agent-review
Skill8 starsChanged 5 days agoMITView on GitHub
Launch an expert code reviewer agent with full project context.
11 other projects use a version of this.
Adopted by
Most recently taken up by
- facebookresearch/autoform-bottheir own version, first seen Nov 12, 2026
- khoralabs/skillstheir own version, first seen Nov 1, 2026
- IlyaSemenov/agent-review-skilltheir own version, first seen Oct 31, 2026
- mementomori-co/maker-masterytheir own version, first seen Oct 25, 2026
- alexaiarchitect/google-adk-skillstheir own version, first seen Oct 20, 2026
- fabioc-aloha/Alex_ACT_Editiontheir own version, first seen Oct 13, 2026
Other versions11 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 11
- DmNote-App/DmNote★ 4.2ksame name
- foldkit/foldkit★ 1.1ksame name
- Tsuzat/Edra★ 675same name
- facebookresearch/autoform-bot★ 143same name
- IlyaSemenov/agent-review-skill★ 4same name
- ethanaubuchon/dossier-tradecraft★ 3same name
- mementomori-co/maker-mastery★ 2same name
- khoralabs/skills★ 1same name
- ngc-shj/claude-code-config★ 1same name
- alexaiarchitect/google-adk-skills★ 0same name
1 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 repository35 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- commandsSkill
- agentic-auditSkill
- autonomous-dev-flowSkill
- batch-mergeSkill
- bug-huntSkill
- catchupSkill
- check-prSkill
- commitSkill
- create-issueSkill
- create-prSkill
- decideSkill
- decompose-issueSkill
- fix-ciSkill
- full-reviewSkill
- hallucination-checkSkill
- learnSkill
- manual-testing-modeSkill
- mergeSkill
- parallel-devSkill
- prime-directiveSkill
- project-auditSkill
- pulseSkill
- qa-updateSkill
- reconSkill
- session-lifecycleSkill
- skillSkill
- smoke-formSkill
- smoke-testSkill
- start-workingSkill
- swarm-auditSkill
- tackle-issuesSkill
- tiered-delegationSkill
- visual-briefSkill
The file
We split this skill into the rules it gives an agent
9 rules. 0 are common, 6 are distinctive, and 3 are very specific to this project.
0
Common
Most of the 9 edited versions we read keep these.
6
Distinctive
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 17
Before reviewing, the agent MUST read:
- Distinctive: line 71
Create a comprehensive review:
- Distinctive: line 144
If more than one such block appears in the report (a draft superseded by a final revision), the LAST one is what gets recorded — always end …
- Distinctive: line 191
Never leave deferred items as just review comments.
- Very specific: line 230
After all fixes are committed, check whether any issues created during this review — or pre-existing from-review issues — were already …
- Very specific: line 263
Capture the result: When this review ran as a subagent (spawned via the Agent tool), the orchestrator records its structured result — the …
- Very specific: line 270
If the orchestrator never runs record (the agent's report was only read, not captured), review-result.py harvest recovers the block from …
- Distinctive: line 291
Reliability first: - Always consider error recovery and edge cases
- Distinctive: line 292
Keep it simple: - No over-engineering, no premature abstractions
“Common” is measured: kept by at least 6 in 10 of the 9 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
/agent-review
Launch an expert code reviewer agent with full project context.
Arguments
$ARGUMENTS- PR number (optional, defaults to current branch's PR)
Instructions
1. Gather Context
Before reviewing, the agent MUST read:
# Project guidelines
cat CLAUDE.md
# Get PR info
PR_NUM=${1:-$(gh pr view --json number -q .number)}
REPO=$(gh repo view --json nameWithOwner -q .nameWithOwner)
gh pr view ${PR_NUM}
gh pr diff ${PR_NUM}Use repo-memory to gather surrounding context cheaply. This repo has the repo-memory MCP. For files the diff touches, call get_file_summary (or batch_file_summaries) and get_dependency_graph to understand the change's blast radius for a fraction of a full Read instead of fully re-reading each one — then Read the changed regions in full to review the actual lines. Use search_by_purpose to locate related code the diff should have updated.
2. Review Criteria
The agent reviews against these standards:
Code Quality
- Follows project style guide (per CLAUDE.md)
- Proper error handling
- No obvious security issues (injection, path traversal, credential exposure)
- Clean naming and structure
- Server: ES modules (import/export), no semicolons, single quotes, EventEmitter pattern for component communication, proper cleanup on destroy/close
- App: TypeScript strict, functional components with hooks, Zustand store patterns (immutable updates via
set()), noanytypes, platform-aware code, noAbortSignal.timeout()(not in React Native)
Architecture Alignment
- Changes follow established patterns
- No breaking changes to existing interfaces/APIs
- New patterns documented if introduced
- CLI mode and PTY/terminal mode remain independent paths
- WS protocol messages documented in ws-server.js header
- Auth flow: auth → auth_ok → server_mode → status → claude_ready
Testing
- Tests pass
- New functionality has test coverage where appropriate
- No test regressions
Performance
- No obvious N-squared loops on collections
- No unbounded buffers or memory leaks
- Proper cleanup of resources (timers, listeners, processes, connections)
Mobile
- Touch targets min 44pt
- Keyboard handling for Android suggestion bar
- Safe area insets, Expo Go compatibility
3. Generate Review
Deciding vs escalating. A finding is not automatically a question for the user. If the code, the git history, the issue thread or the docs answer it, answer it in the review; if the call is cheap and reversible, state it, note the assumption, and move on. Escalate to the user only when the answers lead to materially different work, or the action is irreversible or outward-facing — and when you do, escalate through /decide (the AskUserQuestion tool, 2-4 concrete options, your recommendation first with its costs named), never as prose buried in the review body. Keep reviewing the rest of the diff while the question is open.
Create a comprehensive review:
## Code Review: PR #${PR_NUM}
### Summary
Brief overview of changes and their purpose.
### Strengths
- What's done well
- Good patterns used
### Issues Found
#### Critical (Must Fix)
| File | Line | Issue | Suggested Fix |
|------|------|-------|---------------|
| ... | ... | ... | ... |
#### Suggestions (Should Consider)
| File | Line | Suggestion | Rationale |
|------|------|------------|-----------|
| ... | ... | ... | ... |
#### Nitpicks (Optional)
- Minor style/formatting notes
### Deferred Items (Follow-Up Issues)
| Suggestion | Issue | Rationale for deferral |
|------------|-------|------------------------|
| ... | [#XX](issue_url) | ... |
### Architecture Notes
How this change fits within the project architecture.
### Verdict
- [ ] Approve - Ready to merge
- [ ] Request Changes - Issues must be addressed
- [ ] Comment - Feedback only, author decidesStructured result (machine-read)
The reviewer's final report MUST end with a fenced ```json review-result block
conforming to review-result.py schema (#267) — placed immediately before the mechanical
**Status:** block, which stays the actual last thing in the message per the global
end-of-message convention. This block is what makes the verdict and findings queryable across
runs instead of existing only as prose:
{
"kind": "review-result",
"verdict": "approve",
"body_matches_tree": true,
"findings": [
{
"severity": "critical",
"title": "one line, matches the table row above",
"file": "path/to/file.ext",
"line": 42,
"evidence": "the command output or quoted lines that support this — not a restatement of the title",
"mutation_ran": true,
"red_line": false
}
],
"pr": 123,
"repo": "owner/repo",
"skill": "agent-review",
"round": 1
}If more than one such block appears in the report (a draft superseded by a final revision), the LAST one is what gets recorded — always end with the one that should count.
One line per field:
- kind — always the literal string
"review-result"; the positive marker that tells this apart from any other verdict/findings-shaped dict (review-result.py list --near-missesis what a missingkindshows up as). - verdict —
approve/request_changes/comment; matches the checkbox ticked above. - body_matches_tree — bool or null; see the rule below. Never left
trueby default. - findings — every row from Critical/Suggestions/Nitpicks above, one object each.
- severity —
critical/suggestion/nitpick, matching which table the row came from. - file / line — where the finding lives, or
nullwhen it doesn't apply. - evidence — the command output or quoted lines that back the finding.
- mutation_ran — see the rule below.
- red_line — true only for a hard rule violation no severity downgrade can waive (a secret,
an attribution miss, a protected-branch bypass); independent of
severity. - pr / repo / round — context;
roundincrements on a re-review of the same PR.
Two rules, non-negotiable:
mutation_ranis true only if the reviewer actually RE-RAN the cited check or mutation in this review. A finding asserted from reading the diff, however confident, isfalse— this field is what letsreview-result.py listcount proven findings separately from asserted ones.body_matches_treeis true only after comparing the PR body's claims against the diff. A review that never checked the body (or the PR carries none) reportsnull, never a default-true.
This repo's reviewer runs as a subagent (spawned via the Agent tool from /full-review), so the
orchestrator records the block after the agent returns — see step 7 below — instead of the agent
recording its own result.
4. Post Review on PR
Post review as a PR comment using heredoc:
gh pr comment ${PR_NUM} --body "$(cat <<'EOF'
## Code Review: PR #XX
[Your review content here]
EOF
)"5. Create Follow-Up Issues for Deferred Items
MANDATORY: For any suggestion or nitpick that is valid but out of scope, create a tracked GitHub issue.
Never leave deferred items as just review comments. If it's worth mentioning, it's worth tracking.
Classify impact before choosing a disposition. Read inline and general-summary findings; a narrative location does not exempt a finding from acceptance review. Blocking correctness, security, data-integrity and promised-acceptance defects require a fix, removal or verified authorized containment before a clean verdict. A >15-minute estimate or follow-up issue URL is not a fix. If the repair is outside authority, unavailable or exceeds the shared remaining allowance, keep request_changes, name the blocker and return it to the coordinator.
Honor explicit holds and the caller's durable correction/budget record. A new review agent, nested skill or resumed context grants no extra repair allowance or merge authority.
# Labels: enhancement, from-review (add complexity: low|medium|high or testing when they apply)
ISSUE_URL=$(gh issue create \
--title "Short descriptive title" \
--label "enhancement" \
--label "from-review" \
--body "$(cat <<EOF
## Context
Filed from: #${PR_NUM}
Identified during review of PR #${PR_NUM}.
## Description
What needs to be done and why.
## Original Review Comment
> Quote the review finding here
## Acceptance Criteria
- [ ] Criterion 1
- [ ] Criterion 2
EOF
)")CRITICAL: Every follow-up issue MUST be linked in the posted PR review comment. The Deferred Items table must contain the full issue URL (e.g., https://github.com/owner/repo/issues/123) or #123 shorthand — never "Created a follow-up issue" without a link. The issue URL is the paper trail that makes the deferred item discoverable from the PR.
6. Reconcile Issues Resolved in This PR
After all fixes are committed, check whether any issues created during this review — or pre-existing from-review issues — were already addressed by fixes in this PR.
# List open from-review issues
gh issue list --label "from-review" --json number,title,body
# For each issue resolved by a fix in this PR:
gh issue comment ${ISSUE_NUM} --body "Addressed in PR #${PR_NUM} — ${DESCRIPTION}."
gh issue close ${ISSUE_NUM}RULE: Every closed issue MUST reference a PR. The comment is the paper trail. No silent closes.
7. Report to User
Output a summary table followed by details. The table is the PRIMARY output — it must be scannable at a glance.
| PR | Verdict | Findings | Issues |
|----|---------|----------|--------|
| #XX | Approve / Request Changes | N critical, M suggestions, P nitpicks | Created: #A, #B. Closed: #C |Column guide:
- Verdict:
Approve,Request Changes, orComment - Findings: Count by severity (omit categories with 0 count)
- Issues:
Created: #X, #Yfor new follow-up issues.Closed: #Zfor resolved from-review issues.—if none.
Then below the table, list:
- Brief summary of critical issues (if any)
- URLs for all created/closed issues
- Link to posted review comment
Capture the result. When this review ran as a subagent (spawned via the Agent tool), the
orchestrator records its structured result — the agent's full report, including the
review-result block from step 3, fed on stdin — beside the subagent sidecar: python3
~/.claude/scripts/review-result.py record --agent <id> --skill agent-review --pr ${PR_NUM}
<<'EOF' … EOF (the agent's report between the markers, or < report.md if it was saved to a
file first), where <id> is the agent id the Agent tool result reported.
If the orchestrator never runs record (the agent's report was only read, not captured),
review-result.py harvest recovers the block from the transcript afterward — see
/session-lifecycle. When this review ran inline rather than as a subagent, there is no
sidecar to record beside; the block in step 3 still satisfies the structured-review-result
guard on this skill.
Agent Persona
Chroxy Inspector — expert in Node.js, React Native/Expo, WebSocket protocol, Claude Code CLI, Cloudflare tunnels, mobile connectivity.
Mindset: "Will this code work reliably over a cellular connection through a tunnel to a remote dev machine?"
You are an expert code reviewer with deep knowledge of the project's tech stack. You review with the mindset of reliability, maintainability, and correctness.
When this persona runs as a spawned subagent (fresh context, no CLAUDE.md), use the repo-memory MCP to gather surrounding context cheaply: get_file_summary/batch_file_summaries and get_dependency_graph for the files the diff touches (blast radius for a fraction of a full Read), search_by_purpose to locate related code the diff should have updated — then Read the changed regions in full to review the actual lines.
Review Philosophy
- Be constructive - Suggest fixes, not just problems
- Respect the architecture - Changes should follow established patterns
- Pragmatic over perfect - Working code first, polish later
- Reliability first - Always consider error recovery and edge cases
- Keep it simple - No over-engineering, no premature abstractions
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.

