codex-review
Skill1 starsChanged 6 months agoMITView on GitHub
Run one round of `codex review` against an existing GitHub PR by invoking `~/ai/bin/codex-pr-review.sh`. The review is posted to the PR as a `gh pr review` comment, and the skill reports verdict + finding count + URL back in chat. Use when the user wants a codex pass on a PR they (or another agent) already opened — not for full ship-feature loops, which iterate with auto-fix.
- Deletes or force-pushes
- Commits and pushes
91 other projects use a version of this.
Adopted by
Most recently taken up by
- p-chan/agent-skillstheir own version, first seen Nov 8, 2026
- p-chan/skillstheir own version, first seen Nov 8, 2026
- Smirnov-Labs/claude-skillstheir own version, first seen Nov 8, 2026
- juriwiens/skillstheir own version, first seen Nov 8, 2026
- ryanportfolio/winuse-mcptheir own version, first seen Nov 8, 2026
- fab-codes/agent-workflow-templatetheir own version, first seen Nov 8, 2026
Used with
Projects with it also use addskill, apply-firmware and 10 more.
Other files people use alongside this one.
Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
ryanportfolioSkilladdskill
Create, import, update, or install repo or personal skills: runtime ownership, resources, discovery checks.
Also in 6 projects that use this
ryanportfolioSkillapply-firmware
Bring current folder (plain folder, existing repo, old or partial firmware copy) up to full Harness Firmware in place. Mirror external repo → adopt-repo.
Also in 6 projects that use this
- +10 more from ryanportfolio
Other versions91 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 91
- ClickHouse/ClickHouse★ 51ksame name
- chaseai-yt/claudex-loop★ 3.2ksame name
- morethanwords/tweb★ 2.7ksame name
- alinaqi/maggy★ 711same name
- alinaqi/claude-bootstrap★ 708same name
- ThibautBaissac/rails_ai_agents★ 665same name
- ai-analyst-lab/ai-analyst★ 304same name
- Sateezg/codex-bridge★ 248same name
- sd0xdev/sd0x-dev-flow★ 192same name
- sd0xdev/sd0x-harness★ 192same name
81 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository13 other files it gives its agents.
- add-recipeSkill
- captureSkill
- footySkill
- grill-meSkill
- improve-codebase-architectureSkill
- promote-agents-mdSkill
- repo-initSkill
- synthesiseSkill
- tddSkill
- tracerSkill
- trailsSkill
- typesafe-aiSkill
- video-noteSkill
The file
We split this skill into the rules it gives an agent
11 rules. 0 are common, 6 are distinctive, and 5 are very specific to this project.
0
Common
Most of the 79 edited versions we read keep these.
6
Distinctive
This author's own choices. Worth a look.
5
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 12
Single round: No fix-and-iterate loop. After the review posts, the user (or the calling Claude) decides what to do with the findings.
- Distinctive: line 14
The codex output belongs on the PR, not the terminal: Don't paraphrase or dump the review body back into chat.
- Very specific: line 24
Explicit arg: If the skill was invoked with a PR number (42) or URL (https://github.com/owner/repo/pull/42), use that.
- Distinctive: line 30
If one PR is clearly the obvious candidate, propose it with a single-question confirm: e.g.
- Very specific: line 49
If CURRENT != PR_HEAD, the codex prompt's tool-use ("you may use git, rg, file reads to inspect surrounding code") will read the wrong tree …
- Very specific: line 53
If the user agrees: gh pr checkout "$PR_N" and proceed.
- Distinctive: line 54
If the user declines: proceed and include a one-line warning in the chat report.
- Distinctive: line 74
If the count looks wrong (e.g.
- Very specific: line 91
If the user declined to checkout the PR's branch in step 2, append a second line: *"(reviewed from <CURRENT> — codex's local reads may be …
- Distinctive: line 93
Do not dump the review body, finding bullets, or any prose from the review into chat.
- Very specific: line 99
No fix loop: If the verdict is changes-requested, stop.
“Common” is measured: kept by at least 6 in 10 of the 79 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
codex-review
Runs one round of codex review against a GitHub PR via the shared script at ~/ai/bin/codex-pr-review.sh, then reports the verdict in chat. The review body goes on the PR as the audit trail; chat gets a one-line summary.
Scope
- Single round. No fix-and-iterate loop. After the review posts, the user (or the calling Claude) decides what to do with the findings.
- Existing PRs only. This skill does not create, branch, or push. If there is no PR yet, stop and tell the user.
- The codex output belongs on the PR, not the terminal. Don't paraphrase or dump the review body back into chat.
For the full build-and-iterate flow, use the ship-feature agent instead.
Process
1. Resolve the PR
Two paths:
- Explicit arg. If the skill was invoked with a PR number (
42) or URL (https://github.com/owner/repo/pull/42), use that. No confirmation needed — typing it is the confirmation. Context scan. Otherwise, look at recent conversation context for the most-obvious PR. Signals, in priority order:
- A PR URL mentioned in the last few turns (
https://github.com/.../pull/N) - An explicit
#Nreference where N is clearly a PR (e.g. just-opened PR, recentgh pr view/gh pr createoutput) - The current branch's open PR — check with
gh pr view --json number,title,baseRefName,headRefName,url(no PR-number arg uses the current branch)
If one PR is clearly the obvious candidate, propose it with a single-question confirm: e.g. "Run codex review on PR #42 — "Add foo to bar"?" If nothing obvious surfaces, stop and ask the user for a PR number or URL.
- A PR URL mentioned in the last few turns (
Once resolved, capture:
PR_N=<number>
BASE=$(gh pr view "$PR_N" --json baseRefName -q .baseRefName)
PR_URL=$(gh pr view "$PR_N" --json url -q .url)
PR_HEAD=$(gh pr view "$PR_N" --json headRefName -q .headRefName)2. Check branch state
Get the current branch:
CURRENT=$(git rev-parse --abbrev-ref HEAD)If CURRENT != PR_HEAD, the codex prompt's tool-use ("you may use git, rg, file reads to inspect surrounding code") will read the wrong tree and may produce false positives where the PR introduces new symbols. Surface this and offer to switch:
"You're on
<CURRENT>but the PR is on<PR_HEAD>. Codex's local file reads will see the wrong tree. Rungh pr checkout <PR_N>first?"
- If the user agrees:
gh pr checkout "$PR_N"and proceed. - If the user declines: proceed and include a one-line warning in the chat report.
3. Run the review
VERDICT=$(~/ai/bin/codex-pr-review.sh "$PR_N" "$BASE" 1)The script posts the review to the PR via gh pr review (falling back to gh pr comment on self-PRs) and writes the full review to /tmp/codex-pr${PR_N}-r1.md. Stdout is one of: approved, changes-requested, missing, error.
Round header caveat. The script labels the posted review "round 1/3" because it's shared with the ship-feature agent's bounded loop. For a standalone re-run on the same PR, that header is misleading but harmless — the review body is what matters, and the PR's review history shows the true sequence. Don't try to work around it from here.
4. Count findings
Parse the saved review file for findings count. Findings are bulleted items in the codex review body. A robust enough heuristic:
COUNT=$(grep -cE '^[-*] ' "/tmp/codex-pr${PR_N}-r1.md" || echo 0)If the count looks wrong (e.g. the review is prose-only), fall back to reporting "findings posted on PR" without a number.
5. Report back
A single line in chat:
<verdict> — <COUNT> findings — <PR_URL>Examples:
approved — 0 findings — https://github.com/owner/repo/pull/42changes-requested — 4 findings — https://github.com/owner/repo/pull/42missing — 2 findings — https://github.com/owner/repo/pull/42(codex didn't emit a verdict tag; treat as changes-requested)error — see PR for codex output — https://github.com/owner/repo/pull/42
If the user declined to checkout the PR's branch in step 2, append a second line: "(reviewed from <CURRENT> — codex's local reads may be off-tree)".
Do not dump the review body, finding bullets, or any prose from the review into chat. The user reads it on the PR.
Rules
- Never merge the PR. This skill reviews; it doesn't merge. No
gh pr merge, ever. - No destructive git. No
git push --force,git reset --hard,git clean -f,git branch -D,git checkout --/git restore .. The only git mutation this skill ever performs is the optionalgh pr checkoutin step 2, with explicit user confirmation. - No fix loop. If the verdict is
changes-requested, stop. The calling Claude or the user decides what to do next. - Don't paraphrase the review. The PR has the audit trail; chat gets only the one-line summary.
- British English in any prose (commit messages, comments) — but this skill should rarely produce any prose other than the one-line report.
- If
codex-pr-review.shis missing or non-executable, stop and tell the user. Don't try to reconstruct the review flow inline.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

