Shutdown Dev

Gracefully terminate a dev teammate's session while preserving their working context for future resumption.

When to use

• Team scale-down (e.g. going from 8 → 6 devs after OOM pressure or sprint close)
• Dev assigned to an orphaned branch after their PR was closed
• Sprint retrospective and wrap-up — shutting down devs between sprints
• Memory pressure requiring you to shed agent processes

Pre-shutdown checklist

1. Is the dev in the middle of something? Check their task status:

   # via TaskList, filter for owner == <dev-name>

   If they're mid-implementation on a non-trivial task, prefer rotating them OFF that task first rather than losing the work.

2. Is the user currently talking to this dev? If yes, confirm with the user before shutting down.

3. Do they have uncommitted work in their worktree? Check via their file-locks entries and any in-flight PRs. Warn them to commit before approving shutdown.

Step 1: Send the shutdown_request

SendMessage to: <dev-name>
Type: shutdown_request

{"type": "shutdown_request", "reason": "<reason — e.g. 'Scaling team from 8 to 6 after OOM recovery; your current PR is merged/closed and you have no active task'>. Before approving shutdown: (1) write your working context to plan/agent-context/<dev-name>.md — include active branches, any in-progress thoughts, and anything a successor agent would need to resume. Keep it concise (~100 lines). (2) Commit any uncommitted WIP in your worktrees, or note explicitly that nothing is WIP. (3) Reply with shutdown_response approve=true. Your worktrees and branches stay — another dev can resume them later."}

Critical constraints in the message:

• MUST instruct them to write plan/agent-context/<name>.md FIRST
• MUST tell them to commit uncommitted WIP
• MUST NOT ask them to narrate their context in the chat reply (that bloats the tech lead's context)
• Reply should be a one-line shutdown_response approve=true

Step 2: Wait for response

The dev will either:

• Reply with {"type":"shutdown_response","approve":true,"request_id":"..."} → proceed to step 3
• Reply asking questions (rare) → answer briefly, re-send shutdown_request
• Not respond (stuck process) → see step 5

Step 3: Verify the context file was written

ls -la plan/agent-context/<dev-name>.md
head -20 plan/agent-context/<dev-name>.md

The file should exist, be recent, and have substantive content (not empty).

If the dev skipped this step, re-send the shutdown_request with an explicit "blocked until context file exists" note.

Step 4: Verify the worktree is in a known state

wt=/workspace/.claude/worktrees/<branch-name>
git -C "$wt" diff --stat      # uncommitted changes
git -C "$wt" log --oneline main..HEAD  # unmerged commits

If there are uncommitted changes: write them to the agent-context file as "Suspended Work" with a resume recipe, then commit or drop per the dev's note.

If there are unmerged commits: note the branch name in the context file so a successor can pick up.

Step 5: If the dev is unresponsive

The dev's claude process may be stuck or already dead. Verify:

ps -ef | grep claude | grep -v grep

Options:

1. Wait another 60s — they may be mid-tool-call
2. Send a PAUSE to interrupt the current work, then retry the shutdown_request
3. If clearly dead: manually write their agent-context file based on known state (branch name, task ID, last-known progress from TaskList), then kill the process

Step 6: Commit the context file

If the dev committed it themselves, you're done. If not:

cd /workspace
git add plan/agent-context/<dev-name>.md
git commit -m "docs(agent-context): preserve <dev-name> state before shutdown

✓

<dev-name> shut down during <reason>. Context preserved for resumption.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>"

Step 7: Update team state

• Remove the dev from any active dispatch messaging
• Update plan/method/agent-sessions.md if it tracks live sessions
• Update plan/method/file-locks.md to drop any claims held by the dev

Step 8: Confirm termination

After the dev's process exits (approved shutdown), verify no orphan vitest or subprocess runners remain:

ps -ef | grep -E '(vitest|claude)' | grep -v grep

If leftover processes trace back to the shut-down dev, kill them.

Output

Shutdown complete: <dev-name>
Context saved: plan/agent-context/<dev-name>.md (~NN lines)
Worktree: <branch> preserved, <uncommitted|clean>, <unmerged|merged>
Team size: N → N-1

Notes

• The agent-context file is the handoff channel — without it, a successor agent has to reconstruct everything from issue files and git history.
• Never delete worktrees on shutdown — they may contain recoverable state a successor needs. Cleanup happens at sprint close, not at shutdown.
• Don't shut down devs mid-PR-review — if their PR is open and awaiting CI, wait for CI to finish (merge or close) before shutting down. Otherwise their context file has to describe a PR state that may change.
• Scale-down batches: when shutting down multiple devs, send all the shutdown_requests in parallel (one SendMessage each in a single message) and wait for all approvals before verifying.