技能 修复

${var} — Skill name to repair. If empty, runs systemic triage and picks the worst fixable target. ${var} modifiers: prefix dry-run: to diagnose only without writing a PR (e.g. dry-run:digest).

Today is ${today}. Your task is to diagnose and repair the worst-impact failing or degraded skill — preferring a single shared fix over N per-skill patches when failures cluster.

Phases

PREFLIGHT → TRIAGE → DIAGNOSE → REPAIR → VERIFY → LOG

Stop early at the appropriate exit code if any phase finds nothing actionable.

Exit taxonomy

Pick exactly one before notifying.

1. PREFLIGHT

Bail early with REPAIR_BLOCKED (and notify with the reason) if any of these fails:

• gh auth status succeeds.
• memory/cron-state.json exists and parses as JSON.
• memory/issues/INDEX.md exists. If absent, bootstrap a minimal one (Open + Resolved tables, no rows).
• memory/state/skill-repair-history.json exists. If absent, create {}.

Cooldown / idempotency (skip target with REPAIR_BLOCKED if any matches; don't loop on a fix that didn't take):

• The chosen target appears in memory/state/skill-repair-history.json with last_repair_at within 24h. (Operator can override by deleting the entry.)
• An open PR already exists matching fix/skill-repair-{name}-* — gh pr list --state open --search "head:fix/skill-repair-{name}".
• More than 3 skill-repair PRs already opened in the current UTC day — rate-limit our own PRs.

If ${var} starts with dry-run:, strip the prefix to get the target name and skip the cooldown.

2. TRIAGE

Identify the target. Two paths:

Path A — ${var} set explicitly: repair that skill. Skip step 2's clustering.

Path B — ${var} empty (auto-select):

1. Read memory/issues/INDEX.md. Extract open issues. Skip permanent-limitation.
2. Read memory/cron-state.json. Compute candidates where any of:
   • consecutive_failures >= 2, OR
   • success_rate < 0.5 AND total_runs >= 3, OR
   • last_status == "failed" AND last_failed within 48h, OR
   • last_quality_score <= 2 (degraded output even when "successful").
3. Cluster by error signature. Group candidates by normalized last_error (lowercase, strip timestamps/ids/digits) AND by issue category. If 2+ skills share a signature OR a non-trivial category (api-change, rate-limit, missing-secret, sandbox-limitation):
   • This is systemic. Switch to systemic mode:
     • File or update a single shared issue (affected_skills: [list]) instead of N per-skill issues.
     • If the shared root cause is fixable in one place (e.g., a shared script under scripts/, a CLAUDE.md pattern, a shared config), open one PR addressing that. Otherwise emit REPAIR_DIAGNOSED_NO_FIX with the systemic finding.
     • Exit with REPAIR_OK_SYSTEMIC after step 5.
4. Pick worst single target. Sort: critical issue > high issue > consecutive_failures desc > lowest success_rate > stalest last_success. Skip permanent-limitation and any target whose preflight cooldown blocks it. If nothing remains: REPAIR_NO_TARGETS.

3. DIAGNOSE

Build a diagnostic dossier for the target before touching any file. Sources are independent — each one's status feeds the source-status footer (ok/empty/fail).

a. Skill file: read skills/{name}/SKILL.md. Note frontmatter, declared data sources, env-var references.

b. Cron-state entry: extract last_error, last_failed, last_success, success_rate, consecutive_failures, last_quality_score.

c. Regression hunter: if last_success exists, run

git log --oneline --since="$LAST_SUCCESS" -- skills/{name}/SKILL.md aeon.yml scripts/

Any commit listed is a candidate regression source. If exactly one commit touched the skill file in this window, it is the prime suspect — record its SHA + subject in the dossier.

d. Recent failed runs (last 5, not just 1):

gh run list --workflow=aeon.yml --limit 50 --json databaseId,name,conclusion,createdAt \
  | jq -r '[.[] | select(.name | contains("{name}")) | select(.conclusion=="failure")] | .[0:5]'

For each, prefer gh run view "$RUN_ID" --log-failed (already filtered to failed steps) over the full log; fall back to gh run view "$RUN_ID" --log only if --log-failed returns nothing. Then:

gh api "repos/{owner}/{repo}/actions/runs/$RUN_ID/check-runs" \
  | jq -r '.check_runs[].output.annotations[]? | "\(.path):\(.start_line) \(.annotation_level): \(.message)"'

Annotations give clean error rows; logs give context. Distinguish consistent (same signature 4-5/5 runs → likely deterministic bug, secret, API change) from intermittent (1-2/5 → rate limit, flaky upstream).

e. Logs: search last 3 days of memory/logs/*.md for {name} mentions. Surface any prior diagnoses.

f. Quality history: if memory/skill-health/{name}.json exists, note avg_score trend.

g. Output expectations: if skills/skill-evals/evals.json has an entry for {name}, extract its min_words, required_patterns, forbidden_patterns. A passing run that fails these is quality-regression.

h. Issue: if memory/issues/INDEX.md lists an open issue for this skill, read the file — its category and root_cause short-circuit the playbook lookup below.

4. REPAIR — per-category playbook

Categories follow CLAUDE.md. Pick the most specific category that fits the diagnostic dossier (issue category if present > error-signature pattern match > best inference). Apply the matching playbook.

Risk classification (pick one, gate the PR):

• LOW — clarifying prompt, adding fallback, comment-only changes, single-section edit (< 30 lines diff).
• MED — changes a data source, adds a new env-var reference (must already be in workflow), or modifies output format.
• HIGH — touches aeon.yml, removes existing features, disables a skill, modifies a scripts/*.sh file. HIGH risk PRs must add the label manual-review and must NOT be auto-mergeable (skip auto-merge-friendly framing in the PR body).

Frontmatter integrity check: after editing skills/{name}/SKILL.md, re-read it. Confirm the YAML frontmatter still has name, description, var, tags. If broken, abort the edit and exit REPAIR_BLOCKED.

5. VERIFY — append a verification plan to the PR

Every PR (except REPAIR_DIAGNOSED_NO_FIX) must include a Verification section the operator can execute. Use this template:

## Verification

**Manual trigger:** [Run skill](https://github.com/{owner}/{repo}/actions/workflows/aeon.yml) with `skill={name}` and `var={var}`.

**Expected result:**
- Workflow conclusion: `success`
- Output file matches `{evals.json output_pattern or "memory/logs/${today}.md mentions {name}"}`
- {category-specific signal — e.g. "no `rate limit` strings in run logs" / "produces ≥ {min_words} words" / "annotation count ≤ 0"}

**If still failing after this PR:** delete `memory/state/skill-repair-history.json[{name}]` to remove the cooldown, then re-dispatch `skill-repair` with `var={name}` for a second pass.

Record the chosen verification command in the issue file's ## Repair Attempt section so the next skill-repair run can read prior outcomes.

6. Branch, commit, PR

TODAY="${today}"
BRANCH="fix/skill-repair-{name}-${TODAY}"
git checkout -b "$BRANCH"
git add skills/{name}/SKILL.md  # plus aeon.yml or scripts/* iff in playbook
git commit -m "fix({name}): [one-line root cause → fix]"
git push -u origin "$BRANCH"

gh pr create --title "fix({name}): [short]" --body "$(cat <<'EOF'
## Symptom
[what failed — error signature, run URL]

## Diagnosis
[dossier summary: regression commit if any, consistent vs intermittent, category]

## Root cause
[one paragraph]

## Fix
[what changed and why]

## Risk
LOW | MED | HIGH — [rationale]

## Verification
[copy from step 5]

## Source status
cron_state=ok | issues_index=ok | gh_runs=ok | gh_logs=ok | git_log=ok | check_runs=ok
EOF
)"

If risk is HIGH, also: gh pr edit "$PR_URL" --add-label manual-review.

7. Update issue tracker (memory/issues/)

• If an open issue for this skill exists:
  • Fix applied → set status: resolved, resolved_at: ${today}, fix_pr: <url>. Move row from Open → Resolved in INDEX.md.
  • No fix possible → append ## Repair Attempt — ${today} with the dossier and reason.
• If no issue exists but a real problem was found and fixed → create memory/issues/ISS-{NNN}.md with status already resolved (NNN = next free number from INDEX.md).
• If systemic clustering fired in step 2 → ensure affected_skills: lists every skill matched by the signature.

8. Persist cooldown

Update memory/state/skill-repair-history.json:

{
  "{name}": {
    "last_repair_at": "${today}T...Z",
    "exit_code": "REPAIR_OK_FIXED",
    "fix_pr": "https://github.com/.../pull/N",
    "issue": "ISS-NNN"
  }
}

9. Notify

Send via ./notify (one-paragraph max — verdict line first):

*skill-repair — {EXIT_CODE}*
Target: {name} (or systemic: skill-a, skill-b, ...)
Root cause: [one line]
Fix: [one line] (risk: LOW|MED|HIGH)
PR: {url}  Issue: {ISS-NNN}
Verify: workflow_dispatch skill={name}

10. Log

Append to memory/logs/${today}.md:

### skill-repair
- Exit: {EXIT_CODE}
- Target: {name} (or systemic group)
- Category: {category}
- Diagnosis: [root cause]
- Fix: [what changed] (risk: {LOW|MED|HIGH})
- Regression suspect: {commit SHA or "none in window"}
- Failures observed: {N}/5 recent runs ({consistent|intermittent})
- PR: {url or "—"}
- Issue: {ISS-NNN created|updated|resolved or "—"}
- Source status: cron_state | issues_index | gh_runs | gh_logs | git_log | check_runs

Sandbox note

gh and git work inside the sandbox. The diagnostic curls go through gh api (auth handled). For any external API spec lookup in the api-change playbook, prefer WebFetch over curl — see CLAUDE.md.

Constraints

• One target per run (or one systemic cluster). Never bundle unrelated repairs.
• Minimum-edit principle: keep diffs as small as possible. The original failure mode is rarely "the skill needs a rewrite".
• Never modify secrets, the workflow file (.github/workflows/aeon.yml), or messages.yml.
• Never push to main. Always branch + PR.
• Never auto-merge HIGH-risk PRs. They carry the manual-review label.
• If a skill has been failing > 7 days with no clear root cause and the category is unknown, recommend (in the issue and notify) enabled: false in aeon.yml — but do not apply that change without an explicit operator-approved issue.
• Skip when ${var} matches a skill that has been repaired in the last 24h unless operator clears the cooldown entry. This prevents repair loops on fixes that didn't take.