| name | learn |
| description | Capture one durable lesson from the completed workflow. Reads issue, PR, diff, review, and .context evidence; writes a concise learning file and hands naturally to /merge. |
/learn
Read the shared contracts before output: ../_shared/operator-output.md, ../_shared/attention-budget.md, ../_shared/evidence-record.md, ../_shared/state.md, ../_shared/cognitive-load.md, ../_shared/collaboration.md.
Purpose
Preserve the mechanism worth repeating or avoiding.
Inputs
Current PR, issue, commits, final diff, review comments, .context/ artifacts, or user-provided workflow summary.
Reads
Issue/PR bodies, comments, commits, diff, CI result, .context/ evidence, prior learnings when relevant.
Writes
docs/learnings/YYYY-MM-DD-<slug>.md- learning commit/push when operating on a PR branch
.context/learn.md.context/session-state.md
Process
- Compare intended contract with what shipped.
- Identify one non-obvious lesson about mechanism, incentives, boundary, failure mode, or process.
- Write the shortest durable learning that future agents can apply.
- Include an AGENTS/skill amendment candidate only when it generalizes.
- Record supporting evidence in
.context/learn.md.
- Update session state.
Operator Output
Return the learning file path, up to three evidence bullets, remaining risk, and next action. Suggest /merge when the PR is ready.
Stop Conditions
Stop if the PR/workflow evidence is too incomplete to produce an honest lesson, or if the learning would merely restate the diff.
