code-review-gate-skill runs four review probes in fresh contexts for Claude Code

Many teams run a coding agent on a change. They then ask the same agent to review the work. A new skill argues the review must run in a fresh context. The review has no memory of writing the code. It has four parts. The four parts cannot vote for one another. The repository kyy-logs/code-review-gate-skill went live on 29 October. It had 445 stars on 9 November. That is up from zero 11 days before. The project is released under the MIT licence.

What the four probes do

The README lists four separate review passes. The behavior probe asks one question. Does the changed behavior match its contract and the real path through the code? The test coverage probe asks whether the tests mean something. It also asks whether incremental coverage was measured. The repo guard probe asks whether the change breaks a repository or project rule. The impact probe asks who calls this code, which contracts change, and how wide the blast is. Each probe runs on its own. There is no shared memory.

Every finding shows its evidence

The author sets a strict contract for a finding. Each one must include a location, a summary, the evidence behind it, a severity and a status. The README reads: "high-severity findings must cite the actual method signature and the contract they violate, or a concrete code path."

A hunch is capped at medium. It stays there until someone checks it. The severities are high, medium and low.

The merge gate reads JSON, not prose

The project calls itself a review a team can audit. The README says "every gate decision is rendered from machine-readable JSON."

Counts and blockers come from the JSON files the probes wrote. Nothing in the gate reads English a model might have retyped. A build server can read the same files. It can then apply the same rules. That matters when a review is the only thing between a pull request and merge.

What the author claims against plain AI review

The README states: "No self-grading. The implementation context does not sign off on its own work." The author's point is simple. One chat window reviewing the code it wrote has already committed to the design. A fresh context has not. Four probes with no shared memory also mean a bug has to clear four reviewers, not one.

How to try it

The project targets Claude Code with Agent Skills support. It also targets Codex with custom-agent support. A team already using either can drop the skill into its agent directory. The skill has 445 stars and 6 forks on 9 November. It was created on 29 October. The traction is real for its age. The method is still one person's claim. A team should run it on a recent pull request first. Read the JSON before putting the gate on the main branch.

Source

• GitHub, kyy-logs/code-review-gate-skill