Prototype adversarial review workflow as a cmux skill

[teamleaderleo]

Summary

• add a cmux-review agent skill that turns review into a staged protocol: review brief → intent compliance → independent discovery → triage → adversarial challenge → executable verification → optional repair
• add a machine-readable local review receipt schema for preserving requirements coverage, findings, refutations, evidence, and dispositions without committing review artifacts
• add a Claude /review command that invokes the same repo-owned workflow

Why this first

This intentionally prototypes the review semantics and UX before native model/runtime integration.

cmux already has diff comments, review workspace state, Vault checkpoints/forks, repo review rules, and agent skills. A skill lets the team dogfood:

• whether the review brief actually reduces review time
• whether task/requirement compliance catches "clean implementation of the wrong thing"
• whether the severity/publication threshold keeps noise low
• whether adversarial challenge eliminates enough false positives
• what evidence types actually occur in practice
• which receipt fields survive real review sessions

Those traces can then drive the native review-ledger/socket design instead of freezing a schema from a whiteboard.

Core principle:

Spend compute freely on investigation; spend developer attention reluctantly.

Evidence policy

Model agreement is supporting signal. Reproduction, executable checks, verified call paths, and concrete counterexamples carry more weight.

Generated tests are treated as claims too: a passing generated test becomes strong repair evidence only when the workflow establishes that it actually distinguishes defective and repaired behavior.

Receipt behavior

Review receipts live under the repository's Git metadata path returned by:

git rev-parse --git-path cmux/reviews

They carry exact base/head/diff identity plus review policy/ruleset identity and must never enter source control.

Testing

• parsed the added JSON Schema successfully after the requirement-coverage additions
• Markdown/agent workflow only; no app runtime or build changes

Part of #13510

Summary by CodeRabbit

• New Features
  • Added an adversarial review workflow for evaluating code changes.
  • Added structured review receipts that record findings, evidence, verification status, and review metadata.
  • Added guidance for identifying, challenging, and verifying potential defects before reporting them.

[coderabbitai]

Navigate logical layers of code changes, visualize relationships, and explore their blast radius.

Note

Currently processing new changes in this PR. This may take a few minutes, please wait...

⚙️ Run configuration

Configuration used: Repository: manaflow-ai/cmux/.coderabbit.yaml

Review profile: ASSERTIVE

Plan: Advanced

Run ID: 07aaa1ba-c11f-484b-a573-6445537d66b7

📥 Commits

Reviewing files that changed from the base of the PR and between 9e4b77c and 4d06724.

📒 Files selected for processing (3)

• .claude/commands/review.md
• skills/cmux-review/SKILL.md
• skills/cmux-review/references/review-receipt.schema.json

 _______________________________________________________________________________________
< Analyze workflow to improve concurrency. Exploit concurrency in your user's workflow. >
 ---------------------------------------------------------------------------------------
  \
   \   \
        \ /\
        ( )
      .( o ).

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Commit to this branch
• Create a new PR

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[teamleaderleo]

@codex review
@coderabbitai review
@greptile-apps review

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.

[coderabbitai]

@teamleaderleo I will review the changes in #13517.

⚠️ Action not completed

Pull request is closed.

Note: CodeRabbit is an incremental review system and does not re-review already reviewed commits. This command is applicable only when automatic reviews are paused.