Loop Retrospective

---
name: loop-retrospective
description: Parked, human-invoked skill. Use when a human explicitly asks for a retrospective that turns repeated agent-process friction into durable improvements. Converts repeated blocked categories, check failures, review iterations, or recurring follow-up topics drawn from event logs into skill updates, docs entries, or follow-up task records. Not part of the closeout gate.
metadata:
  area: process-improvement
  status: parked
  side_effects: writes-backend
  credentials: backend-dependent
  runs_scripts: none
---

# Loop retrospective

Loop retrospectives turn observed friction into durable process improvements. The output is never private memory; every retained lesson becomes reviewed text in the repo or a task record a later agent can discover.

This skill is **parked and optional**. It is not part of [[task-closeout]] and does not gate it. Run it only when a human explicitly asks for a retrospective, and only when there is enough recorded signal to work from. Its primary inputs are the local event logs and blocked/needs_context events – not task completion summaries, which are written for a different purpose. Until event logging matures into routine use, this skill will usually have nothing durable to act on; that is expected.

## Workflow

1. **Gather signals.** Read the local `.agenticloop/logs/<TASK-ID>.jsonl` event log entries (when event logging is enabled and files exist), blocked and needs_context events, recorded block categories, review-result events across rounds, and recurring follow-up task topics. Do not reconstruct friction from task completion summaries or copy raw agent exchanges into docs.
2. **Find recurring patterns.** Treat repeated check failures, repeated block categories, repeated review rounds, and repeated follow-up themes as candidates. Do not turn one-off noise into new process. One directly evidenced required-gate bypass, fabrication, or material misordering is the narrow serious-incident exception: create one durable improvement proposal without waiting for repetition.
3. **Choose one durable artifact per pattern.**
   - Update a skill when the pattern is a recurring agent behavior or rationalization.
   - Add a docs entry when the pattern is a durable out-of-scope decision, known limitation, or operational note.
   - File a follow-up task record when the pattern requires implementation or human decision.
4. **Keep provenance visible.** Cite grouping ids when relevant, task ids, artifact references, and command evidence. Do not promote interpretation into fact without evidence.

When the defect belongs to Agentic Loop rather than the target project, keep the
Project Operating Fact as the local current record and use the explicit,
human-confirmed export route:

```text
npx agenticloop improvement propose-toolkit-escalation --input <closed-facts.json> --toolkit-repository <identity> --output .agenticloop/tmp/<name>.toolkit-proposal.json --yes --json
```

The closed input contains only version, command, diagnostic codes, affected
toolkit surface, durable references, and a bounded reproduction outline. The
export redacts local paths/session ids and refuses open-ended fields such as raw
transcripts. It never writes to or authenticates the receiving repository; a
human transfers it and that repository recreates/imports it using its own
durable evidence. Do not invoke this route automatically for one-off warnings.

For the serious-incident exception, use `npx agenticloop improvement new` with
both the audit reference and the erroneous marker/carrier reference. Proposal
creation is collision-safe and proposal-only; it never changes the target surface.

## Anti-pattern

Wrong:

```text
Agents struggled with reviews again. Remember to be more careful next time.
```

Right:

```text
Three related tasks had missing final check output. Add a rationalization row to [[verification-evidence]] countering "the log is too long to paste", and cite the affected task ids in the closeout comment.
```

## Checklist

- Repeated patterns were separated from isolated task noise.
- Every retained pattern became one reviewed artifact: skill row, docs entry, or follow-up task record.
- Each artifact cites task-record, artifact, or command evidence.
- No private memory store or raw transcript dump was used.

## See also

- [[task-closeout]]
- [[blocked-state]]
- [[verification-evidence]]
- [[debugging-before-fixes]]