SuperReview

An open-source code review skill

A review you can
trace back to code.

Give your coding agent a pull request. Get findings with a concrete trigger, supporting evidence, and a clear account of what was reviewed.

Install SuperReview Explore the source

Codex · Claude Code · Your existing agent account

From change to findingSynthetic example
session.py Changed line 2 
  def load_session(storage, tenant, key):
−     return storage.find(tenant=tenant, key=key)
+     return storage.find(key=key)
P1

Session lookup loses the tenant boundary

Trigger
A caller requests a session belonging to another tenant.
Evidence
The changed lookup filters by key alone.
Countercheck
The example’s storage contract has no implicit ownership filter.
Reasoned findingLocal report

Inside your existing workflow

A PR in. An inspectable review out.

  1. 1

    Choose the change

    Use a GitHub PR URL, PR number, or local commit range. The helper pins the base and head so the review has a precise scope.

  2. 2

    Examine the evidence

    The agent traces behavior, checks callers and contracts, and tries to disprove each candidate finding before reporting it.

  3. 3

    Inspect the result

    Read the findings and coverage gaps. Large reviews save checkpoints, so you can resume without starting over.

Get started

One install.
Your next pull request.

Install in your repository, then ask your coding agent to review a pull request. SuperReview works in the current session.

Git · Python 3.9+ · Node.js 22.20+ for npx

Install without Node.js
Choose your agent
npx skills add OmniTensorLabs/SuperReview --skill superreview --agent codex

Then ask your agent

Review PR 42.

A PR URL or “Review this branch against main” works too. The agent can select SuperReview automatically. To choose it explicitly, ask “Use SuperReview to review PR 42.”

Codex: $superreview review PR 42
Claude Code: /superreview 42
Restart your agent if the skill is missing. Reports stay local unless you ask to publish them.

Adopt with evidence

Know what the review establishes.

Visible limits

Validation checks report structure, changed-line locations, and declared coverage. It cannot prove that a finding is correct. Missing context stays visible in the report.

Production guidance

Measure on your codebase

Use the optional Multivon integration to score saved, human-adjudicated reviews. Track supported findings, known defect recall, and unscored runs.

Evaluation walkthrough