菜单
After finishing the current task, run an external Codex CLI review, return its feedback verbatim, and add your own independent assessment of the findings. Use when the user says "Codex review", "ask Codex to review", or runs /codex-review. Codex is a second opinion — not ground truth.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Run a bounded review-remediate loop in which every review round uses the external Codex CLI through the codex-review workflow. Use when asked to have Codex iteratively review and fix uncommitted changes, a branch or commit diff, a pull-request diff, named paths, or another explicit target until no verified finding remains or a safe round limit or blocker is reached.
Run a bounded, evidence-driven review and remediation loop over uncommitted changes, a base-branch merge diff, commits, pull-request diffs, or an explicit custom target. Use when asked to review code iteratively, fix verified findings within the original authorization, repeat review until clean or safely bounded, or provide a final review audit without external reviewers or remote writes.
Compare the current agent's prior user-visible analysis report with one explicitly supplied report file and produce a concise numbered decision table covering material conflicts, useful improvements, and questions needing user judgment. Independently recommend how to resolve conflicts. Use only when the user explicitly invokes this skill with exactly one target report file; never invoke implicitly, from a general comparison request, or without that file.
Compare the current Git worktree with one explicitly supplied peer worktree and identify only material improvements to apply to the current worktree. Use only when the user explicitly invokes this skill with exactly one target worktree path; never invoke implicitly, from a general comparison request, or without that path.
Upgrade all project dependencies across Mix/Hex, npm applications including Phoenix assets and browser extensions, and Tailwind. Use when asked to update, upgrade, or modernize every dependency and repair resulting code, configuration, or test breakage.
Perform one deliberate final pass over completed changes to catch material omissions or inconsistencies before handing off. Use at the end of implementation and verification, especially when code, documentation, tests, or configuration must remain coherent.
| name | codex-review |
| description | After finishing the current task, run an external Codex CLI review, return its feedback verbatim, and add your own independent assessment of the findings. Use when the user says "Codex review", "ask Codex to review", or runs /codex-review. Codex is a second opinion — not ground truth. |
| metadata | {"short-description":"Codex review + independent assessment"} |
| compatibility | Requires `codex` CLI on PATH (`codex login`). Shell + `mktemp` only. |
Finish any in-progress work first, then run Codex on the current branch / working tree. Paste Codex's output verbatim, then give your own independent assessment of each finding (Accept / Partial / Decline). Codex is a second opinion — not ground truth. Never auto-fix unless asked (then apply only Accepted / Partial findings).
Stop if there is nothing plausible to review (no changes on the current branch or working tree, and no named paths outside git).
Only the original task request (not the whole chat). Strip a trailing bare "Codex review". Optional one-line focus from the trigger. Non-git work: name absolute paths in that same prompt text.
Do not send summaries, inventories, diffs, logs, tool traces, or prior
reviews. cd to the repo root (or parent of named paths); Codex discovers
changes on the current branch and working tree itself.
One shell, tool timeout: 600000. Transcript goes to the log; Codex text is
stdout from cat "$out_file". Full user reply is that text plus Assessment.
command -v codex >/dev/null || { echo "codex not on PATH"; exit 1; }
umask 077
scratch="${TMPDIR:-/tmp}/grok-$(id -u)"
mkdir -p "$scratch" && chmod 700 "$scratch"
out_file="$(mktemp "$scratch/out.md.XXXXXX")"
log_file="$(mktemp "$scratch/stream.log.XXXXXX")"
echo "out_file=$out_file"
echo "log_file=$log_file"
cd "/ABS/REPO_ROOT_OR_PATHS_PARENT"
# Inline the real task text into the heredoc body when constructing this
# command. Use a quoted delimiter that does NOT appear as a line in the task
# (default below; change it if the task contains that line). Quoted form
# prevents shell expansion of $ / backticks / quotes in the task.
codex exec review --ephemeral --skip-git-repo-check \
-o "$out_file" - >"$log_file" 2>&1 <<'CODEX_REVIEW_PROMPT_EOF'
You are performing an independent code review.
## Original user request
…inline original task here…
Discover the changes yourself on the current branch and working tree
(git status/diff/log and/or by reading any paths named above).
Do not assume a change list was provided.
## Review criteria
- Focus on missing items, fixes, and real improvements.
- Be unbiased and concise. No praise-only summaries.
- Skip low-value nits. If nothing material, say so briefly.
- **No markdown links.** Reference files as relative `path/to/file.ext:LINE`
(line optional when not applicable), e.g. `src/app.ts:42`. Never use
`file://`, absolute paths, or `[text](url)` link syntax for code locations.
## Output
Review feedback only. No tool chatter or preamble. No markdown links.
CODEX_REVIEW_PROMPT_EOF
if [ -s "$out_file" ]; then
cat "$out_file"
else
echo "FAILED: empty out_file" >&2
tail -n 30 "$log_file" >&2
echo "try: codex doctor / codex login" >&2
fi
rm -f "$out_file" "$log_file"
--base / --uncommitted / --commit (CLI rejects
the combination). Do not pass -m unless the user asked.$out_file (ignore exit code alone). On empty: use the
stderr log tail; do not invent findings.out_file
until non-empty (bounded wait), then cat and rm -f both printed paths.
If still empty when the bound expires: report both printed paths, do not
delete, do not invent findings. Do not inspect or match codex processes.Always both parts — never Codex alone:
## Codex review
<full cat of out_file — zero edits>
## Assessment
For each Codex finding (or the review as a whole if clean), triage:
- **Accept** — valid; short reason grounded in the actual code or task
- **Partial** — partly right; what holds and what does not
- **Decline** — false positive, wrong severity, or out of scope; short reason
Also note anything material Codex missed. Do not restate Codex.
If the user asked to fix: change code only for Accept / Partial findings.
Do not skip Assessment when findings look obvious or empty. A clean Codex pass still needs a brief Accept (confirmed clean) or Decline (you disagree).