Loop Governance And Learning

Use this after a meaningful loop finishes: a bug is fixed, a review changes the direction, a workflow succeeds, an incident exposes a gap, or a repeated failure reveals a pattern. The goal is to convert durable learning into the right artifact so the same lesson does not stay trapped in a transcript.

Not every lesson deserves persistence. Capture reusable knowledge; discard temporary task state.

First Classify The Learning

Name the lesson in one sentence, then classify it:

If the lesson is not reusable, do not persist it. Stale memory is worse than no memory because future agents treat it as fact.

Governance Loop

1. Extract the lesson. Separate the durable pattern from incidental task details, credentials, private names, timestamps, branch names, and one-off state.
2. Choose the substrate. Decide whether the lesson belongs in a test, skill, doc, issue, decision log, monitor, or no artifact.
3. Prefer strengthening an existing artifact. Patch an existing skill, test, or doc when the new lesson refines a known workflow. Create a new artifact only when the gap is recurring, named, and not covered elsewhere.
4. Attach evidence. Link the artifact to source-backed evidence: failing test, reproduction, review comment, incident note, trace, or decision record.
5. Verify the artifact. Run the check, validate the skill, read the rendered doc, or confirm the issue/decision log captures the right next action.
6. Report the change. State what was learned, what artifact changed, why that substrate was chosen, and what was intentionally not persisted.

When To Update A Skill

Use [[skill-curation]] before creating or changing a skill. A skill update is appropriate when:

• the behavior is reusable across more than one task or repository;
• the procedure is non-obvious enough that future agents would otherwise rediscover it;
• the trigger can be named cleanly in the frontmatter description;
• the skill can be public-safe or clearly scoped to its intended audience; and
• the update does not duplicate a sharper existing skill.

Do not create a skill for:

• a single task's progress;
• an environment-specific path, person, host, credential, or queue;
• a vague slogan with no operational steps;
• a temporary workaround; or
• knowledge that belongs in tests or source docs instead.

When To Update Tests Or Evals

Prefer a test, eval, or monitor when the lesson is about observable behavior. Use [[regression-ratchet]] for bugs and [[feature-coverage-not-just-line-coverage]] for user or system promises.

The check should fail on the bad behavior and turn green after the fix. If it cannot fail independently, it is not evidence; it is a restatement of the implementation.

When To Update History Or Docs

Use [[source-history-decision-log]] when the lesson is a rationale: why a shape was chosen, why an alternative was rejected, or what prior failure a line of code protects against.

Use docs when the lesson explains how to operate, configure, or understand a system. Keep docs close to the source they explain, and remove stale statements instead of layering contradictions.

Stop Conditions

Stop with a durable-learning change when:

• the reusable lesson is captured in the right artifact;
• the artifact is verified at the cheapest meaningful layer;
• sensitive or task-local details were removed; and
• the report distinguishes durable learning from temporary state.

Stop without changing durable artifacts when:

• an existing artifact already covers the lesson;
• the lesson is only task progress;
• the evidence is too weak to generalize;
• the user asked not to persist it; or
• the proposed artifact would leak private context.

Output Shape

Lesson: <durable pattern learned>
Artifact: <test, skill, doc, issue, decision log, monitor, or none>
Why this substrate: <short rationale>
Verification: <how the artifact was checked>
Not persisted: <task-local or unsafe details intentionally left out>

Anti-Patterns

• Turning every completed task into memory.
• Storing branch names, PR numbers, local paths, or timestamps as durable facts.
• Creating a new skill when an existing skill only needed one sharper sentence.
• Writing a lesson without evidence from the loop that produced it.
• Capturing private context in public artifacts.
• Treating a transcript summary as a substitute for a test, doc, or decision log.