agent-loop-review-close

name	agent-loop-review-close
description	Review, verify, and close an agent-loop task safely after implementation, including blind-spot review, strict verification, and explicit accepted-risk handling.

Agent Loop Review Close

Use this skill when implementation is done or nearly done and the task needs to be reviewed, verified, and closed in a governed way.

Fast Path

Record execution evidence, review, outcomes, and the learning decision before verify/report/close:

vendor/bin/agent-loop session validation record <task-id> --brief-revision <n> --command "<brief command>" --status passed --exit-code 0 --by <actor>
vendor/bin/agent-loop review blindspots <task-id>
vendor/bin/agent-loop recall log-outcome --root <learning-root> --draft recall/<task-id>/recall-log.draft.json --by <actor> --commit <sha>
vendor/bin/agent-loop session learning decide <task-id> --status no_durable_learning --by <actor>
vendor/bin/agent-loop verify
vendor/bin/agent-loop workflow report <task-id> --changed-file <path>
vendor/bin/agent-loop workflow close <task-id> --status done

Do not skip required evidence or reorder the close gates.

Blind-Spot Review

vendor/bin/agent-loop review blindspots <task-id>

This is a deterministic check, not an LLM call. It writes Markdown/JSON reports under .agent-recall/reviews/ using task, session, and recall artifacts as context. It warns when session notes do not show that review blindspots itself was checked.

Review output is not human approval. It does not approve code. It does not approve durable learning. Human review remains required.

Verify

vendor/bin/agent-loop verify

Checks cross-package consistency: sessions, recall coverage, board, and learning state. Each check prints [OK], [SKIP], or [FAIL].

In CI or repos where all parts of the stack are expected to be present, use --strict to turn baseline skips into failures:

vendor/bin/agent-loop verify --strict

--strict fails when tasks/ or session_plan/ is missing entirely. Use it in expected-complete repos; omit it when the repo only wires part of the stack.

Close

vendor/bin/agent-loop workflow close <task-id> --status done

workflow close is gated: it requires an approved current work brief, passing evidence for every required command in that exact brief revision, recall metadata and explicit outcomes for selected guidance, a blind-spot review, an explicit learning decision, and a passing verify before accepting --status done. If the gate is not satisfied, the command exits with an error describing what is missing.

workflow report is the read-only handoff view. Pass each observed changed path with --changed-file; it deliberately does not run Git or infer scope.

Accepted Risk

If verify fails or a gate is not satisfied and you need to close anyway, accepted risk is explicit and written to disk:

vendor/bin/agent-loop workflow close <task-id> \
  --status done \
  --accept-risk "Manual review by Lars for urgent legacy hotfix."

--accept-risk writes .agent-loop/risks/<task-id>.accepted-risk.md.

Never use --accept-risk as a lazy bypass. It is for situations where the risk is real, understood, and explicitly owned by a named actor. If verification fails and the root cause is not understood, fix it first.

When Verification Fails

1. Read the [FAIL] output to understand what is missing.
2. Fix the underlying issue (re-plan/approve scope, run the missing command, record exact validation evidence, log outcomes, or resolve session state).
3. Record a checkpoint explaining the resolution:

   vendor/bin/agent-loop session checkpoint <task-id> --title "Verify fix" --body "..."
   

4. Re-run verify to confirm it passes.
5. If the failure is understood and fixing it is not feasible before the deadline, accept risk explicitly with --accept-risk and a named actor.

Close Is Not Durable Learning Approval

Closing a task with workflow close is not an approval of durable learning. Findings and learning candidates remain review inputs. Only reviewed decisions become durable guidance. See docs/workflow/learning-boundary.md.

Validation

• Blind-spot review report exists under .agent-recall/reviews/
• required validation evidence is passed and matches the current work-brief revision
• selected guidance has explicit, truthful recall outcomes
• an explicit session learning decision exists
• vendor/bin/agent-loop verify passes (or accepted risk is explicit and named)
• vendor/bin/agent-loop workflow report <task-id> shows no unaccepted scope or evidence gap
• vendor/bin/agent-loop workflow close <task-id> --status done succeeds

Skill Boundary

This skill owns:

• the review and close step of a governed agent-loop task in a consuming repo
• understanding the verify gate and its --strict mode
• understanding accepted-risk as an explicit, named, last-resort path
• knowing that close is not durable learning approval

This skill does not own:

• the task opening step (see agent-loop-task-start)
• recall compilation and L2 context (see agent-loop-l2-context)
• developing agent-loop itself

Example Triggers

• "Close this agent-loop task."
• "Run the review/verify gate."
• "Can I mark this done?"
• "Accept the risk and close."