AI Code Review Triage
Split reviewing AI-written PRs into catching bugs, aligning context, and owning the sign-off.
- Source repo
- laulauraluo/looks-good-to-ai
- Stars
- ★ 108
- Last updated
- 3d ago
- License
- MIT
- Primary language
- Python
- FA score
- 50/100 · Major gaps
At a glance
- How it runs
- Works with
- Platform-specificClaude Code
- Cost
- Free software; you pay for model usage
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- Agent-written PRs are piling up and you need to decide fast which can merge and which need real scrutiny
- Not a fit if
- Teams not on Claude Code (no native Codex, Copilot, or Cursor integration)
- Teams wanting a scanner that auto-merges PRs without human sign-off
- Users reviewing loose code snippets rather than full pull requests
- Maintenance
- Commits in the last 30 dayslast commit 2026-11-18latest release v2.0.0 (2026-11-17)+4.4 ★/day over 7 days
- Source review
- 50/100 · Major gaps
What does this agent do, and when should you use it?
AI Code Review Triage is a Claude Code skill for reviewing pull requests written by AI agents. Instead of one pass of line-by-line reading, it splits review into three jobs: a fresh-context sub-agent hunts for bugs and adds boundary tests or repros; a human checks magnitude, history, neighboring callers, and whether the change matches the original request; and the approver states an explicit verification scope with detection and rollback plans. A Step 0 risk tier decides how much runs — low-risk PRs only get Step 1, while production behavior, data, security, or money changes run all three. The output is a Markdown review report with confirmed defects, context flags, and a recommendation such as request changes. The repository ships SKILL.md, plugin and marketplace manifests, and two runnable walkthroughs covering timeout units and a pagination cap that silently drops rows. It organizes the review and the sign-off decision; it does not replace linters, scanners, or human judgment.
The skill lives in skills/ai-code-review-triage/SKILL.md and is triggered by Claude Code when you ask it to review an AI-generated PR. Execution runs as a flowchart: Step 0 assigns a risk tier; Step 1 hands the diff to a fresh-context sub-agent for an independent pass, which adds boundary tests or a reproduction and marks findings as [CONFIRMED]; Step 2 compares magnitudes, commit history (a change reverted before), neighboring callers, and requirement scope; Step 3 requires the approver to explain what the PR actually does, then records detection signals (login-rate and 401-rate alerts), a rollback path, and what was verified versus not. The result is a Markdown report with a risk label, a bug list, context concerns, and a recommendation. Two example files show the full flow end to end: SESSION_TIMEOUT going from 30 to 30000 so sessions expire in 30 seconds, and a page limit moving from 100 to 1000 so an exporter stops with 150 of 250 rows missing.
- Agent-written PRs are piling up and you need to decide fast which can merge and which need real scrutiny
- Tests pass and the diff looks tidy, but you suspect a unit, boundary, or contract problem underneath
- A request targeted only one user group (e.g. "remember me" users) and you want to confirm the diff doesn't change everyone
- A piece of code was reverted once before and you want to check history and callers such as the mobile client before approving
- You must tell the team exactly what your approval covered and have a rollback ready if it breaks
How do you install or deploy this agent?
As a plugin inside Claude Code:
/plugin marketplace add laulauraluo/ai-code-review-triage
/plugin install ai-code-review-triage@ai-code-review-triageAs a standalone skill copied into your user-level skills directory:
git clone https://github.com/laulauraluo/ai-code-review-triage.git
cp -r ai-code-review-triage/skills/ai-code-review-triage ~/.claude/skills/Prerequisite: an installed and signed-in Claude Code, since model calls run on your Claude account. The repository declares no other dependencies.
How do you use this agent?
No manual invocation is needed — ask Claude Code in plain language and the skill triggers itself:
Review this AI-generated PR before I merge it.I'm about to approve #123. What should I verify first?It returns a review report in Markdown. To see the expected output shape first, read the bundled walkthroughs:
cat examples/session-timeout-walkthrough.md
cat examples/pagination-limit-walkthrough.mdWhat are this agent's strengths and limitations?
- A fresh-context sub-agent must add a test or repro, so the reviewing model isn't grading its own code
- Step 0 risk triage keeps low-risk PRs to a single step and saves the full three-step pass for production, data, security, or money changes
- Context alignment is its own step, targeting magnitude, revert history, hard-coded callers, and requirement scope that line-by-line reading misses
- Every approval must name detection, rollback, and a verification scope, turning sign-off into a bounded commitment
- Two concrete, runnable walkthroughs (timeout units; a pagination cap losing rows) make the value tangible instead of abstract
- Installation and triggering are documented for Claude Code only; Codex, Copilot, and Cursor are described as workflow-compatible but have no native integration
- It explicitly does not replace linters, security scanners, or human review, so you still wire those into Step 1 yourself
- Step 3 depends on the approver actually understanding the PR; that judgment cannot be automated
- Both examples are invented scenarios — no benchmarks or evidence from real large codebases are provided
- At version 1.0.0, there is no evidence yet of how actively the skill and its checks will be maintained
How does this agent compare with similar options?
The README positions it alongside, not against, Claude Code's built-in /code-review and specialized security, performance, and over-engineering checklists, which you run inside Step 1. No other named alternatives appear in the source.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| AI Code Review Triage This agent | 50 · Major gaps | Agent plugin / skillFree + model costs | ★ 108 | 3d ago | Python | Claude Code |
| Kodus | 52 · Major gaps | Hosted serviceFreemium | ★ 1.6k | today | TypeScript | OpenAI API · Claude API |
| Vet Code Change Verifier | 47 · Major gaps | CLIFree + model costs | ★ 474 | 2mo ago | Python | Codex · Claude Code · OpenAI API · Claude API |
| DeepSeek Harness for GitHub | 85 · Good | Chat botFree + model costs | ★ 64 | 20d ago | TypeScript | — |
How does FollowAgents rate this agent?
Why each dimension lost points
The skill is essentially prompt/process documentation with no network, credential, or write permissions requested; Step 1 explicitly mandates an independent sub-agent with a neutral prompt, a good least-privilege practice (least_privilege=2). However, SKILL.md does not require user confirmation before running git log/grep or spawning sub-agents, and gives no data-flow statement about which diffs/code reach sub-agents or external tools; sensitive-data handling and dependency security are unaddressed; rollback appears only as a config-revert example, not a product mechanism — hence 1s. Source attribution is present via README acknowledgments and plugin.json author (2).
The three-step flow is consistent across README and SKILL.md, and risk tiers and report structure are self-consistent (self_consistency=2). Dependencies (Claude Code Agent tool, git, grep, test/lint toolchains) are assumed present with no availability check or degradation path; failure messaging only distinguishes CONFIRMED/PLAUSIBLE and lacks handling for missing tools or unavailable sub-agents — each 1.
The SKILL.md description lists trigger phrases and scenarios and directs plain security/performance checklists to a general skill, giving good boundaries and trigger precision (2 each). Audience/scenario coverage spans low/medium/high risk tiers (2). Environment fit is Claude Code only; README states no native integration with other agents, so environment_fit=1.
README is well structured with install, usage, repo layout, and FAQ links — information architecture and install notes each 2. Naming is consistent across plugin.json, marketplace.json, and SKILL.md (2). Two walkthrough examples exist (2). Known limitations are only scattered, not consolidated (1). LICENSE is full MIT (2). Version 1.0.0 exists but no CHANGELOG (1). Maintenance responsibility rests only on the author field, with no support channel or update policy (1).
The report template is clear and directly usable for review decisions (output_usability=2). It offers genuine marginal value over line-by-line review (2). Cost-benefit is weaker: Step 1 needs an extra sub-agent and Step 2 needs multiple git queries, with no quantified benefit or cost discussion (1).
README claims two runnable walkthroughs but explicitly states the scenarios are invented, so real effectiveness cannot be verified (claim_traceability=1). Cross-source corroboration is limited to README and SKILL.md echoing each other (1). Fact/inference separation is good: CONFIRMED vs PLAUSIBLE is distinguished and examples are labeled fictional (2).
- SKILL.md is truncated at Step 3, so the full sign-off step and its safety constraints cannot be confirmed.
- No statement on whether diffs, code, or repo contents are passed to sub-agents or external tools — data flow is opaque.
- Depends on Claude Code's Agent tool, git, grep, and test toolchains without availability checks or fallback when missing.
- Both walkthrough examples are invented scenarios and cannot serve as evidence of real effectiveness.
- Publisher identity is unverified by FollowAgents; maintenance responsibility and update path are unclear.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v2.0.0 v2.0.0 — Looks Good to AI (formerly AI Code Review Triage)
Release notes excerpt
Looks Good to AI v2.0.0 (formerly AI Code Review Triage) The project has a new name. The repository, skill, plugin and Gemini extension are now looks-good-to-ai. GitHub redirects the old repository URL, but an install under the old name has to be reinstalled. Install ```sh npx skills add laulauraluo/looks-good-to-ai ``` | Agent | Command | |---|---| | Claude Code | /plugin marketplace add laulauraluo/looks-good-to-ai, then /plugin install looks-good-to-ai@looks-good-to-ai | …View on GitHub ↗ -
v1.2.0 v1.2.0 — native installs for Codex, Copilot CLI and Gemini CLI
Release notes excerpt
AI Code Review Triage v1.2.0 Runs in any agent that loads Agent Skills, and now installs natively in four of them. Install ```sh npx skills add laulauraluo/ai-code-review-triage ``` Or with the agent's own installer: | Agent | Command | |---|---| | Claude Code | /plugin marketplace add laulauraluo/ai-code-review-triage, then /plugin install ai-code-review-triage@ai-code-review-triage | | Codex | codex plugin marketplace add laulauraluo/ai-code-review-triage, then codex plugin add ai-code-review-triage@ai-code-review-triage | …View on GitHub ↗ -
v1.0.0 v1.0.0 — three-job review for AI-written PRs
Release notes excerpt
First release. - ai-code-review-triage skill: triage → catch bugs (fresh-context sub-agent) → align context → own the sign-off - Installable as a Claude Code plugin: /plugin marketplace add laulauraluo/ai-code-review-triage - End-to-end example: [session-timeout walkthrough](examples/session-timeout-walkthrough.md) - 中文 READMEView on GitHub ↗
FAQ
Does it merge or approve PRs automatically?
Can it review code written by Codex, Copilot, or Cursor?
Does every PR need all three steps?
How does it relate to linters and security scanners?
/code-review or security checklists inside Step 1 instead of assuming the skill covers them.What does it cost?
Is AI Code Review Triage open source?
Is AI Code Review Triage still maintained?
Who makes AI Code Review Triage?
Thanks — we use this to improve the page.