feat: commands の旧モデル儀式を成果指向に書換（#114、v0.64.0）#129
Conversation
- orchestrate: 固定 4 エージェント直列 + HANDOFF/レポートテンプレ廃止 → メインスレッド実装 + 並列レビュー - build-fix: per-error フルビルド → 根本原因バッチ + 最速検証 + 最後にフル 1 回 - refactor-clean: per-deletion 前後フルテスト → baseline + バッチ + focused tests + バイセクト - learn: 出力を <pattern>/SKILL.md + frontmatter（Use when... 形式）に修正、auto-memory 棲み分け明記 - update-docs: RUNBOOK 生成を実在ソース条件付きに（捏造防止） Closes #114
There was a problem hiding this comment.
Pull request overview
This PR modernizes several /commands/* docs by removing legacy “fixed ritual” workflows aimed at older LLM limitations and rewriting them into outcome-oriented guidance aligned with current Claude Code behavior and the kit’s review/guardrail conventions.
Changes:
- Rewrites
/orchestrateto prefer main-thread execution + parallel independent reviewers, removing fixed agent relay chains and handoff/report templates. - Rewrites
/build-fixand/refactor-cleanto favor root-cause batching, fast checks / focused tests, and a single final full verification step. - Updates
/learnto emit learned skills in.../<pattern-name>/SKILL.mdwith YAML frontmatter; tightens/update-docsto derive RUNBOOK sections only from real operational sources; adds a v0.64.0 changelog entry and adjusts unit tests accordingly.
Reviewed changes
Copilot reviewed 8 out of 8 changed files in this pull request and generated 4 comments.
Show a summary per file
| File | Description |
|---|---|
commands/orchestrate.md |
Removes fixed sequential agent chains/templates; documents parallel review spawning and minimal custom chaining handoff. |
commands/build-fix.md |
Switches guidance from per-error rebuild loops to root-cause batching + fastest verification + one final full build. |
commands/refactor-clean.md |
Replaces per-deletion full-suite ritual with baseline → batch SAFE deletions → focused tests → final full suite + bisect guidance. |
commands/learn.md |
Updates learned skill output to directory + SKILL.md + YAML frontmatter; clarifies auto-memory vs explicit skill promotion. |
commands/update-docs.md |
Limits RUNBOOK generation to sections derivable from real ops sources; generalizes task-definition source detection. |
tests/unit/test-content-cleanup.sh |
Updates assertions to match new /orchestrate wording (no fixed chains; parallel reviewers). |
tests/unit/test-doc-blocker.sh |
Updates allowed learned-skill path example to nested .../pattern/SKILL.md. |
CHANGELOG.md |
Adds release notes for 0.64.0 describing the command rewrites. |
| ## Usage | ||
|
|
||
| `/orchestrate [workflow-type] [task-description]` | ||
|
|
|
|
||
| 5. Stop if: | ||
| - The error count is increasing, or fixes are going in circles | ||
| - Same error persists after 3 attempts |
| 1. Detect the project's task definition source (package.json scripts, | ||
| Makefile, pyproject.toml, etc.) | ||
| - Generate a scripts/tasks reference table | ||
| - Include descriptions from comments |
| - Apply SAFE deletions as a batch, one commit per deletion or logical | ||
| group (easy rollback) |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 594619fbbb
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| Create a skill at `~/.claude/skills/learned/<pattern-name>/SKILL.md` so the | ||
| current skill discovery mechanism (directory + SKILL.md with YAML | ||
| frontmatter) can load it: |
There was a problem hiding this comment.
When users follow /learn, this path creates ~/.claude/skills/learned/<pattern-name>/SKILL.md, i.e. a grouping directory under the personal skills root rather than a skill directory directly under it. Claude Code documents personal skills as ~/.claude/skills/<skill-name>/SKILL.md (https://docs.claude.com/en/docs/claude-code/skills), and the repo's shipped skills all follow that layout, so the generated learned patterns can remain undiscovered/unloaded despite this command saying they are activatable skills. Use a top-level skill directory such as ~/.claude/skills/<pattern-name>/SKILL.md or add a discovery-compatible indexing strategy.
Useful? React with 👍 / 👎.
概要
LLM 性能監査計画 #121 の P3-08（#114）。旧モデルの暴走防止・手抜き対策として書かれた固定手順 5 件を成果指向に書き換える。
変更内容
~/.claude/skills/learned/<pattern-name>/SKILL.md+ YAML frontmatter（when_to_use は規約どおりUse when ...形式）に修正。auto-memory との棲み分けを Notes に明記検証
Use when...規約と不整合）+ low 2 件 → when_to_use 形式と CAUTION 例示の言語中立化を修正済み。残り 1 件（learned スキルは CI のfrontmatter 検査対象外）は既存の構造的制約として記録Closes #114