Reflect

Reflect

Use only the current task, repository artifacts, and task or memory records the user has placed in scope. Do not scan guessed transcript directories or unrelated conversations.

1. Reconstruct the task timeline from requests, decisions, tool use, failures, corrections, and verification.
2. Separate one-off facts from reusable agent behavior. A lesson earns a skill change only when it changes a future decision and cannot be enforced more reliably in code, tests, or configuration.
3. For a complex task, use independent deep, general, and fast review lenses for judgment, tooling, and divergent interpretation. Give them the same evidence and no proposed fix.
4. Accept, reject, or backlog each suggestion with a concrete reason. Prefer editing an existing skill over adding a new one.
5. When the user authorized updates, use $skill-creator, validate the affected skill, and report the exact behavior changed. Otherwise return a reviewable proposal without modifying files.

Do not update Codex memory unless the user explicitly asks for a memory update.