---
name: pre-review-self-check
user-invocable: false
description: |
  Use whenever implementation code has just been written or modified and a completion claim or
  code-review request is imminent. Trigger on phrases like "自己レビュー", "コードの品質を確認",
  "レビュー前にチェック", "self review", "pre-review check", "sanity check the diff". This skill
  is the self-review step the implementing agent runs on its own diff BEFORE dispatching
  `code-review-loop`. It catches obvious problems (naming, complexity, boundaries, error handling,
  dead code, secrets, logging) before the expensive subagent round-trip.
  Do NOT trigger on T0 tasks or when the user explicitly skips review.
  This skill is self-contained and MUST NOT delegate to any `superpowers:*` skill.
required-by:
  - code-review-loop
---

# pre-review-self-check

## Independence

This skill **MUST NOT** invoke or delegate to any `superpowers:*` skill.

## Distinction from `code-review-loop`

| Skill | Reviewer | Inputs | Cost |
|---|---|---|---|
| `spec-coexist:pre-review-self-check` | The implementing agent itself | Diff + checklist | Cheap, fast |
| `spec-coexist:code-review-loop` | A fresh subagent (third-party) | Diff only | Expensive, authoritative |

Findings from this skill **MUST** be fixed locally; do not defer them to the reviewer.

## References

- `references/code-quality-checklist.md` — SOLID, naming, complexity, boundaries, error handling, dead code, secrets, logging.
- `references/self-review-protocol.md` — exact procedure: diff, walk checklist, classify, fix, write evidence.
- `references/red-flags.md` — bilingual rationalization table.

## Scripts

- `scripts/run_self_review.sh [--base <sha>]` — prints diff summary + checklist skeleton.

## Procedure

1. Run `scripts/run_self_review.sh`. HALT if working tree is empty.
2. Read `references/code-quality-checklist.md`; tick every item per changed file.
3. Classify findings (Critical / Important / Minor). Critical + Important should be fixed in-session. Minor may be deferred with rationale.
4. Read `references/red-flags.md`. Reject any matching rationalization.
5. Write evidence via `../_shared/scripts/write_evidence.sh` with `proof-type: self-review`.
6. Only after evidence is written may the caller invoke `spec-coexist:code-review-loop`.

## Hard Constraints

- For T1 tasks, this skill is **RECOMMENDED** before `spec-coexist:code-review-loop` (if review is requested).
- For T2/T3 tasks, the implementing agent **MUST** run this skill before `spec-coexist:code-review-loop`.
- Critical and Important findings **MUST NOT** be forwarded to the reviewer subagent.
- This skill **MUST NOT** spawn a subagent.

## Flow

```mermaid
flowchart TD
    A[Code changes staged] --> B[run_self_review.sh]
    B --> C[Walk checklist]
    C --> D[Classify findings]
    D --> E{Critical/Important<br/>remain?}
    E -- Yes --> F[Fix in-session] --> B
    E -- No --> G[Check red-flags]
    G --> H[write_evidence proof-type: self-review]
    H --> I[Hand off to code-review-loop]
```