Skip to main content

review-loop

Use when the user wants the current code proactively reviewed and fixed (e.g. "review-loop", "/review-loop", "review what I have and fix it"). Self + cross-model review → triage → fix → update PR, loop until clean.

ソース情報

リポジトリ
skyhook-io/skyhook-skills
ソースの最終更新活動
2026年7月1日 12:26
検出された SKILL.md の言語
英語
スター
1
フォーク
0

インストール方法

デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。

ソースファイルを確認

インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。

SKILL.md を表示中

SKILL.md
ソースの指示 · 読み取り専用プレビュー
name
review-loop
description
Use when the user wants the current code proactively reviewed and fixed (e.g. "review-loop", "/review-loop", "review what I have and fix it"). Self + cross-model review → triage → fix → update PR, loop until clean.
metadata
{"short-description":"Proactive self + Claude review → triage → fix → update PR"}
# Review Loop (Codex) Proactively review the current code and fix it, Codex-side — the proactive counterpart to reacting to CI/Bugbot on a PR. Same workflow as Claude's `/review-loop`. **Canonical procedure:** read `~/.codex/skills/skyhook-skills-commands/review-loop.md` and follow it (scope → review → triage → fix → update PR → loop; caps; safety). **Codex translations:** - **Self review** ("/review", "/simple") → YOU review your own diff hard: correctness bugs, security, silent failures, races, logic errors, breaking API — and over-engineering / dead weight (options no caller sets, one-impl interfaces, tests that assert the implementation). Don't ping-pong with prior reviewers. - **Cross-model review** ("/cross-review") → the **claude-review** skill — from Codex the secondary reviewer stays Claude (it reviews your work). Nontrivial changes only — don't ceremony-ize a typo. - **Scenario-sensitive work** (copy, diagnostics/remediation, detector precision, error classification, security/permissions, UI states) → build a scenario ledger before calling the review complete: scenario, expected final behavior or copy, source-of-truth evidence, self-review verdict, Claude status, tests/live proof, open decision. Claude status is either the verdict or `skipped: <reason>` when cross-review was intentionally skipped. "Review loop ran" is not a completion claim. - **Risk / blast radius** → for nontrivial changes, include a practical risk note: affected surface, likely failure mode, mitigation/test proof, and residual risk. Keep it proportional; don't add boilerplate for harmless edits. - **Triage** ("/triage-findings") → inline: read the real code at each `file:line`, **never auto-accept**, cite evidence on every skip. - **Fix** ("/fix-findings") → apply the confirmed Fix verdicts. - **`/qa`** → read the repo's `.claude/commands/qa.md` and follow it (type-check + tests; visual-test default-skip, only warranted self-contained UI changes). - **Update PR** ("/pr") → via `git` + `gh`: stage relevant files only, commit/ amend, push `--force-with-lease`. **Write the body per `~/.codex/skills/skyhook-skills-commands/pr.md`** — depth proportional to the change, motivation + design, **no review-fix trivia; re-derive the narrative, don't append fix bullets each round.** No PR yet ⇒ leave commits local and say so. **Invariants:** triage skeptically and never auto-accept; cross-review only when nontrivial; cap at 3 rounds and report (stop if the same finding survives 2 fix attempts); never on `main`, never merge, `--force-with-lease` only. Narrate phases with the glyph-tagged banners and **end with the run-summary ledger** (per-step counts, triage breakdown, result line, open items) defined in the canonical file. Steps are judgment calls — decide which apply, and **honor free-text steering in the invocation** (e.g. `consult claude`, `no PR`, `review only`, `thorough`, `focus on <area>`) as an override, per the canonical file's directives section.
GitHubで見る