dx(orchestrate): fix stale-review gate and add pr-test evaluation rules to SKILL.md (#12701)

## Changes

### verify-complete.sh
- CHANGES_REQUESTED reviews are now compared against the latest commit
timestamp. If the review was submitted **before** the latest commit, it
is treated as stale and does not block verification.
- Added fail-closed guard: if the `gh pr view` fetch fails, the script
exits 1 (rather than treating missing data as "no blocking reviews")
- Fixed edge case: a `CHANGES_REQUESTED` review with a null
`submittedAt` is now counted as fresh/blocking (previously silently
skipped)
- Combined two separate `gh pr view` calls into one (`--json
commits,reviews`) to reduce API calls and ensure consistency

### SKILL.md (orchestrate skill)
- Added `### /pr-test result evaluation` section with explicit
pass/partial/fail handling table
- **PARTIAL on any headline feature scenario = immediate blocker**:
re-brief the agent, fix, and re-run from scratch. Never approve or
output ORCHESTRATOR:DONE with a PARTIAL headline result.
- Concrete incident callout: PR #12699 S5 (Apply suggestions) was
PARTIAL — AI never output JSON action blocks — but was nearly approved.
This rule prevents recurrence.
- Updated `verify-complete.sh` description throughout to include "no
fresh CHANGES_REQUESTED"
- Added staleness rule documentation: a review only blocks if submitted
*after* the latest commit

## Why

Two separate incidents prompted these changes:

1. **verify-complete.sh false positive**: An automated bot
(autogpt-pr-reviewer) submitted a `CHANGES_REQUESTED` review in April.
An agent then pushed fixing commits. The old script still blocked on the
stale review, preventing the PR from being verified as done.

2. **Missed PARTIAL signal**: PR #12699 had a PARTIAL result on its
headline scenario (S5 Apply button) because the AI emitted direct
builder tool calls instead of JSON action blocks. The orchestrator
nearly approved it. The new SKILL.md rule makes PARTIAL = blocker
explicit.

## Checklist

- [x] I have read the contribution guide
- [x] My changes follow the code style of this project  
- [x] Changes are limited to the scope of this PR (< 20% unrelated
changes)
- [x] All new and existing tests pass

.claude/skills/orchestrate/SKILL.md

@@ -0,0 +1,545 @@
+---
+name: orchestrate
+description: "Meta-agent supervisor that manages a fleet of Claude Code agents running in tmux windows. Auto-discovers spare worktrees, spawns agents, monitors state, kicks idle agents, approves safe confirmations, and recycles worktrees when done. TRIGGER when user asks to supervise agents, run parallel tasks, manage worktrees, check agent status, or orchestrate parallel work."
+user-invocable: true
+argument-hint: "any free text — e.g. 'start 3 agents on X Y Z', 'show status', 'add task: implement feature A', 'stop', 'how many are free?'"
+metadata:
+  author: autogpt-team
+  version: "6.0.0"
+---
+
+# Orchestrate — Agent Fleet Supervisor
+
+One tmux session, N windows — each window is one agent working in its own worktree. Speak naturally; Claude maps your intent to the right scripts.
+
+## Scripts
+
+```bash
+SKILLS_DIR=$(git rev-parse --show-toplevel)/.claude/skills/orchestrate/scripts
+STATE_FILE=~/.claude/orchestrator-state.json
+```
+
+| Script | Purpose |
+|---|---|
+| `find-spare.sh [REPO_ROOT]` | List free worktrees — one `PATH BRANCH` per line |
+| `spawn-agent.sh SESSION PATH SPARE NEW_BRANCH OBJECTIVE [PR_NUMBER] [STEPS...]` | Create window + checkout branch + launch claude + send task. **Stdout: `SESSION:WIN` only** |
+| `recycle-agent.sh WINDOW PATH SPARE_BRANCH` | Kill window + restore spare branch |
+| `run-loop.sh` | **Mechanical babysitter** — idle restart + dialog approval + recycle on ORCHESTRATOR:DONE + supervisor health check + all-done notification |
+| `verify-complete.sh WINDOW` | Verify PR is done: checkpoints ✓ + 0 unresolved threads + CI green + no fresh CHANGES_REQUESTED. Repo auto-derived from state file `.repo` or git remote. |
+| `notify.sh MESSAGE` | Send notification via Discord webhook (env `DISCORD_WEBHOOK_URL` or state `.discord_webhook`), macOS notification center, and stdout |
+| `capacity.sh [REPO_ROOT]` | Print available + in-use worktrees |
+| `status.sh` | Print fleet status + live pane commands |
+| `poll-cycle.sh` | One monitoring cycle — classifies panes, tracks checkpoints, returns JSON action array |
+| `classify-pane.sh WINDOW` | Classify one pane state |
+
+## Supervision model
+
+```
+Orchestrating Claude (this Claude session — IS the supervisor)
+  └── Reads pane output, checks CI, intervenes with targeted guidance
+        run-loop.sh (separate tmux window, every 30s)
+          └── Mechanical only: idle restart, dialog approval, recycle on ORCHESTRATOR:DONE
+```
+
+**You (the orchestrating Claude)** are the supervisor. After spawning agents, stay in this conversation and actively monitor: poll each agent's pane every 2-3 minutes, check CI, nudge stalled agents, and verify completions. Do not spawn a separate supervisor Claude window — it loses context, is hard to observe, and compounds context compression problems.
+
+**run-loop.sh** is the mechanical layer — zero tokens, handles things that need no judgment: restart crashed agents, press Enter on dialogs, recycle completed worktrees (only after `verify-complete.sh` passes).
+
+## Checkpoint protocol
+
+Agents output checkpoints as they complete each required step:
+
+```
+CHECKPOINT:<step-name>
+```
+
+Required steps are passed as args to `spawn-agent.sh` (e.g. `pr-address pr-test`). `run-loop.sh` will not recycle a window until all required checkpoints are found in the pane output. If `verify-complete.sh` fails, the agent is re-briefed automatically.
+
+## Worktree lifecycle
+
+```text
+spare/N branch  →  spawn-agent.sh (--session-id UUID)  →  window + feat/branch + claude running
+                                                                 ↓
+                                               CHECKPOINT:<step> (as steps complete)
+                                                                 ↓
+                                                        ORCHESTRATOR:DONE
+                                                                 ↓
+                                    verify-complete.sh: checkpoints ✓ + 0 threads + CI green + no fresh CHANGES_REQUESTED
+                                                                 ↓
+                                              state → "done", notify, window KEPT OPEN
+                                                                 ↓
+                              user/orchestrator explicitly requests recycle
+                                                                 ↓
+                                         recycle-agent.sh → spare/N (free again)
+```
+
+**Windows are never auto-killed.** The worktree stays on its branch, the session stays alive. The agent is done working but the window, git state, and Claude session are all preserved until you choose to recycle.
+
+**To resume a done or crashed session:**
+```bash
+# Resume by stored session ID (preferred — exact session, full context)
+claude --resume SESSION_ID --permission-mode bypassPermissions
+
+# Or resume most recent session in that worktree directory
+cd /path/to/worktree && claude --continue --permission-mode bypassPermissions
+```
+
+**To manually recycle when ready:**
+```bash
+bash ~/.claude/orchestrator/scripts/recycle-agent.sh SESSION:WIN WORKTREE_PATH spare/N
+# Then update state:
+jq --arg w "SESSION:WIN" '.agents |= map(if .window == $w then .state = "recycled" else . end)' \
+  ~/.claude/orchestrator-state.json > /tmp/orch.tmp && mv /tmp/orch.tmp ~/.claude/orchestrator-state.json
+```
+
+## State file (`~/.claude/orchestrator-state.json`)
+
+Never committed to git. You maintain this file directly using `jq` + atomic writes (`.tmp` → `mv`).
+
+```json
+{
+  "active": true,
+  "tmux_session": "autogpt1",
+  "idle_threshold_seconds": 300,
+  "loop_window": "autogpt1:5",
+  "repo": "Significant-Gravitas/AutoGPT",
+  "discord_webhook": "https://discord.com/api/webhooks/...",
+  "last_poll_at": 0,
+  "agents": [
+    {
+      "window": "autogpt1:3",
+      "worktree": "AutoGPT6",
+      "worktree_path": "/path/to/AutoGPT6",
+      "spare_branch": "spare/6",
+      "branch": "feat/my-feature",
+      "objective": "Implement X and open a PR",
+      "pr_number": "12345",
+      "session_id": "550e8400-e29b-41d4-a716-446655440000",
+      "steps": ["pr-address", "pr-test"],
+      "checkpoints": ["pr-address"],
+      "state": "running",
+      "last_output_hash": "",
+      "last_seen_at": 0,
+      "spawned_at": 0,
+      "idle_since": 0,
+      "revision_count": 0,
+      "last_rebriefed_at": 0
+    }
+  ]
+}
+```
+
+Top-level optional fields:
+- `repo` — GitHub `owner/repo` for CI/thread checks. Auto-derived from git remote if omitted.
+- `discord_webhook` — Discord webhook URL for completion notifications. Also reads `DISCORD_WEBHOOK_URL` env var.
+
+Per-agent fields:
+- `session_id` — UUID passed to `claude --session-id` at spawn; use with `claude --resume UUID` to restore exact session context after a crash or window close.
+- `last_rebriefed_at` — Unix timestamp of last re-brief; enforces 5-min cooldown to prevent spam.
+
+Agent states: `running` | `idle` | `stuck` | `waiting_approval` | `complete` | `done` | `escalated`
+
+`done` means verified complete — window is still open, session still alive, worktree still on task branch. Not recycled yet.
+
+## Serial /pr-test rule
+
+`/pr-test` and `/pr-test --fix` run local Docker + integration tests that use shared ports, a shared database, and shared build caches. **Running two `/pr-test` jobs simultaneously will cause port conflicts and database corruption.**
+
+**Rule: only one `/pr-test` runs at a time. The orchestrator serializes them.**
+
+You (the orchestrating Claude) own the test queue:
+1. Agents do `pr-review` and `pr-address` in parallel — that's safe (they only push code and reply to GitHub).
+2. When a PR needs local testing, add it to your mental queue — don't give agents a `pr-test` step.
+3. Run `/pr-test https://github.com/OWNER/REPO/pull/PR_NUMBER --fix` yourself, sequentially.
+4. Feed results back to the relevant agent via `tmux send-keys`:
+   ```bash
+   tmux send-keys -t SESSION:WIN "Local tests for PR #N: <paste failure output or 'all passed'>. Fix any failures and push, then output ORCHESTRATOR:DONE."
+   sleep 0.3
+   tmux send-keys -t SESSION:WIN Enter
+   ```
+5. Wait for CI to confirm green before marking the agent done.
+
+If multiple PRs need testing at the same time, pick the one furthest along (fewest pending CI checks) and test it first. Only start the next test after the previous one completes.
+
+## Session restore (tested and confirmed)
+
+Agent sessions are saved to disk. To restore a closed or crashed session:
+
+```bash
+# If session_id is in state (preferred):
+NEW_WIN=$(tmux new-window -t SESSION -n WORKTREE_NAME -P -F '#{window_index}')
+tmux send-keys -t "SESSION:${NEW_WIN}" "cd /path/to/worktree && claude --resume SESSION_ID --permission-mode bypassPermissions" Enter
+
+# If no session_id (use --continue for most recent session in that directory):
+tmux send-keys -t "SESSION:${NEW_WIN}" "cd /path/to/worktree && claude --continue --permission-mode bypassPermissions" Enter
+```
+
+`--continue` restores the full conversation history including all tool calls, file edits, and context. The agent resumes exactly where it left off. After restoring, update the window address in the state file:
+
+```bash
+jq --arg old "SESSION:OLD_WIN" --arg new "SESSION:NEW_WIN" \
+  '(.agents[] | select(.window == $old)).window = $new' \
+  ~/.claude/orchestrator-state.json > /tmp/orch.tmp && mv /tmp/orch.tmp ~/.claude/orchestrator-state.json
+```
+
+## Intent → action mapping
+
+Match the user's message to one of these intents:
+
+| The user says something like… | What to do |
+|---|---|
+| "status", "what's running", "show agents" | Run `status.sh` + `capacity.sh`, show output |
+| "how many free", "capacity", "available worktrees" | Run `capacity.sh`, show output |
+| "start N agents on X, Y, Z" or "run these tasks: …" | See **Spawning agents** below |
+| "add task: …", "add one more agent for …" | See **Adding an agent** below |
+| "stop", "shut down", "pause the fleet" | See **Stopping** below |
+| "poll", "check now", "run a cycle" | Run `poll-cycle.sh`, process actions |
+| "recycle window X", "free up autogpt3" | Run `recycle-agent.sh` directly |
+
+When the intent is ambiguous, show capacity first and ask what tasks to run.
+
+## Spawning agents
+
+### 1. Resolve tmux session
+
+```bash
+tmux list-sessions -F "#{session_name}: #{session_windows} windows" 2>/dev/null
+```
+
+Use an existing session. **Never create a tmux session from within Claude** — it becomes a child of Claude's process and dies when the session ends. If no session exists, tell the user to run `tmux new-session -d -s autogpt1` in their terminal first, then re-invoke `/orchestrate`.
+
+### 2. Show available capacity
+
+```bash
+bash $SKILLS_DIR/capacity.sh $(git rev-parse --show-toplevel)
+```
+
+### 3. Collect tasks from the user
+
+For each task, gather:
+- **objective** — what to do (e.g. "implement feature X and open a PR")
+- **branch name** — e.g. `feat/my-feature` (derive from objective if not given)
+- **pr_number** — GitHub PR number if working on an existing PR (for verification)
+- **steps** — required checkpoint names in order (e.g. `pr-address pr-test`) — derive from objective
+
+Ask for `idle_threshold_seconds` only if the user mentions it (default: 300).
+
+Never ask the user to specify a worktree — auto-assign from `find-spare.sh`.
+
+### 4. Spawn one agent per task
+
+```bash
+# Get ordered list of spare worktrees
+SPARE_LIST=$(bash $SKILLS_DIR/find-spare.sh $(git rev-parse --show-toplevel))
+
+# For each task, take the next spare line:
+WORKTREE_PATH=$(echo "$SPARE_LINE" | awk '{print $1}')
+SPARE_BRANCH=$(echo "$SPARE_LINE" | awk '{print $2}')
+
+# With PR number and required steps:
+WINDOW=$(bash $SKILLS_DIR/spawn-agent.sh "$SESSION" "$WORKTREE_PATH" "$SPARE_BRANCH" "$NEW_BRANCH" "$OBJECTIVE" "$PR_NUMBER" "pr-address" "pr-test")
+
+# Without PR (new work):
+WINDOW=$(bash $SKILLS_DIR/spawn-agent.sh "$SESSION" "$WORKTREE_PATH" "$SPARE_BRANCH" "$NEW_BRANCH" "$OBJECTIVE")
+```
+
+Build an agent record and append it to the state file. If the state file doesn't exist yet, initialize it:
+
+```bash
+# Derive repo from git remote (used by verify-complete.sh + supervisor)
+REPO=$(git remote get-url origin 2>/dev/null | sed 's|.*github\.com[:/]||; s|\.git$||' || echo "")
+
+jq -n \
+  --arg session "$SESSION" \
+  --arg repo "$REPO" \
+  --argjson threshold 300 \
+  '{active:true, tmux_session:$session, idle_threshold_seconds:$threshold,
+    repo:$repo, loop_window:null, supervisor_window:null, last_poll_at:0, agents:[]}' \
+  > ~/.claude/orchestrator-state.json
+```
+
+Optionally add a Discord webhook for completion notifications:
+```bash
+jq --arg hook "$DISCORD_WEBHOOK_URL" '.discord_webhook = $hook' ~/.claude/orchestrator-state.json \
+  > /tmp/orch.tmp && mv /tmp/orch.tmp ~/.claude/orchestrator-state.json
+```
+
+`spawn-agent.sh` writes the initial agent record (window, worktree_path, branch, objective, state, etc.) to the state file automatically — **do not append the record again after calling it.** The record already exists and `pr_number`/`steps` are patched in by the script itself.
+
+### 5. Start the mechanical babysitter
+
+```bash
+LOOP_WIN=$(tmux new-window -t "$SESSION" -n "orchestrator" -P -F '#{window_index}')
+LOOP_WINDOW="${SESSION}:${LOOP_WIN}"
+tmux send-keys -t "$LOOP_WINDOW" "bash $SKILLS_DIR/run-loop.sh" Enter
+
+jq --arg w "$LOOP_WINDOW" '.loop_window = $w' ~/.claude/orchestrator-state.json \
+  > /tmp/orch.tmp && mv /tmp/orch.tmp ~/.claude/orchestrator-state.json
+```
+
+### 6. Begin supervising directly in this conversation
+
+You are the supervisor. After spawning, immediately start your first poll loop (see **Supervisor duties** below) and continue every 2-3 minutes. Do NOT spawn a separate supervisor Claude window.
+
+## Adding an agent
+
+Find the next spare worktree, then spawn and append to state — same as steps 2–4 above but for a single task. If no spare worktrees are available, tell the user.
+
+## Supervisor duties (YOUR job, every 2-3 min in this conversation)
+
+You are the supervisor. Run this poll loop directly in your Claude session — not in a separate window.
+
+### Poll loop mechanism
+
+You are reactive — you only act when a tool completes or the user sends a message. To create a self-sustaining poll loop without user involvement:
+
+1. Start each poll with `run_in_background: true` + a sleep before the work:
+   ```bash
+   sleep 120 && tmux capture-pane -t autogpt1:0 -p -S -200 | tail -40
+   # + similar for each active window
+   ```
+2. When the background job notifies you, read the pane output and take action.
+3. Immediately schedule the next background poll — this keeps the loop alive.
+4. Stop scheduling when all agents are done/escalated.
+
+**Never tell the user "I'll poll every 2-3 minutes"** — that does nothing without a trigger. Start the background job instead.
+
+### Each poll: what to check
+
+```bash
+# 1. Read state
+cat ~/.claude/orchestrator-state.json | jq '.agents[] | {window, worktree, branch, state, pr_number, checkpoints}'
+
+# 2. For each running/stuck/idle agent, capture pane
+tmux capture-pane -t SESSION:WIN -p -S -200 | tail -60
+```
+
+For each agent, decide:
+
+| What you see | Action |
+|---|---|
+| Spinner / tools running | Do nothing — agent is working |
+| Idle `❯` prompt, no `ORCHESTRATOR:DONE` | Stalled — send specific nudge with objective from state |
+| Stuck in error loop | Send targeted fix with exact error + solution |
+| Waiting for input / question | Answer and unblock via `tmux send-keys` |
+| CI red | `gh pr checks PR_NUMBER --repo REPO` → tell agent exactly what's failing |
+| Context compacted / agent lost | Send recovery: `cat ~/.claude/orchestrator-state.json | jq '.agents[] | select(.window=="WIN")'` + `gh pr view PR_NUMBER --json title,body` |
+| `ORCHESTRATOR:DONE` in output | Run `verify-complete.sh` — if it fails, re-brief with specific reason |
+
+### Strict ORCHESTRATOR:DONE gate
+
+`verify-complete.sh` handles the main checks automatically (checkpoints, threads, CI green, spawned_at, and CHANGES_REQUESTED). Run it:
+
+**CHANGES_REQUESTED staleness rule**: a `CHANGES_REQUESTED` review only blocks if it was submitted *after* the latest commit. If the latest commit postdates the review, the review is considered stale (feedback already addressed) and does not block. This avoids false negatives when a bot reviewer hasn't re-reviewed after the agent's fixing commits.
+
+```bash
+SKILLS_DIR=~/.claude/orchestrator/scripts
+bash $SKILLS_DIR/verify-complete.sh SESSION:WIN
+```
+
+If it passes → run-loop.sh will recycle the window automatically. No manual action needed.
+If it fails → re-brief the agent with the failure reason. Never manually mark state `done` to bypass this.
+
+### Re-brief a stalled agent
+
+```bash
+OBJ=$(jq -r --arg w SESSION:WIN '.agents[] | select(.window==$w) | .objective' ~/.claude/orchestrator-state.json)
+PR=$(jq -r --arg w SESSION:WIN '.agents[] | select(.window==$w) | .pr_number' ~/.claude/orchestrator-state.json)
+tmux send-keys -t SESSION:WIN "You appear stalled. Your objective: $OBJ. Check: gh pr view $PR --json title,body,headRefName to reorient."
+sleep 0.3
+tmux send-keys -t SESSION:WIN Enter
+```
+
+If `image_path` is set on the agent record, include: "Re-read context at IMAGE_PATH with the Read tool."
+
+## Self-recovery protocol (agents)
+
+spawn-agent.sh automatically includes this instruction in every objective:
+
+> If your context compacts and you lose track of what to do, run:
+> `cat ~/.claude/orchestrator-state.json | jq '.agents[] | select(.window=="SESSION:WIN")'`
+> and `gh pr view PR_NUMBER --json title,body,headRefName` to reorient.
+> Output each completed step as `CHECKPOINT:<step-name>` on its own line.
+
+## Passing images and screenshots to agents
+
+`tmux send-keys` is text-only — you cannot paste a raw image into a pane. To give an agent visual context (screenshots, diagrams, mockups):
+
+1. **Save the image to a temp file** with a stable path:
+   ```bash
+   # If the user drags in a screenshot or you receive a file path:
+   IMAGE_PATH="/tmp/orchestrator-context-$(date +%s).png"
+   cp "$USER_PROVIDED_PATH" "$IMAGE_PATH"
+   ```
+
+2. **Reference the path in the objective string**:
+   ```bash
+   OBJECTIVE="Implement the layout shown in /tmp/orchestrator-context-1234567890.png. Read that image first with the Read tool to understand the design."
+   ```
+
+3. The agent uses its `Read` tool to view the image at startup — Claude Code agents are multimodal and can read image files directly.
+
+**Rule**: always use `/tmp/orchestrator-context-<timestamp>.png` as the naming convention so the supervisor knows what to look for if it needs to re-brief an agent with the same image.
+
+---
+
+## Orchestrator final evaluation (YOU decide, not the script)
+
+`verify-complete.sh` is a gate — it blocks premature marking. But it cannot tell you if the work is actually good. That is YOUR job.
+
+When run-loop marks an agent `pending_evaluation` and you're notified, do all of these before marking done:
+
+### 1. Run /pr-test (required, serialized, use TodoWrite to queue)
+
+`/pr-test` is the only reliable confirmation that the objective is actually met. Run it yourself, not the agent.
+
+**When multiple PRs reach `pending_evaluation` at the same time, use TodoWrite to queue them:**
+```
+- [ ] /pr-test PR #12636 — fix copilot retry logic
+- [ ] /pr-test PR #12699 — builder chat panel
+```
+Run one at a time. Check off as you go.
+
+```
+/pr-test https://github.com/Significant-Gravitas/AutoGPT/pull/PR_NUMBER
+```
+
+**/pr-test can be lazy** — if it gives vague output, re-run with full context:
+
+```
+/pr-test https://github.com/OWNER/REPO/pull/PR_NUMBER
+Context: This PR implements <objective from state file>. Key files: <list>.
+Please verify: <specific behaviors to check>.
+```
+
+Only one `/pr-test` at a time — they share ports and DB.
+
+### /pr-test result evaluation
+
+**PARTIAL on any headline feature scenario is an immediate blocker.** Do not approve, do not mark done, do not let the agent output `ORCHESTRATOR:DONE`.
+
+| `/pr-test` result | Action |
+|---|---|
+| All headline scenarios **PASS** | Proceed to evaluation step 2 |
+| Any headline scenario **PARTIAL** | Re-brief the agent immediately — see below |
+| Any headline scenario **FAIL** | Re-brief the agent immediately |
+
+**What PARTIAL means**: the feature is only partly working. Example: the Apply button never appeared, or the AI returned no action blocks. The agent addressed part of the objective but not all of it.
+
+**When any headline scenario is PARTIAL or FAIL:**
+
+1. Do NOT mark the agent done or accept `ORCHESTRATOR:DONE`
+2. Re-brief the agent with the specific scenario that failed and what was missing:
+   ```bash
+   tmux send-keys -t SESSION:WIN "PARTIAL result on /pr-test — S5 (Apply button) never appeared. The AI must output JSON action blocks for the Apply button to render. Fix this before re-running /pr-test."
+   sleep 0.3
+   tmux send-keys -t SESSION:WIN Enter
+   ```
+3. Set state back to `running`:
+   ```bash
+   jq --arg w "SESSION:WIN" '(.agents[] | select(.window == $w)).state = "running"' \
+     ~/.claude/orchestrator-state.json > /tmp/orch.tmp && mv /tmp/orch.tmp ~/.claude/orchestrator-state.json
+   ```
+4. Wait for new `ORCHESTRATOR:DONE`, then re-run `/pr-test` from scratch
+
+**Rule: only ALL-PASS qualifies for approval.** A mix of PASS + PARTIAL is a failure.
+
+> **Why this matters**: PR #12699 was wrongly approved with S5 PARTIAL — the AI never output JSON action blocks so the Apply button never appeared. The fix was already in the agent's reach but slipped through because PARTIAL was not treated as blocking.
+
+### 2. Do your own evaluation
+
+1. **Read the PR diff and objective** — does the code actually implement what was asked? Is anything obviously missing or half-done?
+2. **Read the resolved threads** — were comments addressed with real fixes, or just dismissed/resolved without changes?
+3. **Check CI run names** — any suspicious retries that shouldn't have passed?
+4. **Check the PR description** — title, summary, test plan complete?
+
+### 3. Decide
+
+- `/pr-test` all scenarios PASS + evaluation looks good → mark `done` in state, tell the user the PR is ready, ask if window should be closed
+- `/pr-test` any scenario PARTIAL or FAIL → re-brief the agent with the specific failing scenario, set state back to `running` (see `/pr-test result evaluation` above)
+- Evaluation finds gaps even with all PASS → re-brief the agent with specific gaps, set state back to `running`
+
+**Never mark done based purely on script output.** You hold the full objective context; the script does not.
+
+```bash
+# Mark done after your positive evaluation:
+jq --arg w "SESSION:WIN" '(.agents[] | select(.window == $w)).state = "done"' \
+  ~/.claude/orchestrator-state.json > /tmp/orch.tmp && mv /tmp/orch.tmp ~/.claude/orchestrator-state.json
+```
+
+## When to stop the fleet
+
+Stop the fleet (`active = false`) when **all** of the following are true:
+
+| Check | How to verify |
+|---|---|
+| All agents are `done` or `escalated` | `jq '[.agents[] | select(.state | test("running\|stuck\|idle\|waiting_approval"))] | length' ~/.claude/orchestrator-state.json` == 0 |
+| All PRs have 0 unresolved review threads | GraphQL `isResolved` check per PR |
+| All PRs have green CI **on a run triggered after the agent's last push** | `gh run list --branch BRANCH --limit 1` timestamp > `spawned_at` in state |
+| No fresh CHANGES_REQUESTED (after latest commit) | `verify-complete.sh` checks this — stale pre-commit reviews are ignored |
+| No agents are `escalated` without human review | If any are escalated, surface to user first |
+
+**Do NOT stop just because agents output `ORCHESTRATOR:DONE`.** That is a signal to verify, not a signal to stop.
+
+**Do stop** if the user explicitly says "stop", "shut down", or "kill everything", even with agents still running.
+
+```bash
+# Graceful stop
+jq '.active = false' ~/.claude/orchestrator-state.json > /tmp/orch.tmp \
+  && mv /tmp/orch.tmp ~/.claude/orchestrator-state.json
+
+LOOP_WINDOW=$(jq -r '.loop_window // ""' ~/.claude/orchestrator-state.json)
+[ -n "$LOOP_WINDOW" ] && tmux kill-window -t "$LOOP_WINDOW" 2>/dev/null || true
+```
+
+Does **not** recycle running worktrees — agents may still be mid-task. Run `capacity.sh` to see what's still in progress.
+
+## tmux send-keys pattern
+
+**Always split long messages into text + Enter as two separate calls with a sleep between them.** If sent as one call (`"text" Enter`), Enter can fire before the full string is buffered into Claude's input — leaving the message stuck as `[Pasted text +N lines]` unsent.
+
+```bash
+# CORRECT — text then Enter separately
+tmux send-keys -t "$WINDOW" "your long message here"
+sleep 0.3
+tmux send-keys -t "$WINDOW" Enter
+
+# WRONG — Enter may fire before text is buffered
+tmux send-keys -t "$WINDOW" "your long message here" Enter
+```
+
+Short single-character sends (`y`, `Down`, empty Enter for dialog approval) are safe to combine since they have no buffering lag.
+
+---
+
+## Protected worktrees
+
+Some worktrees must **never** be used as spare worktrees for agent tasks because they host files critical to the orchestrator itself:
+
+| Worktree | Protected branch | Why |
+|---|---|---|
+| `AutoGPT1` | `dx/orchestrate-skill` | Hosts the orchestrate skill scripts. `recycle-agent.sh` would check out `spare/1`, wiping `.claude/skills/` and breaking all subsequent `spawn-agent.sh` calls. |
+
+**Rule**: when selecting spare worktrees via `find-spare.sh`, skip any worktree whose CURRENT branch matches a protected branch. If you accidentally spawn an agent in a protected worktree, do not let `recycle-agent.sh` run on it — manually restore the branch after the agent finishes.
+
+When `dx/orchestrate-skill` is merged into `dev`, `AutoGPT1` becomes a normal spare again.
+
+---
+
+## Key rules
+
+1. **Scripts do all the heavy lifting** — don't reimplement their logic inline in this file
+2. **Never ask the user to pick a worktree** — auto-assign from `find-spare.sh` output
+3. **Never restart a running agent** — only restart on `idle` kicks (foreground is a shell)
+4. **Auto-dismiss settings dialogs** — if "Enter to confirm" appears, send Down+Enter
+5. **Always `--permission-mode bypassPermissions`** on every spawn
+6. **Escalate after 3 kicks** — mark `escalated`, surface to user
+7. **Atomic state writes** — always write to `.tmp` then `mv`
+8. **Never approve destructive commands** outside the worktree scope — when in doubt, escalate
+9. **Never recycle without verification** — `verify-complete.sh` must pass before recycling
+10. **No TASK.md files** — commit risk; use state file + `gh pr view` for agent context persistence
+11. **Re-brief stalled agents** — read objective from state file + `gh pr view`, send via tmux
+12. **ORCHESTRATOR:DONE is a signal to verify, not to accept** — always run `verify-complete.sh` and check CI run timestamp before recycling
+13. **Protected worktrees** — never use the worktree hosting the skill scripts as a spare
+14. **Images via file path** — save screenshots to `/tmp/orchestrator-context-<ts>.png`, pass path in objective; agents read with the `Read` tool
+15. **Split send-keys** — always separate text and Enter with `sleep 0.3` between calls for long strings

.claude/skills/orchestrate/scripts/capacity.sh

@@ -0,0 +1,43 @@
+#!/usr/bin/env bash
+# capacity.sh — show fleet capacity: available spare worktrees + in-use agents
+#
+# Usage: capacity.sh [REPO_ROOT]
+#   REPO_ROOT defaults to the root worktree of the current git repo.
+#
+# Reads: ~/.claude/orchestrator-state.json (skipped if missing or corrupt)
+
+set -euo pipefail
+
+SCRIPTS_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+STATE_FILE="${ORCHESTRATOR_STATE_FILE:-$HOME/.claude/orchestrator-state.json}"
+REPO_ROOT="${1:-$(git rev-parse --show-toplevel 2>/dev/null || echo "")}"
+
+echo "=== Available (spare) worktrees ==="
+if [ -n "$REPO_ROOT" ]; then
+  SPARE=$("$SCRIPTS_DIR/find-spare.sh" "$REPO_ROOT" 2>/dev/null || echo "")
+else
+  SPARE=$("$SCRIPTS_DIR/find-spare.sh" 2>/dev/null || echo "")
+fi
+
+if [ -z "$SPARE" ]; then
+  echo "  (none)"
+else
+  while IFS= read -r line; do
+    [ -z "$line" ] && continue
+    echo "  ✓ $line"
+  done <<< "$SPARE"
+fi
+
+echo ""
+echo "=== In-use worktrees ==="
+if [ -f "$STATE_FILE" ] && jq -e '.' "$STATE_FILE" >/dev/null 2>&1; then
+  IN_USE=$(jq -r '.agents[] | select(.state != "done") | "  [\(.state)] \(.worktree_path) → \(.branch)"' \
+    "$STATE_FILE" 2>/dev/null || echo "")
+  if [ -n "$IN_USE" ]; then
+    echo "$IN_USE"
+  else
+    echo "  (none)"
+  fi
+else
+  echo "  (no active state file)"
+fi

.claude/skills/orchestrate/scripts/classify-pane.sh

@@ -0,0 +1,85 @@
+#!/usr/bin/env bash
+# classify-pane.sh — Classify the current state of a tmux pane
+#
+# Usage: classify-pane.sh <tmux-target>
+#   tmux-target: e.g. "work:0", "work:1.0"
+#
+# Output (stdout): JSON object:
+#   { "state": "running|idle|waiting_approval|complete", "reason": "...", "pane_cmd": "..." }
+#
+# Exit codes: 0=ok, 1=error (invalid target or tmux window not found)
+
+set -euo pipefail
+
+TARGET="${1:-}"
+
+if [ -z "$TARGET" ]; then
+  echo '{"state":"error","reason":"no target provided","pane_cmd":""}'
+  exit 1
+fi
+
+# Validate tmux target format: session:window or session:window.pane
+if ! [[ "$TARGET" =~ ^[a-zA-Z0-9_.-]+:[a-zA-Z0-9_.-]+(\.[0-9]+)?$ ]]; then
+  echo '{"state":"error","reason":"invalid tmux target format","pane_cmd":""}'
+  exit 1
+fi
+
+# Check session exists (use %%:* to extract session name from session:window)
+if ! tmux list-windows -t "${TARGET%%:*}" &>/dev/null 2>&1; then
+  echo '{"state":"error","reason":"tmux target not found","pane_cmd":""}'
+  exit 1
+fi
+
+# Get the current foreground command in the pane
+PANE_CMD=$(tmux display-message -t "$TARGET" -p '#{pane_current_command}' 2>/dev/null || echo "unknown")
+
+# Capture and strip ANSI codes (use perl for cross-platform compatibility — BSD sed lacks \x1b support)
+RAW=$(tmux capture-pane -t "$TARGET" -p -S -50 2>/dev/null || echo "")
+CLEAN=$(echo "$RAW" | perl -pe 's/\x1b\[[0-9;]*[a-zA-Z]//g; s/\x1b\(B//g; s/\x1b\[\?[0-9]*[hl]//g; s/\r//g' \
+  | grep -v '^[[:space:]]*$' || true)
+
+# --- Check: explicit completion marker ---
+# Must be on its own line (not buried in the objective text sent at spawn time).
+if echo "$CLEAN" | grep -qE "^[[:space:]]*ORCHESTRATOR:DONE[[:space:]]*$"; then
+  jq -n --arg cmd "$PANE_CMD" '{"state":"complete","reason":"ORCHESTRATOR:DONE marker found","pane_cmd":$cmd}'
+  exit 0
+fi
+
+# --- Check: Claude Code approval prompt patterns ---
+LAST_40=$(echo "$CLEAN" | tail -40)
+APPROVAL_PATTERNS=(
+  "Do you want to proceed"
+  "Do you want to make this"
+  "\\[y/n\\]"
+  "\\[Y/n\\]"
+  "\\[n/Y\\]"
+  "Proceed\\?"
+  "Allow this command"
+  "Run bash command"
+  "Allow bash"
+  "Would you like"
+  "Press enter to continue"
+  "Esc to cancel"
+)
+for pattern in "${APPROVAL_PATTERNS[@]}"; do
+  if echo "$LAST_40" | grep -qiE "$pattern"; then
+    jq -n --arg pattern "$pattern" --arg cmd "$PANE_CMD" \
+      '{"state":"waiting_approval","reason":"approval pattern: \($pattern)","pane_cmd":$cmd}'
+    exit 0
+  fi
+done
+
+# --- Check: shell prompt (claude has exited) ---
+# If the foreground process is a shell (not claude/node), the agent has exited
+case "$PANE_CMD" in
+  zsh|bash|fish|sh|dash|tcsh|ksh)
+    jq -n --arg cmd "$PANE_CMD" \
+      '{"state":"idle","reason":"agent exited — shell prompt active","pane_cmd":$cmd}'
+    exit 0
+    ;;
+esac
+
+# Agent is still running (claude/node/python is the foreground process)
+jq -n --arg cmd "$PANE_CMD" \
+  '{"state":"running","reason":"foreground process: \($cmd)","pane_cmd":$cmd}'
+exit 0

.claude/skills/orchestrate/scripts/find-spare.sh

@@ -0,0 +1,24 @@
+#!/usr/bin/env bash
+# find-spare.sh — list worktrees on spare/N branches (free to use)
+#
+# Usage: find-spare.sh [REPO_ROOT]
+#   REPO_ROOT defaults to the root worktree containing the current git repo.
+#
+# Output (stdout): one line per available worktree: "PATH BRANCH"
+#   e.g.: /Users/me/Code/AutoGPT3 spare/3
+
+set -euo pipefail
+
+REPO_ROOT="${1:-$(git rev-parse --show-toplevel 2>/dev/null || echo "")}"
+if [ -z "$REPO_ROOT" ]; then
+  echo "Error: not inside a git repo and no REPO_ROOT provided" >&2
+  exit 1
+fi
+
+git -C "$REPO_ROOT" worktree list --porcelain \
+  | awk '
+      /^worktree / { path = substr($0, 10) }
+      /^branch /   { branch = substr($0, 8); print path " " branch }
+    ' \
+  | { grep -E " refs/heads/spare/[0-9]+$" || true; } \
+  | sed 's|refs/heads/||'

.claude/skills/orchestrate/scripts/notify.sh

@@ -0,0 +1,40 @@
+#!/usr/bin/env bash
+# notify.sh — send a fleet notification message
+#
+# Delivery order (first available wins):
+#   1. Discord webhook — DISCORD_WEBHOOK_URL env var OR state file .discord_webhook
+#   2. macOS notification center — osascript (silent fail if unavailable)
+#   3. Stdout only
+#
+# Usage: notify.sh MESSAGE
+# Exit: always 0 (notification failure must not abort the caller)
+
+MESSAGE="${1:-}"
+[ -z "$MESSAGE" ] && exit 0
+
+STATE_FILE="${ORCHESTRATOR_STATE_FILE:-$HOME/.claude/orchestrator-state.json}"
+
+# --- Resolve Discord webhook ---
+WEBHOOK="${DISCORD_WEBHOOK_URL:-}"
+if [ -z "$WEBHOOK" ] && [ -f "$STATE_FILE" ]; then
+  WEBHOOK=$(jq -r '.discord_webhook // ""' "$STATE_FILE" 2>/dev/null || echo "")
+fi
+
+# --- Discord delivery ---
+if [ -n "$WEBHOOK" ]; then
+  PAYLOAD=$(jq -n --arg msg "$MESSAGE" '{"content": $msg}')
+  curl -s -X POST "$WEBHOOK" \
+    -H "Content-Type: application/json" \
+    -d "$PAYLOAD" > /dev/null 2>&1 || true
+fi
+
+# --- macOS notification center (silent if not macOS or osascript missing) ---
+if command -v osascript &>/dev/null 2>&1; then
+  # Escape single quotes for AppleScript
+  SAFE_MSG=$(echo "$MESSAGE" | sed "s/'/\\\\'/g")
+  osascript -e "display notification \"${SAFE_MSG}\" with title \"Orchestrator\"" 2>/dev/null || true
+fi
+
+# Always print to stdout so run-loop.sh logs it
+echo "$MESSAGE"
+exit 0

.claude/skills/orchestrate/scripts/poll-cycle.sh

@@ -0,0 +1,257 @@
+#!/usr/bin/env bash
+# poll-cycle.sh — Single orchestrator poll cycle
+#
+# Reads ~/.claude/orchestrator-state.json, classifies each agent, updates state,
+# and outputs a JSON array of actions for Claude to take.
+#
+# Usage: poll-cycle.sh
+# Output (stdout): JSON array of action objects
+#   [{ "window": "work:0", "action": "kick|approve|none", "state": "...",
+#      "worktree": "...", "objective": "...", "reason": "..." }]
+#
+# The state file is updated in-place (atomic write via .tmp).
+
+set -euo pipefail
+
+STATE_FILE="${ORCHESTRATOR_STATE_FILE:-$HOME/.claude/orchestrator-state.json}"
+SCRIPTS_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+CLASSIFY="$SCRIPTS_DIR/classify-pane.sh"
+
+# Cross-platform md5: always outputs just the hex digest
+md5_hash() {
+  if command -v md5sum &>/dev/null; then
+    md5sum | awk '{print $1}'
+  else
+    md5 | awk '{print $NF}'
+  fi
+}
+
+# Clean up temp file on any exit (avoids stale .tmp if jq write fails)
+trap 'rm -f "${STATE_FILE}.tmp"' EXIT
+
+# Ensure state file exists
+if [ ! -f "$STATE_FILE" ]; then
+  echo '{"active":false,"agents":[]}' > "$STATE_FILE"
+fi
+
+# Validate JSON upfront before any jq reads that run under set -e.
+# A truncated/corrupt file (e.g. from a SIGKILL mid-write) would otherwise
+# abort the script at the ACTIVE read below without emitting any JSON output.
+if ! jq -e '.' "$STATE_FILE" >/dev/null 2>&1; then
+  echo "State file parse error — check $STATE_FILE" >&2
+  echo "[]"
+  exit 0
+fi
+
+ACTIVE=$(jq -r '.active // false' "$STATE_FILE")
+if [ "$ACTIVE" != "true" ]; then
+  echo "[]"
+  exit 0
+fi
+
+NOW=$(date +%s)
+IDLE_THRESHOLD=$(jq -r '.idle_threshold_seconds // 300' "$STATE_FILE")
+
+ACTIONS="[]"
+UPDATED_AGENTS="[]"
+
+# Read agents as newline-delimited JSON objects.
+# jq exits non-zero when .agents[] has no matches on an empty array, which is valid —
+# so we suppress that exit code and separately validate the file is well-formed JSON.
+if ! AGENTS_JSON=$(jq -e -c '.agents // empty | .[]' "$STATE_FILE" 2>/dev/null); then
+  if ! jq -e '.' "$STATE_FILE" > /dev/null 2>&1; then
+    echo "State file parse error — check $STATE_FILE" >&2
+  fi
+  echo "[]"
+  exit 0
+fi
+
+if [ -z "$AGENTS_JSON" ]; then
+  echo "[]"
+  exit 0
+fi
+
+while IFS= read -r agent; do
+  [ -z "$agent" ] && continue
+
+  # Use // "" defaults so a single malformed field doesn't abort the whole cycle
+  WINDOW=$(echo "$agent"   | jq -r '.window // ""')
+  WORKTREE=$(echo "$agent" | jq -r '.worktree // ""')
+  OBJECTIVE=$(echo "$agent"| jq -r '.objective // ""')
+  STATE=$(echo "$agent"    | jq -r '.state // "running"')
+  LAST_HASH=$(echo "$agent"| jq -r '.last_output_hash // ""')
+  IDLE_SINCE=$(echo "$agent"| jq -r '.idle_since // 0')
+  REVISION_COUNT=$(echo "$agent"| jq -r '.revision_count // 0')
+
+  # Validate window format to prevent tmux target injection.
+  # Allow session:window (numeric or named) and session:window.pane
+  if ! [[ "$WINDOW" =~ ^[a-zA-Z0-9_.-]+:[a-zA-Z0-9_.-]+(\.[0-9]+)?$ ]]; then
+    echo "Skipping agent with invalid window value: $WINDOW" >&2
+    UPDATED_AGENTS=$(echo "$UPDATED_AGENTS" | jq --argjson a "$agent" '. + [$a]')
+    continue
+  fi
+
+  # Pass-through terminal-state agents
+  if [[ "$STATE" == "done" || "$STATE" == "escalated" || "$STATE" == "complete" || "$STATE" == "pending_evaluation" ]]; then
+    UPDATED_AGENTS=$(echo "$UPDATED_AGENTS" | jq --argjson a "$agent" '. + [$a]')
+    continue
+  fi
+
+  # Classify pane.
+  # classify-pane.sh always emits JSON before exit (even on error), so using
+  # "|| echo '...'" would concatenate two JSON objects when it exits non-zero.
+  # Use "|| true" inside the substitution so set -euo pipefail does not abort
+  # the poll cycle when classify exits with a non-zero status code.
+  CLASSIFICATION=$("$CLASSIFY" "$WINDOW" 2>/dev/null || true)
+  [ -z "$CLASSIFICATION" ] && CLASSIFICATION='{"state":"error","reason":"classify failed","pane_cmd":"unknown"}'
+
+  PANE_STATE=$(echo "$CLASSIFICATION" | jq -r '.state')
+  PANE_REASON=$(echo "$CLASSIFICATION" | jq -r '.reason')
+
+  # Capture full pane output once — used for hash (stuck detection) and checkpoint parsing.
+  # Use -S -500 to get the last ~500 lines of scrollback so checkpoints aren't missed.
+  RAW=$(tmux capture-pane -t "$WINDOW" -p -S -500 2>/dev/null || echo "")
+
+  # --- Checkpoint tracking ---
+  # Parse any "CHECKPOINT:<step>" lines the agent has output and merge into state file.
+  # The agent writes these as it completes each required step so verify-complete.sh can gate recycling.
+  EXISTING_CPS=$(echo "$agent" | jq -c '.checkpoints // []')
+  NEW_CHECKPOINTS_JSON="$EXISTING_CPS"
+  if [ -n "$RAW" ]; then
+    FOUND_CPS=$(echo "$RAW" \
+      | grep -oE "CHECKPOINT:[a-zA-Z0-9_-]+" \
+      | sed 's/CHECKPOINT://' \
+      | sort -u \
+      | jq -R . | jq -s . 2>/dev/null || echo "[]")
+    NEW_CHECKPOINTS_JSON=$(jq -n \
+      --argjson existing "$EXISTING_CPS" \
+      --argjson found "$FOUND_CPS" \
+      '($existing + $found) | unique' 2>/dev/null || echo "$EXISTING_CPS")
+  fi
+
+  # Compute content hash for stuck-detection (only for running agents)
+  CURRENT_HASH=""
+  if [[ "$PANE_STATE" == "running" ]] && [ -n "$RAW" ]; then
+    CURRENT_HASH=$(echo "$RAW" | tail -20 | md5_hash)
+  fi
+
+  NEW_STATE="$STATE"
+  NEW_IDLE_SINCE="$IDLE_SINCE"
+  NEW_REVISION_COUNT="$REVISION_COUNT"
+  ACTION="none"
+  REASON="$PANE_REASON"
+
+  case "$PANE_STATE" in
+    complete)
+      # Agent output ORCHESTRATOR:DONE — mark pending_evaluation so orchestrator handles it.
+      # run-loop does NOT verify or notify; orchestrator's background poll picks this up.
+      NEW_STATE="pending_evaluation"
+      ACTION="complete"  # run-loop logs it but takes no action
+      ;;
+    waiting_approval)
+      NEW_STATE="waiting_approval"
+      ACTION="approve"
+      ;;
+    idle)
+      # Agent process has exited — needs restart
+      NEW_STATE="idle"
+      ACTION="kick"
+      REASON="agent exited (shell is foreground)"
+      NEW_REVISION_COUNT=$(( REVISION_COUNT + 1 ))
+      NEW_IDLE_SINCE=$NOW
+      if [ "$NEW_REVISION_COUNT" -ge 3 ]; then
+        NEW_STATE="escalated"
+        ACTION="none"
+        REASON="escalated after ${NEW_REVISION_COUNT} kicks — needs human attention"
+      fi
+      ;;
+    running)
+      # Clear idle_since only when transitioning from idle (agent was kicked and
+      # restarted). Do NOT reset for stuck — idle_since must persist across polls
+      # so STUCK_DURATION can accumulate and trigger escalation.
+      # Also update the local IDLE_SINCE so the hash-stability check below uses
+      # the reset value on this same poll, not the stale kick timestamp.
+      if [[ "$STATE" == "idle" ]]; then
+        NEW_IDLE_SINCE=0
+        IDLE_SINCE=0
+      fi
+      # Check if hash has been stable (agent may be stuck mid-task)
+      if [ -n "$CURRENT_HASH" ] && [ "$CURRENT_HASH" = "$LAST_HASH" ] && [ "$LAST_HASH" != "" ]; then
+        if [ "$IDLE_SINCE" = "0" ] || [ "$IDLE_SINCE" = "null" ]; then
+          NEW_IDLE_SINCE=$NOW
+        else
+          STUCK_DURATION=$(( NOW - IDLE_SINCE ))
+          if [ "$STUCK_DURATION" -gt "$IDLE_THRESHOLD" ]; then
+            NEW_REVISION_COUNT=$(( REVISION_COUNT + 1 ))
+            NEW_IDLE_SINCE=$NOW
+            if [ "$NEW_REVISION_COUNT" -ge 3 ]; then
+              NEW_STATE="escalated"
+              ACTION="none"
+              REASON="escalated after ${NEW_REVISION_COUNT} kicks — needs human attention"
+            else
+              NEW_STATE="stuck"
+              ACTION="kick"
+              REASON="output unchanged for ${STUCK_DURATION}s (threshold: ${IDLE_THRESHOLD}s)"
+            fi
+          fi
+        fi
+      else
+        # Only reset the idle timer when we have a valid hash comparison (pane
+        # capture succeeded). If CURRENT_HASH is empty (tmux capture-pane failed),
+        # preserve existing timers so stuck detection is not inadvertently reset.
+        if [ -n "$CURRENT_HASH" ]; then
+          NEW_STATE="running"
+          NEW_IDLE_SINCE=0
+        fi
+      fi
+      ;;
+    error)
+      REASON="classify error: $PANE_REASON"
+      ;;
+  esac
+
+  # Build updated agent record (ensure idle_since and revision_count are numeric)
+  # Use || true on each jq call so a malformed field skips this agent rather than
+  # aborting the entire poll cycle under set -e.
+  UPDATED_AGENT=$(echo "$agent" | jq \
+    --arg state "$NEW_STATE" \
+    --arg hash "$CURRENT_HASH" \
+    --argjson now "$NOW" \
+    --arg idle_since "$NEW_IDLE_SINCE" \
+    --arg revision_count "$NEW_REVISION_COUNT" \
+    --argjson checkpoints "$NEW_CHECKPOINTS_JSON" \
+    '.state = $state
+     | .last_output_hash = (if $hash == "" then .last_output_hash else $hash end)
+     | .last_seen_at = $now
+     | .idle_since = ($idle_since | tonumber)
+     | .revision_count = ($revision_count | tonumber)
+     | .checkpoints = $checkpoints' 2>/dev/null) || {
+    echo "Warning: failed to build updated agent for window $WINDOW — keeping original" >&2
+    UPDATED_AGENTS=$(echo "$UPDATED_AGENTS" | jq --argjson a "$agent" '. + [$a]')
+    continue
+  }
+
+  UPDATED_AGENTS=$(echo "$UPDATED_AGENTS" | jq --argjson a "$UPDATED_AGENT" '. + [$a]')
+
+  # Add action if needed
+  if [ "$ACTION" != "none" ]; then
+    ACTION_OBJ=$(jq -n \
+      --arg window "$WINDOW" \
+      --arg action "$ACTION" \
+      --arg state "$NEW_STATE" \
+      --arg worktree "$WORKTREE" \
+      --arg objective "$OBJECTIVE" \
+      --arg reason "$REASON" \
+      '{window:$window, action:$action, state:$state, worktree:$worktree, objective:$objective, reason:$reason}')
+    ACTIONS=$(echo "$ACTIONS" | jq --argjson a "$ACTION_OBJ" '. + [$a]')
+  fi
+
+done <<< "$AGENTS_JSON"
+
+# Atomic state file update
+jq --argjson agents "$UPDATED_AGENTS" \
+   --argjson now "$NOW" \
+   '.agents = $agents | .last_poll_at = $now' \
+   "$STATE_FILE" > "${STATE_FILE}.tmp" && mv "${STATE_FILE}.tmp" "$STATE_FILE"
+
+echo "$ACTIONS"

.claude/skills/orchestrate/scripts/recycle-agent.sh

@@ -0,0 +1,32 @@
+#!/usr/bin/env bash
+# recycle-agent.sh — kill a tmux window and restore the worktree to its spare branch
+#
+# Usage: recycle-agent.sh WINDOW WORKTREE_PATH SPARE_BRANCH
+#   WINDOW        — tmux target, e.g. autogpt1:3
+#   WORKTREE_PATH — absolute path to the git worktree
+#   SPARE_BRANCH  — branch to restore, e.g. spare/6
+#
+# Stdout: one status line
+
+set -euo pipefail
+
+if [ $# -lt 3 ]; then
+  echo "Usage: recycle-agent.sh WINDOW WORKTREE_PATH SPARE_BRANCH" >&2
+  exit 1
+fi
+
+WINDOW="$1"
+WORKTREE_PATH="$2"
+SPARE_BRANCH="$3"
+
+# Kill the tmux window (ignore error — may already be gone)
+tmux kill-window -t "$WINDOW" 2>/dev/null || true
+
+# Restore to spare branch: abort any in-progress operation, then clean
+git -C "$WORKTREE_PATH" rebase --abort 2>/dev/null || true
+git -C "$WORKTREE_PATH" merge --abort 2>/dev/null || true
+git -C "$WORKTREE_PATH" reset --hard HEAD 2>/dev/null
+git -C "$WORKTREE_PATH" clean -fd 2>/dev/null
+git -C "$WORKTREE_PATH" checkout "$SPARE_BRANCH"
+
+echo "Recycled: $(basename "$WORKTREE_PATH") → $SPARE_BRANCH (window $WINDOW closed)"

.claude/skills/orchestrate/scripts/run-loop.sh

@@ -0,0 +1,164 @@
+#!/usr/bin/env bash
+# run-loop.sh — Mechanical babysitter for the agent fleet (runs in its own tmux window)
+#
+# Handles ONLY two things that need no intelligence:
+#   idle    → restart claude using --resume SESSION_ID (or --continue) to restore context
+#   approve → auto-approve safe dialogs, press Enter on numbered-option dialogs
+#
+# Everything else — ORCHESTRATOR:DONE, verification, /pr-test, final evaluation,
+# marking done, deciding to close windows — is the orchestrating Claude's job.
+# poll-cycle.sh sets state to pending_evaluation when ORCHESTRATOR:DONE is detected;
+# the orchestrator's background poll loop handles it from there.
+#
+# Usage: run-loop.sh
+# Env:   POLL_INTERVAL (default: 30), ORCHESTRATOR_STATE_FILE
+
+set -euo pipefail
+
+# Copy scripts to a stable location outside the repo so they survive branch
+# checkouts (e.g. recycle-agent.sh switching spare/N back into this worktree
+# would wipe .claude/skills/orchestrate/scripts if the skill only exists on the
+# current branch).
+_ORIGIN_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+STABLE_SCRIPTS_DIR="$HOME/.claude/orchestrator/scripts"
+mkdir -p "$STABLE_SCRIPTS_DIR"
+cp "$_ORIGIN_DIR"/*.sh "$STABLE_SCRIPTS_DIR/"
+chmod +x "$STABLE_SCRIPTS_DIR"/*.sh
+SCRIPTS_DIR="$STABLE_SCRIPTS_DIR"
+
+STATE_FILE="${ORCHESTRATOR_STATE_FILE:-$HOME/.claude/orchestrator-state.json}"
+POLL_INTERVAL="${POLL_INTERVAL:-30}"
+
+# ---------------------------------------------------------------------------
+# update_state WINDOW FIELD VALUE
+# ---------------------------------------------------------------------------
+update_state() {
+  local window="$1" field="$2" value="$3"
+  jq --arg w "$window" --arg f "$field" --arg v "$value" \
+    '.agents |= map(if .window == $w then .[$f] = $v else . end)' \
+    "$STATE_FILE" > "${STATE_FILE}.tmp" && mv "${STATE_FILE}.tmp" "$STATE_FILE"
+}
+
+update_state_int() {
+  local window="$1" field="$2" value="$3"
+  jq --arg w "$window" --arg f "$field" --argjson v "$value" \
+    '.agents |= map(if .window == $w then .[$f] = $v else . end)' \
+    "$STATE_FILE" > "${STATE_FILE}.tmp" && mv "${STATE_FILE}.tmp" "$STATE_FILE"
+}
+
+agent_field() {
+  jq -r --arg w "$1" --arg f "$2" \
+    '.agents[] | select(.window == $w) | .[$f] // ""' \
+    "$STATE_FILE" 2>/dev/null
+}
+
+# ---------------------------------------------------------------------------
+# wait_for_prompt WINDOW — wait up to 60s for Claude's ❯ prompt
+# ---------------------------------------------------------------------------
+wait_for_prompt() {
+  local window="$1"
+  for i in $(seq 1 60); do
+    local cmd pane
+    cmd=$(tmux display-message -t "$window" -p '#{pane_current_command}' 2>/dev/null || echo "")
+    pane=$(tmux capture-pane -t "$window" -p 2>/dev/null || echo "")
+    if echo "$pane" | grep -q "Enter to confirm"; then
+      tmux send-keys -t "$window" Down Enter; sleep 2; continue
+    fi
+    [[ "$cmd" == "node" ]] && echo "$pane" | grep -q "❯" && return 0
+    sleep 1
+  done
+  return 1  # timed out
+}
+
+# ---------------------------------------------------------------------------
+# handle_kick WINDOW STATE — only for idle (crashed) agents, not stuck
+# ---------------------------------------------------------------------------
+handle_kick() {
+  local window="$1" state="$2"
+  [[ "$state" != "idle" ]] && return  # stuck agents handled by supervisor
+
+  local worktree_path session_id
+  worktree_path=$(agent_field "$window" "worktree_path")
+  session_id=$(agent_field "$window" "session_id")
+
+  echo "[$(date +%H:%M:%S)] KICK restart  $window — agent exited, resuming session"
+
+  # Resume the exact session so the agent retains full context — no need to re-send objective
+  if [ -n "$session_id" ]; then
+    tmux send-keys -t "$window" "cd '${worktree_path}' && claude --resume '${session_id}' --permission-mode bypassPermissions" Enter
+  else
+    tmux send-keys -t "$window" "cd '${worktree_path}' && claude --continue --permission-mode bypassPermissions" Enter
+  fi
+
+  wait_for_prompt "$window" || echo "[$(date +%H:%M:%S)] KICK WARNING  $window — timed out waiting for ❯"
+}
+
+# ---------------------------------------------------------------------------
+# handle_approve WINDOW — auto-approve dialogs that need no judgment
+# ---------------------------------------------------------------------------
+handle_approve() {
+  local window="$1"
+  local pane_tail
+  pane_tail=$(tmux capture-pane -t "$window" -p 2>/dev/null | tail -3 || echo "")
+
+  # Settings error dialog at startup
+  if echo "$pane_tail" | grep -q "Enter to confirm"; then
+    echo "[$(date +%H:%M:%S)] APPROVE dialog $window — settings error"
+    tmux send-keys -t "$window" Down Enter
+    return
+  fi
+
+  # Numbered-option dialog (e.g. "Do you want to make this edit?")
+  # ❯ is already on option 1 (Yes) — Enter confirms it
+  if echo "$pane_tail" | grep -qE "❯\s*1\." || echo "$pane_tail" | grep -q "Esc to cancel"; then
+    echo "[$(date +%H:%M:%S)] APPROVE edit   $window"
+    tmux send-keys -t "$window" "" Enter
+    return
+  fi
+
+  # y/n prompt for safe operations
+  if echo "$pane_tail" | grep -qiE "(^git |^npm |^pnpm |^poetry |^pytest|^docker |^make |^cargo |^pip |^yarn |curl .*(localhost|127\.0\.0\.1))"; then
+    echo "[$(date +%H:%M:%S)] APPROVE safe   $window"
+    tmux send-keys -t "$window" "y" Enter
+    return
+  fi
+
+  # Anything else — supervisor handles it, just log
+  echo "[$(date +%H:%M:%S)] APPROVE skip   $window — unknown dialog, supervisor will handle"
+}
+
+# ---------------------------------------------------------------------------
+# Main loop
+# ---------------------------------------------------------------------------
+echo "[$(date +%H:%M:%S)] run-loop started (mechanical only, poll every ${POLL_INTERVAL}s)"
+echo "[$(date +%H:%M:%S)] Supervisor: orchestrating Claude session (not a separate window)"
+echo "---"
+
+while true; do
+  if ! jq -e '.active == true' "$STATE_FILE" >/dev/null 2>&1; then
+    echo "[$(date +%H:%M:%S)] active=false — exiting."
+    exit 0
+  fi
+
+  ACTIONS=$("$SCRIPTS_DIR/poll-cycle.sh" 2>/dev/null || echo "[]")
+  KICKED=0; DONE=0
+
+  while IFS= read -r action; do
+    [ -z "$action" ] && continue
+    WINDOW=$(echo "$action" | jq -r '.window // ""')
+    ACTION=$(echo "$action" | jq -r '.action // ""')
+    STATE=$(echo "$action"  | jq -r '.state // ""')
+
+    case "$ACTION" in
+      kick)     handle_kick "$WINDOW" "$STATE" || true; KICKED=$(( KICKED + 1 )) ;;
+      approve)  handle_approve "$WINDOW" || true ;;
+      complete) DONE=$(( DONE + 1 )) ;;  # poll-cycle already set state=pending_evaluation; orchestrator handles
+    esac
+  done < <(echo "$ACTIONS" | jq -c '.[]' 2>/dev/null || true)
+
+  RUNNING=$(jq '[.agents[] | select(.state | test("running|stuck|waiting_approval|idle"))] | length' \
+    "$STATE_FILE" 2>/dev/null || echo 0)
+
+  echo "[$(date +%H:%M:%S)] Poll — ${RUNNING} running  ${KICKED} kicked  ${DONE} recycled"
+  sleep "$POLL_INTERVAL"
+done

.claude/skills/orchestrate/scripts/spawn-agent.sh

@@ -0,0 +1,122 @@
+#!/usr/bin/env bash
+# spawn-agent.sh — create tmux window, checkout branch, launch claude, send task
+#
+# Usage: spawn-agent.sh SESSION WORKTREE_PATH SPARE_BRANCH NEW_BRANCH OBJECTIVE [PR_NUMBER] [STEPS...]
+#   SESSION       — tmux session name, e.g. autogpt1
+#   WORKTREE_PATH — absolute path to the git worktree
+#   SPARE_BRANCH  — spare branch being replaced, e.g. spare/6 (saved for recycle)
+#   NEW_BRANCH    — task branch to create, e.g. feat/my-feature
+#   OBJECTIVE     — task description sent to the agent
+#   PR_NUMBER     — (optional) GitHub PR number for completion verification
+#   STEPS...      — (optional) required checkpoint names, e.g. pr-address pr-test
+#
+# Stdout: SESSION:WINDOW_INDEX (nothing else — callers rely on this)
+# Exit non-zero on failure.
+
+set -euo pipefail
+
+if [ $# -lt 5 ]; then
+  echo "Usage: spawn-agent.sh SESSION WORKTREE_PATH SPARE_BRANCH NEW_BRANCH OBJECTIVE [PR_NUMBER] [STEPS...]" >&2
+  exit 1
+fi
+
+SESSION="$1"
+WORKTREE_PATH="$2"
+SPARE_BRANCH="$3"
+NEW_BRANCH="$4"
+OBJECTIVE="$5"
+PR_NUMBER="${6:-}"
+STEPS=("${@:7}")
+WORKTREE_NAME=$(basename "$WORKTREE_PATH")
+STATE_FILE="${ORCHESTRATOR_STATE_FILE:-$HOME/.claude/orchestrator-state.json}"
+
+# Generate a stable session ID so this agent's Claude session can always be resumed:
+#   claude --resume $SESSION_ID --permission-mode bypassPermissions
+SESSION_ID=$(uuidgen 2>/dev/null || python3 -c "import uuid; print(uuid.uuid4())")
+
+# Create (or switch to) the task branch
+git -C "$WORKTREE_PATH" checkout -b "$NEW_BRANCH" 2>/dev/null \
+  || git -C "$WORKTREE_PATH" checkout "$NEW_BRANCH"
+
+# Open a new named tmux window; capture its numeric index
+WIN_IDX=$(tmux new-window -t "$SESSION" -n "$WORKTREE_NAME" -P -F '#{window_index}')
+WINDOW="${SESSION}:${WIN_IDX}"
+
+# Append the initial agent record to the state file so subsequent jq updates find it.
+# This must happen before the pr_number/steps update below.
+if [ -f "$STATE_FILE" ]; then
+  NOW=$(date +%s)
+  jq --arg window "$WINDOW" \
+     --arg worktree "$WORKTREE_NAME" \
+     --arg worktree_path "$WORKTREE_PATH" \
+     --arg spare_branch "$SPARE_BRANCH" \
+     --arg branch "$NEW_BRANCH" \
+     --arg objective "$OBJECTIVE" \
+     --arg session_id "$SESSION_ID" \
+     --argjson now "$NOW" \
+     '.agents += [{
+       "window": $window,
+       "worktree": $worktree,
+       "worktree_path": $worktree_path,
+       "spare_branch": $spare_branch,
+       "branch": $branch,
+       "objective": $objective,
+       "session_id": $session_id,
+       "state": "running",
+       "checkpoints": [],
+       "last_output_hash": "",
+       "last_seen_at": $now,
+       "spawned_at": $now,
+       "idle_since": 0,
+       "revision_count": 0,
+       "last_rebriefed_at": 0
+     }]' "$STATE_FILE" > "${STATE_FILE}.tmp" && mv "${STATE_FILE}.tmp" "$STATE_FILE"
+fi
+
+# Store pr_number + steps in state file if provided (enables verify-complete.sh).
+# The agent record was appended above so the jq select now finds it.
+if [ -n "$PR_NUMBER" ] && [ -f "$STATE_FILE" ]; then
+  if [ "${#STEPS[@]}" -gt 0 ]; then
+    STEPS_JSON=$(printf '%s\n' "${STEPS[@]}" | jq -R . | jq -s .)
+  else
+    STEPS_JSON='[]'
+  fi
+  jq --arg w "$WINDOW" --arg pr "$PR_NUMBER" --argjson steps "$STEPS_JSON" \
+    '.agents |= map(if .window == $w then . + {pr_number: $pr, steps: $steps, checkpoints: []} else . end)' \
+    "$STATE_FILE" > "${STATE_FILE}.tmp" && mv "${STATE_FILE}.tmp" "$STATE_FILE"
+fi
+
+# Launch claude with a stable session ID so it can always be resumed after a crash:
+#   claude --resume SESSION_ID --permission-mode bypassPermissions
+tmux send-keys -t "$WINDOW" "cd '${WORKTREE_PATH}' && claude --permission-mode bypassPermissions --session-id '${SESSION_ID}'" Enter
+
+# Wait up to 60s for claude to be fully interactive:
+# both pane_current_command == 'node' AND the '❯' prompt is visible.
+PROMPT_FOUND=false
+for i in $(seq 1 60); do
+  CMD=$(tmux display-message -t "$WINDOW" -p '#{pane_current_command}' 2>/dev/null || echo "")
+  PANE=$(tmux capture-pane -t "$WINDOW" -p 2>/dev/null || echo "")
+  if echo "$PANE" | grep -q "Enter to confirm"; then
+    tmux send-keys -t "$WINDOW" Down Enter
+    sleep 2
+    continue
+  fi
+  if [[ "$CMD" == "node" ]] && echo "$PANE" | grep -q "❯"; then
+    PROMPT_FOUND=true
+    break
+  fi
+  sleep 1
+done
+
+if ! $PROMPT_FOUND; then
+  echo "[spawn-agent] WARNING: timed out waiting for ❯ prompt on $WINDOW — sending objective anyway" >&2
+fi
+
+# Send the task. Split text and Enter — if combined, Enter can fire before the string
+# is fully buffered, leaving the message stuck as "[Pasted text +N lines]" unsent.
+tmux send-keys -t "$WINDOW" "${OBJECTIVE} Output each completed step as CHECKPOINT:<step-name>. When ALL steps are done, output ORCHESTRATOR:DONE on its own line."
+sleep 0.3
+tmux send-keys -t "$WINDOW" Enter
+
+# Only output the window address — nothing else (callers parse this)
+echo "$WINDOW"

.claude/skills/orchestrate/scripts/status.sh

@@ -0,0 +1,43 @@
+#!/usr/bin/env bash
+# status.sh — print orchestrator status: state file summary + live tmux pane commands
+#
+# Usage: status.sh
+# Reads: ~/.claude/orchestrator-state.json
+
+set -euo pipefail
+
+STATE_FILE="${ORCHESTRATOR_STATE_FILE:-$HOME/.claude/orchestrator-state.json}"
+
+if [ ! -f "$STATE_FILE" ] || ! jq -e '.' "$STATE_FILE" >/dev/null 2>&1; then
+  echo "No orchestrator state found at $STATE_FILE"
+  exit 0
+fi
+
+# Header: active status, session, thresholds, last poll
+jq -r '
+  "=== Orchestrator [\(if .active then "RUNNING" else "STOPPED" end)] ===",
+  "Session: \(.tmux_session // "unknown")  |  Idle threshold: \(.idle_threshold_seconds // 300)s",
+  "Last poll: \(if (.last_poll_at // 0) == 0 then "never" else (.last_poll_at | strftime("%H:%M:%S")) end)",
+  ""
+' "$STATE_FILE"
+
+# Each agent: state, window, worktree/branch, truncated objective
+AGENT_COUNT=$(jq '.agents | length' "$STATE_FILE")
+if [ "$AGENT_COUNT" -eq 0 ]; then
+  echo "  (no agents registered)"
+else
+  jq -r '
+    .agents[] |
+    "  [\(.state | ascii_upcase)] \(.window)  \(.worktree)/\(.branch)",
+    "    \(.objective // "" | .[0:70])"
+  ' "$STATE_FILE"
+fi
+
+echo ""
+
+# Live pane_current_command for non-done agents
+while IFS= read -r WINDOW; do
+  [ -z "$WINDOW" ] && continue
+  CMD=$(tmux display-message -t "$WINDOW" -p '#{pane_current_command}' 2>/dev/null || echo "unreachable")
+  echo "  $WINDOW live: $CMD"
+done < <(jq -r '.agents[] | select(.state != "done") | .window' "$STATE_FILE" 2>/dev/null || true)

.claude/skills/orchestrate/scripts/verify-complete.sh

@@ -0,0 +1,180 @@
+#!/usr/bin/env bash
+# verify-complete.sh — verify a PR task is truly done before marking the agent done
+#
+# Check order matters:
+#   1. Checkpoints — did the agent do all required steps?
+#   2. CI complete — no pending (bots post comments AFTER their check runs, must wait)
+#   3. CI passing — no failures (agent must fix before done)
+#   4. spawned_at — a new CI run was triggered after agent spawned (proves real work)
+#   5. Unresolved threads — checked AFTER CI so bot-posted comments are included
+#   6. CHANGES_REQUESTED — checked AFTER CI so bot reviews are included
+#
+# Usage: verify-complete.sh WINDOW
+# Exit 0 = verified complete; exit 1 = not complete (stderr has reason)
+
+set -euo pipefail
+
+WINDOW="$1"
+STATE_FILE="${ORCHESTRATOR_STATE_FILE:-$HOME/.claude/orchestrator-state.json}"
+
+PR_NUMBER=$(jq -r --arg w "$WINDOW" '.agents[] | select(.window == $w) | .pr_number // ""' "$STATE_FILE" 2>/dev/null)
+STEPS=$(jq -r --arg w "$WINDOW" '.agents[] | select(.window == $w) | .steps // [] | .[]' "$STATE_FILE" 2>/dev/null || true)
+CHECKPOINTS=$(jq -r --arg w "$WINDOW" '.agents[] | select(.window == $w) | .checkpoints // [] | .[]' "$STATE_FILE" 2>/dev/null || true)
+WORKTREE_PATH=$(jq -r --arg w "$WINDOW" '.agents[] | select(.window == $w) | .worktree_path // ""' "$STATE_FILE" 2>/dev/null)
+BRANCH=$(jq -r --arg w "$WINDOW" '.agents[] | select(.window == $w) | .branch // ""' "$STATE_FILE" 2>/dev/null)
+SPAWNED_AT=$(jq -r --arg w "$WINDOW" '.agents[] | select(.window == $w) | .spawned_at // "0"' "$STATE_FILE" 2>/dev/null || echo "0")
+
+# No PR number = cannot verify
+if [ -z "$PR_NUMBER" ]; then
+  echo "NOT COMPLETE: no pr_number in state — set pr_number or mark done manually" >&2
+  exit 1
+fi
+
+# --- Check 1: all required steps are checkpointed ---
+MISSING=""
+while IFS= read -r step; do
+  [ -z "$step" ] && continue
+  if ! echo "$CHECKPOINTS" | grep -qFx "$step"; then
+    MISSING="$MISSING $step"
+  fi
+done <<< "$STEPS"
+
+if [ -n "$MISSING" ]; then
+  echo "NOT COMPLETE: missing checkpoints:$MISSING on PR #$PR_NUMBER" >&2
+  exit 1
+fi
+
+# Resolve repo for all GitHub checks below
+REPO=$(jq -r '.repo // ""' "$STATE_FILE" 2>/dev/null || echo "")
+if [ -z "$REPO" ] && [ -n "$WORKTREE_PATH" ] && [ -d "$WORKTREE_PATH" ]; then
+  REPO=$(git -C "$WORKTREE_PATH" remote get-url origin 2>/dev/null \
+    | sed 's|.*github\.com[:/]||; s|\.git$||' || echo "")
+fi
+
+if [ -z "$REPO" ]; then
+  echo "Warning: cannot resolve repo — skipping CI/thread checks" >&2
+  echo "VERIFIED: PR #$PR_NUMBER — checkpoints ✓ (CI/thread checks skipped — no repo)"
+  exit 0
+fi
+
+CI_BUCKETS=$(gh pr checks "$PR_NUMBER" --repo "$REPO" --json bucket 2>/dev/null || echo "[]")
+
+# --- Check 2: CI fully complete — no pending checks ---
+# Pending checks MUST finish before we check threads/reviews:
+# bots (Seer, Check PR Status, etc.) post comments and CHANGES_REQUESTED AFTER their CI check runs.
+PENDING=$(echo "$CI_BUCKETS" | jq '[.[] | select(.bucket == "pending")] | length' 2>/dev/null || echo "0")
+if [ "$PENDING" -gt 0 ]; then
+  PENDING_NAMES=$(gh pr checks "$PR_NUMBER" --repo "$REPO" --json bucket,name 2>/dev/null \
+    | jq -r '[.[] | select(.bucket == "pending") | .name] | join(", ")' 2>/dev/null || echo "unknown")
+  echo "NOT COMPLETE: $PENDING CI checks still pending on PR #$PR_NUMBER ($PENDING_NAMES)" >&2
+  exit 1
+fi
+
+# --- Check 3: CI passing — no failures ---
+FAILING=$(echo "$CI_BUCKETS" | jq '[.[] | select(.bucket == "fail")] | length' 2>/dev/null || echo "0")
+if [ "$FAILING" -gt 0 ]; then
+  FAILING_NAMES=$(gh pr checks "$PR_NUMBER" --repo "$REPO" --json bucket,name 2>/dev/null \
+    | jq -r '[.[] | select(.bucket == "fail") | .name] | join(", ")' 2>/dev/null || echo "unknown")
+  echo "NOT COMPLETE: $FAILING failing CI checks on PR #$PR_NUMBER ($FAILING_NAMES)" >&2
+  exit 1
+fi
+
+# --- Check 4: a new CI run was triggered AFTER the agent spawned ---
+if [ -n "$BRANCH" ] && [ "${SPAWNED_AT:-0}" -gt 0 ]; then
+  LATEST_RUN_AT=$(gh run list --repo "$REPO" --branch "$BRANCH" \
+    --json createdAt --limit 1 2>/dev/null | jq -r '.[0].createdAt // ""')
+  if [ -n "$LATEST_RUN_AT" ]; then
+    if date --version >/dev/null 2>&1; then
+      LATEST_RUN_EPOCH=$(date -d "$LATEST_RUN_AT" "+%s" 2>/dev/null || echo "0")
+    else
+      LATEST_RUN_EPOCH=$(TZ=UTC date -j -f "%Y-%m-%dT%H:%M:%SZ" "$LATEST_RUN_AT" "+%s" 2>/dev/null || echo "0")
+    fi
+    if [ "$LATEST_RUN_EPOCH" -le "$SPAWNED_AT" ]; then
+      echo "NOT COMPLETE: latest CI run on $BRANCH predates agent spawn — agent may not have pushed yet" >&2
+      exit 1
+    fi
+  fi
+fi
+
+OWNER=$(echo "$REPO" | cut -d/ -f1)
+REPONAME=$(echo "$REPO" | cut -d/ -f2)
+
+# --- Check 5: no unresolved review threads (checked AFTER CI — bots post after their check) ---
+UNRESOLVED=$(gh api graphql -f query="
+  { repository(owner: \"${OWNER}\", name: \"${REPONAME}\") {
+      pullRequest(number: ${PR_NUMBER}) {
+        reviewThreads(first: 50) { nodes { isResolved } }
+      }
+    }
+  }
+" --jq '[.data.repository.pullRequest.reviewThreads.nodes[] | select(.isResolved == false)] | length' 2>/dev/null || echo "0")
+
+if [ "$UNRESOLVED" -gt 0 ]; then
+  echo "NOT COMPLETE: $UNRESOLVED unresolved review threads on PR #$PR_NUMBER" >&2
+  exit 1
+fi
+
+# --- Check 6: no CHANGES_REQUESTED (checked AFTER CI — bots post reviews after their check) ---
+# A CHANGES_REQUESTED review is stale if the latest commit was pushed AFTER the review was submitted.
+# Stale reviews (pre-dating the fixing commits) should not block verification.
+#
+# Fetch commits and latestReviews in a single call and fail closed — if gh fails,
+# treat that as NOT COMPLETE rather than silently passing.
+# Use latestReviews (not reviews) so each reviewer's latest state is used — superseded
+# CHANGES_REQUESTED entries are automatically excluded when the reviewer later approved.
+# Note: we intentionally use committedDate (not PR updatedAt) because updatedAt changes on any
+# PR activity (bot comments, label changes) which would create false negatives.
+PR_REVIEW_METADATA=$(gh pr view "$PR_NUMBER" --repo "$REPO" \
+  --json commits,latestReviews 2>/dev/null) || {
+  echo "NOT COMPLETE: unable to fetch PR review metadata for PR #$PR_NUMBER" >&2
+  exit 1
+}
+
+LATEST_COMMIT_DATE=$(jq -r '.commits[-1].committedDate // ""' <<< "$PR_REVIEW_METADATA")
+CHANGES_REQUESTED_REVIEWS=$(jq '[.latestReviews[]? | select(.state == "CHANGES_REQUESTED")]' <<< "$PR_REVIEW_METADATA")
+
+BLOCKING_CHANGES_REQUESTED=0
+BLOCKING_REQUESTERS=""
+
+if [ -n "$LATEST_COMMIT_DATE" ] && [ "$(echo "$CHANGES_REQUESTED_REVIEWS" | jq length)" -gt 0 ]; then
+  if date --version >/dev/null 2>&1; then
+    LATEST_COMMIT_EPOCH=$(date -d "$LATEST_COMMIT_DATE" "+%s" 2>/dev/null || echo "0")
+  else
+    LATEST_COMMIT_EPOCH=$(TZ=UTC date -j -f "%Y-%m-%dT%H:%M:%SZ" "$LATEST_COMMIT_DATE" "+%s" 2>/dev/null || echo "0")
+  fi
+
+  while IFS= read -r review; do
+    [ -z "$review" ] && continue
+    REVIEW_DATE=$(echo "$review" | jq -r '.submittedAt // ""')
+    REVIEWER=$(echo "$review" | jq -r '.author.login // "unknown"')
+    if [ -z "$REVIEW_DATE" ]; then
+      # No submission date — treat as fresh (conservative: blocks verification)
+      BLOCKING_CHANGES_REQUESTED=$(( BLOCKING_CHANGES_REQUESTED + 1 ))
+      BLOCKING_REQUESTERS="${BLOCKING_REQUESTERS:+$BLOCKING_REQUESTERS, }${REVIEWER}"
+    else
+      if date --version >/dev/null 2>&1; then
+        REVIEW_EPOCH=$(date -d "$REVIEW_DATE" "+%s" 2>/dev/null || echo "0")
+      else
+        REVIEW_EPOCH=$(TZ=UTC date -j -f "%Y-%m-%dT%H:%M:%SZ" "$REVIEW_DATE" "+%s" 2>/dev/null || echo "0")
+      fi
+      if [ "$REVIEW_EPOCH" -gt "$LATEST_COMMIT_EPOCH" ]; then
+        # Review was submitted AFTER latest commit — still fresh, blocks verification
+        BLOCKING_CHANGES_REQUESTED=$(( BLOCKING_CHANGES_REQUESTED + 1 ))
+        BLOCKING_REQUESTERS="${BLOCKING_REQUESTERS:+$BLOCKING_REQUESTERS, }${REVIEWER}"
+      fi
+      # Review submitted BEFORE latest commit — stale, skip
+    fi
+  done <<< "$(echo "$CHANGES_REQUESTED_REVIEWS" | jq -c '.[]')"
+else
+  # No commit date or no changes_requested — check raw count as fallback
+  BLOCKING_CHANGES_REQUESTED=$(echo "$CHANGES_REQUESTED_REVIEWS" | jq length 2>/dev/null || echo "0")
+  BLOCKING_REQUESTERS=$(echo "$CHANGES_REQUESTED_REVIEWS" | jq -r '[.[].author.login] | join(", ")' 2>/dev/null || echo "unknown")
+fi
+
+if [ "$BLOCKING_CHANGES_REQUESTED" -gt 0 ]; then
+  echo "NOT COMPLETE: CHANGES_REQUESTED (after latest commit) from ${BLOCKING_REQUESTERS} on PR #$PR_NUMBER" >&2
+  exit 1
+fi
+
+echo "VERIFIED: PR #$PR_NUMBER — checkpoints ✓, CI complete + green, 0 unresolved threads, no CHANGES_REQUESTED"
+exit 0

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

@@ -90,10 +90,12 @@ Address comments **one at a time**: fix → commit → push → inline reply →
 2. Commit and push the fix
 3. Reply **inline** (not as a new top-level comment) referencing the fixing commit — this is what resolves the conversation for bot reviewers (coderabbitai, sentry):
 
+Use a **markdown commit link** so GitHub renders it as a clickable reference. Get the full SHA with `git rev-parse HEAD` after committing:
+
 | Comment type | How to reply |
 |---|---|
-| Inline review (`pulls/{N}/comments`) | `gh api repos/Significant-Gravitas/AutoGPT/pulls/{N}/comments/{ID}/replies -f body="🤖 Fixed in <commit-sha>: <description>"` |
-| Conversation (`issues/{N}/comments`) | `gh api repos/Significant-Gravitas/AutoGPT/issues/{N}/comments -f body="🤖 Fixed in <commit-sha>: <description>"` |
+| Inline review (`pulls/{N}/comments`) | `gh api repos/Significant-Gravitas/AutoGPT/pulls/{N}/comments/{ID}/replies -f body="🤖 Fixed in [abc1234](https://github.com/Significant-Gravitas/AutoGPT/commit/FULL_SHA): <description>"` |
+| Conversation (`issues/{N}/comments`) | `gh api repos/Significant-Gravitas/AutoGPT/issues/{N}/comments -f body="🤖 Fixed in [abc1234](https://github.com/Significant-Gravitas/AutoGPT/commit/FULL_SHA): <description>"` |
 
 ## Codecov coverage
 

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

@@ -530,19 +530,9 @@ After showing all screenshots, output a **detailed** summary table:
 # but Homebrew bash is 5.x; Linux typically has bash 5.x). If running on Bash <4, use a
 # plain variable with a lookup function instead.
 declare -A SCREENSHOT_EXPLANATIONS=(
-  # Each explanation MUST answer three things:
-  #   1. FLOW: Which test scenario / user journey is this part of?
-  #   2. STEPS: What exact actions were taken to reach this state?
-  #   3. EVIDENCE: What does this screenshot prove (pass/fail/data)?
-  #
-  # Good example:
-  #   ["03-cost-log-after-run.png"]="Flow: LLM block cost tracking. Steps: Logged in as tester@gmail.com → ran 'Cost Test Agent' → waited for COMPLETED status. Evidence: PlatformCostLog table shows 1 new row with cost_microdollars=1234 and correct user_id."
-  #
-  # Bad example (too vague — never do this):
-  #   ["03-cost-log.png"]="Shows the cost log table."
-  ["01-login-page.png"]="Flow: Login flow. Steps: Opened /login. Evidence: Login page renders with email/password fields and SSO options visible."
-  ["02-builder-with-block.png"]="Flow: Block execution. Steps: Logged in → /build → added LLM block. Evidence: Builder canvas shows block connected to trigger, ready to run."
-  # ... one entry per screenshot using the flow/steps/evidence format above
+  ["01-login-page.png"]="Shows the login page loaded successfully with SSO options visible."
+  ["02-builder-with-block.png"]="The builder canvas displays the newly added block connected to the trigger."
+  # ... one entry per screenshot, using the same explanations you showed the user above
 )
 
 TEST_RESULTS_TABLE="| 1 | Login flow | PASS | N/A | 01-login-before.png, 02-login-after.png |
@@ -557,8 +547,7 @@ Upload screenshots to the PR using the GitHub Git API (no local git operations
 
 **This step is MANDATORY. Every test run MUST post a PR comment with screenshots. No exceptions.**
 
-> **CRITICAL — NEVER post a bare directory link like `https://github.com/.../tree/...`.**
-> Every screenshot MUST appear as `![name](raw_url)` inline in the PR comment so reviewers can see them without clicking any links. After posting, the verification step below greps the comment for `![` tags and exits 1 if none are found — the test run is considered incomplete until this passes.
+**CRITICAL — NEVER post a bare directory link like `https://github.com/.../tree/...`.** Every screenshot MUST appear as `![name](raw_url)` inline in the PR comment so reviewers can see them without clicking any links. After posting, the verification step below greps the comment for `![` tags and exits 1 if none are found — the test run is considered incomplete until this passes.
 
 ```bash
 # Upload screenshots via GitHub Git API (creates blobs, tree, commit, and ref remotely)
@@ -595,11 +584,11 @@ for img in "${SCREENSHOT_FILES[@]}"; do
 done
 TREE_JSON+=']'
 
-# Step 2: Create tree, commit (with parent), and branch ref
+# Step 2: Create tree, commit, and branch ref
 TREE_SHA=$(echo "$TREE_JSON" | jq -c '{tree: .}' | gh api "repos/${REPO}/git/trees" --input - --jq '.sha')
 
-# Resolve existing branch tip as parent (avoids orphan commits on repeat runs)
-PARENT_SHA=$(gh api "repos/${REPO}/git/refs/heads/${SCREENSHOTS_BRANCH}" --jq '.object.sha' 2>/dev/null || true)
+# Resolve parent commit so screenshots are chained, not orphan root commits
+PARENT_SHA=$(gh api "repos/${REPO}/git/refs/heads/${SCREENSHOTS_BRANCH}" --jq '.object.sha' 2>/dev/null || echo "")
 if [ -n "$PARENT_SHA" ]; then
   COMMIT_SHA=$(gh api "repos/${REPO}/git/commits" \
     -f message="test: add E2E test screenshots for PR #${PR_NUMBER}" \
@@ -607,7 +596,6 @@ if [ -n "$PARENT_SHA" ]; then
     -f "parents[]=$PARENT_SHA" \
     --jq '.sha')
 else
-  # First commit on this branch — no parent
   COMMIT_SHA=$(gh api "repos/${REPO}/git/commits" \
     -f message="test: add E2E test screenshots for PR #${PR_NUMBER}" \
     -f tree="$TREE_SHA" \
@@ -618,7 +606,7 @@ gh api "repos/${REPO}/git/refs" \
   -f ref="refs/heads/${SCREENSHOTS_BRANCH}" \
   -f sha="$COMMIT_SHA" 2>/dev/null \
   || gh api "repos/${REPO}/git/refs/heads/${SCREENSHOTS_BRANCH}" \
-    -X PATCH -f sha="$COMMIT_SHA" -f force=true
+    -X PATCH -f sha="$COMMIT_SHA" -F force=true
 ```
 
 Then post the comment with **inline images AND explanations for each screenshot**:
@@ -682,122 +670,122 @@ ${IMAGE_MARKDOWN}
 ${FAILED_SECTION}
 INNEREOF
 
-POSTED_BODY=$(gh api "repos/${REPO}/issues/$PR_NUMBER/comments" -F body=@"$COMMENT_FILE" --jq '.body')
+gh api "repos/${REPO}/issues/$PR_NUMBER/comments" -F body=@"$COMMENT_FILE"
 rm -f "$COMMENT_FILE"
+
+# Verify the posted comment contains inline images — exit 1 if none found
+# Use separate --paginate + jq pipe: --jq applies per-page, not to the full list
+LAST_COMMENT=$(gh api "repos/${REPO}/issues/$PR_NUMBER/comments" --paginate 2>/dev/null | jq -r '.[-1].body // ""')
+if ! echo "$LAST_COMMENT" | grep -q '!\['; then
+  echo "ERROR: Posted comment contains no inline images (![). Bare directory links are not acceptable." >&2
+  exit 1
+fi
+echo "✓ Inline images verified in posted comment"
 ```
 
 **The PR comment MUST include:**
 1. A summary table of all scenarios with PASS/FAIL and before/after API evidence
 2. Every successfully uploaded screenshot rendered inline; any failed uploads listed with manual attachment instructions
-3. A structured explanation below each screenshot covering: **Flow** (which scenario), **Steps** (exact actions taken to reach this state), **Evidence** (what this proves — pass/fail/data values). A bare "shows the page" caption is not acceptable.
+3. A 1-2 sentence explanation below each screenshot describing what it proves
 
 This approach uses the GitHub Git API to create blobs, trees, commits, and refs entirely server-side. No local `git checkout` or `git push` — safe for worktrees and won't interfere with the PR branch.
 
-**Verify inline rendering after posting — this is required, not optional:**
+## Step 8: Evaluate and post a formal PR review
 
+After the test comment is posted, evaluate whether the run was thorough enough to make a merge decision, then post a formal GitHub review (approve or request changes). **This step is mandatory — every test run MUST end with a formal review decision.**
+
+### Evaluation criteria
+
+Re-read the PR description:
 ```bash
-# 1. Confirm the posted comment body contains inline image markdown syntax
-if ! echo "$POSTED_BODY" | grep -q '!\['; then
-  echo "❌ FAIL: No inline image tags in posted comment body. Re-check IMAGE_MARKDOWN and re-post."
-  exit 1
-fi
-
-# 2. Verify at least one raw URL actually resolves (catches wrong branch name, wrong path, etc.)
-FIRST_IMG_URL=$(echo "$POSTED_BODY" | grep -o 'https://raw.githubusercontent.com[^)]*' | head -1)
-if [ -n "$FIRST_IMG_URL" ]; then
-  HTTP_STATUS=$(curl -s -o /dev/null -w "%{http_code}" --max-time 10 "$FIRST_IMG_URL")
-  if [ "$HTTP_STATUS" = "200" ]; then
-    echo "✅ Inline images confirmed and raw URL resolves (HTTP 200)"
-  else
-    echo "❌ FAIL: Raw image URL returned HTTP $HTTP_STATUS — images will not render inline."
-    echo "   URL: $FIRST_IMG_URL"
-    echo "   Check branch name, path, and that the push succeeded."
-    exit 1
-  fi
-else
-  echo "⚠️  Could not extract a raw URL from the comment — verify manually."
-fi
+gh pr view "$PR_NUMBER" --json body --jq '.body' --repo "$REPO"
 ```
 
-## Step 8: Evaluate test completeness and post a GitHub review
+Score the run against each criterion:
 
-After posting the PR comment, evaluate whether the test run actually covered everything it needed to. This is NOT a rubber-stamp — be critical. Then post a formal GitHub review so the PR author and reviewers can see the verdict.
+| Criterion | Pass condition |
+|-----------|---------------|
+| **Coverage** | Every feature/change described in the PR has at least one test scenario |
+| **All scenarios pass** | No FAIL rows in the results table |
+| **Negative tests** | At least one failure-path test per feature (invalid input, unauthorized, edge case) |
+| **Before/after evidence** | Every state-changing API call has before/after values logged |
+| **Screenshots are meaningful** | Screenshots show the actual state change, not just a loading spinner or blank page |
+| **No regressions** | Existing core flows (login, agent create/run) still work |
 
-### 8a. Evaluate against the test plan
+### Decision logic
 
-Re-read `$RESULTS_DIR/test-plan.md` (written in Step 2) and `$RESULTS_DIR/test-report.md` (written in Step 5). For each scenario in the plan, answer:
+```
+ALL criteria pass                            → APPROVE
+Any scenario FAIL or missing PR feature      → REQUEST_CHANGES (list gaps)
+Evidence weak (no before/after, vague shots) → REQUEST_CHANGES (list what's missing)
+```
 
-> **Note:** `test-report.md` is written in Step 5. If it doesn't exist, write it before proceeding here — see the Step 5 template. Do not skip evaluation because the file is missing; create it from your notes instead.
-
-| Question | Pass criteria |
-|----------|--------------|
-| Was it tested? | Explicit steps were executed, not just described |
-| Is there screenshot evidence? | At least one before/after screenshot per scenario |
-| Did the core feature work correctly? | Expected state matches actual state |
-| Were negative cases tested? | At least one failure/rejection case per feature |
-| Was DB/API state verified (not just UI)? | Raw API response or DB query confirms state change |
-
-Build a verdict:
-- **APPROVE** — every scenario tested, evidence present, no bugs found or all bugs are minor/known
-- **REQUEST_CHANGES** — one or more: untested scenarios, missing evidence, bugs found, data not verified
-
-### 8b. Post the GitHub review
+### Post the review
 
 ```bash
-EVAL_FILE=$(mktemp)
+REVIEW_FILE=$(mktemp)
 
-# === STEP A: Write header ===
-cat > "$EVAL_FILE" << 'ENDEVAL'
-## 🧪 Test Evaluation
+# Count results
+PASS_COUNT=$(echo "$TEST_RESULTS_TABLE" | grep -c "PASS" || true)
+FAIL_COUNT=$(echo "$TEST_RESULTS_TABLE" | grep -c "FAIL" || true)
+TOTAL=$(( PASS_COUNT + FAIL_COUNT ))
 
-### Coverage checklist
-ENDEVAL
+# List any coverage gaps found during evaluation (populate this array as you assess)
+# e.g. COVERAGE_GAPS=("PR claims to add X but no test covers it")
+COVERAGE_GAPS=()
+```
 
-# === STEP B: Append ONE line per scenario — do this BEFORE calculating verdict ===
-# Format: "- ✅ **Scenario N – name**: <what was done and verified>"
-#      or "- ❌ **Scenario N – name**: <what is missing or broken>"
-# Examples:
-#   echo "- ✅ **Scenario 1 – Login flow**: tested, screenshot evidence present, auth token verified via API" >> "$EVAL_FILE"
-#   echo "- ❌ **Scenario 3 – Cost logging**: NOT verified in DB — UI showed entry but raw SQL query was skipped" >> "$EVAL_FILE"
-#
-# !!! IMPORTANT: append ALL scenario lines here before proceeding to STEP C !!!
+**If APPROVING** — all criteria met, zero failures, full coverage:
 
-# === STEP C: Derive verdict from the checklist — runs AFTER all lines are appended ===
-FAIL_COUNT=$(grep -c "^- ❌" "$EVAL_FILE" || true)
-if [ "$FAIL_COUNT" -eq 0 ]; then
-  VERDICT="APPROVE"
-else
-  VERDICT="REQUEST_CHANGES"
-fi
+```bash
+cat > "$REVIEW_FILE" <<REVIEWEOF
+## E2E Test Evaluation — APPROVED
 
-# === STEP D: Append verdict section ===
-cat >> "$EVAL_FILE" << ENDVERDICT
+**Results:** ${PASS_COUNT}/${TOTAL} scenarios passed.
 
-### Verdict
-ENDVERDICT
+**Coverage:** All features described in the PR were exercised.
 
-if [ "$VERDICT" = "APPROVE" ]; then
-  echo "✅ All scenarios covered with evidence. No blocking issues found." >> "$EVAL_FILE"
-else
-  echo "❌ $FAIL_COUNT scenario(s) incomplete or have confirmed bugs. See ❌ items above." >> "$EVAL_FILE"
-  echo "" >> "$EVAL_FILE"
-  echo "**Required before merge:** address each ❌ item above." >> "$EVAL_FILE"
-fi
+**Evidence:** Before/after API values logged for all state-changing operations; screenshots show meaningful state transitions.
 
-# === STEP E: Post the review ===
-gh api "repos/${REPO}/pulls/$PR_NUMBER/reviews" \
-  --method POST \
-  -f body="$(cat "$EVAL_FILE")" \
-  -f event="$VERDICT"
+**Negative tests:** Failure paths tested for each feature.
 
-rm -f "$EVAL_FILE"
+No regressions observed on core flows.
+REVIEWEOF
+
+gh pr review "$PR_NUMBER" --repo "$REPO" --approve --body "$(cat "$REVIEW_FILE")"
+echo "✅ PR approved"
+```
+
+**If REQUESTING CHANGES** — any failure, coverage gap, or missing evidence:
+
+```bash
+FAIL_LIST=$(echo "$TEST_RESULTS_TABLE" | grep "FAIL" | awk -F'|' '{print "- Scenario" $2 "failed"}' || true)
+
+cat > "$REVIEW_FILE" <<REVIEWEOF
+## E2E Test Evaluation — Changes Requested
+
+**Results:** ${PASS_COUNT}/${TOTAL} scenarios passed, ${FAIL_COUNT} failed.
+
+### Required before merge
+
+${FAIL_LIST}
+$(for gap in "${COVERAGE_GAPS[@]}"; do echo "- $gap"; done)
+
+Please fix the above and re-run the E2E tests.
+REVIEWEOF
+
+gh pr review "$PR_NUMBER" --repo "$REPO" --request-changes --body "$(cat "$REVIEW_FILE")"
+echo "❌ Changes requested"
+```
+
+```bash
+rm -f "$REVIEW_FILE"
 ```
 
 **Rules:**
-- Never auto-approve without checking every scenario in the test plan
-- `REQUEST_CHANGES` if ANY scenario is untested, lacks DB/API evidence, or has a confirmed bug
-- The evaluation body must list every scenario explicitly (✅ or ❌) — not just the failures
-- If you find new bugs during evaluation, add them to the request-changes body and (if `--fix` flag is set) fix them before posting
+- In `--fix` mode, fix all failures before posting the review — the review reflects the final state after fixes
+- Never approve if any scenario failed, even if it seems like a flake — rerun that scenario first
+- Never request changes for issues already fixed in this run
 
 ## Fix mode (--fix flag)