skill-retro – skill Khác cho agent AI

skill-retro — grade the agent setup from real sessions

> Privacy is the contract. Everything runs locally. Transcripts are condensed,

> secret-redacted, chmod-0600, and never uploaded — the only shareable artifact is

> the report the user chooses to share.

Run from the repo being graded. Every artifact goes to one fresh scratch dir, never

into the repo:

RUN="$(mktemp -d "${TMPDIR:-/tmp}/skill-retro-XXXXXXXX")"
python scripts/collect_sessions.py --out "$RUN"          # 1 — harvest + redact

1 — Collect. Scans Claude Code project-history JSONL and Codex rollouts,

discovers repo skills (.claude/skills, .agents/skills, .codex/skills, plugin

layouts), detects skill usage (Skill invocations, slash commands, SKILL.md paths), samples

newest-first, and writes redacted transcripts. Read $RUN/inventory.json: if

sessions_sampled is 0, tell the user there is nothing recent to score (suggest

--days 90 or --repo) and stop. skills_found 0 is fine — the report becomes a

case for creating skills. To grade headless runs (for example skill evals saved as

stream-json), add --sessions-dir <dir>.

2 — Score. `python scripts/score_aggregator.py --inventory "$RUN/inventory.json"

--emit-template > "$RUN/session_scores.json"`. Read each transcript in

$RUN/transcripts/ and judge it against both rubrics — scorers/efficiency.md

and scorers/code-quality.md. Fill the template with a **label from the rubric's

table** and a 1–3 sentence reason citing transcript specifics. Never invent numeric

scores — the aggregator derives them from labels. Use insufficient_evidence when

a transcript shows no judgeable diff. Also write 1–5 top_findings: the most

impactful cross-session patterns, concrete and specific.

3 — Draft edits. Follow references/skill_edit_governance.md (the filing bar:

would a competent agent with the current instructions still fail this way?). For

each suggestion that clears it, write the full improved SKILL.md to

$RUN/proposed/<skill>/SKILL.md, produce diff -u <current> <proposed>, and record

it in $RUN/suggestions.json citing the sampled session id(s) that motivated it.

Zero suggestions is a valid success — say why per finding. Never modify the user's

real skill files in this step.

4 — Aggregate (the gate). `python scripts/score_aggregator.py --inventory

"$RUN/inventory.json" --scores "$RUN/session_scores.json" --suggestions

"$RUN/suggestions.json"`. It validates labels against the rubric tables, refuses

scores for unsampled sessions, requires substantive reasons, rejects suggestions

that cite no scored session, computes `overall = 0.5·efficiency +

0.35·code_quality + 0.15·skill_coverage, and writes report.json`. **Exit 4 is a

stop**: fix what it names and re-run; never hand-edit report.json around it.

5 — Render + tell. python scripts/render_report.py --report "$RUN/report.json"

→ one self-contained report.html (no JS, no CDN, dark-mode + print-to-PDF). Then

tell the user the grade and the top findings in text, link

file://$RUN/report.html, and ask whether to apply the proposed diffs to their

real skills — apply only on an explicit yes, skill by skill.

Hard rules

1. Never upload transcripts, session files, or any excerpt. Local only.

2. Labels only, from the rubric tables. The aggregator owns all arithmetic.

3. Every suggestion traces to a scored session — or it is dropped. Generic best practice is not evidence.

4. Zero suggestions is a success, not a failure to report around.

5. Exit 4 from the aggregator is a stop, not an error to swallow or bypass.

6. Never touch the user's real skill files without an explicit per-skill yes; proposed edits live under $RUN/proposed/.

7. A proposed skill edit follows write-a-skill discipline — trigger phrase in the description, smallest change that expresses the rule, replace over append.

Scripts

| Script | Role | Exit codes |

|---|---|---|

| scripts/collect_sessions.py | Harvest Claude Code + Codex sessions, redact secrets, sample, inventory | 0 · 3 bad input |

| scripts/score_aggregator.py | Validate labels/reasons/suggestions, compute grade, emit report.json | 0 · 2 warnings · 3 bad input · 4 validation failure |

| scripts/render_report.py | report.json → single self-contained report.html | 0 · 3 bad input |

All support --help, --output json, and --sample (no real history needed).

References and assets

• [scorers/efficiency.md](scorers/efficiency.md) · [scorers/code-quality.md](scorers/code-quality.md) — the two rubrics, preserved verbatim from upstream
• [references/transcript_scoring_canon.md](references/transcript_scoring_canon.md) — why rubric-anchored LLM judging works and where it fails (7 sources)
• [references/session_mining_privacy.md](references/session_mining_privacy.md) — the local-only contract, redaction pattern canon (7 sources)
• [references/skill_edit_governance.md](references/skill_edit_governance.md) — the filing bar for proposing skill edits (7 sources)
• [assets/session_scores.example.json](assets/session_scores.example.json) · [assets/suggestions.example.json](assets/suggestions.example.json) · [assets/report.example.json](assets/report.example.json) — the three handoff shapes