How does this gate differ from just asking Claude to review my code in a chat?

It plugs into your git workflow: a chat assistant can only comment, while this gate can stop a git push when it finds a confident high-severity bug. The block/warn/pass verdict is decided by auditable code, not by the model, and the review runs on your own Claude subscription inside Claude Code.

What are the technical requirements for running the Claude Code Review Gate?

The skill is specifically designed for the Claude Code CLI environment and requires Python 3 and Git to be installed on your local machine to manage the hook integration.

Will this skill block my workflow with too many false positives?

Two steps keep false alarms down: one reviewer looks at the whole change set, then a second subagent that never saw its reasoning tries to disprove each blocking finding and drops it only on direct counter-evidence. Only confident high-severity findings block; everything else is reported and the push goes through. OCR_ADVISORY=1 makes it advisory-only.

What happens if the review process fails or times out during a push?

If the AI analysis fails due to a timeout or crash, the gate defaults to a 'fail-closed' state, meaning it will block the push to ensure no unverified code is accidentally moved to your remote repository.

What exactly is included?

The full Claude Code plugin (review skill, reviewer/filter/resolver subagents, hooks and scripts) bundled with a SKILL.md that walks you through installing it. The SKILL.md also documents exactly what the plugin runs, reads and writes.

claude-code-review-gate: blocking AI code review

Example session with this skill installed

Push my branch, but block it if the review finds any high-severity bugs.

Read your context and instructions
Compiled the claude-code-review-gate

I ran the review gate on your unpushed commits. It found a confident high-severity bug, so the push was blocked:

$ git push
[review-gate] review-gate blocked this push (high-severity issues):
  [high] db.py:6 - SQL injection: `user_input` is concatenated directly into the
         query string; an attacker can pass `' OR '1'='1` or `'; DROP TABLE users; --`.
  [medium] db.py:10-11 - the sqlite3 connection is never closed; callers leak the handle.

Fix the issues, or bypass with: git push --no-verify

Finding	Severity	Effect on the push
SQL injection, db.py:6	high	blocks
Connection never closed, db.py:10-11	medium	reported only

Next steps

Switch the query to a parameterized one (cursor.execute("... WHERE name = ?", (user_input,))).
Close the connection with a with block.
Push again: the gate re-reviews and returns verdict: pass, so the push goes through.

Connects securely to your tools. The creator never sees your data.

About this skill

Most AI reviewers comment. This one can say no.

claude-code-review-gate reviews your diff, staged changes or a whole repo, and gates every git push. Confident high-severity findings block the push; everything else is printed and the push goes through. Review runs inside Claude Code on your own Claude subscription, with no API key and no third-party binary.

What it does

On-demand review of the working diff, staged changes, or a full-file scan with /review-gate:review.
A blocking push gate that reviews every unpushed commit once per push. By default it gates pushes made through Claude Code; an optional installer extends it to every push (terminal and IDE too).
One reviewer for the whole change set, so cross-file bugs are visible: a symbol renamed in one file and stale in another, or a guard removed in A but still assumed by B.
An independent falsify pass: a second subagent that never saw the reviewer's reasoning gets only the diff and the findings, and drops a finding only with direct counter-evidence.
A deterministic verdict (block / warn / pass) decided by auditable code, not the model's discretion.
Fails closed on timeout, crash or unparseable output: a broken gate blocks rather than waving changes through.

Built for long, interruptible work

Checkpointed, content-addressed chunks mean a crash or usage limit costs at most one chunk. A per-file ledger makes a no-change re-push cost zero model calls, and a resolver accepts "fixed" only when the code it quotes exists at the tip.

Escape hatches

git push --no-verify skips one push, OCR_ADVISORY=1 prints findings without blocking, and OCR_FAIL_OPEN=1 is the emergency bypass.

Backed by 389 tests, with the review methodology adapted from open-code-review.

How to use

Skills on Agensi run inside your AI agent. Copy the prompt, paste it into your agent, and it connects to Agensi and installs this skill for you. If a payment is needed, your agent shows it to you first.

Reviews

No reviews yet

Be one of the first to try it. Every listed skill passes our trust checks below.

Security scanned

Passed our 8-point scan before listing

Fresh listing

Recently published to Agensi

Money-back guarantee

Doesn't do what its page promises? You get your money back.

claude-code-review-gate: blocking AI code review