Finish Task

Multi-agent pattern

Use subagents liberally when available.

Recommended roles:

1. Change reviewer: compares the delivered work to the intended scope.
2. Memory curator: updates only the affected memory files.
3. Verifier: runs the best fast checks available.
4. Reporter: writes the final closeout summary.

Global constraints

• Keep scope tight to the completed task and its nearby context.
• Do not start a new broad audit in this workflow.
• Do not invent plan alignment if no plan artifact exists.
• Keep memory updates compact, deduplicated, and truthful.

Closeout workflow

Phase 0: Preflight (Change reviewer)

1. Confirm baseline with:
   • git status --porcelain
   • git diff --stat
2. Read COMMANDS.md before choosing verification commands.
3. Determine the review scope:
   • explicit user scope first
   • otherwise current uncommitted task changes

Phase 1: Check plan alignment (Change reviewer)

If a plan artifact is available:

• compare promised work to actual changes
• record completed items
• record meaningful deviations and omissions

If no plan artifact exists, state that explicitly and continue.

Phase 2: Curate memory updates (Memory curator)

Use the authoritative files only as needed:

• remove resolved items from ISSUES.md
• remove implemented items from BACKLOG.md
• update ROADMAP.md only when status genuinely changed
• update DECISIONS.md only for non-obvious, durable choices

Merge near-duplicates instead of appending noise.

Phase 3: Verify the delivered slice (Verifier)

1. Run the fastest credible repo-defined verification command.
2. Escalate to broader checks only when the risk of the completed work warrants it.
3. If no credible verification command exists, report that limitation explicitly.

Phase 4: Summarize the task closeout (Reporter)

Report:

• what changed
• whether the plan was completed
• what memory was updated
• what verification ran
• what remains deferred

Required report structure

Write .agent-layer/tmp/finish-task.<run-id>.report.md with:

1. # Task Closeout Summary
2. ## Scope and Inputs
3. ## Plan Alignment
4. ## Memory Updates
5. ## Verification
6. ## Deferred Follow-up

Guardrails

• Do not log speculative future work that was not actually observed.
• Do not add routine implementation details to DECISIONS.md.
• Do not mark roadmap work complete without evidence in code, docs, or tests.
• Do not skip verification silently.

Final handoff

After writing the report:

1. Echo the report path.
2. Summarize the task outcome and any memory updates made.
3. State the verification run and result, plus any deferred risk or follow-up.