Skill Session Review

Skill Session Review

In the commands below, <skill-dir> is this skill's base directory, announced when the skill loads. Substitute the literal path. On Windows, py -3 may replace python.

Review how a named existing skill was followed in real sessions. Persist a validated Markdown and self-contained HTML report package. Do not edit the target skill. Do not call qiaomu-meta to apply changes.

Hard gates

• Do not write inside the target skill directory.
• Do not emit diff.patch or edit SKILL.md of the target.
• Do not run qiaomu-meta as part of this skill.
• The only permitted side-effect types are two report artifacts, the repo-root .gitignore exact line through its independent helper, the governed .input/<name>.json through its input manager, and opening the generated HTML through open_report.py. The report-subtree helpers never modify .gitignore.
• Permitted side-effect types are not authorization. Before any one is used, follow the exact preview and confirmation gates below. The named report package, each replacement, and .gitignore are separate confirmations.
• The HTML report must be self-contained, with no external resource references or vendor promotion.
• Do not print full private chats in the conversation. The report uses short excerpts only.

Workflow

1. Resolve the target. Path → that instance. Name → if more than one SKILL.md, list paths and stop.
2. Resolve the current repo root: git rev-parse --show-toplevel or an explicit root the user gave.
3. Scan sessions:

python "<skill-dir>/scripts/scan_invocations.py" --skill-name <name> [--skill-path <abs>] --scope global --repo-root <abs>

Use --scope cwd only when the user asked to limit to this repository. Read invocation signals. Treat loaded and available as coverage, not as required-change evidence.

4. Apply the zero-sample gate before reading private session slices. When all four stores are missing-store, output unrated: no-session-stores plus the four-platform coverage/counts and stop. When stores are available but no session is invoked, output unrated: no-invoked-sessions plus the same bounded counts and stop. Do not read slices, construct review JSON, call a helper, write a file, calculate a mean/ratio/overall/grade, or open a browser on either branch.

5. With at least one invoked session, read the target SKILL.md (read-only) and invoked session slices. Score every invoked session with the review scorecard, and fill findings per the finding contract. Set language from the user's request language and set skill_name to the same canonical basename later passed as --name. Fill the declared aggregate, including overall and grade; helpers revalidate and recompute it with Decimal.

6. Prepare authorization before running a side-effect helper:

   • If reports/skill-session-review/ is not ignored, independently preview the complete candidate repo-root .gitignore, exact path, and create or replace effect. Obtain explicit confirmation before running ensure_report_ignore.py. Replacement also requires the current SHA-256.

     python "<skill-dir>/scripts/ensure_report_ignore.py" --repo-root <abs> [--replace --expected-sha256 <gitignore-sha256>]
     

     Pass the complete candidate .gitignore through raw stdin. If the exact ignore line is already effective, the helper returns unchanged without a write or an authorization expansion.

   • Display the confirmed repo root, canonical name, exact .input/<name>.json, <name>.md, and <name>.html paths, plus exactly this effect sequence: create input → create Markdown → create HTML → proof-gated remove input → open HTML. Use AskUserQuestion or an equivalent explicit interaction to obtain one confirmation for this named report package. Freeze the authorization snapshot; any root, name, path, or effect drift invalidates it and requires a new preview and confirmation.

   • The package confirmation authorizes no-clobber creation, proof-gated input removal, and one open only. If any input or report target already exists, stop and separately show its exact path, current SHA-256, and replacement effect. Obtain a separate explicit confirmation for each replacement.

7. After confirmation, persist the validated JSON, render one artifact per writer call, prove both artifacts before removing the input, then open HTML:

python "<skill-dir>/scripts/manage_review_input.py" create --repo-root <abs> --name <name>
python "<skill-dir>/scripts/write_session_review.py" --repo-root <abs> --name <name> --format markdown --review-json <abs-input-json>
python "<skill-dir>/scripts/write_session_review.py" --repo-root <abs> --name <name> --format html --review-json <abs-input-json>
python "<skill-dir>/scripts/manage_review_input.py" remove --repo-root <abs> --name <name> --expected-sha256 <input-sha256> --artifact-sha256 markdown=<md-sha256> --artifact-sha256 html=<html-sha256>
python "<skill-dir>/scripts/open_report.py" --repo-root <abs> --name <name>

Pass the complete review JSON to input-manager create through raw stdin; only input-manager create and replace accept that payload. The report writer accepts neither report stdin nor the former --input path. The input manager, report writer, and opener use only canonical --name; the scanner in step 3 intentionally keeps --skill-name. Template and lifecycle: report contract. Do not git add.

Input and report creation is no-clobber. A confirmed replacement uses replace --expected-sha256 for the input manager or --replace --expected-sha256 for a report writer. Both report calls reuse and revalidate the same JSON. If either fails, retain the input and successful artifact, then retry only the failed format. If the target exists, repeat its separate replacement preview and confirmation. Remove the input only after both report hashes and the current input hash form complete proof. The remove helper acquires input, Markdown, and HTML destination leases in that order, revalidates every proof while all three are held, and binds deletion to the proved input identity. Lease contention or a late input/artifact replacement fails closed and preserves the unproved object; never bypass that result with a manual path deletion. Open HTML only after both reports succeed and input cleanup succeeds.

8. Chat output: one conclusion line, both report paths, the HTML file:// URI, and one text fence with a filled handoff prompt. If open_report.py returns "opened": false, say that the browser could not be opened automatically and tell the user to open the file:// URI manually. Do not paste the SSR table. Do not write 见上一条消息.

Routing

• Usage counts or unused-skill deletion: not this skill.
• Resume a previous chat: resume-claude / resume-codex / resume-cursor.
• Optimize CLAUDE.md or AGENTS.md: claude-context-improver / codex-context-improver.
• Create or rewrite a skill from materials: qiaomu-meta, in a later turn using the handoff prompt.