code-ultrareview

これはレビュー時に保存された抜粋です。完全で最新の内容は外部ソースを確認してください。

---
name: code-ultrareview
description: Eight-axis judgment code review for the current diff — Correctness, Simplification, Tests, Documentation, Style, Intent, Design/API, Performance (+ Coherence on metadata changes). Five-phase pipeline scope → deterministic tool battery (npx/uvx-preferred, zero-install for the JS + Python majority) → 8 parallel LLM axis reviewers → Haiku validators on sub-80 findings (verbatim rubric, ≥80 threshold) → synthesis with no-silent-drop + Conventional Comments JSONL. Every report closes with "What I did NOT check" (security → /security-review, runtime perf, flaky detection). Opt-in flags `--verify-build`, `--mutation-test`, `--reconcile`, `--apply-safe`. Public-skill posture — zero auto-install, graceful skip on missing native tools. Invoke before a commit or PR — "review my changes", "deep review", "did I miss anything", "check before I commit", "audit this PR", "drift / gaps / blind spots".
when_to_use: 'User-invoked before commit or PR; runs the full 8-axis fan-out — no tiers. Defers security to /security-review (link in every report); defers runtime performance and benchmarks (explicit non-goal). Distinct from Anthropic''s remote /ultrareview command.'
argument-hint: "[-b <ref>] [--repo-kind <kind>] [--reconcile <input>] [--verify-build] [--mutation-test] [--apply-safe] [--include-prose] [--axes <list>] [--preflight] [-s] [-S]"
license: MIT
compatibility: "Optimized for Claude Code; degrades gracefully on any agent implementing the Agent Skills standard."
allowed-tools: Read, Grep, Glob, Bash, Task, WebFetch
disable-model-invocation: true
metadata:
  author: coroboros
  sources:
    - github.com/anthropics/claude-plugins-official/tree/main/plugins/code-review
    - github.com/anthropics/claude-plugins-official/tree/main/plugins/code-simplifier
    - code.claude.com/docs/en/code-review
    - code.claude.com/docs/en/ultrareview
    - github.com/mattpocock/skills/tree/main/skills/engineering/improve-codebase-architecture
---

# Code ultrareview

<!-- canonical:adversarial-verification:start -->
## Critical — Adversarial verification

These rules govern how this skill trusts its own output — apply them whenever it verifies a claim, a defect, a source, or a decision before acting on it.

- Refute by default. Treat each non-trivial finding as unproven until a fresh-context check fails to refute it — the context that produced a claim cannot reliably clear it.
- No silent drop. Every finding flips the conclusion, is refuted in writing, or is filed as a risk or open question. A finding that vanishes without a verdict is a defect.
- Don't re-litigate settled facts. Spend adversarial effort on load-bearing or contested claims; let established facts pass. Over-refutation manufactures false doubt — it does not add rigor.
- Stay selective and cost-aware. Scale verification to the stakes; reversible, low-impact work gets a light touch, not a full adversarial sweep.
- Concede only to a strong rebuttal. A weak counter folds into the finding or gets filed; it does not overturn it.
<!-- canonical:adversarial-verification:end -->

<!-- canonical:execution-discipline:start -->
## Important — Engineering discipline

These rules govern how this skill changes code — apply them whenever it writes, edits, or proposes a fix.

- Minimal scope. Only what's directly requested or clearly necessary — no extra files, no abstraction for one use, no configurability nobody asked for, no error handling for states that can't happen. Validate at system boundaries; trust internal code.
- General solution, not the test cases. Implement the real logic for all valid inputs; never hard-code to inputs or bolt on workaround scripts to make a test pass. Tests verify the solution; they don't define it. A test is wrong? Say so — don't bend correct code to a broken test.
- Investigate before claiming. Never speculate about code you haven't opened; read the referenced file before answering. Ground every claim in what you actually read, not a