retro-promote

1. Find what repeats

Run saw retro. It writes a proposal in docs/proposals/ for each check that failed in retro.repeat_threshold or more tasks (default 2). A failure seen once is not a pattern: leave it.

2. Root cause (one per proposal)

Read the failing logs in .saw/logs/ and pick one:

• missing context: the agent didn't know a fact that was written down somewhere
• poor instruction: a skill, playbook or AGENTS.md line was unclear or wrong
• missing tool: no command or check exists for this
• missing test: nothing caught it before review
• reasoning error: the facts were there and the agent got it wrong

Write it under Root cause, citing the log lines.

3. Fix, most enforceable first

1. test or lint rule, with an error message that says how to fix it
2. hook
3. skill or playbook edit
4. one AGENTS.md line (last resort: it costs tokens in every session)

4. Prove it

1. Add evals/regress-<check>/ that reproduces the failure (see evals/README.md).
2. saw eval --save-baseline before the fix, saw eval after. The fix lands only if no case gets worse.
3. Record both results under Eval case.

5. Approval

Under autonomy.actions.self_improve = "propose" (the default), show the user the proposal and wait for a yes. Never edit global files (edit_global) or commit without their go-ahead. When it merges, set status: promoted and promoted: <date>; the dashboard's promotion_effect panel then tracks whether the failure stopped. A rejected fix gets status: rejected and a reason.