Code Review Gate: a review that has to show its evidence
Code Review Gate is a review skill by GitHub user kyy-logs for Claude Code and Codex. Four probes read your change in separate contexts, every finding has to cite where and why, and a script decides from their JSON records whether the change may merge.
npx -y skills add kyy-logs/code-review-gate-skill --agent claude-code --skill code-review-gate --yesType /code-review-gate after Claude Code or Codex has changed your code, and the skill reviews the diff before you commit. The agent that wrote the code doesn't get to approve it: three probes, four with the optional one, each read the change in a fresh context and write their own record. GitHub user kyy-logs publishes it under the MIT license. Anthropic doesn't publish it. For how skills load, see what Claude skills are.
The four probes
Three run on every review and one is optional, per the README's probe list:
- behavior-review checks whether the changed code still does what its contract says.
- test-coverage asks whether the tests prove anything, and measures new-line coverage when a report exists.
- repo-guard looks for breaks of your project's own rules.
- impact-analysis (optional) lists who calls the changed code and which contracts move with it.
The skill never edits the code it reviews, and it doesn't need a build or a test run to start. If tests failed or never ran, the review still happens and the report says so.
What counts as a finding
Each finding carries a location, a summary, evidence and a severity of high, medium or low. A high finding must name a method signature and the contract it breaks, or a concrete code path; a hunch nobody verified is capped at medium. Before the report closes, every high finding has to end as reconciled or marked a false positive.
The merge is blocked when a required probe errors, a high finding stays open, or coverage you supplied falls below 85% of lines, the default target. Edit the code after the review and the gate answers stale-review until you run it again.
How to install it
Run this inside your project; for Codex, swap claude-code for codex:
npx -y skills add kyy-logs/code-review-gate-skill --agent claude-code --skill code-review-gate --yes
It lands in .claude/skills/code-review-gate/. The scripts need Bash, Git, Python 3.9+ with PyYAML, and Node.js 22.20+ for the installer. A Chinese version installs as code-review-gate-zh, one language per project.
Example request
Use code-review-gate on commits [base commit]..[head commit]. The acceptance criteria are: [what the change must do]. Project rules are in [path, e.g. .code-review/rules/].
Leave out the commit range and it reviews your staged, unstaged and untracked changes. You get code-review.md plus one JSON file per probe in docs/work/active/, with the verdict on whether anything blocks the merge.
Limits
- Same model, every probe. The README calls it separation of duties, not an independent judge: every probe runs on the model you're already using.
- A pass is not approval. In the README's words, "A passing gate is not human approval, CI, or proof that no issues remain."
- The built-in rules lean Java. repo-guard scans .java, .ts, .tsx and .vue files, with Java and Spring rules; other stacks need rules of your own, and a project rule package replaces the built-in one rather than adding to it.
When to use it
Use it on a change an agent wrote that you'll merge without reading every line, when you want a record of what was checked. For a quick bug pass on a branch or a pull request with no gate or JSON records, use Claude Code's own /code-review instead.
Sources
- kyy-logs/code-review-gate-skill README on GitHub, accessed November 2026
- code-review-gate SKILL.md on GitHub, accessed November 2026