技能 Evals

${var} — Skill name to evaluate. If empty, evaluates all skills in evals.json.

Today is ${today}. Read memory/MEMORY.md for context.

This skill exists to catch quality regressions between runs — not just to re-state a snapshot of the latest output. The lede is what changed since last eval and what to do about it, not a flat pass/fail table.

Steps

1. Load inputs

• skills/skill-evals/evals.json — assertion manifest (read with jq; if parse fails, retry once after 5s, then exit SKILL_EVALS_ERROR).
• aeon.yml — registered skills, enabled flags, cron schedules.
• memory/cron-state.json — total_runs, success_rate, last_quality_score, last_failed, last_success per skill.
• memory/issues/INDEX.md — currently open issues (used to dedupe issue filing in step 5b).
• Most recent prior eval at articles/skill-evals-*.md (sorted descending, excluding today's). If none exist, mark prior_run as BOOTSTRAP — every result is NEW_*.

If evals.json is missing or has zero skills keys, run ./scripts/eval-audit --stubs to scaffold a starter spec and exit SKILL_EVALS_BOOTSTRAP with a notify telling the operator to commit the stub.

2. Run coverage audit (delegated)

Call ./scripts/eval-audit --json and parse:

• summary.coverage_pct, summary.covered, summary.uncovered_enabled, summary.uncovered_disabled
• uncovered_enabled[].skill and .inferred_pattern — these are the spec-gap candidates surfaced in the Action Queue.

Do not re-implement coverage detection in prose. If the script fails, fall back to the in-memory check (compare evals.json keys to aeon.yml enabled skills) and mark eval-audit=fail in the source-status footer.

3. Determine eval scope

• If ${var} is set → evaluate only that one skill (skip if not in evals.json and notify "skill-evals: ${var} has no spec — add an entry to evals.json").
• Otherwise evaluate every skill in evals.json.

4. For each skill in scope, run checks

a. Find latest output: glob output_pattern, sort descending by filename. If empty → status NO_OUTPUT, root_cause no_file_match, skip remaining checks.

b. Empty/stale: if file size is 0 bytes → FAIL, root_cause empty_file. If file mtime is older than 2× expected_cadence (derived from the skill's cron in aeon.yml; fall back to 14 days if cron is workflow_dispatch or unparseable) → STALE, root_cause stale_file. Stale outputs still run their assertions but are reported as STALE so dashboard noise is correct.

c. Word count: count words; fail if < min_words → root_cause word_count (record actual vs threshold).

d. Required patterns: for each pattern (pipe-separated alternatives), grep with -E. Missing → root_cause missing_pattern:<pattern>.

e. Forbidden patterns: any match → root_cause forbidden_pattern:<pattern>.

f. Numeric checks: for each entry, extract first regex match. Outside [min, max] → root_cause numeric_oob:<label>. If no match found and entry has skip_if_not_found: true, skip; otherwise WARN with root_cause numeric_missing:<label>.

g. Quality cross-check: read memory/skill-health/{skill}.json. If avg_score < 2.5 → status QUALITY_DEGRADED, root_cause quality_score:<avg>. If 2.5 ≤ avg_score < 3.5, record as note (no status change). If file missing, record quality=unknown.

Final status precedence: NO_OUTPUT > FAIL > STALE > QUALITY_DEGRADED > WARN > PASS.

5. Diff vs prior eval (the lede)

Parse the prior eval article's results table. For each skill produce one of:

• NEW_FAIL — was PASS/STALE/WARN, now FAIL/QUALITY_DEGRADED/NO_OUTPUT

• FIXED — was failing, now PASS

• STILL_FAIL — was failing, still failing (carry the issue ID forward)

• NEW_PASS — wasn't in prior (newly added to evals.json)

• NEW_NO_COVERAGE — covered prior, no eval entry now (rare; usually means evals.json edit)

• STABLE — same status both runs

  a. Issue filing. For every NEW_FAIL and NEW_QUALITY_DEGRADED:

  • Check memory/issues/INDEX.md — if an open issue already names this skill in the title, skip (avoid duplicates).
  • Else write memory/issues/ISS-{NNN}.md with frontmatter:

    ---
    id: ISS-{NNN}
    title: {skill}: {root_cause_short}
    status: open
    severity: {high if NEW_FAIL, medium if QUALITY_DEGRADED}
    category: {map root_cause: missing_pattern→prompt-bug, forbidden_pattern→prompt-bug, numeric_oob→quality-regression, word_count→quality-regression, stale_file→missing-secret-or-cron, empty_file→quality-regression, quality_score→quality-regression, no_file_match→missing-secret-or-cron}
    detected_by: skill-evals
    detected_at: {ISO timestamp}
    affected_skills: [{skill}]
    root_cause: {full root_cause string}
    ---
    
    {one-paragraph context with file path, expected vs actual, link to article}
    

  • {NNN} = next free 3-digit ID (scan memory/issues/ISS-*.md, take max + 1, zero-pad).
  • Append a row to memory/issues/INDEX.md Open table.

  b. Issue closing. For every FIXED: scan memory/issues/ISS-*.md for an open issue whose affected_skills contains this skill and detected_by: skill-evals; flip status: resolved, set resolved_at, move row from Open → Resolved table in INDEX.md. (Don't touch issues filed by other detectors.)

6. Compute verdict

One-line verdict, picked by precedence:

1. SKILL_EVALS_REGRESSED — any NEW_FAIL exists
2. SKILL_EVALS_QUALITY_DROP — any NEW_QUALITY_DEGRADED (no NEW_FAIL)
3. SKILL_EVALS_RECOVERED — FIXED ≥ 1 and zero new failures
4. SKILL_EVALS_COVERAGE_CLIFF — coverage_pct dropped ≥ 10 points vs last run, or absolute coverage_pct < 25 and last run was ≥ 25
5. SKILL_EVALS_OK — all stable, all green

7. Build the Action Queue

A short, ordered, concrete checklist at the top of the article. Cap at 8 items. Each item is one line, naming a specific skill and a specific next step:

• Patch (regex/threshold tweaks): Patch evals.json:{skill} — {root_cause}
• Investigate (FAIL with no obvious fix): Investigate {skill} — {root_cause} (ISS-{NNN})
• Re-run (NO_OUTPUT, no recent dispatch): Dispatch {skill} — no output in {N} days
• Add spec (uncovered enabled): Add evals.json entry for {skill} — pattern: {inferred_pattern} (one line per uncovered enabled skill, max 5; if more, summarize "+N more — see Coverage Gaps")

If the queue is empty, write Action Queue: none — all green.

8. Write the article

Path: articles/skill-evals-${today}.md. Skeleton:

# Skill Evals — ${today}

**Verdict:** {VERDICT}
**Coverage:** {covered}/{enabled_total} ({coverage_pct}%) {↑↓ vs prior or "(first run)"}
**Diff:** {N_NEW_FAIL} new fail · {N_FIXED} fixed · {N_STILL_FAIL} still failing · {N_STABLE} stable

## Action Queue
1. ...
2. ...

## Regressions (NEW_FAIL + NEW_QUALITY_DEGRADED)
| Skill | Status | Root cause | Issue |
|-------|--------|------------|-------|
| ... | NEW_FAIL | missing_pattern:stars | ISS-014 |

## Recovered (FIXED)
| Skill | Was | Now |
|-------|-----|-----|

## Still Failing
| Skill | Status | Root cause | Issue | Failing since |
|-------|--------|------------|-------|---------------|

## Full Results
| Skill | Status | Diff | Root cause | Quality | Words | Last output |
|-------|--------|------|------------|---------|-------|-------------|

## Coverage Gaps (enabled in aeon.yml, missing from evals.json)
- {skill} — inferred pattern: `{inferred_pattern}`

## Sources
- evals.json={ok|fail} · cron-state={ok|fail} · skill-health={ok|empty|fail} · eval-audit={ok|fail} · prior-article={ok|none}

Omit empty sections (no Recovered section if zero FIXED, etc.). Keep the Coverage Gaps section bounded to 10 lines max — overflow into a +N more summary.

9. Notify (gated)

Only call ./notify when one of the following holds:

• Verdict is SKILL_EVALS_REGRESSED, SKILL_EVALS_QUALITY_DROP, or SKILL_EVALS_COVERAGE_CLIFF
• Verdict is SKILL_EVALS_RECOVERED (good news worth a ping)

Stay silent on SKILL_EVALS_OK (still write the article + log entry; just don't ping). This trains the operator that a notification means action is needed.

Notify body (concise, soul-voice):

*Skill Evals — {VERDICT}*
{N_NEW_FAIL} new fail · {N_FIXED} fixed · coverage {coverage_pct}%
Top action: {action_queue[0]}
Article: articles/skill-evals-${today}.md

If N_NEW_FAIL > 0, append the first 3 regressions as {skill}: {root_cause} lines.

10. Log

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

### skill-evals
- Verdict: {VERDICT}
- Diff: {N_NEW_FAIL} new fail / {N_FIXED} fixed / {N_STILL_FAIL} still failing / {N_STABLE} stable
- Coverage: {covered}/{enabled_total} ({coverage_pct}%)
- Issues filed: [list ISS-IDs]
- Issues closed: [list ISS-IDs]
- Action queue head: {action_queue[0] or "none"}

Sandbox note

All inputs are local files (evals.json, aeon.yml, memory/*, articles/*, scripts/eval-audit). No outbound HTTP — no fallback needed. ./scripts/eval-audit is a local bash script and uses jq; if jq is missing (rare on GH Actions ubuntu runners), the script will exit non-zero — mark eval-audit=fail in the source-status footer and continue with the in-memory coverage check.

Constraints

• Never overwrite a prior issue file. Always allocate a fresh ISS-{NNN} number.
• Never close an issue this skill didn't file (only detected_by: skill-evals issues are closeable here).
• Don't notify when verdict is SKILL_EVALS_OK — silence is the correct signal on a green week.
• Preserve the assertion schema (output_pattern, min_words, required_patterns, forbidden_patterns, numeric_checks) — additions allowed (skip_if_not_found, expected_cadence), removals are breaking.
• Cap Coverage Gaps and Action Queue sections to keep the article scannable; the article is read by humans, not just machines.