Skip to main content

codex-autoreview

Run a slim Codex-native closeout review for Apple Agent Kit branches and PRs before commit, PR, or merge.

来源信息

仓库
neonwatty/apple-agent-kit
最近来源活动
2026年6月28日 20:36
检测到的 SKILL.md 语言
英语
星标
1
Forks
0

安装方式

默认使用会先检查来源的 Prompt;你也可以切换为直接命令,或下载本地副本。

检查来源文件

决定是否安装前,请先阅读 SKILL.md,以及 SkillsMP 当前展示的配套文件。

正在显示 SKILL.md

SKILL.md
来源说明 · 只读预览
name
codex-autoreview
description
Run a slim Codex-native closeout review for Apple Agent Kit branches and PRs before commit, PR, or merge.
# Codex Autoreview Use this skill when the user asks for autoreview, Codex review, PR closeout review, or a rigorous review before commit, push, PR, or merge. ## Contract This is code review, not approval routing. Review output is advisory until verified against the real files. Before reviewing, freeze: - original request or issue; - base branch or PR base; - intended behavior and owner boundary; - changed files and non-test scope; - validator commands already run; - privacy and hardware actions explicitly out of scope. Classify findings as: - `in-scope blocker`: introduced by this branch, concrete, and fixable without changing the PR purpose; - `follow-up`: real but outside this branch's owner boundary; - `stop-and-escalate`: requires a new public contract, private adapter decision, credentials, hardware dispatch, app launch, screenshots, physical-device automation, or release-process change. Fix only in-scope blockers. After any fix, rerun focused validators and repeat review only if the reviewed diff materially changed. ## Apple Agent Kit Checks For public repo changes, confirm: - `python3 scripts/validate-release.py` passes; - `git diff --check` passes; - no private runner labels, hostnames, local paths, device IDs, bundle IDs, signing teams, screenshots, secrets, raw logs, private command lines, or repo-specific adapter details appear in public files; - schemas, examples, docs, and validators agree; - public workflows do not run private hardware/app automation. For private adapter changes, do not run hardware actions unless the user explicitly approved that specific dispatch. Planning and static review are allowed; install, app launch, UI smoke, screenshots, product build/test, physical-device automation, and Codex-session bridge execution are gated. ## Output Lead with findings, ordered by severity, using file and line references. Then list validators run and end with one verdict: - `ready` - `needs-fix` - `blocked` If there are no issues, say so clearly and mention any residual risk or unavailable proof.
在 GitHub 查看