mspec-review

name	mspec-review
description	self-review step of mspec workflow — sanity-check all artifacts via subagent
when_to_use	User runs /mspec:review, or workflow auto-continues to self-review

Procedure

1. Run mspec status --change <change-dir> --json first.
2. Run mspec continue --change <change-dir> --json to fetch subagent_prompt.
3. Read all requires (checklist.md, design.md, architecture-overview.md, quickstart.md).
4. Invoke mspec-self-reviewer subagent (Task tool) with subagent_prompt.
5. Append the subagent's review notes as a ## Self-Review section to design.md (no new files; produces: []).
6. Update Constitution Check tables across artifacts where the reviewer flagged gaps.
7. Run mspec validate --change <change-dir>.
8. block: true — stop and ask the user to run /mspec:continue.

Observation (Agent Experience Log)

After the subagent completes (step 4), count [blocker] lines in the review output, then record the run:

mspec agent-run record self-review \
  --change <change-name> \
  --bytes <sum-of-input-artifact-bytes> \
  --artifacts <space-separated-artifact-paths> \
  --edits <count-of-[blocker]-lines>

Example: mspec agent-run record self-review --change 2026-05-25-my-feature --bytes 22180 --artifacts checklist.md design.md tasks.md --edits 2

This appends one JSONL entry to changes/<change>/.agent-runs.jsonl. Do NOT include prompt text or file contents.

Verification (C2)

• mspec validate --change <change> — アーティファクト整合性チェック
• mspec anchor check --change <change> — アンカー解決確認
• self-review の [blocker] 指摘がゼロになっているか確認

Learning (C3)

このスキルの実行で発生した学習候補を記録する:

<!-- LEARNING: <パターン説明> | source: <FR-ID> | confidence: low|medium|high -->

mspec learn コマンドが archive 済み changes からこれらを収集してpost-condition候補をproposeする。