Full Bash scripting is allowed (loops, conditionals, pipes, functions, etc.).
+Safety comes from OS-level isolation (bubblewrap): only system dirs visible
+read-only, writable workspace only, clean env, no network.
+
+Requires bubblewrap (``bwrap``) — the tool is disabled when bwrap is not
+available (e.g. macOS development).
+"""
+
+import logging
+from typing import Any
+
+from backend.api.features.chat.model import ChatSession
+from backend.api.features.chat.tools.base import BaseTool
+from backend.api.features.chat.tools.models import (
+    BashExecResponse,
+    ErrorResponse,
+    ToolResponseBase,
+)
+from backend.api.features.chat.tools.sandbox import (
+    get_workspace_dir,
+    has_full_sandbox,
+    run_sandboxed,
+)
+
+logger = logging.getLogger(__name__)
+
+
+class BashExecTool(BaseTool):
+    """Execute Bash commands in a bubblewrap sandbox."""
+
+    @property
+    def name(self) -> str:
+        return "bash_exec"
+
+    @property
+    def description(self) -> str:
+        if not has_full_sandbox():
+            return (
+                "Bash execution is DISABLED — bubblewrap sandbox is not "
+                "available on this platform. Do not call this tool."
+            )
+        return (
+            "Execute a Bash command or script in a bubblewrap sandbox. "
+            "Full Bash scripting is supported (loops, conditionals, pipes, "
+            "functions, etc.). "
+            "The sandbox shares the same working directory as the SDK Read/Write "
+            "tools — files created by either are accessible to both. "
+            "SECURITY: Only system directories (/usr, /bin, /lib, /etc) are "
+            "visible read-only, the per-session workspace is the only writable "
+            "path, environment variables are wiped (no secrets), all network "
+            "access is blocked at the kernel level, and resource limits are "
+            "enforced (max 64 processes, 512MB memory, 50MB file size). "
+            "Application code, configs, and other directories are NOT accessible. "
+            "To fetch web content, use the web_fetch tool instead. "
+            "Execution is killed after the timeout (default 30s, max 120s). "
+            "Returns stdout and stderr. "
+            "Useful for file manipulation, data processing with Unix tools "
+            "(grep, awk, sed, jq, etc.), and running shell scripts."
+        )
+
+    @property
+    def parameters(self) -> dict[str, Any]:
+        return {
+            "type": "object",
+            "properties": {
+                "command": {
+                    "type": "string",
+                    "description": "Bash command or script to execute.",
+                },
+                "timeout": {
+                    "type": "integer",
+                    "description": (
+                        "Max execution time in seconds (default 30, max 120)."
+                    ),
+                    "default": 30,
+                },
+            },
+            "required": ["command"],
+        }
+
+    @property
+    def requires_auth(self) -> bool:
+        return False
+
+    async def _execute(
+        self,
+        user_id: str | None,
+        session: ChatSession,
+        **kwargs: Any,
+    ) -> ToolResponseBase:
+        session_id = session.session_id if session else None
+
+        if not has_full_sandbox():
+            return ErrorResponse(
+                message="bash_exec requires bubblewrap sandbox (Linux only).",
+                error="sandbox_unavailable",
+                session_id=session_id,
+            )
+
+        command: str = (kwargs.get("command") or "").strip()
+        timeout: int = kwargs.get("timeout", 30)
+
+        if not command:
+            return ErrorResponse(
+                message="No command provided.",
+                error="empty_command",
+                session_id=session_id,
+            )
+
+        workspace = get_workspace_dir(session_id or "default")
+
+        stdout, stderr, exit_code, timed_out = await run_sandboxed(
+            command=["bash", "-c", command],
+            cwd=workspace,
+            timeout=timeout,
+        )
+
+        return BashExecResponse(
+            message=(
+                "Execution timed out"
+                if timed_out
+                else f"Command executed (exit {exit_code})"
+            ),
+            stdout=stdout,
+            stderr=stderr,
+            exit_code=exit_code,
+            timed_out=timed_out,
+            session_id=session_id,
+        )

autogpt_platform/backend/backend/api/features/chat/tools/check_operation_status.py

@@ -0,0 +1,127 @@
+"""CheckOperationStatusTool — query the status of a long-running operation."""
+
+import logging
+from typing import Any
+
+from backend.api.features.chat.model import ChatSession
+from backend.api.features.chat.tools.base import BaseTool
+from backend.api.features.chat.tools.models import (
+    ErrorResponse,
+    ResponseType,
+    ToolResponseBase,
+)
+
+logger = logging.getLogger(__name__)
+
+
+class OperationStatusResponse(ToolResponseBase):
+    """Response for check_operation_status tool."""
+
+    type: ResponseType = ResponseType.OPERATION_STATUS
+    task_id: str
+    operation_id: str
+    status: str  # "running", "completed", "failed"
+    tool_name: str | None = None
+    message: str = ""
+
+
+class CheckOperationStatusTool(BaseTool):
+    """Check the status of a long-running operation (create_agent, edit_agent, etc.).
+
+    The CoPilot uses this tool to report back to the user whether an
+    operation that was started earlier has completed, failed, or is still
+    running.
+    """
+
+    @property
+    def name(self) -> str:
+        return "check_operation_status"
+
+    @property
+    def description(self) -> str:
+        return (
+            "Check the current status of a long-running operation such as "
+            "create_agent or edit_agent. Accepts either an operation_id or "
+            "task_id from a previous operation_started response. "
+            "Returns the current status: running, completed, or failed."
+        )
+
+    @property
+    def parameters(self) -> dict[str, Any]:
+        return {
+            "type": "object",
+            "properties": {
+                "operation_id": {
+                    "type": "string",
+                    "description": (
+                        "The operation_id from an operation_started response."
+                    ),
+                },
+                "task_id": {
+                    "type": "string",
+                    "description": (
+                        "The task_id from an operation_started response. "
+                        "Used as fallback if operation_id is not provided."
+                    ),
+                },
+            },
+            "required": [],
+        }
+
+    @property
+    def requires_auth(self) -> bool:
+        return False
+
+    async def _execute(
+        self,
+        user_id: str | None,
+        session: ChatSession,
+        **kwargs,
+    ) -> ToolResponseBase:
+        from backend.api.features.chat import stream_registry
+
+        operation_id = (kwargs.get("operation_id") or "").strip()
+        task_id = (kwargs.get("task_id") or "").strip()
+
+        if not operation_id and not task_id:
+            return ErrorResponse(
+                message="Please provide an operation_id or task_id.",
+                error="missing_parameter",
+            )
+
+        task = None
+        if operation_id:
+            task = await stream_registry.find_task_by_operation_id(operation_id)
+        if task is None and task_id:
+            task = await stream_registry.get_task(task_id)
+
+        if task is None:
+            # Task not in Redis — it may have already expired (TTL).
+            # Check conversation history for the result instead.
+            return ErrorResponse(
+                message=(
+                    "Operation not found — it may have already completed and "
+                    "expired from the status tracker. Check the conversation "
+                    "history for the result."
+                ),
+                error="not_found",
+            )
+
+        status_messages = {
+            "running": (
+                f"The {task.tool_name or 'operation'} is still running. "
+                "Please wait for it to complete."
+            ),
+            "completed": (
+                f"The {task.tool_name or 'operation'} has completed successfully."
+            ),
+            "failed": f"The {task.tool_name or 'operation'} has failed.",
+        }
+
+        return OperationStatusResponse(
+            task_id=task.task_id,
+            operation_id=task.operation_id,
+            status=task.status,
+            tool_name=task.tool_name,
+            message=status_messages.get(task.status, f"Status: {task.status}"),
+        )

autogpt_platform/backend/backend/api/features/chat/tools/find_block.py

@@ -7,7 +7,6 @@ from backend.api.features.chat.model import ChatSession
 from backend.api.features.chat.tools.base import BaseTool, ToolResponseBase
 from backend.api.features.chat.tools.models import (
     BlockInfoSummary,
-    BlockInputFieldInfo,
     BlockListResponse,
     ErrorResponse,
     NoResultsResponse,
@@ -55,7 +54,8 @@ class FindBlockTool(BaseTool):
             "Blocks are reusable components that perform specific tasks like "
             "sending emails, making API calls, processing text, etc. "
             "IMPORTANT: Use this tool FIRST to get the block's 'id' before calling run_block. "
-            "The response includes each block's id, required_inputs, and input_schema."
+            "The response includes each block's id, name, and description. "
+            "Call run_block with the block's id **with no inputs** to see detailed inputs/outputs and execute it."
         )
 
     @property
@@ -124,7 +124,7 @@ class FindBlockTool(BaseTool):
                     session_id=session_id,
                 )
 
-            # Enrich results with full block information
+            # Enrich results with block information
             blocks: list[BlockInfoSummary] = []
             for result in results:
                 block_id = result["content_id"]
@@ -141,65 +141,12 @@ class FindBlockTool(BaseTool):
                 ):
                     continue
 
-                # Get input/output schemas
-                input_schema = {}
-                output_schema = {}
-                try:
-                    input_schema = block.input_schema.jsonschema()
-                except Exception as e:
-                    logger.debug(
-                        "Failed to generate input schema for block %s: %s",
-                        block_id,
-                        e,
-                    )
-                try:
-                    output_schema = block.output_schema.jsonschema()
-                except Exception as e:
-                    logger.debug(
-                        "Failed to generate output schema for block %s: %s",
-                        block_id,
-                        e,
-                    )
-
-                # Get categories from block instance
-                categories = []
-                if hasattr(block, "categories") and block.categories:
-                    categories = [cat.value for cat in block.categories]
-
-                # Extract required inputs for easier use
-                required_inputs: list[BlockInputFieldInfo] = []
-                if input_schema:
-                    properties = input_schema.get("properties", {})
-                    required_fields = set(input_schema.get("required", []))
-                    # Get credential field names to exclude from required inputs
-                    credentials_fields = set(
-                        block.input_schema.get_credentials_fields().keys()
-                    )
-
-                    for field_name, field_schema in properties.items():
-                        # Skip credential fields - they're handled separately
-                        if field_name in credentials_fields:
-                            continue
-
-                        required_inputs.append(
-                            BlockInputFieldInfo(
-                                name=field_name,
-                                type=field_schema.get("type", "string"),
-                                description=field_schema.get("description", ""),
-                                required=field_name in required_fields,
-                                default=field_schema.get("default"),
-                            )
-                        )
-
                 blocks.append(
                     BlockInfoSummary(
                         id=block_id,
                         name=block.name,
                         description=block.description or "",
-                        categories=categories,
+                        categories=[c.value for c in block.categories],
-                        input_schema=input_schema,
-                        output_schema=output_schema,
-                        required_inputs=required_inputs,
                     )
                 )
 
@@ -228,8 +175,7 @@ class FindBlockTool(BaseTool):
             return BlockListResponse(
                 message=(
                     f"Found {len(blocks)} block(s) matching '{query}'. "
-                    "To execute a block, use run_block with the block's 'id' field "
+                    "To see a block's inputs/outputs and execute it, use run_block with the block's 'id' - providing no inputs."
-                    "and provide 'input_data' matching the block's input_schema."
                 ),
                 blocks=blocks,
                 count=len(blocks),

autogpt_platform/backend/backend/api/features/chat/tools/find_block_test.py

@@ -18,7 +18,13 @@ _TEST_USER_ID = "test-user-find-block"
 
 
 def make_mock_block(
-    block_id: str, name: str, block_type: BlockType, disabled: bool = False
+    block_id: str,
+    name: str,
+    block_type: BlockType,
+    disabled: bool = False,
+    input_schema: dict | None = None,
+    output_schema: dict | None = None,
+    credentials_fields: dict | None = None,
 ):
     """Create a mock block for testing."""
     mock = MagicMock()
@@ -28,10 +34,13 @@ def make_mock_block(
     mock.block_type = block_type
     mock.disabled = disabled
     mock.input_schema = MagicMock()
-    mock.input_schema.jsonschema.return_value = {"properties": {}, "required": []}
+    mock.input_schema.jsonschema.return_value = input_schema or {
-    mock.input_schema.get_credentials_fields.return_value = {}
+        "properties": {},
+        "required": [],
+    }
+    mock.input_schema.get_credentials_fields.return_value = credentials_fields or {}
     mock.output_schema = MagicMock()
-    mock.output_schema.jsonschema.return_value = {}
+    mock.output_schema.jsonschema.return_value = output_schema or {}
     mock.categories = []
     return mock
 
@@ -137,3 +146,241 @@ class TestFindBlockFiltering:
         assert isinstance(response, BlockListResponse)
         assert len(response.blocks) == 1
         assert response.blocks[0].id == "normal-block-id"
+
+    @pytest.mark.asyncio(loop_scope="session")
+    async def test_response_size_average_chars_per_block(self):
+        """Measure average chars per block in the serialized response."""
+        session = make_session(user_id=_TEST_USER_ID)
+
+        # Realistic block definitions modeled after real blocks
+        block_defs = [
+            {
+                "id": "http-block-id",
+                "name": "Send Web Request",
+                "input_schema": {
+                    "properties": {
+                        "url": {
+                            "type": "string",
+                            "description": "The URL to send the request to",
+                        },
+                        "method": {
+                            "type": "string",
+                            "description": "The HTTP method to use",
+                        },
+                        "headers": {
+                            "type": "object",
+                            "description": "Headers to include in the request",
+                        },
+                        "json_format": {
+                            "type": "boolean",
+                            "description": "If true, send the body as JSON",
+                        },
+                        "body": {
+                            "type": "object",
+                            "description": "Form/JSON body payload",
+                        },
+                        "credentials": {
+                            "type": "object",
+                            "description": "HTTP credentials",
+                        },
+                    },
+                    "required": ["url", "method"],
+                },
+                "output_schema": {
+                    "properties": {
+                        "response": {
+                            "type": "object",
+                            "description": "The response from the server",
+                        },
+                        "client_error": {
+                            "type": "object",
+                            "description": "Errors on 4xx status codes",
+                        },
+                        "server_error": {
+                            "type": "object",
+                            "description": "Errors on 5xx status codes",
+                        },
+                        "error": {
+                            "type": "string",
+                            "description": "Errors for all other exceptions",
+                        },
+                    },
+                },
+                "credentials_fields": {"credentials": True},
+            },
+            {
+                "id": "email-block-id",
+                "name": "Send Email",
+                "input_schema": {
+                    "properties": {
+                        "to_email": {
+                            "type": "string",
+                            "description": "Recipient email address",
+                        },
+                        "subject": {
+                            "type": "string",
+                            "description": "Subject of the email",
+                        },
+                        "body": {
+                            "type": "string",
+                            "description": "Body of the email",
+                        },
+                        "config": {
+                            "type": "object",
+                            "description": "SMTP Config",
+                        },
+                        "credentials": {
+                            "type": "object",
+                            "description": "SMTP credentials",
+                        },
+                    },
+                    "required": ["to_email", "subject", "body", "credentials"],
+                },
+                "output_schema": {
+                    "properties": {
+                        "status": {
+                            "type": "string",
+                            "description": "Status of the email sending operation",
+                        },
+                        "error": {
+                            "type": "string",
+                            "description": "Error message if sending failed",
+                        },
+                    },
+                },
+                "credentials_fields": {"credentials": True},
+            },
+            {
+                "id": "claude-code-block-id",
+                "name": "Claude Code",
+                "input_schema": {
+                    "properties": {
+                        "e2b_credentials": {
+                            "type": "object",
+                            "description": "API key for E2B platform",
+                        },
+                        "anthropic_credentials": {
+                            "type": "object",
+                            "description": "API key for Anthropic",
+                        },
+                        "prompt": {
+                            "type": "string",
+                            "description": "Task or instruction for Claude Code",
+                        },
+                        "timeout": {
+                            "type": "integer",
+                            "description": "Sandbox timeout in seconds",
+                        },
+                        "setup_commands": {
+                            "type": "array",
+                            "description": "Shell commands to run before execution",
+                        },
+                        "working_directory": {
+                            "type": "string",
+                            "description": "Working directory for Claude Code",
+                        },
+                        "session_id": {
+                            "type": "string",
+                            "description": "Session ID to resume a conversation",
+                        },
+                        "sandbox_id": {
+                            "type": "string",
+                            "description": "Sandbox ID to reconnect to",
+                        },
+                        "conversation_history": {
+                            "type": "string",
+                            "description": "Previous conversation history",
+                        },
+                        "dispose_sandbox": {
+                            "type": "boolean",
+                            "description": "Whether to dispose sandbox after execution",
+                        },
+                    },
+                    "required": [
+                        "e2b_credentials",
+                        "anthropic_credentials",
+                        "prompt",
+                    ],
+                },
+                "output_schema": {
+                    "properties": {
+                        "response": {
+                            "type": "string",
+                            "description": "Output from Claude Code execution",
+                        },
+                        "files": {
+                            "type": "array",
+                            "description": "Files created/modified by Claude Code",
+                        },
+                        "conversation_history": {
+                            "type": "string",
+                            "description": "Full conversation history",
+                        },
+                        "session_id": {
+                            "type": "string",
+                            "description": "Session ID for this conversation",
+                        },
+                        "sandbox_id": {
+                            "type": "string",
+                            "description": "ID of the sandbox instance",
+                        },
+                        "error": {
+                            "type": "string",
+                            "description": "Error message if execution failed",
+                        },
+                    },
+                },
+                "credentials_fields": {
+                    "e2b_credentials": True,
+                    "anthropic_credentials": True,
+                },
+            },
+        ]
+
+        search_results = [
+            {"content_id": d["id"], "score": 0.9 - i * 0.1}
+            for i, d in enumerate(block_defs)
+        ]
+        mock_blocks = {
+            d["id"]: make_mock_block(
+                block_id=d["id"],
+                name=d["name"],
+                block_type=BlockType.STANDARD,
+                input_schema=d["input_schema"],
+                output_schema=d["output_schema"],
+                credentials_fields=d["credentials_fields"],
+            )
+            for d in block_defs
+        }
+
+        with patch(
+            "backend.api.features.chat.tools.find_block.unified_hybrid_search",
+            new_callable=AsyncMock,
+            return_value=(search_results, len(search_results)),
+        ), patch(
+            "backend.api.features.chat.tools.find_block.get_block",
+            side_effect=lambda bid: mock_blocks.get(bid),
+        ):
+            tool = FindBlockTool()
+            response = await tool._execute(
+                user_id=_TEST_USER_ID, session=session, query="test"
+            )
+
+        assert isinstance(response, BlockListResponse)
+        assert response.count == len(block_defs)
+
+        total_chars = len(response.model_dump_json())
+        avg_chars = total_chars // response.count
+
+        # Print for visibility in test output
+        print(f"\nTotal response size: {total_chars} chars")
+        print(f"Number of blocks: {response.count}")
+        print(f"Average chars per block: {avg_chars}")
+
+        # The old response was ~90K for 10 blocks (~9K per block).
+        # Previous optimization reduced it to ~1.5K per block (no raw JSON schemas).
+        # Now with only id/name/description, we expect ~300 chars per block.
+        assert avg_chars < 500, (
+            f"Average chars per block ({avg_chars}) exceeds 500. "
+            f"Total response: {total_chars} chars for {response.count} blocks."
+        )

autogpt_platform/backend/backend/api/features/chat/tools/models.py

@@ -25,6 +25,7 @@ class ResponseType(str, Enum):
     AGENT_SAVED = "agent_saved"
     CLARIFICATION_NEEDED = "clarification_needed"
     BLOCK_LIST = "block_list"
+    BLOCK_DETAILS = "block_details"
     BLOCK_OUTPUT = "block_output"
     DOC_SEARCH_RESULTS = "doc_search_results"
     DOC_PAGE = "doc_page"
@@ -40,6 +41,12 @@ class ResponseType(str, Enum):
     OPERATION_IN_PROGRESS = "operation_in_progress"
     # Input validation
     INPUT_VALIDATION_ERROR = "input_validation_error"
+    # Web fetch
+    WEB_FETCH = "web_fetch"
+    # Code execution
+    BASH_EXEC = "bash_exec"
+    # Operation status check
+    OPERATION_STATUS = "operation_status"
 
 
 # Base response model
@@ -335,11 +342,17 @@ class BlockInfoSummary(BaseModel):
     name: str
     description: str
     categories: list[str]
-    input_schema: dict[str, Any]
+    input_schema: dict[str, Any] = Field(
-    output_schema: dict[str, Any]
+        default_factory=dict,
+        description="Full JSON schema for block inputs",
+    )
+    output_schema: dict[str, Any] = Field(
+        default_factory=dict,
+        description="Full JSON schema for block outputs",
+    )
     required_inputs: list[BlockInputFieldInfo] = Field(
         default_factory=list,
-        description="List of required input fields for this block",
+        description="List of input fields for this block",
     )
 
 
@@ -352,10 +365,29 @@ class BlockListResponse(ToolResponseBase):
     query: str
     usage_hint: str = Field(
         default="To execute a block, call run_block with block_id set to the block's "
-        "'id' field and input_data containing the required fields from input_schema."
+        "'id' field and input_data containing the fields listed in required_inputs."
     )
 
 
+class BlockDetails(BaseModel):
+    """Detailed block information."""
+
+    id: str
+    name: str
+    description: str
+    inputs: dict[str, Any] = {}
+    outputs: dict[str, Any] = {}
+    credentials: list[CredentialsMetaInput] = []
+
+
+class BlockDetailsResponse(ToolResponseBase):
+    """Response for block details (first run_block attempt)."""
+
+    type: ResponseType = ResponseType.BLOCK_DETAILS
+    block: BlockDetails
+    user_authenticated: bool = False
+
+
 class BlockOutputResponse(ToolResponseBase):
     """Response for run_block tool."""
 
@@ -421,3 +453,24 @@ class AsyncProcessingResponse(ToolResponseBase):
     status: str = "accepted"  # Must be "accepted" for detection
     operation_id: str | None = None
     task_id: str | None = None
+
+
+class WebFetchResponse(ToolResponseBase):
+    """Response for web_fetch tool."""
+
+    type: ResponseType = ResponseType.WEB_FETCH
+    url: str
+    status_code: int
+    content_type: str
+    content: str
+    truncated: bool = False
+
+
+class BashExecResponse(ToolResponseBase):
+    """Response for bash_exec tool."""
+
+    type: ResponseType = ResponseType.BASH_EXEC
+    stdout: str
+    stderr: str
+    exit_code: int
+    timed_out: bool = False

autogpt_platform/backend/backend/api/features/chat/tools/run_block.py

@@ -23,8 +23,11 @@ from backend.util.exceptions import BlockError
 from .base import BaseTool
 from .helpers import get_inputs_from_schema
 from .models import (
+    BlockDetails,
+    BlockDetailsResponse,
     BlockOutputResponse,
     ErrorResponse,
+    InputValidationErrorResponse,
     SetupInfo,
     SetupRequirementsResponse,
     ToolResponseBase,
@@ -51,8 +54,8 @@ class RunBlockTool(BaseTool):
             "Execute a specific block with the provided input data. "
             "IMPORTANT: You MUST call find_block first to get the block's 'id' - "
             "do NOT guess or make up block IDs. "
-            "Use the 'id' from find_block results and provide input_data "
+            "On first attempt (without input_data), returns detailed schema showing "
-            "matching the block's required_inputs."
+            "required inputs and outputs. Then call again with proper input_data to execute."
         )
 
     @property
@@ -67,11 +70,19 @@ class RunBlockTool(BaseTool):
                         "NEVER guess this - always get it from find_block first."
                     ),
                 },
+                "block_name": {
+                    "type": "string",
+                    "description": (
+                        "The block's human-readable name from find_block results. "
+                        "Used for display purposes in the UI."
+                    ),
+                },
                 "input_data": {
                     "type": "object",
                     "description": (
-                        "Input values for the block. Use the 'required_inputs' field "
+                        "Input values for the block. "
-                        "from find_block to see what fields are needed."
+                        "First call with empty {} to see the block's schema, "
+                        "then call again with proper values to execute."
                     ),
                 },
             },
@@ -156,6 +167,34 @@ class RunBlockTool(BaseTool):
             await self._resolve_block_credentials(user_id, block, input_data)
         )
 
+        # Get block schemas for details/validation
+        try:
+            input_schema: dict[str, Any] = block.input_schema.jsonschema()
+        except Exception as e:
+            logger.warning(
+                "Failed to generate input schema for block %s: %s",
+                block_id,
+                e,
+            )
+            return ErrorResponse(
+                message=f"Block '{block.name}' has an invalid input schema",
+                error=str(e),
+                session_id=session_id,
+            )
+        try:
+            output_schema: dict[str, Any] = block.output_schema.jsonschema()
+        except Exception as e:
+            logger.warning(
+                "Failed to generate output schema for block %s: %s",
+                block_id,
+                e,
+            )
+            return ErrorResponse(
+                message=f"Block '{block.name}' has an invalid output schema",
+                error=str(e),
+                session_id=session_id,
+            )
+
         if missing_credentials:
             # Return setup requirements response with missing credentials
             credentials_fields_info = block.input_schema.get_credentials_fields_info()
@@ -188,6 +227,53 @@ class RunBlockTool(BaseTool):
                 graph_version=None,
             )
 
+        # Check if this is a first attempt (required inputs missing)
+        # Return block details so user can see what inputs are needed
+        credentials_fields = set(block.input_schema.get_credentials_fields().keys())
+        required_keys = set(input_schema.get("required", []))
+        required_non_credential_keys = required_keys - credentials_fields
+        provided_input_keys = set(input_data.keys()) - credentials_fields
+
+        # Check for unknown input fields
+        valid_fields = (
+            set(input_schema.get("properties", {}).keys()) - credentials_fields
+        )
+        unrecognized_fields = provided_input_keys - valid_fields
+        if unrecognized_fields:
+            return InputValidationErrorResponse(
+                message=(
+                    f"Unknown input field(s) provided: {', '.join(sorted(unrecognized_fields))}. "
+                    f"Block was not executed. Please use the correct field names from the schema."
+                ),
+                session_id=session_id,
+                unrecognized_fields=sorted(unrecognized_fields),
+                inputs=input_schema,
+            )
+
+        # Show details when not all required non-credential inputs are provided
+        if not (required_non_credential_keys <= provided_input_keys):
+            # Get credentials info for the response
+            credentials_meta = []
+            for field_name, cred_meta in matched_credentials.items():
+                credentials_meta.append(cred_meta)
+
+            return BlockDetailsResponse(
+                message=(
+                    f"Block '{block.name}' details. "
+                    "Provide input_data matching the inputs schema to execute the block."
+                ),
+                session_id=session_id,
+                block=BlockDetails(
+                    id=block_id,
+                    name=block.name,
+                    description=block.description or "",
+                    inputs=input_schema,
+                    outputs=output_schema,
+                    credentials=credentials_meta,
+                ),
+                user_authenticated=True,
+            )
+
         try:
             # Get or create user's workspace for CoPilot file operations
             workspace = await get_or_create_workspace(user_id)

autogpt_platform/backend/backend/api/features/chat/tools/run_block_test.py

@@ -1,10 +1,15 @@
-"""Tests for block execution guards in RunBlockTool."""
+"""Tests for block execution guards and input validation in RunBlockTool."""
 
-from unittest.mock import MagicMock, patch
+from unittest.mock import AsyncMock, MagicMock, patch
 
 import pytest
 
-from backend.api.features.chat.tools.models import ErrorResponse
+from backend.api.features.chat.tools.models import (
+    BlockDetailsResponse,
+    BlockOutputResponse,
+    ErrorResponse,
+    InputValidationErrorResponse,
+)
 from backend.api.features.chat.tools.run_block import RunBlockTool
 from backend.blocks._base import BlockType
 
@@ -28,6 +33,39 @@ def make_mock_block(
     return mock
 
 
+def make_mock_block_with_schema(
+    block_id: str,
+    name: str,
+    input_properties: dict,
+    required_fields: list[str],
+    output_properties: dict | None = None,
+):
+    """Create a mock block with a defined input/output schema for validation tests."""
+    mock = MagicMock()
+    mock.id = block_id
+    mock.name = name
+    mock.block_type = BlockType.STANDARD
+    mock.disabled = False
+    mock.description = f"Test block: {name}"
+
+    input_schema = {
+        "properties": input_properties,
+        "required": required_fields,
+    }
+    mock.input_schema = MagicMock()
+    mock.input_schema.jsonschema.return_value = input_schema
+    mock.input_schema.get_credentials_fields_info.return_value = {}
+    mock.input_schema.get_credentials_fields.return_value = {}
+
+    output_schema = {
+        "properties": output_properties or {"result": {"type": "string"}},
+    }
+    mock.output_schema = MagicMock()
+    mock.output_schema.jsonschema.return_value = output_schema
+
+    return mock
+
+
 class TestRunBlockFiltering:
     """Tests for block execution guards in RunBlockTool."""
 
@@ -104,3 +142,221 @@ class TestRunBlockFiltering:
         # (may be other errors like missing credentials, but not the exclusion guard)
         if isinstance(response, ErrorResponse):
             assert "cannot be run directly in CoPilot" not in response.message
+
+
+class TestRunBlockInputValidation:
+    """Tests for input field validation in RunBlockTool.
+
+    run_block rejects unknown input field names with InputValidationErrorResponse,
+    preventing silent failures where incorrect keys would be ignored and the block
+    would execute with default values instead of the caller's intended values.
+    """
+
+    @pytest.mark.asyncio(loop_scope="session")
+    async def test_unknown_input_fields_are_rejected(self):
+        """run_block rejects unknown input fields instead of silently ignoring them.
+
+        Scenario: The AI Text Generator block has a field called 'model' (for LLM model
+        selection), but the LLM calling the tool guesses wrong and sends 'LLM_Model'
+        instead. The block should reject the request and return the valid schema.
+        """
+        session = make_session(user_id=_TEST_USER_ID)
+
+        mock_block = make_mock_block_with_schema(
+            block_id="ai-text-gen-id",
+            name="AI Text Generator",
+            input_properties={
+                "prompt": {"type": "string", "description": "The prompt to send"},
+                "model": {
+                    "type": "string",
+                    "description": "The LLM model to use",
+                    "default": "gpt-4o-mini",
+                },
+                "sys_prompt": {
+                    "type": "string",
+                    "description": "System prompt",
+                    "default": "",
+                },
+            },
+            required_fields=["prompt"],
+            output_properties={"response": {"type": "string"}},
+        )
+
+        with patch(
+            "backend.api.features.chat.tools.run_block.get_block",
+            return_value=mock_block,
+        ):
+            tool = RunBlockTool()
+
+            # Provide 'prompt' (correct) but 'LLM_Model' instead of 'model' (wrong key)
+            response = await tool._execute(
+                user_id=_TEST_USER_ID,
+                session=session,
+                block_id="ai-text-gen-id",
+                input_data={
+                    "prompt": "Write a haiku about coding",
+                    "LLM_Model": "claude-opus-4-6",  # WRONG KEY - should be 'model'
+                },
+            )
+
+        assert isinstance(response, InputValidationErrorResponse)
+        assert "LLM_Model" in response.unrecognized_fields
+        assert "Block was not executed" in response.message
+        assert "inputs" in response.model_dump()  # valid schema included
+
+    @pytest.mark.asyncio(loop_scope="session")
+    async def test_multiple_wrong_keys_are_all_reported(self):
+        """All unrecognized field names are reported in a single error response."""
+        session = make_session(user_id=_TEST_USER_ID)
+
+        mock_block = make_mock_block_with_schema(
+            block_id="ai-text-gen-id",
+            name="AI Text Generator",
+            input_properties={
+                "prompt": {"type": "string"},
+                "model": {"type": "string", "default": "gpt-4o-mini"},
+                "sys_prompt": {"type": "string", "default": ""},
+                "retry": {"type": "integer", "default": 3},
+            },
+            required_fields=["prompt"],
+        )
+
+        with patch(
+            "backend.api.features.chat.tools.run_block.get_block",
+            return_value=mock_block,
+        ):
+            tool = RunBlockTool()
+
+            response = await tool._execute(
+                user_id=_TEST_USER_ID,
+                session=session,
+                block_id="ai-text-gen-id",
+                input_data={
+                    "prompt": "Hello",  # correct
+                    "llm_model": "claude-opus-4-6",  # WRONG - should be 'model'
+                    "system_prompt": "Be helpful",  # WRONG - should be 'sys_prompt'
+                    "retries": 5,  # WRONG - should be 'retry'
+                },
+            )
+
+        assert isinstance(response, InputValidationErrorResponse)
+        assert set(response.unrecognized_fields) == {
+            "llm_model",
+            "system_prompt",
+            "retries",
+        }
+        assert "Block was not executed" in response.message
+
+    @pytest.mark.asyncio(loop_scope="session")
+    async def test_unknown_fields_rejected_even_with_missing_required(self):
+        """Unknown fields are caught before the missing-required-fields check."""
+        session = make_session(user_id=_TEST_USER_ID)
+
+        mock_block = make_mock_block_with_schema(
+            block_id="ai-text-gen-id",
+            name="AI Text Generator",
+            input_properties={
+                "prompt": {"type": "string"},
+                "model": {"type": "string", "default": "gpt-4o-mini"},
+            },
+            required_fields=["prompt"],
+        )
+
+        with patch(
+            "backend.api.features.chat.tools.run_block.get_block",
+            return_value=mock_block,
+        ):
+            tool = RunBlockTool()
+
+            # 'prompt' is missing AND 'LLM_Model' is an unknown field
+            response = await tool._execute(
+                user_id=_TEST_USER_ID,
+                session=session,
+                block_id="ai-text-gen-id",
+                input_data={
+                    "LLM_Model": "claude-opus-4-6",  # wrong key, and 'prompt' is missing
+                },
+            )
+
+        # Unknown fields are caught first
+        assert isinstance(response, InputValidationErrorResponse)
+        assert "LLM_Model" in response.unrecognized_fields
+
+    @pytest.mark.asyncio(loop_scope="session")
+    async def test_correct_inputs_still_execute(self):
+        """Correct input field names pass validation and the block executes."""
+        session = make_session(user_id=_TEST_USER_ID)
+
+        mock_block = make_mock_block_with_schema(
+            block_id="ai-text-gen-id",
+            name="AI Text Generator",
+            input_properties={
+                "prompt": {"type": "string"},
+                "model": {"type": "string", "default": "gpt-4o-mini"},
+            },
+            required_fields=["prompt"],
+        )
+
+        async def mock_execute(input_data, **kwargs):
+            yield "response", "Generated text"
+
+        mock_block.execute = mock_execute
+
+        with (
+            patch(
+                "backend.api.features.chat.tools.run_block.get_block",
+                return_value=mock_block,
+            ),
+            patch(
+                "backend.api.features.chat.tools.run_block.get_or_create_workspace",
+                new_callable=AsyncMock,
+                return_value=MagicMock(id="test-workspace-id"),
+            ),
+        ):
+            tool = RunBlockTool()
+
+            response = await tool._execute(
+                user_id=_TEST_USER_ID,
+                session=session,
+                block_id="ai-text-gen-id",
+                input_data={
+                    "prompt": "Write a haiku",
+                    "model": "gpt-4o-mini",  # correct field name
+                },
+            )
+
+        assert isinstance(response, BlockOutputResponse)
+        assert response.success is True
+
+    @pytest.mark.asyncio(loop_scope="session")
+    async def test_missing_required_fields_returns_details(self):
+        """Missing required fields returns BlockDetailsResponse with schema."""
+        session = make_session(user_id=_TEST_USER_ID)
+
+        mock_block = make_mock_block_with_schema(
+            block_id="ai-text-gen-id",
+            name="AI Text Generator",
+            input_properties={
+                "prompt": {"type": "string"},
+                "model": {"type": "string", "default": "gpt-4o-mini"},
+            },
+            required_fields=["prompt"],
+        )
+
+        with patch(
+            "backend.api.features.chat.tools.run_block.get_block",
+            return_value=mock_block,
+        ):
+            tool = RunBlockTool()
+
+            # Only provide valid optional field, missing required 'prompt'
+            response = await tool._execute(
+                user_id=_TEST_USER_ID,
+                session=session,
+                block_id="ai-text-gen-id",
+                input_data={
+                    "model": "gpt-4o-mini",  # valid but optional
+                },
+            )
+
+        assert isinstance(response, BlockDetailsResponse)

autogpt_platform/backend/backend/api/features/chat/tools/sandbox.py

@@ -0,0 +1,265 @@
+"""Sandbox execution utilities for code execution tools.
+
+Provides filesystem + network isolated command execution using **bubblewrap**
+(``bwrap``): whitelist-only filesystem (only system dirs visible read-only),
+writable workspace only, clean environment, network blocked.
+
+Tools that call :func:`run_sandboxed` must first check :func:`has_full_sandbox`
+and refuse to run if bubblewrap is not available.
+"""
+
+import asyncio
+import logging
+import os
+import platform
+import shutil
+
+logger = logging.getLogger(__name__)
+
+_DEFAULT_TIMEOUT = 30
+_MAX_TIMEOUT = 120
+
+
+# ---------------------------------------------------------------------------
+# Sandbox capability detection (cached at first call)
+# ---------------------------------------------------------------------------
+
+_BWRAP_AVAILABLE: bool | None = None
+
+
+def has_full_sandbox() -> bool:
+    """Return True if bubblewrap is available (filesystem + network isolation).
+
+    On non-Linux platforms (macOS), always returns False.
+    """
+    global _BWRAP_AVAILABLE
+    if _BWRAP_AVAILABLE is None:
+        _BWRAP_AVAILABLE = (
+            platform.system() == "Linux" and shutil.which("bwrap") is not None
+        )
+    return _BWRAP_AVAILABLE
+
+
+WORKSPACE_PREFIX = "/tmp/copilot-"
+
+
+def make_session_path(session_id: str) -> str:
+    """Build a sanitized, session-specific path under :data:`WORKSPACE_PREFIX`.
+
+    Shared by both the SDK working-directory setup and the sandbox tools so
+    they always resolve to the same directory for a given session.
+
+    Steps:
+        1. Strip all characters except ``[A-Za-z0-9-]``.
+        2. Construct ``/tmp/copilot-<safe_id>``.
+        3. Validate via ``os.path.normpath`` + ``startswith`` (CodeQL-recognised
+           sanitizer) to prevent path traversal.
+
+    Raises:
+        ValueError: If the resulting path escapes the prefix.
+    """
+    import re
+
+    safe_id = re.sub(r"[^A-Za-z0-9-]", "", session_id)
+    if not safe_id:
+        safe_id = "default"
+    path = os.path.normpath(f"{WORKSPACE_PREFIX}{safe_id}")
+    if not path.startswith(WORKSPACE_PREFIX):
+        raise ValueError(f"Session path escaped prefix: {path}")
+    return path
+
+
+def get_workspace_dir(session_id: str) -> str:
+    """Get or create the workspace directory for a session.
+
+    Uses :func:`make_session_path` — the same path the SDK uses — so that
+    bash_exec shares the workspace with the SDK file tools.
+    """
+    workspace = make_session_path(session_id)
+    os.makedirs(workspace, exist_ok=True)
+    return workspace
+
+
+# ---------------------------------------------------------------------------
+# Bubblewrap command builder
+# ---------------------------------------------------------------------------
+
+# System directories mounted read-only inside the sandbox.
+# ONLY these are visible — /app, /root, /home, /opt, /var etc. are NOT accessible.
+_SYSTEM_RO_BINDS = [
+    "/usr",  # binaries, libraries, Python interpreter
+    "/etc",  # system config: ld.so, locale, passwd, alternatives
+]
+
+# Compat paths: symlinks to /usr/* on modern Debian, real dirs on older systems.
+# On Debian 13 these are symlinks (e.g. /bin -> usr/bin).  bwrap --ro-bind
+# can't create a symlink target, so we detect and use --symlink instead.
+# /lib64 is critical: the ELF dynamic linker lives at /lib64/ld-linux-x86-64.so.2.
+_COMPAT_PATHS = [
+    ("/bin", "usr/bin"),  # -> /usr/bin on Debian 13
+    ("/sbin", "usr/sbin"),  # -> /usr/sbin on Debian 13
+    ("/lib", "usr/lib"),  # -> /usr/lib on Debian 13
+    ("/lib64", "usr/lib64"),  # 64-bit libraries / ELF interpreter
+]
+
+# Resource limits to prevent fork bombs, memory exhaustion, and disk abuse.
+# Applied via ulimit inside the sandbox before exec'ing the user command.
+_RESOURCE_LIMITS = (
+    "ulimit -u 64"  # max 64 processes  (prevents fork bombs)
+    " -v 524288"  # 512 MB virtual memory
+    " -f 51200"  # 50 MB max file size  (1024-byte blocks)
+    " -n 256"  # 256 open file descriptors
+    " 2>/dev/null"
+)
+
+
+def _build_bwrap_command(
+    command: list[str], cwd: str, env: dict[str, str]
+) -> list[str]:
+    """Build a bubblewrap command with strict filesystem + network isolation.
+
+    Security model:
+    - **Whitelist-only filesystem**: only system directories (``/usr``, ``/etc``,
+      ``/bin``, ``/lib``) are mounted read-only.  Application code (``/app``),
+      home directories, ``/var``, ``/opt``, etc. are NOT accessible at all.
+    - **Writable workspace only**: the per-session workspace is the sole
+      writable path.
+    - **Clean environment**: ``--clearenv`` wipes all inherited env vars.
+      Only the explicitly-passed safe env vars are set inside the sandbox.
+    - **Network isolation**: ``--unshare-net`` blocks all network access.
+    - **Resource limits**: ulimit caps on processes (64), memory (512MB),
+      file size (50MB), and open FDs (256) to prevent fork bombs and abuse.
+    - **New session**: prevents terminal control escape.
+    - **Die with parent**: prevents orphaned sandbox processes.
+    """
+    cmd = [
+        "bwrap",
+        # Create a new user namespace so bwrap can set up sandboxing
+        # inside unprivileged Docker containers (no CAP_SYS_ADMIN needed).
+        "--unshare-user",
+        # Wipe all inherited environment variables (API keys, secrets, etc.)
+        "--clearenv",
+    ]
+
+    # Set only the safe env vars inside the sandbox
+    for key, value in env.items():
+        cmd.extend(["--setenv", key, value])
+
+    # System directories: read-only
+    for path in _SYSTEM_RO_BINDS:
+        cmd.extend(["--ro-bind", path, path])
+
+    # Compat paths: use --symlink when host path is a symlink (Debian 13),
+    # --ro-bind when it's a real directory (older distros).
+    for path, symlink_target in _COMPAT_PATHS:
+        if os.path.islink(path):
+            cmd.extend(["--symlink", symlink_target, path])
+        elif os.path.exists(path):
+            cmd.extend(["--ro-bind", path, path])
+
+    # Wrap the user command with resource limits:
+    #   sh -c 'ulimit ...; exec "$@"' -- <original command>
+    # `exec "$@"` replaces the shell so there's no extra process overhead,
+    # and properly handles arguments with spaces.
+    limited_command = [
+        "sh",
+        "-c",
+        f'{_RESOURCE_LIMITS}; exec "$@"',
+        "--",
+        *command,
+    ]
+
+    cmd.extend(
+        [
+            # Fresh virtual filesystems
+            "--dev",
+            "/dev",
+            "--proc",
+            "/proc",
+            "--tmpfs",
+            "/tmp",
+            # Workspace bind AFTER --tmpfs /tmp so it's visible through the tmpfs.
+            # (workspace lives under /tmp/copilot-<session>)
+            "--bind",
+            cwd,
+            cwd,
+            # Isolation
+            "--unshare-net",
+            "--die-with-parent",
+            "--new-session",
+            "--chdir",
+            cwd,
+            "--",
+            *limited_command,
+        ]
+    )
+
+    return cmd
+
+
+# ---------------------------------------------------------------------------
+# Public API
+# ---------------------------------------------------------------------------
+
+
+async def run_sandboxed(
+    command: list[str],
+    cwd: str,
+    timeout: int = _DEFAULT_TIMEOUT,
+    env: dict[str, str] | None = None,
+) -> tuple[str, str, int, bool]:
+    """Run a command inside a bubblewrap sandbox.
+
+    Callers **must** check :func:`has_full_sandbox` before calling this
+    function.  If bubblewrap is not available, this function raises
+    :class:`RuntimeError` rather than running unsandboxed.
+
+    Returns:
+        (stdout, stderr, exit_code, timed_out)
+    """
+    if not has_full_sandbox():
+        raise RuntimeError(
+            "run_sandboxed() requires bubblewrap but bwrap is not available. "
+            "Callers must check has_full_sandbox() before calling this function."
+        )
+
+    timeout = min(max(timeout, 1), _MAX_TIMEOUT)
+
+    safe_env = {
+        "PATH": "/usr/local/bin:/usr/bin:/bin",
+        "HOME": cwd,
+        "TMPDIR": cwd,
+        "LANG": "en_US.UTF-8",
+        "PYTHONDONTWRITEBYTECODE": "1",
+        "PYTHONIOENCODING": "utf-8",
+    }
+    if env:
+        safe_env.update(env)
+
+    full_command = _build_bwrap_command(command, cwd, safe_env)
+
+    try:
+        proc = await asyncio.create_subprocess_exec(
+            *full_command,
+            stdout=asyncio.subprocess.PIPE,
+            stderr=asyncio.subprocess.PIPE,
+            cwd=cwd,
+            env=safe_env,
+        )
+
+        try:
+            stdout_bytes, stderr_bytes = await asyncio.wait_for(
+                proc.communicate(), timeout=timeout
+            )
+            stdout = stdout_bytes.decode("utf-8", errors="replace")
+            stderr = stderr_bytes.decode("utf-8", errors="replace")
+            return stdout, stderr, proc.returncode or 0, False
+        except asyncio.TimeoutError:
+            proc.kill()
+            await proc.communicate()
+            return "", f"Execution timed out after {timeout}s", -1, True
+
+    except RuntimeError:
+        raise
+    except Exception as e:
+        return "", f"Sandbox error: {e}", -1, False

autogpt_platform/backend/backend/api/features/chat/tools/test_run_block_details.py

@@ -0,0 +1,153 @@
+"""Tests for BlockDetailsResponse in RunBlockTool."""
+
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+
+from backend.api.features.chat.tools.models import BlockDetailsResponse
+from backend.api.features.chat.tools.run_block import RunBlockTool
+from backend.blocks._base import BlockType
+from backend.data.model import CredentialsMetaInput
+from backend.integrations.providers import ProviderName
+
+from ._test_data import make_session
+
+_TEST_USER_ID = "test-user-run-block-details"
+
+
+def make_mock_block_with_inputs(
+    block_id: str, name: str, description: str = "Test description"
+):
+    """Create a mock block with input/output schemas for testing."""
+    mock = MagicMock()
+    mock.id = block_id
+    mock.name = name
+    mock.description = description
+    mock.block_type = BlockType.STANDARD
+    mock.disabled = False
+
+    # Input schema with non-credential fields
+    mock.input_schema = MagicMock()
+    mock.input_schema.jsonschema.return_value = {
+        "properties": {
+            "url": {"type": "string", "description": "URL to fetch"},
+            "method": {"type": "string", "description": "HTTP method"},
+        },
+        "required": ["url"],
+    }
+    mock.input_schema.get_credentials_fields.return_value = {}
+    mock.input_schema.get_credentials_fields_info.return_value = {}
+
+    # Output schema
+    mock.output_schema = MagicMock()
+    mock.output_schema.jsonschema.return_value = {
+        "properties": {
+            "response": {"type": "object", "description": "HTTP response"},
+            "error": {"type": "string", "description": "Error message"},
+        }
+    }
+
+    return mock
+
+
+@pytest.mark.asyncio(loop_scope="session")
+async def test_run_block_returns_details_when_no_input_provided():
+    """When run_block is called without input_data, it should return BlockDetailsResponse."""
+    session = make_session(user_id=_TEST_USER_ID)
+
+    # Create a block with inputs
+    http_block = make_mock_block_with_inputs(
+        "http-block-id", "HTTP Request", "Send HTTP requests"
+    )
+
+    with patch(
+        "backend.api.features.chat.tools.run_block.get_block",
+        return_value=http_block,
+    ):
+        # Mock credentials check to return no missing credentials
+        with patch.object(
+            RunBlockTool,
+            "_resolve_block_credentials",
+            new_callable=AsyncMock,
+            return_value=({}, []),  # (matched_credentials, missing_credentials)
+        ):
+            tool = RunBlockTool()
+            response = await tool._execute(
+                user_id=_TEST_USER_ID,
+                session=session,
+                block_id="http-block-id",
+                input_data={},  # Empty input data
+            )
+
+    # Should return BlockDetailsResponse showing the schema
+    assert isinstance(response, BlockDetailsResponse)
+    assert response.block.id == "http-block-id"
+    assert response.block.name == "HTTP Request"
+    assert response.block.description == "Send HTTP requests"
+    assert "url" in response.block.inputs["properties"]
+    assert "method" in response.block.inputs["properties"]
+    assert "response" in response.block.outputs["properties"]
+    assert response.user_authenticated is True
+
+
+@pytest.mark.asyncio(loop_scope="session")
+async def test_run_block_returns_details_when_only_credentials_provided():
+    """When only credentials are provided (no actual input), should return details."""
+    session = make_session(user_id=_TEST_USER_ID)
+
+    # Create a block with both credential and non-credential inputs
+    mock = MagicMock()
+    mock.id = "api-block-id"
+    mock.name = "API Call"
+    mock.description = "Make API calls"
+    mock.block_type = BlockType.STANDARD
+    mock.disabled = False
+
+    mock.input_schema = MagicMock()
+    mock.input_schema.jsonschema.return_value = {
+        "properties": {
+            "credentials": {"type": "object", "description": "API credentials"},
+            "endpoint": {"type": "string", "description": "API endpoint"},
+        },
+        "required": ["credentials", "endpoint"],
+    }
+    mock.input_schema.get_credentials_fields.return_value = {"credentials": True}
+    mock.input_schema.get_credentials_fields_info.return_value = {}
+
+    mock.output_schema = MagicMock()
+    mock.output_schema.jsonschema.return_value = {
+        "properties": {"result": {"type": "object"}}
+    }
+
+    with patch(
+        "backend.api.features.chat.tools.run_block.get_block",
+        return_value=mock,
+    ):
+        with patch.object(
+            RunBlockTool,
+            "_resolve_block_credentials",
+            new_callable=AsyncMock,
+            return_value=(
+                {
+                    "credentials": CredentialsMetaInput(
+                        id="cred-id",
+                        provider=ProviderName("test_provider"),
+                        type="api_key",
+                        title="Test Credential",
+                    )
+                },
+                [],
+            ),
+        ):
+            tool = RunBlockTool()
+            response = await tool._execute(
+                user_id=_TEST_USER_ID,
+                session=session,
+                block_id="api-block-id",
+                input_data={"credentials": {"some": "cred"}},  # Only credential
+            )
+
+    # Should return details because no non-credential inputs provided
+    assert isinstance(response, BlockDetailsResponse)
+    assert response.block.id == "api-block-id"
+    assert response.block.name == "API Call"

autogpt_platform/backend/backend/api/features/chat/tools/web_fetch.py

@@ -0,0 +1,151 @@
+"""Web fetch tool — safely retrieve public web page content."""
+
+import logging
+from typing import Any
+
+import aiohttp
+import html2text
+
+from backend.api.features.chat.model import ChatSession
+from backend.api.features.chat.tools.base import BaseTool
+from backend.api.features.chat.tools.models import (
+    ErrorResponse,
+    ToolResponseBase,
+    WebFetchResponse,
+)
+from backend.util.request import Requests
+
+logger = logging.getLogger(__name__)
+
+# Limits
+_MAX_CONTENT_BYTES = 102_400  # 100 KB download cap
+_REQUEST_TIMEOUT = aiohttp.ClientTimeout(total=15)
+
+# Content types we'll read as text
+_TEXT_CONTENT_TYPES = {
+    "text/html",
+    "text/plain",
+    "text/xml",
+    "text/csv",
+    "text/markdown",
+    "application/json",
+    "application/xml",
+    "application/xhtml+xml",
+    "application/rss+xml",
+    "application/atom+xml",
+}
+
+
+def _is_text_content(content_type: str) -> bool:
+    base = content_type.split(";")[0].strip().lower()
+    return base in _TEXT_CONTENT_TYPES or base.startswith("text/")
+
+
+def _html_to_text(html: str) -> str:
+    h = html2text.HTML2Text()
+    h.ignore_links = False
+    h.ignore_images = True
+    h.body_width = 0
+    return h.handle(html)
+
+
+class WebFetchTool(BaseTool):
+    """Safely fetch content from a public URL using SSRF-protected HTTP."""
+
+    @property
+    def name(self) -> str:
+        return "web_fetch"
+
+    @property
+    def description(self) -> str:
+        return (
+            "Fetch the content of a public web page by URL. "
+            "Returns readable text extracted from HTML by default. "
+            "Useful for reading documentation, articles, and API responses. "
+            "Only supports HTTP/HTTPS GET requests to public URLs "
+            "(private/internal network addresses are blocked)."
+        )
+
+    @property
+    def parameters(self) -> dict[str, Any]:
+        return {
+            "type": "object",
+            "properties": {
+                "url": {
+                    "type": "string",
+                    "description": "The public HTTP/HTTPS URL to fetch.",
+                },
+                "extract_text": {
+                    "type": "boolean",
+                    "description": (
+                        "If true (default), extract readable text from HTML. "
+                        "If false, return raw content."
+                    ),
+                    "default": True,
+                },
+            },
+            "required": ["url"],
+        }
+
+    @property
+    def requires_auth(self) -> bool:
+        return False
+
+    async def _execute(
+        self,
+        user_id: str | None,
+        session: ChatSession,
+        **kwargs: Any,
+    ) -> ToolResponseBase:
+        url: str = (kwargs.get("url") or "").strip()
+        extract_text: bool = kwargs.get("extract_text", True)
+        session_id = session.session_id if session else None
+
+        if not url:
+            return ErrorResponse(
+                message="Please provide a URL to fetch.",
+                error="missing_url",
+                session_id=session_id,
+            )
+
+        try:
+            client = Requests(raise_for_status=False, retry_max_attempts=1)
+            response = await client.get(url, timeout=_REQUEST_TIMEOUT)
+        except ValueError as e:
+            # validate_url raises ValueError for SSRF / blocked IPs
+            return ErrorResponse(
+                message=f"URL blocked: {e}",
+                error="url_blocked",
+                session_id=session_id,
+            )
+        except Exception as e:
+            logger.warning(f"[web_fetch] Request failed for {url}: {e}")
+            return ErrorResponse(
+                message=f"Failed to fetch URL: {e}",
+                error="fetch_failed",
+                session_id=session_id,
+            )
+
+        content_type = response.headers.get("content-type", "")
+        if not _is_text_content(content_type):
+            return ErrorResponse(
+                message=f"Non-text content type: {content_type.split(';')[0]}",
+                error="unsupported_content_type",
+                session_id=session_id,
+            )
+
+        raw = response.content[:_MAX_CONTENT_BYTES]
+        text = raw.decode("utf-8", errors="replace")
+
+        if extract_text and "html" in content_type.lower():
+            text = _html_to_text(text)
+
+        return WebFetchResponse(
+            message=f"Fetched {url}",
+            url=response.url,
+            status_code=response.status,
+            content_type=content_type.split(";")[0].strip(),
+            content=text,
+            truncated=False,
+            session_id=session_id,
+        )

autogpt_platform/backend/backend/api/features/chat/tools/workspace_files.py

@@ -88,7 +88,9 @@ class ListWorkspaceFilesTool(BaseTool):
     @property
     def description(self) -> str:
         return (
-            "List files in the user's workspace. "
+            "List files in the user's persistent workspace (cloud storage). "
+            "These files survive across sessions. "
+            "For ephemeral session files, use the SDK Read/Glob tools instead. "
             "Returns file names, paths, sizes, and metadata. "
             "Optionally filter by path prefix."
         )
@@ -204,7 +206,9 @@ class ReadWorkspaceFileTool(BaseTool):
     @property
     def description(self) -> str:
         return (
-            "Read a file from the user's workspace. "
+            "Read a file from the user's persistent workspace (cloud storage). "
+            "These files survive across sessions. "
+            "For ephemeral session files, use the SDK Read tool instead. "
             "Specify either file_id or path to identify the file. "
             "For small text files, returns content directly. "
             "For large or binary files, returns metadata and a download URL. "
@@ -378,7 +382,9 @@ class WriteWorkspaceFileTool(BaseTool):
     @property
     def description(self) -> str:
         return (
-            "Write or create a file in the user's workspace. "
+            "Write or create a file in the user's persistent workspace (cloud storage). "
+            "These files survive across sessions. "
+            "For ephemeral session files, use the SDK Write tool instead. "
             "Provide the content as a base64-encoded string. "
             f"Maximum file size is {Config().max_file_size_mb}MB. "
             "Files are saved to the current session's folder by default. "
@@ -523,7 +529,7 @@ class DeleteWorkspaceFileTool(BaseTool):
     @property
     def description(self) -> str:
         return (
-            "Delete a file from the user's workspace. "
+            "Delete a file from the user's persistent workspace (cloud storage). "
             "Specify either file_id or path to identify the file. "
             "Paths are scoped to the current session by default. "
             "Use /sessions/<session_id>/... for cross-session access."

autogpt_platform/backend/backend/blocks/claude_code.py

@@ -1,10 +1,10 @@
 import json
 import shlex
 import uuid
-from typing import Literal, Optional
+from typing import TYPE_CHECKING, Literal, Optional
 
 from e2b import AsyncSandbox as BaseAsyncSandbox
-from pydantic import BaseModel, SecretStr
+from pydantic import SecretStr
 
 from backend.blocks._base import (
     Block,
@@ -20,6 +20,13 @@ from backend.data.model import (
     SchemaField,
 )
 from backend.integrations.providers import ProviderName
+from backend.util.sandbox_files import (
+    SandboxFileOutput,
+    extract_and_store_sandbox_files,
+)
+
+if TYPE_CHECKING:
+    from backend.executor.utils import ExecutionContext
 
 
 class ClaudeCodeExecutionError(Exception):
@@ -174,22 +181,15 @@ class ClaudeCodeBlock(Block):
             advanced=True,
         )
 
-    class FileOutput(BaseModel):
-        """A file extracted from the sandbox."""
-
-        path: str
-        relative_path: str  # Path relative to working directory (for GitHub, etc.)
-        name: str
-        content: str
-
     class Output(BlockSchemaOutput):
         response: str = SchemaField(
             description="The output/response from Claude Code execution"
         )
-        files: list["ClaudeCodeBlock.FileOutput"] = SchemaField(
+        files: list[SandboxFileOutput] = SchemaField(
             description=(
                 "List of text files created/modified by Claude Code during this execution. "
-                "Each file has 'path', 'relative_path', 'name', and 'content' fields."
+                "Each file has 'path', 'relative_path', 'name', 'content', and 'workspace_ref' fields. "
+                "workspace_ref contains a workspace:// URI if the file was stored to workspace."
             )
         )
         conversation_history: str = SchemaField(
@@ -252,6 +252,7 @@ class ClaudeCodeBlock(Block):
                             "relative_path": "index.html",
                             "name": "index.html",
                             "content": "<html>Hello World</html>",
+                            "workspace_ref": None,
                         }
                     ],
                 ),
@@ -267,11 +268,12 @@ class ClaudeCodeBlock(Block):
                 "execute_claude_code": lambda *args, **kwargs: (
                     "Created index.html with hello world content",  # response
                     [
-                        ClaudeCodeBlock.FileOutput(
+                        SandboxFileOutput(
                             path="/home/user/index.html",
                             relative_path="index.html",
                             name="index.html",
                             content="<html>Hello World</html>",
+                            workspace_ref=None,
                         )
                     ],  # files
                     "User: Create a hello world HTML file\n"
@@ -294,7 +296,8 @@ class ClaudeCodeBlock(Block):
         existing_sandbox_id: str,
         conversation_history: str,
         dispose_sandbox: bool,
-    ) -> tuple[str, list["ClaudeCodeBlock.FileOutput"], str, str, str]:
+        execution_context: "ExecutionContext",
+    ) -> tuple[str, list[SandboxFileOutput], str, str, str]:
         """
         Execute Claude Code in an E2B sandbox.
 
@@ -449,14 +452,18 @@ class ClaudeCodeBlock(Block):
                 else:
                     new_conversation_history = turn_entry
 
-            # Extract files created/modified during this run
+            # Extract files created/modified during this run and store to workspace
-            files = await self._extract_files(
+            sandbox_files = await extract_and_store_sandbox_files(
-                sandbox, working_directory, start_timestamp
+                sandbox=sandbox,
+                working_directory=working_directory,
+                execution_context=execution_context,
+                since_timestamp=start_timestamp,
+                text_only=True,
             )
 
             return (
                 response,
-                files,
+                sandbox_files,  # Already SandboxFileOutput objects
                 new_conversation_history,
                 current_session_id,
                 sandbox_id,
@@ -471,140 +478,6 @@ class ClaudeCodeBlock(Block):
             if dispose_sandbox and sandbox:
                 await sandbox.kill()
 
-    async def _extract_files(
-        self,
-        sandbox: BaseAsyncSandbox,
-        working_directory: str,
-        since_timestamp: str | None = None,
-    ) -> list["ClaudeCodeBlock.FileOutput"]:
-        """
-        Extract text files created/modified during this Claude Code execution.
-
-        Args:
-            sandbox: The E2B sandbox instance
-            working_directory: Directory to search for files
-            since_timestamp: ISO timestamp - only return files modified after this time
-
-        Returns:
-            List of FileOutput objects with path, relative_path, name, and content
-        """
-        files: list[ClaudeCodeBlock.FileOutput] = []
-
-        # Text file extensions we can safely read as text
-        text_extensions = {
-            ".txt",
-            ".md",
-            ".html",
-            ".htm",
-            ".css",
-            ".js",
-            ".ts",
-            ".jsx",
-            ".tsx",
-            ".json",
-            ".xml",
-            ".yaml",
-            ".yml",
-            ".toml",
-            ".ini",
-            ".cfg",
-            ".conf",
-            ".py",
-            ".rb",
-            ".php",
-            ".java",
-            ".c",
-            ".cpp",
-            ".h",
-            ".hpp",
-            ".cs",
-            ".go",
-            ".rs",
-            ".swift",
-            ".kt",
-            ".scala",
-            ".sh",
-            ".bash",
-            ".zsh",
-            ".sql",
-            ".graphql",
-            ".env",
-            ".gitignore",
-            ".dockerfile",
-            "Dockerfile",
-            ".vue",
-            ".svelte",
-            ".astro",
-            ".mdx",
-            ".rst",
-            ".tex",
-            ".csv",
-            ".log",
-        }
-
-        try:
-            # List files recursively using find command
-            # Exclude node_modules and .git directories, but allow hidden files
-            # like .env and .gitignore (they're filtered by text_extensions later)
-            # Filter by timestamp to only get files created/modified during this run
-            safe_working_dir = shlex.quote(working_directory)
-            timestamp_filter = ""
-            if since_timestamp:
-                timestamp_filter = f"-newermt {shlex.quote(since_timestamp)} "
-            find_result = await sandbox.commands.run(
-                f"find {safe_working_dir} -type f "
-                f"{timestamp_filter}"
-                f"-not -path '*/node_modules/*' "
-                f"-not -path '*/.git/*' "
-                f"2>/dev/null"
-            )
-
-            if find_result.stdout:
-                for file_path in find_result.stdout.strip().split("\n"):
-                    if not file_path:
-                        continue
-
-                    # Check if it's a text file we can read
-                    is_text = any(
-                        file_path.endswith(ext) for ext in text_extensions
-                    ) or file_path.endswith("Dockerfile")
-
-                    if is_text:
-                        try:
-                            content = await sandbox.files.read(file_path)
-                            # Handle bytes or string
-                            if isinstance(content, bytes):
-                                content = content.decode("utf-8", errors="replace")
-
-                            # Extract filename from path
-                            file_name = file_path.split("/")[-1]
-
-                            # Calculate relative path by stripping working directory
-                            relative_path = file_path
-                            if file_path.startswith(working_directory):
-                                relative_path = file_path[len(working_directory) :]
-                                # Remove leading slash if present
-                                if relative_path.startswith("/"):
-                                    relative_path = relative_path[1:]
-
-                            files.append(
-                                ClaudeCodeBlock.FileOutput(
-                                    path=file_path,
-                                    relative_path=relative_path,
-                                    name=file_name,
-                                    content=content,
-                                )
-                            )
-                        except Exception:
-                            # Skip files that can't be read
-                            pass
-
-        except Exception:
-            # If file extraction fails, return empty results
-            pass
-
-        return files
-
     def _escape_prompt(self, prompt: str) -> str:
         """Escape the prompt for safe shell execution."""
         # Use single quotes and escape any single quotes in the prompt
@@ -617,6 +490,7 @@ class ClaudeCodeBlock(Block):
         *,
         e2b_credentials: APIKeyCredentials,
         anthropic_credentials: APIKeyCredentials,
+        execution_context: "ExecutionContext",
         **kwargs,
     ) -> BlockOutput:
         try:
@@ -637,6 +511,7 @@ class ClaudeCodeBlock(Block):
                 existing_sandbox_id=input_data.sandbox_id,
                 conversation_history=input_data.conversation_history,
                 dispose_sandbox=input_data.dispose_sandbox,
+                execution_context=execution_context,
             )
 
             yield "response", response

autogpt_platform/backend/backend/blocks/code_executor.py

@@ -1,5 +1,5 @@
 from enum import Enum
-from typing import Any, Literal, Optional
+from typing import TYPE_CHECKING, Any, Literal, Optional
 
 from e2b_code_interpreter import AsyncSandbox
 from e2b_code_interpreter import Result as E2BExecutionResult
@@ -20,6 +20,13 @@ from backend.data.model import (
     SchemaField,
 )
 from backend.integrations.providers import ProviderName
+from backend.util.sandbox_files import (
+    SandboxFileOutput,
+    extract_and_store_sandbox_files,
+)
+
+if TYPE_CHECKING:
+    from backend.executor.utils import ExecutionContext
 
 TEST_CREDENTIALS = APIKeyCredentials(
     id="01234567-89ab-cdef-0123-456789abcdef",
@@ -85,6 +92,9 @@ class CodeExecutionResult(MainCodeExecutionResult):
 class BaseE2BExecutorMixin:
     """Shared implementation methods for E2B executor blocks."""
 
+    # Default working directory in E2B sandboxes
+    WORKING_DIR = "/home/user"
+
     async def execute_code(
         self,
         api_key: str,
@@ -95,14 +105,21 @@ class BaseE2BExecutorMixin:
         timeout: Optional[int] = None,
         sandbox_id: Optional[str] = None,
         dispose_sandbox: bool = False,
+        execution_context: Optional["ExecutionContext"] = None,
+        extract_files: bool = False,
     ):
         """
         Unified code execution method that handles all three use cases:
         1. Create new sandbox and execute (ExecuteCodeBlock)
         2. Create new sandbox, execute, and return sandbox_id (InstantiateCodeSandboxBlock)
         3. Connect to existing sandbox and execute (ExecuteCodeStepBlock)
+
+        Args:
+            extract_files: If True and execution_context provided, extract files
+                           created/modified during execution and store to workspace.
         """  # noqa
         sandbox = None
+        files: list[SandboxFileOutput] = []
         try:
             if sandbox_id:
                 # Connect to existing sandbox (ExecuteCodeStepBlock case)
@@ -118,6 +135,12 @@ class BaseE2BExecutorMixin:
                     for cmd in setup_commands:
                         await sandbox.commands.run(cmd)
 
+            # Capture timestamp before execution to scope file extraction
+            start_timestamp = None
+            if extract_files:
+                ts_result = await sandbox.commands.run("date -u +%Y-%m-%dT%H:%M:%S")
+                start_timestamp = ts_result.stdout.strip() if ts_result.stdout else None
+
             # Execute the code
             execution = await sandbox.run_code(
                 code,
@@ -133,7 +156,24 @@ class BaseE2BExecutorMixin:
             stdout_logs = "".join(execution.logs.stdout)
             stderr_logs = "".join(execution.logs.stderr)
 
-            return results, text_output, stdout_logs, stderr_logs, sandbox.sandbox_id
+            # Extract files created/modified during this execution
+            if extract_files and execution_context:
+                files = await extract_and_store_sandbox_files(
+                    sandbox=sandbox,
+                    working_directory=self.WORKING_DIR,
+                    execution_context=execution_context,
+                    since_timestamp=start_timestamp,
+                    text_only=False,  # Include binary files too
+                )
+
+            return (
+                results,
+                text_output,
+                stdout_logs,
+                stderr_logs,
+                sandbox.sandbox_id,
+                files,
+            )
         finally:
             # Dispose of sandbox if requested to reduce usage costs
             if dispose_sandbox and sandbox:
@@ -238,6 +278,12 @@ class ExecuteCodeBlock(Block, BaseE2BExecutorMixin):
             description="Standard output logs from execution"
         )
         stderr_logs: str = SchemaField(description="Standard error logs from execution")
+        files: list[SandboxFileOutput] = SchemaField(
+            description=(
+                "Files created or modified during execution. "
+                "Each file has path, name, content, and workspace_ref (if stored)."
+            ),
+        )
 
     def __init__(self):
         super().__init__(
@@ -259,23 +305,30 @@ class ExecuteCodeBlock(Block, BaseE2BExecutorMixin):
                 ("results", []),
                 ("response", "Hello World"),
                 ("stdout_logs", "Hello World\n"),
+                ("files", []),
             ],
             test_mock={
-                "execute_code": lambda api_key, code, language, template_id, setup_commands, timeout, dispose_sandbox: (  # noqa
+                "execute_code": lambda api_key, code, language, template_id, setup_commands, timeout, dispose_sandbox, execution_context, extract_files: (  # noqa
                     [],  # results
                     "Hello World",  # text_output
                     "Hello World\n",  # stdout_logs
                     "",  # stderr_logs
                     "sandbox_id",  # sandbox_id
+                    [],  # files
                 ),
             },
         )
 
     async def run(
-        self, input_data: Input, *, credentials: APIKeyCredentials, **kwargs
+        self,
+        input_data: Input,
+        *,
+        credentials: APIKeyCredentials,
+        execution_context: "ExecutionContext",
+        **kwargs,
     ) -> BlockOutput:
         try:
-            results, text_output, stdout, stderr, _ = await self.execute_code(
+            results, text_output, stdout, stderr, _, files = await self.execute_code(
                 api_key=credentials.api_key.get_secret_value(),
                 code=input_data.code,
                 language=input_data.language,
@@ -283,6 +336,8 @@ class ExecuteCodeBlock(Block, BaseE2BExecutorMixin):
                 setup_commands=input_data.setup_commands,
                 timeout=input_data.timeout,
                 dispose_sandbox=input_data.dispose_sandbox,
+                execution_context=execution_context,
+                extract_files=True,
             )
 
             # Determine result object shape & filter out empty formats
@@ -296,6 +351,8 @@ class ExecuteCodeBlock(Block, BaseE2BExecutorMixin):
                 yield "stdout_logs", stdout
             if stderr:
                 yield "stderr_logs", stderr
+            # Always yield files (empty list if none)
+            yield "files", [f.model_dump() for f in files]
         except Exception as e:
             yield "error", str(e)
 
@@ -393,6 +450,7 @@ class InstantiateCodeSandboxBlock(Block, BaseE2BExecutorMixin):
                     "Hello World\n",  # stdout_logs
                     "",  # stderr_logs
                     "sandbox_id",  # sandbox_id
+                    [],  # files
                 ),
             },
         )
@@ -401,7 +459,7 @@ class InstantiateCodeSandboxBlock(Block, BaseE2BExecutorMixin):
         self, input_data: Input, *, credentials: APIKeyCredentials, **kwargs
     ) -> BlockOutput:
         try:
-            _, text_output, stdout, stderr, sandbox_id = await self.execute_code(
+            _, text_output, stdout, stderr, sandbox_id, _ = await self.execute_code(
                 api_key=credentials.api_key.get_secret_value(),
                 code=input_data.setup_code,
                 language=input_data.language,
@@ -500,6 +558,7 @@ class ExecuteCodeStepBlock(Block, BaseE2BExecutorMixin):
                     "Hello World\n",  # stdout_logs
                     "",  # stderr_logs
                     sandbox_id,  # sandbox_id
+                    [],  # files
                 ),
             },
         )
@@ -508,7 +567,7 @@ class ExecuteCodeStepBlock(Block, BaseE2BExecutorMixin):
         self, input_data: Input, *, credentials: APIKeyCredentials, **kwargs
     ) -> BlockOutput:
         try:
-            results, text_output, stdout, stderr, _ = await self.execute_code(
+            results, text_output, stdout, stderr, _, _ = await self.execute_code(
                 api_key=credentials.api_key.get_secret_value(),
                 code=input_data.step_code,
                 language=input_data.language,

autogpt_platform/backend/backend/util/feature_flag.py

@@ -38,6 +38,7 @@ class Flag(str, Enum):
     AGENT_ACTIVITY = "agent-activity"
     ENABLE_PLATFORM_PAYMENT = "enable-platform-payment"
     CHAT = "chat"
+    COPILOT_SDK = "copilot-sdk"
 
 
 def is_configured() -> bool:

autogpt_platform/backend/backend/util/sandbox_files.py

@@ -0,0 +1,288 @@
+"""
+Shared utilities for extracting and storing files from E2B sandboxes.
+
+This module provides common file extraction and workspace storage functionality
+for blocks that run code in E2B sandboxes (Claude Code, Code Executor, etc.).
+"""
+
+import base64
+import logging
+import mimetypes
+import shlex
+from dataclasses import dataclass
+from typing import TYPE_CHECKING
+
+from pydantic import BaseModel
+
+from backend.util.file import store_media_file
+from backend.util.type import MediaFileType
+
+if TYPE_CHECKING:
+    from e2b import AsyncSandbox as BaseAsyncSandbox
+
+    from backend.executor.utils import ExecutionContext
+
+logger = logging.getLogger(__name__)
+
+# Text file extensions that can be safely read and stored as text
+TEXT_EXTENSIONS = {
+    ".txt",
+    ".md",
+    ".html",
+    ".htm",
+    ".css",
+    ".js",
+    ".ts",
+    ".jsx",
+    ".tsx",
+    ".json",
+    ".xml",
+    ".yaml",
+    ".yml",
+    ".toml",
+    ".ini",
+    ".cfg",
+    ".conf",
+    ".py",
+    ".rb",
+    ".php",
+    ".java",
+    ".c",
+    ".cpp",
+    ".h",
+    ".hpp",
+    ".cs",
+    ".go",
+    ".rs",
+    ".swift",
+    ".kt",
+    ".scala",
+    ".sh",
+    ".bash",
+    ".zsh",
+    ".sql",
+    ".graphql",
+    ".env",
+    ".gitignore",
+    ".dockerfile",
+    "Dockerfile",
+    ".vue",
+    ".svelte",
+    ".astro",
+    ".mdx",
+    ".rst",
+    ".tex",
+    ".csv",
+    ".log",
+}
+
+
+class SandboxFileOutput(BaseModel):
+    """A file extracted from a sandbox and optionally stored in workspace."""
+
+    path: str
+    """Full path in the sandbox."""
+
+    relative_path: str
+    """Path relative to the working directory."""
+
+    name: str
+    """Filename only."""
+
+    content: str
+    """File content as text (for backward compatibility)."""
+
+    workspace_ref: str | None = None
+    """Workspace reference (workspace://{id}#mime) if stored, None otherwise."""
+
+
+@dataclass
+class ExtractedFile:
+    """Internal representation of an extracted file before storage."""
+
+    path: str
+    relative_path: str
+    name: str
+    content: bytes
+    is_text: bool
+
+
+async def extract_sandbox_files(
+    sandbox: "BaseAsyncSandbox",
+    working_directory: str,
+    since_timestamp: str | None = None,
+    text_only: bool = True,
+) -> list[ExtractedFile]:
+    """
+    Extract files from an E2B sandbox.
+
+    Args:
+        sandbox: The E2B sandbox instance
+        working_directory: Directory to search for files
+        since_timestamp: ISO timestamp - only return files modified after this time
+        text_only: If True, only extract text files (default). If False, extract all files.
+
+    Returns:
+        List of ExtractedFile objects with path, content, and metadata
+    """
+    files: list[ExtractedFile] = []
+
+    try:
+        # Build find command
+        safe_working_dir = shlex.quote(working_directory)
+        timestamp_filter = ""
+        if since_timestamp:
+            timestamp_filter = f"-newermt {shlex.quote(since_timestamp)} "
+
+        find_result = await sandbox.commands.run(
+            f"find {safe_working_dir} -type f "
+            f"{timestamp_filter}"
+            f"-not -path '*/node_modules/*' "
+            f"-not -path '*/.git/*' "
+            f"2>/dev/null"
+        )
+
+        if not find_result.stdout:
+            return files
+
+        for file_path in find_result.stdout.strip().split("\n"):
+            if not file_path:
+                continue
+
+            # Check if it's a text file
+            is_text = any(file_path.endswith(ext) for ext in TEXT_EXTENSIONS)
+
+            # Skip non-text files if text_only mode
+            if text_only and not is_text:
+                continue
+
+            try:
+                # Read file content as bytes
+                content = await sandbox.files.read(file_path, format="bytes")
+                if isinstance(content, str):
+                    content = content.encode("utf-8")
+                elif isinstance(content, bytearray):
+                    content = bytes(content)
+
+                # Extract filename from path
+                file_name = file_path.split("/")[-1]
+
+                # Calculate relative path
+                relative_path = file_path
+                if file_path.startswith(working_directory):
+                    relative_path = file_path[len(working_directory) :]
+                    if relative_path.startswith("/"):
+                        relative_path = relative_path[1:]
+
+                files.append(
+                    ExtractedFile(
+                        path=file_path,
+                        relative_path=relative_path,
+                        name=file_name,
+                        content=content,
+                        is_text=is_text,
+                    )
+                )
+            except Exception as e:
+                logger.debug(f"Failed to read file {file_path}: {e}")
+                continue
+
+    except Exception as e:
+        logger.warning(f"File extraction failed: {e}")
+
+    return files
+
+
+async def store_sandbox_files(
+    extracted_files: list[ExtractedFile],
+    execution_context: "ExecutionContext",
+) -> list[SandboxFileOutput]:
+    """
+    Store extracted sandbox files to workspace and return output objects.
+
+    Args:
+        extracted_files: List of files extracted from sandbox
+        execution_context: Execution context for workspace storage
+
+    Returns:
+        List of SandboxFileOutput objects with workspace refs
+    """
+    outputs: list[SandboxFileOutput] = []
+
+    for file in extracted_files:
+        # Decode content for text files (for backward compat content field)
+        if file.is_text:
+            try:
+                content_str = file.content.decode("utf-8", errors="replace")
+            except Exception:
+                content_str = ""
+        else:
+            content_str = f"[Binary file: {len(file.content)} bytes]"
+
+        # Build data URI (needed for storage and as binary fallback)
+        mime_type = mimetypes.guess_type(file.name)[0] or "application/octet-stream"
+        data_uri = f"data:{mime_type};base64,{base64.b64encode(file.content).decode()}"
+
+        # Try to store in workspace
+        workspace_ref: str | None = None
+        try:
+            result = await store_media_file(
+                file=MediaFileType(data_uri),
+                execution_context=execution_context,
+                return_format="for_block_output",
+            )
+            if result.startswith("workspace://"):
+                workspace_ref = result
+            elif not file.is_text:
+                # Non-workspace context (graph execution): store_media_file
+                # returned a data URI — use it as content so binary data isn't lost.
+                content_str = result
+        except Exception as e:
+            logger.warning(f"Failed to store file {file.name} to workspace: {e}")
+            # For binary files, fall back to data URI to prevent data loss
+            if not file.is_text:
+                content_str = data_uri
+
+        outputs.append(
+            SandboxFileOutput(
+                path=file.path,
+                relative_path=file.relative_path,
+                name=file.name,
+                content=content_str,
+                workspace_ref=workspace_ref,
+            )
+        )
+
+    return outputs
+
+
+async def extract_and_store_sandbox_files(
+    sandbox: "BaseAsyncSandbox",
+    working_directory: str,
+    execution_context: "ExecutionContext",
+    since_timestamp: str | None = None,
+    text_only: bool = True,
+) -> list[SandboxFileOutput]:
+    """
+    Extract files from sandbox and store them in workspace.
+
+    This is the main entry point combining extraction and storage.
+
+    Args:
+        sandbox: The E2B sandbox instance
+        working_directory: Directory to search for files
+        execution_context: Execution context for workspace storage
+        since_timestamp: ISO timestamp - only return files modified after this time
+        text_only: If True, only extract text files
+
+    Returns:
+        List of SandboxFileOutput objects with content and workspace refs
+    """
+    extracted = await extract_sandbox_files(
+        sandbox=sandbox,
+        working_directory=working_directory,
+        since_timestamp=since_timestamp,
+        text_only=text_only,
+    )
+
+    return await store_sandbox_files(extracted, execution_context)

autogpt_platform/backend/backend/util/settings.py

@@ -368,6 +368,10 @@ class Config(UpdateTrackingModel["Config"], BaseSettings):
         default=600,
         description="The timeout in seconds for Agent Generator service requests (includes retries for rate limits)",
     )
+    agentgenerator_use_dummy: bool = Field(
+        default=False,
+        description="Use dummy agent generator responses for testing (bypasses external service)",
+    )
 
     enable_example_blocks: bool = Field(
         default=False,

autogpt_platform/backend/poetry.lock

@@ -441,14 +441,14 @@ develop = true
 colorama = "^0.4.6"
 cryptography = "^46.0"
 expiringdict = "^1.2.2"
-fastapi = "^0.128.0"
+fastapi = "^0.128.7"
 google-cloud-logging = "^3.13.0"
-launchdarkly-server-sdk = "^9.14.1"
+launchdarkly-server-sdk = "^9.15.0"
 pydantic = "^2.12.5"
 pydantic-settings = "^2.12.0"
 pyjwt = {version = "^2.11.0", extras = ["crypto"]}
 redis = "^6.2.0"
-supabase = "^2.27.2"
+supabase = "^2.28.0"
 uvicorn = "^0.40.0"
 
 [package.source]
@@ -897,6 +897,29 @@ files = [
     {file = "charset_normalizer-3.4.4.tar.gz", hash = "sha256:94537985111c35f28720e43603b8e7b43a6ecfb2ce1d3058bbe955b73404e21a"},
 ]
 
+[[package]]
+name = "claude-agent-sdk"
+version = "0.1.35"
+description = "Python SDK for Claude Code"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "claude_agent_sdk-0.1.35-py3-none-macosx_11_0_arm64.whl", hash = "sha256:df67f4deade77b16a9678b3a626c176498e40417f33b04beda9628287f375591"},
+    {file = "claude_agent_sdk-0.1.35-py3-none-manylinux_2_17_aarch64.whl", hash = "sha256:14963944f55ded7c8ed518feebfa5b4284aa6dd8d81aeff2e5b21a962ce65097"},
+    {file = "claude_agent_sdk-0.1.35-py3-none-manylinux_2_17_x86_64.whl", hash = "sha256:84344dcc535d179c1fc8a11c6f34c37c3b583447bdf09d869effb26514fd7a65"},
+    {file = "claude_agent_sdk-0.1.35-py3-none-win_amd64.whl", hash = "sha256:1b3d54b47448c93f6f372acd4d1757f047c3c1e8ef5804be7a1e3e53e2c79a5f"},
+    {file = "claude_agent_sdk-0.1.35.tar.gz", hash = "sha256:0f98e2b3c71ca85abfc042e7a35c648df88e87fda41c52e6779ef7b038dcbb52"},
+]
+
+[package.dependencies]
+anyio = ">=4.0.0"
+mcp = ">=0.1.0"
+typing-extensions = {version = ">=4.0.0", markers = "python_version < \"3.11\""}
+
+[package.extras]
+dev = ["anyio[trio] (>=4.0.0)", "mypy (>=1.0.0)", "pytest (>=7.0.0)", "pytest-asyncio (>=0.20.0)", "pytest-cov (>=4.0.0)", "ruff (>=0.1.0)"]
+
 [[package]]
 name = "cleo"
 version = "2.1.0"
@@ -1382,14 +1405,14 @@ tzdata = "*"
 
 [[package]]
 name = "fastapi"
-version = "0.128.6"
+version = "0.128.7"
 description = "FastAPI framework, high performance, easy to learn, fast to code, ready for production"
 optional = false
 python-versions = ">=3.9"
 groups = ["main"]
 files = [
-    {file = "fastapi-0.128.6-py3-none-any.whl", hash = "sha256:bb1c1ef87d6086a7132d0ab60869d6f1ee67283b20fbf84ec0003bd335099509"},
+    {file = "fastapi-0.128.7-py3-none-any.whl", hash = "sha256:6bd9bd31cb7047465f2d3fa3ba3f33b0870b17d4eaf7cdb36d1576ab060ad662"},
-    {file = "fastapi-0.128.6.tar.gz", hash = "sha256:0cb3946557e792d731b26a42b04912f16367e3c3135ea8290f620e234f2b604f"},
+    {file = "fastapi-0.128.7.tar.gz", hash = "sha256:783c273416995486c155ad2c0e2b45905dedfaf20b9ef8d9f6a9124670639a24"},
 ]
 
 [package.dependencies]
@@ -2593,6 +2616,18 @@ http2 = ["h2 (>=3,<5)"]
 socks = ["socksio (==1.*)"]
 zstd = ["zstandard (>=0.18.0)"]
 
+[[package]]
+name = "httpx-sse"
+version = "0.4.3"
+description = "Consume Server-Sent Event (SSE) messages with HTTPX."
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "httpx_sse-0.4.3-py3-none-any.whl", hash = "sha256:0ac1c9fe3c0afad2e0ebb25a934a59f4c7823b60792691f779fad2c5568830fc"},
+    {file = "httpx_sse-0.4.3.tar.gz", hash = "sha256:9b1ed0127459a66014aec3c56bebd93da3c1bc8bb6618c8082039a44889a755d"},
+]
+
 [[package]]
 name = "huggingface-hub"
 version = "1.4.1"
@@ -3117,14 +3152,14 @@ urllib3 = ">=1.26.0,<3"
 
 [[package]]
 name = "launchdarkly-server-sdk"
-version = "9.14.1"
+version = "9.15.0"
 description = "LaunchDarkly SDK for Python"
 optional = false
-python-versions = ">=3.9"
+python-versions = ">=3.10"
 groups = ["main"]
 files = [
-    {file = "launchdarkly_server_sdk-9.14.1-py3-none-any.whl", hash = "sha256:a9e2bd9ecdef845cd631ae0d4334a1115e5b44257c42eb2349492be4bac7815c"},
+    {file = "launchdarkly_server_sdk-9.15.0-py3-none-any.whl", hash = "sha256:c267e29bfa3fb5e2a06a208448ada6ed5557a2924979b8d79c970b45d227c668"},
-    {file = "launchdarkly_server_sdk-9.14.1.tar.gz", hash = "sha256:1df44baf0a0efa74d8c1dad7a00592b98bce7d19edded7f770da8dbc49922213"},
+    {file = "launchdarkly_server_sdk-9.15.0.tar.gz", hash = "sha256:f31441b74bc1a69c381db57c33116509e407a2612628ad6dff0a7dbb39d5020b"},
 ]
 
 [package.dependencies]
@@ -3310,6 +3345,39 @@ files = [
     {file = "mccabe-0.7.0.tar.gz", hash = "sha256:348e0240c33b60bbdf4e523192ef919f28cb2c3d7d5c7794f74009290f236325"},
 ]
 
+[[package]]
+name = "mcp"
+version = "1.26.0"
+description = "Model Context Protocol SDK"
+optional = false
+python-versions = ">=3.10"
+groups = ["main"]
+files = [
+    {file = "mcp-1.26.0-py3-none-any.whl", hash = "sha256:904a21c33c25aa98ddbeb47273033c435e595bbacfdb177f4bd87f6dceebe1ca"},
+    {file = "mcp-1.26.0.tar.gz", hash = "sha256:db6e2ef491eecc1a0d93711a76f28dec2e05999f93afd48795da1c1137142c66"},
+]
+
+[package.dependencies]
+anyio = ">=4.5"
+httpx = ">=0.27.1"
+httpx-sse = ">=0.4"
+jsonschema = ">=4.20.0"
+pydantic = ">=2.11.0,<3.0.0"
+pydantic-settings = ">=2.5.2"
+pyjwt = {version = ">=2.10.1", extras = ["crypto"]}
+python-multipart = ">=0.0.9"
+pywin32 = {version = ">=310", markers = "sys_platform == \"win32\""}
+sse-starlette = ">=1.6.1"
+starlette = ">=0.27"
+typing-extensions = ">=4.9.0"
+typing-inspection = ">=0.4.1"
+uvicorn = {version = ">=0.31.1", markers = "sys_platform != \"emscripten\""}
+
+[package.extras]
+cli = ["python-dotenv (>=1.0.0)", "typer (>=0.16.0)"]
+rich = ["rich (>=13.9.4)"]
+ws = ["websockets (>=15.0.1)"]
+
 [[package]]
 name = "mdurl"
 version = "0.1.2"
@@ -4728,14 +4796,14 @@ tests = ["coverage-conditional-plugin (>=0.9.0)", "portalocker[redis]", "pytest
 
 [[package]]
 name = "postgrest"
-version = "2.27.3"
+version = "2.28.0"
 description = "PostgREST client for Python. This library provides an ORM interface to PostgREST."
 optional = false
 python-versions = ">=3.9"
 groups = ["main"]
 files = [
-    {file = "postgrest-2.27.3-py3-none-any.whl", hash = "sha256:ed79123af7127edd78d538bfe8351d277e45b1a36994a4dbf57ae27dde87a7b7"},
+    {file = "postgrest-2.28.0-py3-none-any.whl", hash = "sha256:7bca2f24dd1a1bf8a3d586c7482aba6cd41662da6733045fad585b63b7f7df75"},
-    {file = "postgrest-2.27.3.tar.gz", hash = "sha256:c2e2679addfc8eaab23197bad7ddaee6cbb4cbe8c483ebd2d2e5219543037cc3"},
+    {file = "postgrest-2.28.0.tar.gz", hash = "sha256:c36b38646d25ea4255321d3d924ce70f8d20ec7799cb42c1221d6a818d4f6515"},
 ]
 
 [package.dependencies]
@@ -5994,7 +6062,7 @@ description = "Python for Window Extensions"
 optional = false
 python-versions = "*"
 groups = ["main"]
-markers = "platform_system == \"Windows\""
+markers = "sys_platform == \"win32\" or platform_system == \"Windows\""
 files = [
     {file = "pywin32-311-cp310-cp310-win32.whl", hash = "sha256:d03ff496d2a0cd4a5893504789d4a15399133fe82517455e78bad62efbb7f0a3"},
     {file = "pywin32-311-cp310-cp310-win_amd64.whl", hash = "sha256:797c2772017851984b97180b0bebe4b620bb86328e8a884bb626156295a63b3b"},
@@ -6260,14 +6328,14 @@ all = ["numpy"]
 
 [[package]]
 name = "realtime"
-version = "2.27.3"
+version = "2.28.0"
 description = ""
 optional = false
 python-versions = ">=3.9"
 groups = ["main"]
 files = [
-    {file = "realtime-2.27.3-py3-none-any.whl", hash = "sha256:f571115f86988e33c41c895cb3fba2eaa1b693aeaede3617288f44274ca90f43"},
+    {file = "realtime-2.28.0-py3-none-any.whl", hash = "sha256:db1bd59bab9b1fcc9f9d3b1a073bed35bf4994d720e6751f10031a58d57a3836"},
-    {file = "realtime-2.27.3.tar.gz", hash = "sha256:02b082243107656a5ef3fb63e8e2ab4c40bc199abb45adb8a42ed63f089a1041"},
+    {file = "realtime-2.28.0.tar.gz", hash = "sha256:d18cedcebd6a8f22fcd509bc767f639761eb218b7b2b6f14fc4205b6259b50fc"},
 ]
 
 [package.dependencies]
@@ -6974,6 +7042,28 @@ postgresql-psycopgbinary = ["psycopg[binary] (>=3.0.7)"]
 pymysql = ["pymysql"]
 sqlcipher = ["sqlcipher3_binary"]
 
+[[package]]
+name = "sse-starlette"
+version = "3.2.0"
+description = "SSE plugin for Starlette"
+optional = false
+python-versions = ">=3.9"
+groups = ["main"]
+files = [
+    {file = "sse_starlette-3.2.0-py3-none-any.whl", hash = "sha256:5876954bd51920fc2cd51baee47a080eb88a37b5b784e615abb0b283f801cdbf"},
+    {file = "sse_starlette-3.2.0.tar.gz", hash = "sha256:8127594edfb51abe44eac9c49e59b0b01f1039d0c7461c6fd91d4e03b70da422"},
+]
+
+[package.dependencies]
+anyio = ">=4.7.0"
+starlette = ">=0.49.1"
+
+[package.extras]
+daphne = ["daphne (>=4.2.0)"]
+examples = ["aiosqlite (>=0.21.0)", "fastapi (>=0.115.12)", "sqlalchemy[asyncio] (>=2.0.41)", "uvicorn (>=0.34.0)"]
+granian = ["granian (>=2.3.1)"]
+uvicorn = ["uvicorn (>=0.34.0)"]
+
 [[package]]
 name = "stagehand"
 version = "0.5.9"
@@ -7024,14 +7114,14 @@ full = ["httpx (>=0.27.0,<0.29.0)", "itsdangerous", "jinja2", "python-multipart
 
 [[package]]
 name = "storage3"
-version = "2.27.3"
+version = "2.28.0"
 description = "Supabase Storage client for Python."
 optional = false
 python-versions = ">=3.9"
 groups = ["main"]
 files = [
-    {file = "storage3-2.27.3-py3-none-any.whl", hash = "sha256:11a05b7da84bccabeeea12d940bca3760cf63fe6ca441868677335cfe4fdfbe0"},
+    {file = "storage3-2.28.0-py3-none-any.whl", hash = "sha256:ecb50efd2ac71dabbdf97e99ad346eafa630c4c627a8e5a138ceb5fbbadae716"},
-    {file = "storage3-2.27.3.tar.gz", hash = "sha256:dc1a4a010cf36d5482c5cb6c1c28fc5f00e23284342b89e4ae43b5eae8501ddb"},
+    {file = "storage3-2.28.0.tar.gz", hash = "sha256:bc1d008aff67de7a0f2bd867baee7aadbcdb6f78f5a310b4f7a38e8c13c19865"},
 ]
 
 [package.dependencies]
@@ -7091,35 +7181,35 @@ typing-extensions = {version = ">=4.5.0", markers = "python_version >= \"3.7\""}
 
 [[package]]
 name = "supabase"
-version = "2.27.3"
+version = "2.28.0"
 description = "Supabase client for Python."
 optional = false
 python-versions = ">=3.9"
 groups = ["main"]
 files = [
-    {file = "supabase-2.27.3-py3-none-any.whl", hash = "sha256:082a74642fcf9954693f1ce8c251baf23e4bda26ffdbc8dcd4c99c82e60d69ff"},
+    {file = "supabase-2.28.0-py3-none-any.whl", hash = "sha256:42776971c7d0ccca16034df1ab96a31c50228eb1eb19da4249ad2f756fc20272"},
-    {file = "supabase-2.27.3.tar.gz", hash = "sha256:5e5a348232ac4315c1032ddd687278f0b982465471f0cbb52bca7e6a66495ff3"},
+    {file = "supabase-2.28.0.tar.gz", hash = "sha256:aea299aaab2a2eed3c57e0be7fc035c6807214194cce795a3575add20268ece1"},
 ]
 
 [package.dependencies]
 httpx = ">=0.26,<0.29"
-postgrest = "2.27.3"
+postgrest = "2.28.0"
-realtime = "2.27.3"
+realtime = "2.28.0"
-storage3 = "2.27.3"
+storage3 = "2.28.0"
-supabase-auth = "2.27.3"
+supabase-auth = "2.28.0"
-supabase-functions = "2.27.3"
+supabase-functions = "2.28.0"
 yarl = ">=1.22.0"
 
 [[package]]
 name = "supabase-auth"
-version = "2.27.3"
+version = "2.28.0"
 description = "Python Client Library for Supabase Auth"
 optional = false
 python-versions = ">=3.9"
 groups = ["main"]
 files = [
-    {file = "supabase_auth-2.27.3-py3-none-any.whl", hash = "sha256:82a4262eaad85383319d394dab0eea11fcf3ebd774062aef8ea3874ae2f02579"},
+    {file = "supabase_auth-2.28.0-py3-none-any.whl", hash = "sha256:2ac85026cc285054c7fa6d41924f3a333e9ec298c013e5b5e1754039ba7caec9"},
-    {file = "supabase_auth-2.27.3.tar.gz", hash = "sha256:39894d4bc60b6f23b5cff4d0d7d4c1659e5d69563cadf014d4896f780ca8ca78"},
+    {file = "supabase_auth-2.28.0.tar.gz", hash = "sha256:2bb8f18ff39934e44b28f10918db965659f3735cd6fbfcc022fe0b82dbf8233e"},
 ]
 
 [package.dependencies]
@@ -7129,14 +7219,14 @@ pyjwt = {version = ">=2.10.1", extras = ["crypto"]}
 
 [[package]]
 name = "supabase-functions"
-version = "2.27.3"
+version = "2.28.0"
 description = "Library for Supabase Functions"
 optional = false
 python-versions = ">=3.9"
 groups = ["main"]
 files = [
-    {file = "supabase_functions-2.27.3-py3-none-any.whl", hash = "sha256:9d14a931d49ede1c6cf5fbfceb11c44061535ba1c3f310f15384964d86a83d9e"},
+    {file = "supabase_functions-2.28.0-py3-none-any.whl", hash = "sha256:30bf2d586f8df285faf0621bb5d5bb3ec3157234fc820553ca156f009475e4ae"},
-    {file = "supabase_functions-2.27.3.tar.gz", hash = "sha256:e954f1646da8ca6e7e16accef58d0884a5f97b25956ee98e7d4927a210ed92f9"},
+    {file = "supabase_functions-2.28.0.tar.gz", hash = "sha256:db3dddfc37aca5858819eb461130968473bd8c75bd284581013958526dac718b"},
 ]
 
 [package.dependencies]
@@ -8440,4 +8530,4 @@ cffi = ["cffi (>=1.17,<2.0) ; platform_python_implementation != \"PyPy\" and pyt
 [metadata]
 lock-version = "2.1"
 python-versions = ">=3.10,<3.14"
-content-hash = "c06e96ad49388ba7a46786e9ea55ea2c1a57408e15613237b4bee40a592a12af"
+content-hash = "55e095de555482f0fe47de7695f390fe93e7bcf739b31c391b2e5e3c3d938ae3"

autogpt_platform/backend/pyproject.toml

@@ -16,6 +16,7 @@ anthropic = "^0.79.0"
 apscheduler = "^3.11.1"
 autogpt-libs = { path = "../autogpt_libs", develop = true }
 bleach = { extras = ["css"], version = "^6.2.0" }
+claude-agent-sdk = "^0.1.0"
 click = "^8.2.0"
 cryptography = "^46.0"
 discord-py = "^2.5.2"
@@ -65,7 +66,7 @@ sentry-sdk = {extras = ["anthropic", "fastapi", "launchdarkly", "openai", "sqlal
 sqlalchemy = "^2.0.40"
 strenum = "^0.4.9"
 stripe = "^11.5.0"
-supabase = "2.27.3"
+supabase = "2.28.0"
 tenacity = "^9.1.4"
 todoist-api-python = "^2.1.7"
 tweepy = "^4.16.0"

autogpt_platform/backend/test/agent_generator/test_service.py

@@ -25,6 +25,7 @@ class TestServiceConfiguration:
         """Test that external service is not configured when host is empty."""
         mock_settings = MagicMock()
         mock_settings.config.agentgenerator_host = ""
+        mock_settings.config.agentgenerator_use_dummy = False
 
         with patch.object(service, "_get_settings", return_value=mock_settings):
             assert service.is_external_service_configured() is False

autogpt_platform/backend/test/chat/test_security_hooks.py

@@ -0,0 +1,133 @@
+"""Tests for SDK security hooks — workspace paths, tool access, and deny messages.
+
+These are pure unit tests with no external dependencies (no SDK, no DB, no server).
+They validate that the security hooks correctly block unauthorized paths,
+tool access, and dangerous input patterns.
+
+Note: Bash command validation was removed — the SDK built-in Bash tool is not in
+allowed_tools, and the bash_exec MCP tool has kernel-level network isolation
+(unshare --net) making command-level parsing unnecessary.
+"""
+
+from backend.api.features.chat.sdk.security_hooks import (
+    _validate_tool_access,
+    _validate_workspace_path,
+)
+
+SDK_CWD = "/tmp/copilot-test-session"
+
+
+def _is_denied(result: dict) -> bool:
+    hook = result.get("hookSpecificOutput", {})
+    return hook.get("permissionDecision") == "deny"
+
+
+def _reason(result: dict) -> str:
+    return result.get("hookSpecificOutput", {}).get("permissionDecisionReason", "")
+
+
+# ============================================================
+# Workspace path validation (Read, Write, Edit, etc.)
+# ============================================================
+
+
+class TestWorkspacePathValidation:
+    def test_path_in_workspace(self):
+        result = _validate_workspace_path(
+            "Read", {"file_path": f"{SDK_CWD}/file.txt"}, SDK_CWD
+        )
+        assert not _is_denied(result)
+
+    def test_path_outside_workspace(self):
+        result = _validate_workspace_path("Read", {"file_path": "/etc/passwd"}, SDK_CWD)
+        assert _is_denied(result)
+
+    def test_tool_results_allowed(self):
+        result = _validate_workspace_path(
+            "Read",
+            {"file_path": "~/.claude/projects/abc/tool-results/out.txt"},
+            SDK_CWD,
+        )
+        assert not _is_denied(result)
+
+    def test_claude_settings_blocked(self):
+        result = _validate_workspace_path(
+            "Read", {"file_path": "~/.claude/settings.json"}, SDK_CWD
+        )
+        assert _is_denied(result)
+
+    def test_claude_projects_without_tool_results(self):
+        result = _validate_workspace_path(
+            "Read", {"file_path": "~/.claude/projects/abc/credentials.json"}, SDK_CWD
+        )
+        assert _is_denied(result)
+
+    def test_no_path_allowed(self):
+        """Glob/Grep without path defaults to cwd — should be allowed."""
+        result = _validate_workspace_path("Grep", {"pattern": "foo"}, SDK_CWD)
+        assert not _is_denied(result)
+
+    def test_path_traversal_with_dotdot(self):
+        result = _validate_workspace_path(
+            "Read", {"file_path": f"{SDK_CWD}/../../../etc/passwd"}, SDK_CWD
+        )
+        assert _is_denied(result)
+
+
+# ============================================================
+# Tool access validation
+# ============================================================
+
+
+class TestToolAccessValidation:
+    def test_blocked_tools(self):
+        for tool in ("bash", "shell", "exec", "terminal", "command"):
+            result = _validate_tool_access(tool, {})
+            assert _is_denied(result), f"Tool '{tool}' should be blocked"
+
+    def test_bash_builtin_blocked(self):
+        """SDK built-in Bash (capital) is blocked as defence-in-depth."""
+        result = _validate_tool_access("Bash", {"command": "echo hello"}, SDK_CWD)
+        assert _is_denied(result)
+        assert "Bash" in _reason(result)
+
+    def test_workspace_tools_delegate(self):
+        result = _validate_tool_access(
+            "Read", {"file_path": f"{SDK_CWD}/file.txt"}, SDK_CWD
+        )
+        assert not _is_denied(result)
+
+    def test_dangerous_pattern_blocked(self):
+        result = _validate_tool_access("SomeUnknownTool", {"data": "sudo rm -rf /"})
+        assert _is_denied(result)
+
+    def test_safe_unknown_tool_allowed(self):
+        result = _validate_tool_access("SomeSafeTool", {"data": "hello world"})
+        assert not _is_denied(result)
+
+
+# ============================================================
+# Deny message quality (ntindle feedback)
+# ============================================================
+
+
+class TestDenyMessageClarity:
+    """Deny messages must include [SECURITY] and 'cannot be bypassed'
+    so the model knows the restriction is enforced, not a suggestion."""
+
+    def test_blocked_tool_message(self):
+        reason = _reason(_validate_tool_access("bash", {}))
+        assert "[SECURITY]" in reason
+        assert "cannot be bypassed" in reason
+
+    def test_bash_builtin_blocked_message(self):
+        reason = _reason(_validate_tool_access("Bash", {"command": "echo hello"}))
+        assert "[SECURITY]" in reason
+        assert "cannot be bypassed" in reason
+
+    def test_workspace_path_message(self):
+        reason = _reason(
+            _validate_workspace_path("Read", {"file_path": "/etc/passwd"}, SDK_CWD)
+        )
+        assert "[SECURITY]" in reason
+        assert "cannot be bypassed" in reason

autogpt_platform/backend/test/chat/test_transcript.py

@@ -0,0 +1,255 @@
+"""Unit tests for JSONL transcript management utilities."""
+
+import json
+import os
+
+from backend.api.features.chat.sdk.transcript import (
+    STRIPPABLE_TYPES,
+    read_transcript_file,
+    strip_progress_entries,
+    validate_transcript,
+    write_transcript_to_tempfile,
+)
+
+
+def _make_jsonl(*entries: dict) -> str:
+    return "\n".join(json.dumps(e) for e in entries) + "\n"
+
+
+# --- Fixtures ---
+
+
+METADATA_LINE = {"type": "queue-operation", "subtype": "create"}
+FILE_HISTORY = {"type": "file-history-snapshot", "files": []}
+USER_MSG = {"type": "user", "uuid": "u1", "message": {"role": "user", "content": "hi"}}
+ASST_MSG = {
+    "type": "assistant",
+    "uuid": "a1",
+    "parentUuid": "u1",
+    "message": {"role": "assistant", "content": "hello"},
+}
+PROGRESS_ENTRY = {
+    "type": "progress",
+    "uuid": "p1",
+    "parentUuid": "u1",
+    "data": {"type": "bash_progress", "stdout": "running..."},
+}
+
+VALID_TRANSCRIPT = _make_jsonl(METADATA_LINE, FILE_HISTORY, USER_MSG, ASST_MSG)
+
+
+# --- read_transcript_file ---
+
+
+class TestReadTranscriptFile:
+    def test_returns_content_for_valid_file(self, tmp_path):
+        path = tmp_path / "session.jsonl"
+        path.write_text(VALID_TRANSCRIPT)
+        result = read_transcript_file(str(path))
+        assert result is not None
+        assert "user" in result
+
+    def test_returns_none_for_missing_file(self):
+        assert read_transcript_file("/nonexistent/path.jsonl") is None
+
+    def test_returns_none_for_empty_path(self):
+        assert read_transcript_file("") is None
+
+    def test_returns_none_for_empty_file(self, tmp_path):
+        path = tmp_path / "empty.jsonl"
+        path.write_text("")
+        assert read_transcript_file(str(path)) is None
+
+    def test_returns_none_for_metadata_only(self, tmp_path):
+        content = _make_jsonl(METADATA_LINE, FILE_HISTORY)
+        path = tmp_path / "meta.jsonl"
+        path.write_text(content)
+        assert read_transcript_file(str(path)) is None
+
+    def test_returns_none_for_invalid_json(self, tmp_path):
+        path = tmp_path / "bad.jsonl"
+        path.write_text("not json\n{}\n{}\n")
+        assert read_transcript_file(str(path)) is None
+
+    def test_no_size_limit(self, tmp_path):
+        """Large files are accepted — bucket storage has no size limit."""
+        big_content = {"type": "user", "uuid": "u9", "data": "x" * 1_000_000}
+        content = _make_jsonl(METADATA_LINE, FILE_HISTORY, big_content, ASST_MSG)
+        path = tmp_path / "big.jsonl"
+        path.write_text(content)
+        result = read_transcript_file(str(path))
+        assert result is not None
+
+
+# --- write_transcript_to_tempfile ---
+
+
+class TestWriteTranscriptToTempfile:
+    """Tests use /tmp/copilot-* paths to satisfy the sandbox prefix check."""
+
+    def test_writes_file_and_returns_path(self):
+        cwd = "/tmp/copilot-test-write"
+        try:
+            result = write_transcript_to_tempfile(
+                VALID_TRANSCRIPT, "sess-1234-abcd", cwd
+            )
+            assert result is not None
+            assert os.path.isfile(result)
+            assert result.endswith(".jsonl")
+            with open(result) as f:
+                assert f.read() == VALID_TRANSCRIPT
+        finally:
+            import shutil
+
+            shutil.rmtree(cwd, ignore_errors=True)
+
+    def test_creates_parent_directory(self):
+        cwd = "/tmp/copilot-test-mkdir"
+        try:
+            result = write_transcript_to_tempfile(VALID_TRANSCRIPT, "sess-1234", cwd)
+            assert result is not None
+            assert os.path.isdir(cwd)
+        finally:
+            import shutil
+
+            shutil.rmtree(cwd, ignore_errors=True)
+
+    def test_uses_session_id_prefix(self):
+        cwd = "/tmp/copilot-test-prefix"
+        try:
+            result = write_transcript_to_tempfile(
+                VALID_TRANSCRIPT, "abcdef12-rest", cwd
+            )
+            assert result is not None
+            assert "abcdef12" in os.path.basename(result)
+        finally:
+            import shutil
+
+            shutil.rmtree(cwd, ignore_errors=True)
+
+    def test_rejects_cwd_outside_sandbox(self, tmp_path):
+        cwd = str(tmp_path / "not-copilot")
+        result = write_transcript_to_tempfile(VALID_TRANSCRIPT, "sess-1234", cwd)
+        assert result is None
+
+
+# --- validate_transcript ---
+
+
+class TestValidateTranscript:
+    def test_valid_transcript(self):
+        assert validate_transcript(VALID_TRANSCRIPT) is True
+
+    def test_none_content(self):
+        assert validate_transcript(None) is False
+
+    def test_empty_content(self):
+        assert validate_transcript("") is False
+
+    def test_metadata_only(self):
+        content = _make_jsonl(METADATA_LINE, FILE_HISTORY)
+        assert validate_transcript(content) is False
+
+    def test_user_only_no_assistant(self):
+        content = _make_jsonl(METADATA_LINE, FILE_HISTORY, USER_MSG)
+        assert validate_transcript(content) is False
+
+    def test_assistant_only_no_user(self):
+        content = _make_jsonl(METADATA_LINE, FILE_HISTORY, ASST_MSG)
+        assert validate_transcript(content) is False
+
+    def test_invalid_json_returns_false(self):
+        assert validate_transcript("not json\n{}\n{}\n") is False
+
+
+# --- strip_progress_entries ---
+
+
+class TestStripProgressEntries:
+    def test_strips_all_strippable_types(self):
+        """All STRIPPABLE_TYPES are removed from the output."""
+        entries = [
+            USER_MSG,
+            {"type": "progress", "uuid": "p1", "parentUuid": "u1"},
+            {"type": "file-history-snapshot", "files": []},
+            {"type": "queue-operation", "subtype": "create"},
+            {"type": "summary", "text": "..."},
+            {"type": "pr-link", "url": "..."},
+            ASST_MSG,
+        ]
+        result = strip_progress_entries(_make_jsonl(*entries))
+        result_types = {json.loads(line)["type"] for line in result.strip().split("\n")}
+        assert result_types == {"user", "assistant"}
+        for stype in STRIPPABLE_TYPES:
+            assert stype not in result_types
+
+    def test_reparents_children_of_stripped_entries(self):
+        """An assistant message whose parent is a progress entry gets reparented."""
+        progress = {
+            "type": "progress",
+            "uuid": "p1",
+            "parentUuid": "u1",
+            "data": {"type": "bash_progress"},
+        }
+        asst = {
+            "type": "assistant",
+            "uuid": "a1",
+            "parentUuid": "p1",  # Points to progress
+            "message": {"role": "assistant", "content": "done"},
+        }
+        content = _make_jsonl(USER_MSG, progress, asst)
+        result = strip_progress_entries(content)
+        lines = [json.loads(line) for line in result.strip().split("\n")]
+
+        asst_entry = next(e for e in lines if e["type"] == "assistant")
+        # Should be reparented to u1 (the user message)
+        assert asst_entry["parentUuid"] == "u1"
+
+    def test_reparents_through_chain(self):
+        """Reparenting walks through multiple stripped entries."""
+        p1 = {"type": "progress", "uuid": "p1", "parentUuid": "u1"}
+        p2 = {"type": "progress", "uuid": "p2", "parentUuid": "p1"}
+        p3 = {"type": "progress", "uuid": "p3", "parentUuid": "p2"}
+        asst = {
+            "type": "assistant",
+            "uuid": "a1",
+            "parentUuid": "p3",  # 3 levels deep
+            "message": {"role": "assistant", "content": "done"},
+        }
+        content = _make_jsonl(USER_MSG, p1, p2, p3, asst)
+        result = strip_progress_entries(content)
+        lines = [json.loads(line) for line in result.strip().split("\n")]
+
+        asst_entry = next(e for e in lines if e["type"] == "assistant")
+        assert asst_entry["parentUuid"] == "u1"
+
+    def test_preserves_non_strippable_entries(self):
+        """User, assistant, and system entries are preserved."""
+        system = {"type": "system", "uuid": "s1", "message": "prompt"}
+        content = _make_jsonl(system, USER_MSG, ASST_MSG)
+        result = strip_progress_entries(content)
+        result_types = [json.loads(line)["type"] for line in result.strip().split("\n")]
+        assert result_types == ["system", "user", "assistant"]
+
+    def test_empty_input(self):
+        result = strip_progress_entries("")
+        # Should return just a newline (empty content stripped)
+        assert result.strip() == ""
+
+    def test_no_strippable_entries(self):
+        """When there's nothing to strip, output matches input structure."""
+        content = _make_jsonl(USER_MSG, ASST_MSG)
+        result = strip_progress_entries(content)
+        result_lines = result.strip().split("\n")
+        assert len(result_lines) == 2
+
+    def test_handles_entries_without_uuid(self):
+        """Entries without uuid field are handled gracefully."""
+        no_uuid = {"type": "queue-operation", "subtype": "create"}
+        content = _make_jsonl(no_uuid, USER_MSG, ASST_MSG)
+        result = strip_progress_entries(content)
+        result_types = [json.loads(line)["type"] for line in result.strip().split("\n")]
+        # queue-operation is strippable
+        assert "queue-operation" not in result_types
+        assert "user" in result_types
+        assert "assistant" in result_types

autogpt_platform/docker-compose.platform.yml

@@ -37,7 +37,7 @@ services:
       context: ../
       dockerfile: autogpt_platform/backend/Dockerfile
       target: migrate
-    command: ["sh", "-c", "poetry run prisma generate && poetry run gen-prisma-stub && poetry run prisma migrate deploy"]
+    command: ["sh", "-c", "prisma generate && python3 gen_prisma_types_stub.py && prisma migrate deploy"]
     develop:
       watch:
         - path: ./
@@ -56,7 +56,7 @@ services:
       test:
         [
           "CMD-SHELL",
-          "poetry run prisma migrate status | grep -q 'No pending migrations' || exit 1",
+          "prisma migrate status | grep -q 'No pending migrations' || exit 1",
         ]
       interval: 30s
       timeout: 10s

autogpt_platform/frontend/instrumentation-client.ts

@@ -22,6 +22,11 @@ Sentry.init({
 
   enabled: shouldEnable,
 
+  // Suppress cross-origin stylesheet errors from Sentry Replay (rrweb)
+  // serializing DOM snapshots with cross-origin stylesheets
+  // (e.g., from browser extensions or CDN-loaded CSS)
+  ignoreErrors: [/Not allowed to access cross-origin stylesheet/],
+
   // Add optional integrations for additional features
   integrations: [
     Sentry.captureConsoleIntegration(),

autogpt_platform/frontend/src/app/(platform)/copilot/components/ChatMessagesContainer/ChatMessagesContainer.tsx

@@ -20,6 +20,7 @@ import { FindBlocksTool } from "../../tools/FindBlocks/FindBlocks";
 import { RunAgentTool } from "../../tools/RunAgent/RunAgent";
 import { RunBlockTool } from "../../tools/RunBlock/RunBlock";
 import { SearchDocsTool } from "../../tools/SearchDocs/SearchDocs";
+import { GenericTool } from "../../tools/GenericTool/GenericTool";
 import { ViewAgentOutputTool } from "../../tools/ViewAgentOutput/ViewAgentOutput";
 
 // ---------------------------------------------------------------------------
@@ -159,7 +160,7 @@ export const ChatMessagesContainer = ({
 
   return (
     <Conversation className="min-h-0 flex-1">
-      <ConversationContent className="flex min-h-screen flex-1 flex-col gap-6 px-3 py-6">
+      <ConversationContent className="flex flex-1 flex-col gap-6 px-3 py-6">
         {isLoading && messages.length === 0 && (
           <div className="flex min-h-full flex-1 items-center justify-center">
             <LoadingSpinner className="text-neutral-600" />
@@ -255,6 +256,16 @@ export const ChatMessagesContainer = ({
                         />
                       );
                     default:
+                      // Render a generic tool indicator for SDK built-in
+                      // tools (Read, Glob, Grep, etc.) or any unrecognized tool
+                      if (part.type.startsWith("tool-")) {
+                        return (
+                          <GenericTool
+                            key={`${message.id}-${i}`}
+                            part={part as ToolUIPart}
+                          />
+                        );
+                      }
                       return null;
                   }
                 })}

autogpt_platform/frontend/src/app/(platform)/copilot/hooks/Untitled

@@ -1,10 +0,0 @@
-import { parseAsString, useQueryState } from "nuqs";
-
-export function useCopilotSessionId() {
-  const [urlSessionId, setUrlSessionId] = useQueryState(
-    "sessionId",
-    parseAsString,
-  );
-
-  return { urlSessionId, setUrlSessionId };
-}

autogpt_platform/frontend/src/app/(platform)/copilot/hooks/useLongRunningToolPolling.ts

@@ -0,0 +1,126 @@
+import { getGetV2GetSessionQueryKey } from "@/app/api/__generated__/endpoints/chat/chat";
+import { useQueryClient } from "@tanstack/react-query";
+import type { UIDataTypes, UIMessage, UITools } from "ai";
+import { useCallback, useEffect, useRef } from "react";
+import { convertChatSessionMessagesToUiMessages } from "../helpers/convertChatSessionToUiMessages";
+
+const OPERATING_TYPES = new Set([
+  "operation_started",
+  "operation_pending",
+  "operation_in_progress",
+]);
+
+const POLL_INTERVAL_MS = 1_500;
+
+/**
+ * Detects whether any message contains a tool part whose output indicates
+ * a long-running operation is still in progress.
+ */
+function hasOperatingTool(
+  messages: UIMessage<unknown, UIDataTypes, UITools>[],
+) {
+  for (const msg of messages) {
+    for (const part of msg.parts) {
+      if (!part.type.startsWith("tool-")) continue;
+      const toolPart = part as { output?: unknown };
+      if (!toolPart.output) continue;
+      const output =
+        typeof toolPart.output === "string"
+          ? safeParse(toolPart.output)
+          : toolPart.output;
+      if (
+        output &&
+        typeof output === "object" &&
+        "type" in output &&
+        OPERATING_TYPES.has((output as { type: string }).type)
+      ) {
+        return true;
+      }
+    }
+  }
+  return false;
+}
+
+function safeParse(value: string): unknown {
+  try {
+    return JSON.parse(value);
+  } catch {
+    return null;
+  }
+}
+
+/**
+ * Polls the session endpoint while any tool is in an "operating" state
+ * (operation_started / operation_pending / operation_in_progress).
+ *
+ * When the session data shows the tool output has changed (e.g. to
+ * agent_saved), it calls `setMessages` with the updated messages.
+ */
+export function useLongRunningToolPolling(
+  sessionId: string | null,
+  messages: UIMessage<unknown, UIDataTypes, UITools>[],
+  setMessages: (
+    updater: (
+      prev: UIMessage<unknown, UIDataTypes, UITools>[],
+    ) => UIMessage<unknown, UIDataTypes, UITools>[],
+  ) => void,
+) {
+  const queryClient = useQueryClient();
+  const intervalRef = useRef<ReturnType<typeof setInterval> | null>(null);
+
+  const stopPolling = useCallback(() => {
+    if (intervalRef.current) {
+      clearInterval(intervalRef.current);
+      intervalRef.current = null;
+    }
+  }, []);
+
+  const poll = useCallback(async () => {
+    if (!sessionId) return;
+
+    // Invalidate the query cache so the next fetch gets fresh data
+    await queryClient.invalidateQueries({
+      queryKey: getGetV2GetSessionQueryKey(sessionId),
+    });
+
+    // Fetch fresh session data
+    const data = queryClient.getQueryData<{
+      status: number;
+      data: { messages?: unknown[] };
+    }>(getGetV2GetSessionQueryKey(sessionId));
+
+    if (data?.status !== 200 || !data.data.messages) return;
+
+    const freshMessages = convertChatSessionMessagesToUiMessages(
+      sessionId,
+      data.data.messages,
+    );
+
+    if (!freshMessages || freshMessages.length === 0) return;
+
+    // Update when the long-running tool completed
+    if (!hasOperatingTool(freshMessages)) {
+      setMessages(() => freshMessages);
+      stopPolling();
+    }
+  }, [sessionId, queryClient, setMessages, stopPolling]);
+
+  useEffect(() => {
+    const shouldPoll = hasOperatingTool(messages);
+
+    // Always clear any previous interval first so we never leak timers
+    // when the effect re-runs due to dependency changes (e.g. messages
+    // updating as the LLM streams text after the tool call).
+    stopPolling();
+
+    if (shouldPoll && sessionId) {
+      intervalRef.current = setInterval(() => {
+        poll();
+      }, POLL_INTERVAL_MS);
+    }
+
+    return () => {
+      stopPolling();
+    };
+  }, [messages, sessionId, poll, stopPolling]);
+}

autogpt_platform/frontend/src/app/(platform)/copilot/tools/CreateAgent/CreateAgent.tsx

@@ -1,24 +1,30 @@
 "use client";
 
-import { WarningDiamondIcon } from "@phosphor-icons/react";
+import { Button } from "@/components/atoms/Button/Button";
+import { Text } from "@/components/atoms/Text/Text";
+import {
+  BookOpenIcon,
+  CheckFatIcon,
+  PencilSimpleIcon,
+  WarningDiamondIcon,
+} from "@phosphor-icons/react";
 import type { ToolUIPart } from "ai";
+import NextLink from "next/link";
 import { useCopilotChatActions } from "../../components/CopilotChatActionsProvider/useCopilotChatActions";
 import { MorphingTextAnimation } from "../../components/MorphingTextAnimation/MorphingTextAnimation";
-import { ProgressBar } from "../../components/ProgressBar/ProgressBar";
 import {
   ContentCardDescription,
   ContentCodeBlock,
   ContentGrid,
   ContentHint,
-  ContentLink,
   ContentMessage,
 } from "../../components/ToolAccordion/AccordionContent";
 import { ToolAccordion } from "../../components/ToolAccordion/ToolAccordion";
-import { useAsymptoticProgress } from "../../hooks/useAsymptoticProgress";
 import {
   ClarificationQuestionsCard,
   ClarifyingQuestion,
 } from "./components/ClarificationQuestionsCard";
+import { MiniGame } from "./components/MiniGame/MiniGame";
 import {
   AccordionIcon,
   formatMaybeJson,
@@ -52,7 +58,7 @@ function getAccordionMeta(output: CreateAgentToolOutput) {
   const icon = <AccordionIcon />;
 
   if (isAgentSavedOutput(output)) {
-    return { icon, title: output.agent_name };
+    return { icon, title: output.agent_name, expanded: true };
   }
   if (isAgentPreviewOutput(output)) {
     return {
@@ -78,6 +84,7 @@ function getAccordionMeta(output: CreateAgentToolOutput) {
     return {
       icon,
       title: "Creating agent, this may take a few minutes. Sit back and relax.",
+      expanded: true,
     };
   }
   return {
@@ -107,8 +114,6 @@ export function CreateAgentTool({ part }: Props) {
       isOperationPendingOutput(output) ||
       isOperationInProgressOutput(output));
 
-  const progress = useAsymptoticProgress(isOperating);
-
   const hasExpandableContent =
     part.state === "output-available" &&
     !!output &&
@@ -152,31 +157,53 @@ export function CreateAgentTool({ part }: Props) {
         <ToolAccordion {...getAccordionMeta(output)}>
           {isOperating && (
             <ContentGrid>
-              <ProgressBar value={progress} className="max-w-[280px]" />
+              <MiniGame />
               <ContentHint>
-                This could take a few minutes, grab a coffee ☕
+                This could take a few minutes — play while you wait!
               </ContentHint>
             </ContentGrid>
           )}
 
           {isAgentSavedOutput(output) && (
-            <ContentGrid>
+            <div className="rounded-xl border border-border/60 bg-card p-4 shadow-sm">
-              <ContentMessage>{output.message}</ContentMessage>
+              <div className="flex items-baseline gap-2">
-              <div className="flex flex-wrap gap-2">
+                <CheckFatIcon
-                <ContentLink href={output.library_agent_link}>
+                  size={18}
-                  Open in library
+                  weight="regular"
-                </ContentLink>
+                  className="relative top-1 text-green-500"
-                <ContentLink href={output.agent_page_link}>
+                />
-                  Open in builder
+                <Text
-                </ContentLink>
+                  variant="body-medium"
+                  className="text-blacks mb-2 text-[16px]"
+                >
+                  {output.message}
+                </Text>
               </div>
-              <ContentCodeBlock>
+              <div className="mt-3 flex flex-wrap gap-4">
-                {truncateText(
+                <Button variant="outline" size="small">
-                  formatMaybeJson({ agent_id: output.agent_id }),
+                  <NextLink
-                  800,
+                    href={output.library_agent_link}
-                )}
+                    className="inline-flex items-center gap-1.5"
-              </ContentCodeBlock>
+                    target="_blank"
-            </ContentGrid>
+                    rel="noopener noreferrer"
+                  >
+                    <BookOpenIcon size={14} weight="regular" />
+                    Open in library
+                  </NextLink>
+                </Button>
+                <Button variant="outline" size="small">
+                  <NextLink
+                    href={output.agent_page_link}
+                    target="_blank"
+                    rel="noopener noreferrer"
+                    className="inline-flex items-center gap-1.5"
+                  >
+                    <PencilSimpleIcon size={14} weight="regular" />
+                    Open in builder
+                  </NextLink>
+                </Button>
+              </div>
+            </div>
           )}
 
           {isAgentPreviewOutput(output) && (

autogpt_platform/frontend/src/app/(platform)/copilot/tools/CreateAgent/components/MiniGame/MiniGame.tsx

@@ -0,0 +1,21 @@
+"use client";
+
+import { useMiniGame } from "./useMiniGame";
+
+export function MiniGame() {
+  const { canvasRef } = useMiniGame();
+
+  return (
+    <div
+      className="w-full overflow-hidden rounded-md bg-background text-foreground"
+      style={{ border: "1px solid #d17fff" }}
+    >
+      <canvas
+        ref={canvasRef}
+        tabIndex={0}
+        className="block w-full outline-none"
+        style={{ imageRendering: "pixelated" }}
+      />
+    </div>
+  );
+}

autogpt_platform/frontend/src/app/(platform)/copilot/tools/CreateAgent/components/MiniGame/useMiniGame.ts

@@ -0,0 +1,579 @@
+import { useEffect, useRef } from "react";
+
+/* ------------------------------------------------------------------ */
+/*  Constants                                                         */
+/* ------------------------------------------------------------------ */
+
+const CANVAS_HEIGHT = 150;
+const GRAVITY = 0.55;
+const JUMP_FORCE = -9.5;
+const BASE_SPEED = 3;
+const SPEED_INCREMENT = 0.0008;
+const SPAWN_MIN = 70;
+const SPAWN_MAX = 130;
+const CHAR_SIZE = 18;
+const CHAR_X = 50;
+const GROUND_PAD = 20;
+const STORAGE_KEY = "copilot-minigame-highscore";
+
+// Colors
+const COLOR_BG = "#E8EAF6";
+const COLOR_CHAR = "#263238";
+const COLOR_BOSS = "#F50057";
+
+// Boss
+const BOSS_SIZE = 36;
+const BOSS_ENTER_SPEED = 2;
+const BOSS_LEAVE_SPEED = 3;
+const BOSS_SHOOT_COOLDOWN = 90;
+const BOSS_SHOTS_TO_EVADE = 5;
+const BOSS_INTERVAL = 20; // every N score
+const PROJ_SPEED = 4.5;
+const PROJ_SIZE = 12;
+
+/* ------------------------------------------------------------------ */
+/*  Types                                                             */
+/* ------------------------------------------------------------------ */
+
+interface Obstacle {
+  x: number;
+  width: number;
+  height: number;
+  scored: boolean;
+}
+
+interface Projectile {
+  x: number;
+  y: number;
+  speed: number;
+  evaded: boolean;
+  type: "low" | "high";
+}
+
+interface BossState {
+  phase: "inactive" | "entering" | "fighting" | "leaving";
+  x: number;
+  targetX: number;
+  shotsEvaded: number;
+  cooldown: number;
+  projectiles: Projectile[];
+  bob: number;
+}
+
+interface GameState {
+  charY: number;
+  vy: number;
+  obstacles: Obstacle[];
+  score: number;
+  highScore: number;
+  speed: number;
+  frame: number;
+  nextSpawn: number;
+  running: boolean;
+  over: boolean;
+  groundY: number;
+  boss: BossState;
+  bossThreshold: number;
+}
+
+/* ------------------------------------------------------------------ */
+/*  Helpers                                                           */
+/* ------------------------------------------------------------------ */
+
+function randInt(min: number, max: number) {
+  return Math.floor(Math.random() * (max - min + 1)) + min;
+}
+
+function readHighScore(): number {
+  try {
+    return parseInt(localStorage.getItem(STORAGE_KEY) || "0", 10) || 0;
+  } catch {
+    return 0;
+  }
+}
+
+function writeHighScore(score: number) {
+  try {
+    localStorage.setItem(STORAGE_KEY, String(score));
+  } catch {
+    /* noop */
+  }
+}
+
+function makeBoss(): BossState {
+  return {
+    phase: "inactive",
+    x: 0,
+    targetX: 0,
+    shotsEvaded: 0,
+    cooldown: 0,
+    projectiles: [],
+    bob: 0,
+  };
+}
+
+function makeState(groundY: number): GameState {
+  return {
+    charY: groundY - CHAR_SIZE,
+    vy: 0,
+    obstacles: [],
+    score: 0,
+    highScore: readHighScore(),
+    speed: BASE_SPEED,
+    frame: 0,
+    nextSpawn: randInt(SPAWN_MIN, SPAWN_MAX),
+    running: false,
+    over: false,
+    groundY,
+    boss: makeBoss(),
+    bossThreshold: BOSS_INTERVAL,
+  };
+}
+
+function gameOver(s: GameState) {
+  s.running = false;
+  s.over = true;
+  if (s.score > s.highScore) {
+    s.highScore = s.score;
+    writeHighScore(s.score);
+  }
+}
+
+/* ------------------------------------------------------------------ */
+/*  Projectile collision — shared between fighting & leaving phases   */
+/* ------------------------------------------------------------------ */
+
+/** Returns true if the player died. */
+function tickProjectiles(s: GameState): boolean {
+  const boss = s.boss;
+
+  for (const p of boss.projectiles) {
+    p.x -= p.speed;
+
+    if (!p.evaded && p.x + PROJ_SIZE < CHAR_X) {
+      p.evaded = true;
+      boss.shotsEvaded++;
+    }
+
+    // Collision
+    if (
+      !p.evaded &&
+      CHAR_X + CHAR_SIZE > p.x &&
+      CHAR_X < p.x + PROJ_SIZE &&
+      s.charY + CHAR_SIZE > p.y &&
+      s.charY < p.y + PROJ_SIZE
+    ) {
+      gameOver(s);
+      return true;
+    }
+  }
+
+  boss.projectiles = boss.projectiles.filter((p) => p.x + PROJ_SIZE > -20);
+  return false;
+}
+
+/* ------------------------------------------------------------------ */
+/*  Update                                                            */
+/* ------------------------------------------------------------------ */
+
+function update(s: GameState, canvasWidth: number) {
+  if (!s.running) return;
+
+  s.frame++;
+
+  // Speed only ramps during regular play
+  if (s.boss.phase === "inactive") {
+    s.speed = BASE_SPEED + s.frame * SPEED_INCREMENT;
+  }
+
+  // ---- Character physics (always active) ---- //
+  s.vy += GRAVITY;
+  s.charY += s.vy;
+  if (s.charY + CHAR_SIZE >= s.groundY) {
+    s.charY = s.groundY - CHAR_SIZE;
+    s.vy = 0;
+  }
+
+  // ---- Trigger boss ---- //
+  if (s.boss.phase === "inactive" && s.score >= s.bossThreshold) {
+    s.boss.phase = "entering";
+    s.boss.x = canvasWidth + 10;
+    s.boss.targetX = canvasWidth - BOSS_SIZE - 40;
+    s.boss.shotsEvaded = 0;
+    s.boss.cooldown = BOSS_SHOOT_COOLDOWN;
+    s.boss.projectiles = [];
+    s.obstacles = [];
+  }
+
+  // ---- Boss: entering ---- //
+  if (s.boss.phase === "entering") {
+    s.boss.bob = Math.sin(s.frame * 0.05) * 3;
+    s.boss.x -= BOSS_ENTER_SPEED;
+    if (s.boss.x <= s.boss.targetX) {
+      s.boss.x = s.boss.targetX;
+      s.boss.phase = "fighting";
+    }
+    return; // no obstacles while entering
+  }
+
+  // ---- Boss: fighting ---- //
+  if (s.boss.phase === "fighting") {
+    s.boss.bob = Math.sin(s.frame * 0.05) * 3;
+
+    // Shoot
+    s.boss.cooldown--;
+    if (s.boss.cooldown <= 0) {
+      const isLow = Math.random() < 0.5;
+      s.boss.projectiles.push({
+        x: s.boss.x - PROJ_SIZE,
+        y: isLow ? s.groundY - 14 : s.groundY - 70,
+        speed: PROJ_SPEED,
+        evaded: false,
+        type: isLow ? "low" : "high",
+      });
+      s.boss.cooldown = BOSS_SHOOT_COOLDOWN;
+    }
+
+    if (tickProjectiles(s)) return;
+
+    // Boss defeated?
+    if (s.boss.shotsEvaded >= BOSS_SHOTS_TO_EVADE) {
+      s.boss.phase = "leaving";
+      s.score += 5; // bonus
+      s.bossThreshold = s.score + BOSS_INTERVAL;
+    }
+    return;
+  }
+
+  // ---- Boss: leaving ---- //
+  if (s.boss.phase === "leaving") {
+    s.boss.bob = Math.sin(s.frame * 0.05) * 3;
+    s.boss.x += BOSS_LEAVE_SPEED;
+
+    // Still check in-flight projectiles
+    if (tickProjectiles(s)) return;
+
+    if (s.boss.x > canvasWidth + 50) {
+      s.boss = makeBoss();
+      s.nextSpawn = s.frame + randInt(SPAWN_MIN / 2, SPAWN_MAX / 2);
+    }
+    return;
+  }
+
+  // ---- Regular obstacle play ---- //
+  if (s.frame >= s.nextSpawn) {
+    s.obstacles.push({
+      x: canvasWidth + 10,
+      width: randInt(10, 16),
+      height: randInt(20, 48),
+      scored: false,
+    });
+    s.nextSpawn = s.frame + randInt(SPAWN_MIN, SPAWN_MAX);
+  }
+
+  for (const o of s.obstacles) {
+    o.x -= s.speed;
+    if (!o.scored && o.x + o.width < CHAR_X) {
+      o.scored = true;
+      s.score++;
+    }
+  }
+
+  s.obstacles = s.obstacles.filter((o) => o.x + o.width > -20);
+
+  for (const o of s.obstacles) {
+    const oY = s.groundY - o.height;
+    if (
+      CHAR_X + CHAR_SIZE > o.x &&
+      CHAR_X < o.x + o.width &&
+      s.charY + CHAR_SIZE > oY
+    ) {
+      gameOver(s);
+      return;
+    }
+  }
+}
+
+/* ------------------------------------------------------------------ */
+/*  Drawing                                                           */
+/* ------------------------------------------------------------------ */
+
+function drawBoss(ctx: CanvasRenderingContext2D, s: GameState, bg: string) {
+  const bx = s.boss.x;
+  const by = s.groundY - BOSS_SIZE + s.boss.bob;
+
+  // Body
+  ctx.save();
+  ctx.fillStyle = COLOR_BOSS;
+  ctx.globalAlpha = 0.9;
+  ctx.beginPath();
+  ctx.roundRect(bx, by, BOSS_SIZE, BOSS_SIZE, 4);
+  ctx.fill();
+  ctx.restore();
+
+  // Eyes
+  ctx.save();
+  ctx.fillStyle = bg;
+  const eyeY = by + 13;
+  ctx.beginPath();
+  ctx.arc(bx + 10, eyeY, 4, 0, Math.PI * 2);
+  ctx.fill();
+  ctx.beginPath();
+  ctx.arc(bx + 26, eyeY, 4, 0, Math.PI * 2);
+  ctx.fill();
+  ctx.restore();
+
+  // Angry eyebrows
+  ctx.save();
+  ctx.strokeStyle = bg;
+  ctx.lineWidth = 2;
+  ctx.beginPath();
+  ctx.moveTo(bx + 5, eyeY - 7);
+  ctx.lineTo(bx + 14, eyeY - 4);
+  ctx.stroke();
+  ctx.beginPath();
+  ctx.moveTo(bx + 31, eyeY - 7);
+  ctx.lineTo(bx + 22, eyeY - 4);
+  ctx.stroke();
+  ctx.restore();
+
+  // Zigzag mouth
+  ctx.save();
+  ctx.strokeStyle = bg;
+  ctx.lineWidth = 1.5;
+  ctx.beginPath();
+  ctx.moveTo(bx + 10, by + 27);
+  ctx.lineTo(bx + 14, by + 24);
+  ctx.lineTo(bx + 18, by + 27);
+  ctx.lineTo(bx + 22, by + 24);
+  ctx.lineTo(bx + 26, by + 27);
+  ctx.stroke();
+  ctx.restore();
+}
+
+function drawProjectiles(ctx: CanvasRenderingContext2D, boss: BossState) {
+  ctx.save();
+  ctx.fillStyle = COLOR_BOSS;
+  ctx.globalAlpha = 0.8;
+  for (const p of boss.projectiles) {
+    if (p.evaded) continue;
+    ctx.beginPath();
+    ctx.arc(
+      p.x + PROJ_SIZE / 2,
+      p.y + PROJ_SIZE / 2,
+      PROJ_SIZE / 2,
+      0,
+      Math.PI * 2,
+    );
+    ctx.fill();
+  }
+  ctx.restore();
+}
+
+function draw(
+  ctx: CanvasRenderingContext2D,
+  s: GameState,
+  w: number,
+  h: number,
+  fg: string,
+  started: boolean,
+) {
+  ctx.fillStyle = COLOR_BG;
+  ctx.fillRect(0, 0, w, h);
+
+  // Ground
+  ctx.save();
+  ctx.strokeStyle = fg;
+  ctx.globalAlpha = 0.15;
+  ctx.setLineDash([4, 4]);
+  ctx.beginPath();
+  ctx.moveTo(0, s.groundY);
+  ctx.lineTo(w, s.groundY);
+  ctx.stroke();
+  ctx.restore();
+
+  // Character
+  ctx.save();
+  ctx.fillStyle = COLOR_CHAR;
+  ctx.globalAlpha = 0.85;
+  ctx.beginPath();
+  ctx.roundRect(CHAR_X, s.charY, CHAR_SIZE, CHAR_SIZE, 3);
+  ctx.fill();
+  ctx.restore();
+
+  // Eyes
+  ctx.save();
+  ctx.fillStyle = COLOR_BG;
+  ctx.beginPath();
+  ctx.arc(CHAR_X + 6, s.charY + 7, 2.5, 0, Math.PI * 2);
+  ctx.fill();
+  ctx.beginPath();
+  ctx.arc(CHAR_X + 12, s.charY + 7, 2.5, 0, Math.PI * 2);
+  ctx.fill();
+  ctx.restore();
+
+  // Obstacles
+  ctx.save();
+  ctx.fillStyle = fg;
+  ctx.globalAlpha = 0.55;
+  for (const o of s.obstacles) {
+    ctx.fillRect(o.x, s.groundY - o.height, o.width, o.height);
+  }
+  ctx.restore();
+
+  // Boss + projectiles
+  if (s.boss.phase !== "inactive") {
+    drawBoss(ctx, s, COLOR_BG);
+    drawProjectiles(ctx, s.boss);
+  }
+
+  // Score HUD
+  ctx.save();
+  ctx.fillStyle = fg;
+  ctx.globalAlpha = 0.5;
+  ctx.font = "bold 11px monospace";
+  ctx.textAlign = "right";
+  ctx.fillText(`Score: ${s.score}`, w - 12, 20);
+  ctx.fillText(`Best: ${s.highScore}`, w - 12, 34);
+  if (s.boss.phase === "fighting") {
+    ctx.fillText(
+      `Evade: ${s.boss.shotsEvaded}/${BOSS_SHOTS_TO_EVADE}`,
+      w - 12,
+      48,
+    );
+  }
+  ctx.restore();
+
+  // Prompts
+  if (!started && !s.running && !s.over) {
+    ctx.save();
+    ctx.fillStyle = fg;
+    ctx.globalAlpha = 0.5;
+    ctx.font = "12px sans-serif";
+    ctx.textAlign = "center";
+    ctx.fillText("Click or press Space to play while you wait", w / 2, h / 2);
+    ctx.restore();
+  }
+
+  if (s.over) {
+    ctx.save();
+    ctx.fillStyle = fg;
+    ctx.globalAlpha = 0.7;
+    ctx.font = "bold 13px sans-serif";
+    ctx.textAlign = "center";
+    ctx.fillText("Game Over", w / 2, h / 2 - 8);
+    ctx.font = "11px sans-serif";
+    ctx.fillText("Click or Space to restart", w / 2, h / 2 + 10);
+    ctx.restore();
+  }
+}
+
+/* ------------------------------------------------------------------ */
+/*  Hook                                                              */
+/* ------------------------------------------------------------------ */
+
+export function useMiniGame() {
+  const canvasRef = useRef<HTMLCanvasElement>(null);
+  const stateRef = useRef<GameState | null>(null);
+  const rafRef = useRef(0);
+  const startedRef = useRef(false);
+
+  useEffect(() => {
+    const canvas = canvasRef.current;
+    if (!canvas) return;
+
+    const container = canvas.parentElement;
+    if (container) {
+      canvas.width = container.clientWidth;
+      canvas.height = CANVAS_HEIGHT;
+    }
+
+    const groundY = canvas.height - GROUND_PAD;
+    stateRef.current = makeState(groundY);
+
+    const style = getComputedStyle(canvas);
+    let fg = style.color || "#71717a";
+
+    // -------------------------------------------------------------- //
+    //  Jump                                                           //
+    // -------------------------------------------------------------- //
+    function jump() {
+      const s = stateRef.current;
+      if (!s) return;
+
+      if (s.over) {
+        const hs = s.highScore;
+        const gy = s.groundY;
+        stateRef.current = makeState(gy);
+        stateRef.current.highScore = hs;
+        stateRef.current.running = true;
+        startedRef.current = true;
+        return;
+      }
+
+      if (!s.running) {
+        s.running = true;
+        startedRef.current = true;
+        return;
+      }
+
+      // Only jump when on the ground
+      if (s.charY + CHAR_SIZE >= s.groundY) {
+        s.vy = JUMP_FORCE;
+      }
+    }
+
+    function onKey(e: KeyboardEvent) {
+      if (e.code === "Space" || e.key === " ") {
+        e.preventDefault();
+        jump();
+      }
+    }
+
+    function onClick() {
+      canvas?.focus();
+      jump();
+    }
+
+    // -------------------------------------------------------------- //
+    //  Loop                                                           //
+    // -------------------------------------------------------------- //
+    function loop() {
+      const s = stateRef.current;
+      if (!canvas || !s) return;
+      const ctx = canvas.getContext("2d");
+      if (!ctx) return;
+
+      update(s, canvas.width);
+      draw(ctx, s, canvas.width, canvas.height, fg, startedRef.current);
+      rafRef.current = requestAnimationFrame(loop);
+    }
+
+    rafRef.current = requestAnimationFrame(loop);
+
+    canvas.addEventListener("click", onClick);
+    canvas.addEventListener("keydown", onKey);
+
+    const observer = new ResizeObserver((entries) => {
+      for (const entry of entries) {
+        canvas.width = entry.contentRect.width;
+        canvas.height = CANVAS_HEIGHT;
+        if (stateRef.current) {
+          stateRef.current.groundY = canvas.height - GROUND_PAD;
+        }
+        const cs = getComputedStyle(canvas);
+        fg = cs.color || fg;
+      }
+    });
+    if (container) observer.observe(container);
+
+    return () => {
+      cancelAnimationFrame(rafRef.current);
+      canvas.removeEventListener("click", onClick);
+      canvas.removeEventListener("keydown", onKey);
+      observer.disconnect();
+    };
+  }, []);
+
+  return { canvasRef };
+}

autogpt_platform/frontend/src/app/(platform)/copilot/tools/GenericTool/GenericTool.tsx

@@ -0,0 +1,63 @@
+"use client";
+
+import { ToolUIPart } from "ai";
+import { GearIcon } from "@phosphor-icons/react";
+import { MorphingTextAnimation } from "../../components/MorphingTextAnimation/MorphingTextAnimation";
+
+interface Props {
+  part: ToolUIPart;
+}
+
+function extractToolName(part: ToolUIPart): string {
+  // ToolUIPart.type is "tool-{name}", extract the name portion.
+  return part.type.replace(/^tool-/, "");
+}
+
+function formatToolName(name: string): string {
+  // "search_docs" → "Search docs", "Read" → "Read"
+  return name.replace(/_/g, " ").replace(/^\w/, (c) => c.toUpperCase());
+}
+
+function getAnimationText(part: ToolUIPart): string {
+  const label = formatToolName(extractToolName(part));
+
+  switch (part.state) {
+    case "input-streaming":
+    case "input-available":
+      return `Running ${label}…`;
+    case "output-available":
+      return `${label} completed`;
+    case "output-error":
+      return `${label} failed`;
+    default:
+      return `Running ${label}…`;
+  }
+}
+
+export function GenericTool({ part }: Props) {
+  const isStreaming =
+    part.state === "input-streaming" || part.state === "input-available";
+  const isError = part.state === "output-error";
+
+  return (
+    <div className="py-2">
+      <div className="flex items-center gap-2 text-sm text-muted-foreground">
+        <GearIcon
+          size={14}
+          weight="regular"
+          className={
+            isError
+              ? "text-red-500"
+              : isStreaming
+                ? "animate-spin text-neutral-500"
+                : "text-neutral-400"
+          }
+        />
+        <MorphingTextAnimation
+          text={getAnimationText(part)}
+          className={isError ? "text-red-500" : undefined}
+        />
+      </div>
+    </div>
+  );
+}

autogpt_platform/frontend/src/app/(platform)/copilot/tools/RunBlock/RunBlock.tsx

@@ -3,6 +3,7 @@
 import type { ToolUIPart } from "ai";
 import { MorphingTextAnimation } from "../../components/MorphingTextAnimation/MorphingTextAnimation";
 import { ToolAccordion } from "../../components/ToolAccordion/ToolAccordion";
+import { BlockDetailsCard } from "./components/BlockDetailsCard/BlockDetailsCard";
 import { BlockOutputCard } from "./components/BlockOutputCard/BlockOutputCard";
 import { ErrorCard } from "./components/ErrorCard/ErrorCard";
 import { SetupRequirementsCard } from "./components/SetupRequirementsCard/SetupRequirementsCard";
@@ -11,6 +12,7 @@ import {
   getAnimationText,
   getRunBlockToolOutput,
   isRunBlockBlockOutput,
+  isRunBlockDetailsOutput,
   isRunBlockErrorOutput,
   isRunBlockSetupRequirementsOutput,
   ToolIcon,
@@ -41,6 +43,7 @@ export function RunBlockTool({ part }: Props) {
     part.state === "output-available" &&
     !!output &&
     (isRunBlockBlockOutput(output) ||
+      isRunBlockDetailsOutput(output) ||
       isRunBlockSetupRequirementsOutput(output) ||
       isRunBlockErrorOutput(output));
 
@@ -58,6 +61,10 @@ export function RunBlockTool({ part }: Props) {
         <ToolAccordion {...getAccordionMeta(output)}>
           {isRunBlockBlockOutput(output) && <BlockOutputCard output={output} />}
 
+          {isRunBlockDetailsOutput(output) && (
+            <BlockDetailsCard output={output} />
+          )}
+
           {isRunBlockSetupRequirementsOutput(output) && (
             <SetupRequirementsCard output={output} />
           )}

autogpt_platform/frontend/src/app/(platform)/copilot/tools/RunBlock/components/BlockDetailsCard/BlockDetailsCard.stories.tsx

@@ -0,0 +1,188 @@
+import type { Meta, StoryObj } from "@storybook/nextjs";
+import { ResponseType } from "@/app/api/__generated__/models/responseType";
+import type { BlockDetailsResponse } from "../../helpers";
+import { BlockDetailsCard } from "./BlockDetailsCard";
+
+const meta: Meta<typeof BlockDetailsCard> = {
+  title: "Copilot/RunBlock/BlockDetailsCard",
+  component: BlockDetailsCard,
+  parameters: {
+    layout: "centered",
+  },
+  tags: ["autodocs"],
+  decorators: [
+    (Story) => (
+      <div style={{ maxWidth: 480 }}>
+        <Story />
+      </div>
+    ),
+  ],
+};
+
+export default meta;
+type Story = StoryObj<typeof meta>;
+
+const baseBlock: BlockDetailsResponse = {
+  type: ResponseType.block_details,
+  message:
+    "Here are the details for the GetWeather block. Provide the required inputs to run it.",
+  session_id: "session-123",
+  user_authenticated: true,
+  block: {
+    id: "block-abc-123",
+    name: "GetWeather",
+    description: "Fetches current weather data for a given location.",
+    inputs: {
+      type: "object",
+      properties: {
+        location: {
+          title: "Location",
+          type: "string",
+          description:
+            "City name or coordinates (e.g. 'London' or '51.5,-0.1')",
+        },
+        units: {
+          title: "Units",
+          type: "string",
+          description: "Temperature units: 'metric' or 'imperial'",
+        },
+      },
+      required: ["location"],
+    },
+    outputs: {
+      type: "object",
+      properties: {
+        temperature: {
+          title: "Temperature",
+          type: "number",
+          description: "Current temperature in the requested units",
+        },
+        condition: {
+          title: "Condition",
+          type: "string",
+          description: "Weather condition description (e.g. 'Sunny', 'Rain')",
+        },
+      },
+    },
+    credentials: [],
+  },
+};
+
+export const Default: Story = {
+  args: {
+    output: baseBlock,
+  },
+};
+
+export const InputsOnly: Story = {
+  args: {
+    output: {
+      ...baseBlock,
+      message: "This block requires inputs. No outputs are defined.",
+      block: {
+        ...baseBlock.block,
+        outputs: {},
+      },
+    },
+  },
+};
+
+export const OutputsOnly: Story = {
+  args: {
+    output: {
+      ...baseBlock,
+      message: "This block has no required inputs.",
+      block: {
+        ...baseBlock.block,
+        inputs: {},
+      },
+    },
+  },
+};
+
+export const ManyFields: Story = {
+  args: {
+    output: {
+      ...baseBlock,
+      message: "Block with many input and output fields.",
+      block: {
+        ...baseBlock.block,
+        name: "SendEmail",
+        description: "Sends an email via SMTP.",
+        inputs: {
+          type: "object",
+          properties: {
+            to: {
+              title: "To",
+              type: "string",
+              description: "Recipient email address",
+            },
+            subject: {
+              title: "Subject",
+              type: "string",
+              description: "Email subject line",
+            },
+            body: {
+              title: "Body",
+              type: "string",
+              description: "Email body content",
+            },
+            cc: {
+              title: "CC",
+              type: "string",
+              description: "CC recipients (comma-separated)",
+            },
+            bcc: {
+              title: "BCC",
+              type: "string",
+              description: "BCC recipients (comma-separated)",
+            },
+          },
+          required: ["to", "subject", "body"],
+        },
+        outputs: {
+          type: "object",
+          properties: {
+            message_id: {
+              title: "Message ID",
+              type: "string",
+              description: "Unique ID of the sent email",
+            },
+            status: {
+              title: "Status",
+              type: "string",
+              description: "Delivery status",
+            },
+          },
+        },
+      },
+    },
+  },
+};
+
+export const NoFieldDescriptions: Story = {
+  args: {
+    output: {
+      ...baseBlock,
+      message: "Fields without descriptions.",
+      block: {
+        ...baseBlock.block,
+        name: "SimpleBlock",
+        inputs: {
+          type: "object",
+          properties: {
+            input_a: { title: "Input A", type: "string" },
+            input_b: { title: "Input B", type: "number" },
+          },
+          required: ["input_a"],
+        },
+        outputs: {
+          type: "object",
+          properties: {
+            result: { title: "Result", type: "string" },
+          },
+        },
+      },
+    },
+  },
+};

autogpt_platform/frontend/src/app/(platform)/copilot/tools/RunBlock/components/BlockDetailsCard/BlockDetailsCard.tsx

@@ -0,0 +1,103 @@
+"use client";
+
+import type { BlockDetailsResponse } from "../../helpers";
+import {
+  ContentBadge,
+  ContentCard,
+  ContentCardDescription,
+  ContentCardTitle,
+  ContentGrid,
+  ContentMessage,
+} from "../../../../components/ToolAccordion/AccordionContent";
+
+interface Props {
+  output: BlockDetailsResponse;
+}
+
+function SchemaFieldList({
+  title,
+  properties,
+  required,
+}: {
+  title: string;
+  properties: Record<string, unknown>;
+  required?: string[];
+}) {
+  const entries = Object.entries(properties);
+  if (entries.length === 0) return null;
+
+  const requiredSet = new Set(required ?? []);
+
+  return (
+    <ContentCard>
+      <ContentCardTitle className="text-xs">{title}</ContentCardTitle>
+      <div className="mt-2 grid gap-2">
+        {entries.map(([name, schema]) => {
+          const field = schema as Record<string, unknown> | undefined;
+          const fieldTitle =
+            typeof field?.title === "string" ? field.title : name;
+          const fieldType =
+            typeof field?.type === "string" ? field.type : "unknown";
+          const description =
+            typeof field?.description === "string"
+              ? field.description
+              : undefined;
+
+          return (
+            <div key={name} className="rounded-xl border p-2">
+              <div className="flex items-center justify-between gap-2">
+                <ContentCardTitle className="text-xs">
+                  {fieldTitle}
+                </ContentCardTitle>
+                <div className="flex gap-1">
+                  <ContentBadge>{fieldType}</ContentBadge>
+                  {requiredSet.has(name) && (
+                    <ContentBadge>Required</ContentBadge>
+                  )}
+                </div>
+              </div>
+              {description && (
+                <ContentCardDescription className="mt-1 text-xs">
+                  {description}
+                </ContentCardDescription>
+              )}
+            </div>
+          );
+        })}
+      </div>
+    </ContentCard>
+  );
+}
+
+export function BlockDetailsCard({ output }: Props) {
+  const inputs = output.block.inputs as {
+    properties?: Record<string, unknown>;
+    required?: string[];
+  } | null;
+  const outputs = output.block.outputs as {
+    properties?: Record<string, unknown>;
+    required?: string[];
+  } | null;
+
+  return (
+    <ContentGrid>
+      <ContentMessage>{output.message}</ContentMessage>
+
+      {inputs?.properties && Object.keys(inputs.properties).length > 0 && (
+        <SchemaFieldList
+          title="Inputs"
+          properties={inputs.properties}
+          required={inputs.required}
+        />
+      )}
+
+      {outputs?.properties && Object.keys(outputs.properties).length > 0 && (
+        <SchemaFieldList
+          title="Outputs"
+          properties={outputs.properties}
+          required={outputs.required}
+        />
+      )}
+    </ContentGrid>
+  );
+}

autogpt_platform/frontend/src/app/(platform)/copilot/tools/RunBlock/helpers.tsx

@@ -10,18 +10,37 @@ import {
 import type { ToolUIPart } from "ai";
 import { OrbitLoader } from "../../components/OrbitLoader/OrbitLoader";
 
+/** Block details returned on first run_block attempt (before input_data provided). */
+export interface BlockDetailsResponse {
+  type: typeof ResponseType.block_details;
+  message: string;
+  session_id?: string | null;
+  block: {
+    id: string;
+    name: string;
+    description: string;
+    inputs: Record<string, unknown>;
+    outputs: Record<string, unknown>;
+    credentials: unknown[];
+  };
+  user_authenticated: boolean;
+}
+
 export interface RunBlockInput {
   block_id?: string;
+  block_name?: string;
   input_data?: Record<string, unknown>;
 }
 
 export type RunBlockToolOutput =
   | SetupRequirementsResponse
+  | BlockDetailsResponse
   | BlockOutputResponse
   | ErrorResponse;
 
 const RUN_BLOCK_OUTPUT_TYPES = new Set<string>([
   ResponseType.setup_requirements,
+  ResponseType.block_details,
   ResponseType.block_output,
   ResponseType.error,
 ]);
@@ -35,6 +54,15 @@ export function isRunBlockSetupRequirementsOutput(
   );
 }
 
+export function isRunBlockDetailsOutput(
+  output: RunBlockToolOutput,
+): output is BlockDetailsResponse {
+  return (
+    output.type === ResponseType.block_details ||
+    ("block" in output && typeof output.block === "object")
+  );
+}
+
 export function isRunBlockBlockOutput(
   output: RunBlockToolOutput,
 ): output is BlockOutputResponse {
@@ -64,6 +92,7 @@ function parseOutput(output: unknown): RunBlockToolOutput | null {
       return output as RunBlockToolOutput;
     }
     if ("block_id" in output) return output as BlockOutputResponse;
+    if ("block" in output) return output as BlockDetailsResponse;
     if ("setup_info" in output) return output as SetupRequirementsResponse;
     if ("error" in output || "details" in output)
       return output as ErrorResponse;
@@ -84,17 +113,25 @@ export function getAnimationText(part: {
   output?: unknown;
 }): string {
   const input = part.input as RunBlockInput | undefined;
+  const blockName = input?.block_name?.trim();
   const blockId = input?.block_id?.trim();
-  const blockText = blockId ? ` "${blockId}"` : "";
+  // Prefer block_name if available, otherwise fall back to block_id
+  const blockText = blockName
+    ? ` "${blockName}"`
+    : blockId
+      ? ` "${blockId}"`
+      : "";
 
   switch (part.state) {
     case "input-streaming":
     case "input-available":
-      return `Running the block${blockText}`;
+      return `Running${blockText}`;
     case "output-available": {
       const output = parseOutput(part.output);
-      if (!output) return `Running the block${blockText}`;
+      if (!output) return `Running${blockText}`;
       if (isRunBlockBlockOutput(output)) return `Ran "${output.block_name}"`;
+      if (isRunBlockDetailsOutput(output))
+        return `Details for "${output.block.name}"`;
       if (isRunBlockSetupRequirementsOutput(output)) {
         return `Setup needed for "${output.setup_info.agent_name}"`;
       }
@@ -158,6 +195,21 @@ export function getAccordionMeta(output: RunBlockToolOutput): {
     };
   }
 
+  if (isRunBlockDetailsOutput(output)) {
+    const inputKeys = Object.keys(
+      (output.block.inputs as { properties?: Record<string, unknown> })
+        ?.properties ?? {},
+    );
+    return {
+      icon,
+      title: output.block.name,
+      description:
+        inputKeys.length > 0
+          ? `${inputKeys.length} input field${inputKeys.length === 1 ? "" : "s"} available`
+          : output.message,
+    };
+  }
+
   if (isRunBlockSetupRequirementsOutput(output)) {
     const missingCredsCount = Object.keys(
       (output.setup_info.user_readiness?.missing_credentials ?? {}) as Record<

autogpt_platform/frontend/src/app/(platform)/copilot/useCopilotPage.ts

@@ -1,10 +1,14 @@
 import { useGetV2ListSessions } from "@/app/api/__generated__/endpoints/chat/chat";
+import { toast } from "@/components/molecules/Toast/use-toast";
 import { useBreakpoint } from "@/lib/hooks/useBreakpoint";
 import { useSupabase } from "@/lib/supabase/hooks/useSupabase";
 import { useChat } from "@ai-sdk/react";
 import { DefaultChatTransport } from "ai";
-import { useEffect, useMemo, useState } from "react";
+import { useEffect, useMemo, useRef, useState } from "react";
 import { useChatSession } from "./useChatSession";
+import { useLongRunningToolPolling } from "./hooks/useLongRunningToolPolling";
+
+const STREAM_START_TIMEOUT_MS = 12_000;
 
 export function useCopilotPage() {
   const { isUserLoading, isLoggedIn } = useSupabase();
@@ -52,6 +56,24 @@ export function useCopilotPage() {
     transport: transport ?? undefined,
   });
 
+  // Abort the stream if the backend doesn't start sending data within 12s.
+  const stopRef = useRef(stop);
+  stopRef.current = stop;
+  useEffect(() => {
+    if (status !== "submitted") return;
+
+    const timer = setTimeout(() => {
+      stopRef.current();
+      toast({
+        title: "Stream timed out",
+        description: "The server took too long to respond. Please try again.",
+        variant: "destructive",
+      });
+    }, STREAM_START_TIMEOUT_MS);
+
+    return () => clearTimeout(timer);
+  }, [status]);
+
   useEffect(() => {
     if (!hydratedMessages || hydratedMessages.length === 0) return;
     setMessages((prev) => {
@@ -60,6 +82,11 @@ export function useCopilotPage() {
     });
   }, [hydratedMessages, setMessages]);
 
+  // Poll session endpoint when a long-running tool (create_agent, edit_agent)
+  // is in progress. When the backend completes, the session data will contain
+  // the final tool output — this hook detects the change and updates messages.
+  useLongRunningToolPolling(sessionId, messages, setMessages);
+
   // Clear messages when session is null
   useEffect(() => {
     if (!sessionId) setMessages([]);

autogpt_platform/frontend/src/app/(platform)/library/agents/[id]/components/NewAgentLibraryView/components/sidebar/SidebarRunsList/components/ScheduleListItem.tsx

@@ -29,6 +29,7 @@ export function ScheduleListItem({
       description={formatDistanceToNow(schedule.next_run_time, {
         addSuffix: true,
       })}
+      descriptionTitle={new Date(schedule.next_run_time).toString()}
       onClick={onClick}
       selected={selected}
       icon={

autogpt_platform/frontend/src/app/(platform)/library/agents/[id]/components/NewAgentLibraryView/components/sidebar/SidebarRunsList/components/SidebarItemCard.tsx

@@ -7,6 +7,7 @@ import React from "react";
 interface Props {
   title: string;
   description?: string;
+  descriptionTitle?: string;
   icon?: React.ReactNode;
   selected?: boolean;
   onClick?: () => void;
@@ -16,6 +17,7 @@ interface Props {
 export function SidebarItemCard({
   title,
   description,
+  descriptionTitle,
   icon,
   selected,
   onClick,
@@ -38,7 +40,11 @@ export function SidebarItemCard({
           >
             {title}
           </Text>
-          <Text variant="body" className="leading-tight !text-zinc-500">
+          <Text
+            variant="body"
+            className="leading-tight !text-zinc-500"
+            title={descriptionTitle}
+          >
             {description}
           </Text>
         </div>

autogpt_platform/frontend/src/app/(platform)/library/agents/[id]/components/NewAgentLibraryView/components/sidebar/SidebarRunsList/components/TaskListItem.tsx

@@ -81,6 +81,9 @@ export function TaskListItem({
           ? formatDistanceToNow(run.started_at, { addSuffix: true })
           : "—"
       }
+      descriptionTitle={
+        run.started_at ? new Date(run.started_at).toString() : undefined
+      }
       onClick={onClick}
       selected={selected}
       actions={

autogpt_platform/frontend/src/app/api/openapi.json

@@ -1053,6 +1053,7 @@
                       "$ref": "#/components/schemas/ClarificationNeededResponse"
                     },
                     { "$ref": "#/components/schemas/BlockListResponse" },
+                    { "$ref": "#/components/schemas/BlockDetailsResponse" },
                     { "$ref": "#/components/schemas/BlockOutputResponse" },
                     { "$ref": "#/components/schemas/DocSearchResultsResponse" },
                     { "$ref": "#/components/schemas/DocPageResponse" },
@@ -6958,6 +6959,58 @@
         "enum": ["run", "byte", "second"],
         "title": "BlockCostType"
       },
+      "BlockDetails": {
+        "properties": {
+          "id": { "type": "string", "title": "Id" },
+          "name": { "type": "string", "title": "Name" },
+          "description": { "type": "string", "title": "Description" },
+          "inputs": {
+            "additionalProperties": true,
+            "type": "object",
+            "title": "Inputs",
+            "default": {}
+          },
+          "outputs": {
+            "additionalProperties": true,
+            "type": "object",
+            "title": "Outputs",
+            "default": {}
+          },
+          "credentials": {
+            "items": { "$ref": "#/components/schemas/CredentialsMetaInput" },
+            "type": "array",
+            "title": "Credentials",
+            "default": []
+          }
+        },
+        "type": "object",
+        "required": ["id", "name", "description"],
+        "title": "BlockDetails",
+        "description": "Detailed block information."
+      },
+      "BlockDetailsResponse": {
+        "properties": {
+          "type": {
+            "$ref": "#/components/schemas/ResponseType",
+            "default": "block_details"
+          },
+          "message": { "type": "string", "title": "Message" },
+          "session_id": {
+            "anyOf": [{ "type": "string" }, { "type": "null" }],
+            "title": "Session Id"
+          },
+          "block": { "$ref": "#/components/schemas/BlockDetails" },
+          "user_authenticated": {
+            "type": "boolean",
+            "title": "User Authenticated",
+            "default": false
+          }
+        },
+        "type": "object",
+        "required": ["message", "block"],
+        "title": "BlockDetailsResponse",
+        "description": "Response for block details (first run_block attempt)."
+      },
       "BlockInfo": {
         "properties": {
           "id": { "type": "string", "title": "Id" },
@@ -7022,29 +7075,24 @@
           "input_schema": {
             "additionalProperties": true,
             "type": "object",
-            "title": "Input Schema"
+            "title": "Input Schema",
+            "description": "Full JSON schema for block inputs"
           },
           "output_schema": {
             "additionalProperties": true,
             "type": "object",
-            "title": "Output Schema"
+            "title": "Output Schema",
+            "description": "Full JSON schema for block outputs"
           },
           "required_inputs": {
             "items": { "$ref": "#/components/schemas/BlockInputFieldInfo" },
             "type": "array",
             "title": "Required Inputs",
-            "description": "List of required input fields for this block"
+            "description": "List of input fields for this block"
           }
         },
         "type": "object",
-        "required": [
+        "required": ["id", "name", "description", "categories"],
-          "id",
-          "name",
-          "description",
-          "categories",
-          "input_schema",
-          "output_schema"
-        ],
         "title": "BlockInfoSummary",
         "description": "Summary of a block for search results."
       },
@@ -7090,7 +7138,7 @@
           "usage_hint": {
             "type": "string",
             "title": "Usage Hint",
-            "default": "To execute a block, call run_block with block_id set to the block's 'id' field and input_data containing the required fields from input_schema."
+            "default": "To execute a block, call run_block with block_id set to the block's 'id' field and input_data containing the fields listed in required_inputs."
           }
         },
         "type": "object",
@@ -10484,6 +10532,7 @@
           "agent_saved",
           "clarification_needed",
           "block_list",
+          "block_details",
           "block_output",
           "doc_search_results",
           "doc_page",
@@ -10495,7 +10544,10 @@
           "operation_started",
           "operation_pending",
           "operation_in_progress",
-          "input_validation_error"
+          "input_validation_error",
+          "web_fetch",
+          "bash_exec",
+          "operation_status"
         ],
         "title": "ResponseType",
         "description": "Types of tool responses."

autogpt_platform/frontend/src/app/globals.css

@@ -180,3 +180,14 @@ body[data-google-picker-open="true"] [data-dialog-content] {
   z-index: 1 !important;
   pointer-events: none !important;
 }
+
+/* CoPilot chat table styling — remove left/right borders, increase padding */
+[data-streamdown="table-wrapper"] table {
+  border-left: none;
+  border-right: none;
+}
+
+[data-streamdown="table-wrapper"] th,
+[data-streamdown="table-wrapper"] td {
+  padding: 0.875rem 1rem; /* py-3.5 px-4 */
+}

autogpt_platform/frontend/src/components/contextual/CredentialsInput/components/APIKeyCredentialsModal/APIKeyCredentialsModal.tsx

@@ -30,6 +30,7 @@ export function APIKeyCredentialsModal({
   const {
     form,
     isLoading,
+    isSubmitting,
     supportsApiKey,
     providerName,
     schemaDescription,
@@ -138,7 +139,12 @@ export function APIKeyCredentialsModal({
                 />
               )}
             />
-            <Button type="submit" className="min-w-68">
+            <Button
+              type="submit"
+              className="min-w-68"
+              loading={isSubmitting}
+              disabled={isSubmitting}
+            >
               Add API Key
             </Button>
           </form>

autogpt_platform/frontend/src/components/contextual/CredentialsInput/components/APIKeyCredentialsModal/useAPIKeyCredentialsModal.ts

@@ -4,6 +4,7 @@ import {
   CredentialsMetaInput,
 } from "@/lib/autogpt-server-api/types";
 import { zodResolver } from "@hookform/resolvers/zod";
+import { useState } from "react";
 import { useForm, type UseFormReturn } from "react-hook-form";
 import { z } from "zod";
 
@@ -26,6 +27,7 @@ export function useAPIKeyCredentialsModal({
 }: Args): {
   form: UseFormReturn<APIKeyFormValues>;
   isLoading: boolean;
+  isSubmitting: boolean;
   supportsApiKey: boolean;
   provider?: string;
   providerName?: string;
@@ -33,6 +35,7 @@ export function useAPIKeyCredentialsModal({
   onSubmit: (values: APIKeyFormValues) => Promise<void>;
 } {
   const credentials = useCredentials(schema, siblingInputs);
+  const [isSubmitting, setIsSubmitting] = useState(false);
 
   const formSchema = z.object({
     apiKey: z.string().min(1, "API Key is required"),
@@ -40,48 +43,42 @@ export function useAPIKeyCredentialsModal({
     expiresAt: z.string().optional(),
   });
 
-  function getDefaultExpirationDate(): string {
-    const tomorrow = new Date();
-    tomorrow.setDate(tomorrow.getDate() + 1);
-    tomorrow.setHours(0, 0, 0, 0);
-    const year = tomorrow.getFullYear();
-    const month = String(tomorrow.getMonth() + 1).padStart(2, "0");
-    const day = String(tomorrow.getDate()).padStart(2, "0");
-    const hours = String(tomorrow.getHours()).padStart(2, "0");
-    const minutes = String(tomorrow.getMinutes()).padStart(2, "0");
-    return `${year}-${month}-${day}T${hours}:${minutes}`;
-  }
-
   const form = useForm<APIKeyFormValues>({
     resolver: zodResolver(formSchema),
     defaultValues: {
       apiKey: "",
       title: "",
-      expiresAt: getDefaultExpirationDate(),
+      expiresAt: "",
     },
   });
 
   async function onSubmit(values: APIKeyFormValues) {
     if (!credentials || credentials.isLoading) return;
-    const expiresAt = values.expiresAt
+    setIsSubmitting(true);
-      ? new Date(values.expiresAt).getTime() / 1000
+    try {
-      : undefined;
+      const expiresAt = values.expiresAt
-    const newCredentials = await credentials.createAPIKeyCredentials({
+        ? new Date(values.expiresAt).getTime() / 1000
-      api_key: values.apiKey,
+        : undefined;
-      title: values.title,
+      const newCredentials = await credentials.createAPIKeyCredentials({
-      expires_at: expiresAt,
+        api_key: values.apiKey,
-    });
+        title: values.title,
-    onCredentialsCreate({
+        expires_at: expiresAt,
-      provider: credentials.provider,
+      });
-      id: newCredentials.id,
+      onCredentialsCreate({
-      type: "api_key",
+        provider: credentials.provider,
-      title: newCredentials.title,
+        id: newCredentials.id,
-    });
+        type: "api_key",
+        title: newCredentials.title,
+      });
+    } finally {
+      setIsSubmitting(false);
+    }
   }
 
   return {
     form,
     isLoading: !credentials || credentials.isLoading,
+    isSubmitting,
     supportsApiKey: !!credentials?.supportsApiKey,
     provider: credentials?.provider,
     providerName:

autogpt_platform/frontend/src/components/contextual/OutputRenderers/renderers/MarkdownRenderer.tsx

@@ -226,7 +226,7 @@ function renderMarkdown(
           table: ({ children, ...props }) => (
             <div className="my-4 overflow-x-auto">
               <table
-                className="min-w-full divide-y divide-gray-200 rounded-lg border border-gray-200 dark:divide-gray-700 dark:border-gray-700"
+                className="min-w-full divide-y divide-gray-200 border-y border-gray-200 dark:divide-gray-700 dark:border-gray-700"
                 {...props}
               >
                 {children}
@@ -235,7 +235,7 @@ function renderMarkdown(
           ),
           th: ({ children, ...props }) => (
             <th
-              className="bg-gray-50 px-4 py-3 text-left text-xs font-semibold uppercase tracking-wider text-gray-700 dark:bg-gray-800 dark:text-gray-300"
+              className="bg-gray-50 px-4 py-3.5 text-left text-xs font-semibold uppercase tracking-wider text-gray-700 dark:bg-gray-800 dark:text-gray-300"
               {...props}
             >
               {children}
@@ -243,7 +243,7 @@ function renderMarkdown(
           ),
           td: ({ children, ...props }) => (
             <td
-              className="border-t border-gray-200 px-4 py-3 text-sm text-gray-600 dark:border-gray-700 dark:text-gray-400"
+              className="border-t border-gray-200 px-4 py-3.5 text-sm text-gray-600 dark:border-gray-700 dark:text-gray-400"
               {...props}
             >
               {children}

docs/integrations/block-integrations/llm.md

@@ -563,7 +563,7 @@ The block supports conversation continuation through three mechanisms:
 |--------|-------------|------|
 | error | Error message if execution failed | str |
 | response | The output/response from Claude Code execution | str |
-| files | List of text files created/modified by Claude Code during this execution. Each file has 'path', 'relative_path', 'name', and 'content' fields. | List[FileOutput] |
+| files | List of text files created/modified by Claude Code during this execution. Each file has 'path', 'relative_path', 'name', 'content', and 'workspace_ref' fields. workspace_ref contains a workspace:// URI if the file was stored to workspace. | List[SandboxFileOutput] |
 | conversation_history | Full conversation history including this turn. Pass this to conversation_history input to continue on a fresh sandbox if the previous sandbox timed out. | str |
 | session_id | Session ID for this conversation. Pass this back along with sandbox_id to continue the conversation. | str |
 | sandbox_id | ID of the sandbox instance. Pass this back along with session_id to continue the conversation. This is None if dispose_sandbox was True (sandbox was disposed). | str |

docs/integrations/block-integrations/misc.md

@@ -215,6 +215,7 @@ The sandbox includes pip and npm pre-installed. Set timeout to limit execution t
 | response | Text output (if any) of the main execution result | str |
 | stdout_logs | Standard output logs from execution | str |
 | stderr_logs | Standard error logs from execution | str |
+| files | Files created or modified during execution. Each file has path, name, content, and workspace_ref (if stored). | List[SandboxFileOutput] |
 
 ### Possible use case
 <!-- MANUAL: use_case -->