Improve while doing the work

When a concrete bump exposes reusable knowledge, preserve the evidence before continuing. Prefer completing the user's task; capture a short notice when a separate improvement would interrupt it. Do not create a skill for generic advice or a one-off fact. A reproducible failure can justify repair without waiting for recurrence.

Inspect the relevant existing skill or tool first. Choose:

• extend_skill: useful guidance missing from an applicable skill;
• refresh_skill: instructions conflict with verified current behavior;
• create_skill: a reusable workflow has no suitable home;
• repair_tool: executable behavior or an interface is wrong;
• create_tool or compose_tools: a demonstrated capability gap;
• update_behavior: an existing local behavioral instruction needs a scoped correction.

Use adaptation action notice with kind, path, title, rationale, evidence_indices, and acceptance. Evidence indices refer to the current episode's recorded evidence. Include expected behavior, a relevant consumer regression and a counterexample in the acceptance requirements. Record the corrected command or observed result, not only your explanation. Example: a skill selects the wrong interpreter; retain the failed invocation and a successful comparison using the verified interpreter.

Read opportunities to inspect notices and source freshness. A notice is an agent-authored hypothesis, not approval or verified effectiveness. Its observation template starts a separate episode scoped to the target. Freeze an executable suite, build isolation, edit the candidate, evaluate and inspect its diff. Requirements in a notice do not replace evaluator checks. Adding a source file requires explicit scope in that episode; capturing a notice never expands the current candidate's scope.

For skills, test applicability and wrong-trigger cases as well as the instructions' commands. Markdown substring tests establish wording only, not behavioral transfer. For tools, retain contract, dependency, original failure and consumer checks. New guidance remains a reviewed candidate until exact-candidate activation is requested. Use verified task-local findings immediately where appropriate, while keeping shared skill installation and the current task's loaded skill version distinct.