Agent审计

${var} — Target repo to scan forks of (e.g. "owner/aeon"). If empty, reads memory/watched-repos.md and uses the first entry.

Today is ${today}. Generate a weekly leaderboard of which Aeon skills the configured fleet is actually running — and what upstream should do about it.

Why this version

Upstream aeon.yml ships with effectively only heartbeat: enabled: true. Every fresh fork inherits that default. If we score "skills enabled" across all active forks, the leaderboard is a tautology (heartbeat always wins, everything else hovers near 0) and operator learns nothing. The lever is to score against configured forks only — forks whose aeon.yml diverges from upstream defaults — and to convert the result into three actionable recommendations: which skills to promote, which fleet patterns to copy upstream, which skills to sunset.

Steps

1. Determine the target repo

If ${var} is set, use that as TARGET_REPO. Otherwise read memory/watched-repos.md and use the first non-comment, non-empty line. If neither yields a value, log SKILL_LEADERBOARD_NO_TARGET to memory/logs/${today}.md and stop (no notification).

2. Snapshot upstream defaults

Read this running instance's local aeon.yml once. Build UPSTREAM_DEFAULTS: a dict {skill_name -> {enabled, model_or_null, var_or_empty, schedule_or_null}} covering every skill entry under the skills: block. Also build UPSTREAM_SKILLS: the set of skill directory names from skills/ (use ls skills/). These are the comparison baselines.

3. Fetch active forks

CUTOFF=$(date -u -d "30 days ago" +%Y-%m-%dT%H:%M:%SZ 2>/dev/null || date -u -v-30d +%Y-%m-%dT%H:%M:%SZ)
gh api "repos/${TARGET_REPO}/forks?per_page=100" --paginate \
  --jq "[.[] | select(.pushed_at > \"$CUTOFF\") | {owner: .owner.login, full_name: .full_name, pushed_at, stargazers_count, created_at}]"

If zero active forks: log SKILL_LEADERBOARD_NO_FORKS and stop (no notification).

4. Per-fork single-call enumeration

For each active fork, run one recursive git-tree call to enumerate files (cheaper than per-path contents):

gh api "repos/${FORK_FULL}/git/trees/HEAD?recursive=1" --jq '[.tree[] | select(.type == "blob") | .path]'

Handle errors:

• 404 / 409 (empty repo): mark status: "no_tree", skip aeon.yml and skills/ extraction, continue.
• 403 with X-RateLimit-Remaining: 0: sleep 60s, retry once. If still failing, mark status: "rate_limited" and continue with partial fleet.

Then fetch the fork's aeon.yml only if the tree contains it:

gh api "repos/${FORK_FULL}/contents/aeon.yml" --jq '.content' | base64 -d

If the path is in the tree but the contents call 404s, mark status: "yml_unreadable" and continue.

Extract from each readable aeon.yml:

• For every skill entry under skills: — enabled, model (if set), var (if set), schedule (if differs from any upstream default for that skill).
• Detect "fork-only skills": directory names under skills/ in the fork's tree that are NOT in UPSTREAM_SKILLS.

5. Classify each fork

For each fork, compute a divergence signal vector vs UPSTREAM_DEFAULTS:

• enabled_diff: count of skills where the fork's enabled value differs from upstream default
• var_set: count of skills with var: set to a non-empty value where upstream's was empty
• model_override: count of skills with a model: value differing from upstream
• schedule_override: count of skills with a schedule: value differing from upstream
• fork_only_skills: count from step 4

Tier the fork:

• CONFIGURED: any of the above is ≥1. (i.e., the fork actively diverged from defaults)
• TEMPLATE: aeon.yml is readable but every diff signal is 0. Untouched template — exclude from leaderboard math.
• UNREADABLE: no_tree / no aeon.yml / yml_unreadable / rate_limited. Tracked in the source-status footer.

6. Aggregate against the CONFIGURED denominator

Let N_CONFIGURED = count of forks tiered CONFIGURED. If N_CONFIGURED < 2: log SKILL_LEADERBOARD_TEMPLATE_FLEET with the active/template/unreadable counts, write a stub article noting the conversion rate, and skip the notification (no signal worth pushing). Stop.

For each skill name (union of upstream skills and fork-only skills) compute:

• forks_enabled: number of CONFIGURED forks where it's enabled: true
• pct_of_configured: forks_enabled / N_CONFIGURED
• with_var: count of CONFIGURED forks that override var:
• with_model: count that override model:
• with_schedule: count that override schedule:
• customization_depth per-fork-instance: enabled (1) + var (1) + model (1) + schedule (1) → sum across forks; this becomes the tiebreaker
• is_fork_only: true if the skill name is in some fork's tree but not in UPSTREAM_SKILLS

Rank by (forks_enabled desc, customization_depth desc, name asc).

7. Load prior snapshot (week-over-week)

Read memory/topics/skill-leaderboard-state.json if it exists. Schema:

{
  "last_run": "YYYY-MM-DD",
  "n_active_forks": N,
  "n_configured": N,
  "ranking": [{"skill": "name", "forks_enabled": N, "rank": N}, ...]
}

If the file exists and last_run is within the last 14 days, compute:

• Rising: skills that moved up ≥3 ranks
• Falling: skills that moved down ≥3 ranks
• New entries: skills now ranked that weren't last run
• Dropouts: skills last run that aren't ranked now (forks_enabled went to 0)

If the file is missing or stale (>14 days), set deltas to "first ranked snapshot — no comparison".

8. Compute the three actionable categories

Every tier below is a heuristic — operator overrides take precedence. Thresholds are starting points, not hard rules; the operator's call always wins. When in doubt, classify as Match rather than forcing Promote or Sunset.

• Consensus skills: pct_of_configured > 0.50 (heuristic — operator overrides take precedence). The fleet has converged on these — upstream should treat them as canonical examples and ensure they're well-documented.
• Promote candidates: pct_of_configured ≥ 0.25 AND upstream default is enabled: false AND the skill is not a workflow_dispatch-only skill (heuristic — operator overrides take precedence). The fleet found these worth running; upstream may want to flip the default or feature them more prominently.
• Match candidates: skills where ≥2 CONFIGURED forks override model: to the same value (e.g., both pick claude-sonnet-4-6) (heuristic — operator overrides take precedence). The fleet has independently found a cheaper model sufficient — upstream should consider matching the override.
• Sunset candidates: skills present in UPSTREAM_SKILLS with forks_enabled == 0 AND with_var == 0 AND not a meta/dev skill (heuristic — operator overrides take precedence; skip skills tagged meta or dev — those are operator-tools, fork adoption isn't the point). Review for removal or better discoverability.
• Fleet-only skills: any is_fork_only: true skill enabled in ≥1 fork. Surface for review — the fleet built something upstream doesn't have.

9. Write the article

To articles/skill-leaderboard-${today}.md:

# Skill Leaderboard — ${today}

**Verdict:** ${one-line verdict — see step 10 format below}

*Scanned ${N_ACTIVE} active forks of ${TARGET_REPO} (pushed in last 30 days). ${N_CONFIGURED} are configured (aeon.yml diverges from upstream defaults). Leaderboard scored against the configured ${N_CONFIGURED}.*

## Top Skills (configured fleet)

| Rank | Skill | Forks | % Configured | var | model | sched | Δ vs last week |
|------|-------|-------|--------------|-----|-------|-------|----------------|
| 1 | name | N | XX% | N | N | N | — / ↑N / ↓N / NEW |
| ... |

(Top 15. If <15 ranked, list all.)

## What the fleet is telling us

### Promote
${list of Promote candidates with one-line "why" each, OR "none this week"}

### Match
${list of Match candidates: "skill X — N forks override model to claude-sonnet-4-6", OR "none this week"}

### Sunset (review for removal or better docs)
${list of Sunset candidates, capped at 5, OR "none — every shipped skill has at least one configured-fork enable"}

### Fleet-only skills
${list of fork-only skill names with the fork that built each, OR "none this week"}

## Week-over-week

${"First ranked snapshot — no comparison" OR list of Rising / Falling / New / Dropouts}

## Fleet composition

| Tier | Count | % |
|------|-------|---|
| Configured | N_CONFIGURED | XX% |
| Template (untouched aeon.yml) | N_TEMPLATE | XX% |
| Unreadable (no tree / no yml / rate-limited) | N_UNREADABLE | XX% |
| **Total active forks** | N_ACTIVE | 100% |

## Source status

- Trees fetched: N_TREES_OK / N_ACTIVE
- aeon.yml readable: (N_CONFIGURED + N_TEMPLATE) / N_ACTIVE
- Rate-limited: N_RATE_LIMITED
- Fork-only skill files inspected: N_FORK_ONLY_FILES

---
*Source: GitHub API — forks of ${TARGET_REPO}. Methodology: a fork counts as "configured" if its `aeon.yml` differs from upstream defaults on `enabled`, `model`, `var`, or `schedule` for any skill. Untouched templates are excluded from leaderboard math.*

10. Build the verdict line

Pick the strongest single claim, in this priority:

1. If a Promote candidate exists with pct_of_configured ≥ 0.40: "${N_CONFIGURED} configured forks; ${skill} hit ${pct}% — promote candidate"
2. Else if any Rising skill moved ≥5 ranks: "${skill} jumped from rank ${old} to rank ${new} this week"
3. Else if a Fleet-only skill exists: "${fork_owner}/aeon shipped ${skill} — not in upstream"
4. Else if any Match candidate exists: "${N} forks independently override ${skill} to ${model} — consider matching"
5. Else: "Configured-fleet conversion rate: ${N_CONFIGURED}/${N_ACTIVE} (${pct}%); top: ${skill} (${N} forks)"

11. Send notification

Via ./notify:

*Skill Leaderboard — ${today}*
${verdict_line}

Top 5 across ${N_CONFIGURED} configured forks (of ${N_ACTIVE} active):
1. ${skill} — N forks (XX%) ${rising_arrow_or_blank}
2. ${skill} — N forks (XX%) ${rising_arrow_or_blank}
3. ${skill} — N forks (XX%) ${rising_arrow_or_blank}
4. ${skill} — N forks (XX%) ${rising_arrow_or_blank}
5. ${skill} — N forks (XX%) ${rising_arrow_or_blank}

${one of: "Promote: ${skill} (XX% adoption)" | "Match: ${N} forks override ${skill} → ${model}" | "Fleet-only: ${owner}/${skill}" | omit if none of the above}

Full report: https://github.com/${GITHUB_REPOSITORY}/blob/main/articles/skill-leaderboard-${today}.md

Use the $GITHUB_REPOSITORY env var (GitHub Actions sets it to owner/repo) to build the URL — NOT the watched repo. The article lives in this running instance's repo.

Notification is sent only when N_CONFIGURED >= 2 (gated in step 6). Otherwise the run is silent.

12. Persist the snapshot

Write memory/topics/skill-leaderboard-state.json:

{
  "last_run": "${today}",
  "target_repo": "${TARGET_REPO}",
  "n_active_forks": N_ACTIVE,
  "n_configured": N_CONFIGURED,
  "n_template": N_TEMPLATE,
  "n_unreadable": N_UNREADABLE,
  "ranking": [
    {"skill": "name", "forks_enabled": N, "pct_of_configured": 0.NN, "rank": N, "customization_depth": N, "is_fork_only": false}
  ]
}

Overwrite each run. This is the source for next week's deltas — do not depend on parsing the prior article (format may shift, the JSON is the contract).

13. Log

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

## Skill Leaderboard
- **Active forks scanned:** N (of M total)
- **Configured forks:** N (XX% conversion rate)
- **Template forks:** N
- **Unreadable forks:** N
- **Top skill:** ${skill} (N forks, XX%)
- **Verdict:** ${verdict_line}
- **Promote/Match/Sunset/Fleet-only:** counts
- **Notification sent:** yes/no
- **Status:** SKILL_LEADERBOARD_OK | SKILL_LEADERBOARD_TEMPLATE_FLEET | SKILL_LEADERBOARD_NO_FORKS | SKILL_LEADERBOARD_NO_TARGET

Sandbox note

All GitHub API calls use gh api which handles auth internally — no env-var expansion in headers needed. If gh api returns 403 with X-RateLimit-Remaining: 0, back off 60s and retry once; on continued failure, record status: "rate_limited" for that fork and proceed with partial fleet (the verdict and source-status footer surface the gap). No new env vars or secrets required beyond the default GITHUB_TOKEN.

Constraints

• Never send the notification if N_CONFIGURED < 2 — the leaderboard is meaningless without a configured denominator and trains the operator to ignore.
• Never count heartbeat enabled-counts as signal in the verdict (every CONFIGURED fork inherits it from upstream default; it's a tautology). It can still appear in the table.
• Do not parse last week's article for week-over-week — use memory/topics/skill-leaderboard-state.json only.
• Skills tagged meta or dev are excluded from the Sunset list (operator-tools, fork adoption is not the success metric).
• Skills with schedule: "workflow_dispatch" are excluded from Promote (on-demand by design — adoption % is misleading).