adversarial-review-gate

adversarial-review

Multi-agent adversarial review roundtable for coding agents.

The roundtable organizes an adversarial debate between specialist review seats. Each seat inspects material through an independent lens and challenges findings from other seats. A finding survives only when it defeats counter-arguments. A clean-context judge delivers the final ruling.

When to Convene

Use the roundtable when an incorrect decision causes high rework costs:

• Specifications and architecture designs
• Multi-step implementation plans
• Complex code changes and pull requests
• Root-cause investigations for difficult bugs

Do not use the roundtable for one-line changes or routine questions.

Supported Hosts

The installer configures adversarial-review for these agent hosts:

• Claude Code (claude-code)
• OpenCode (opencode)
• Codex (codex)
• Gemini CLI (gemini)

Install

The npm package is adversarial-review-gate. The package gives one command: adversarial-review.

CAUTION: With npx, always type the full package name adversarial-review-gate. The npm name adversarial-review alone belongs to a different package.

Install the command globally:

npm install -g adversarial-review-gate
adversarial-review --version

You can also run each command without an install:

npx adversarial-review-gate --version

The same package is on GitHub Packages as @louisphamdev/adversarial-review-gate.

For Claude Code, you can use the plugin instead:

/plugin marketplace add louisphamdev/adversarial-review
/plugin install adversarial-review@adversarial-review

The examples below use the global command. If you use npx, write npx adversarial-review-gate in place of adversarial-review.

To install the skill for your host:

adversarial-review install --host claude-code

To install the skill for a different host:

adversarial-review install --host opencode

To inspect your configuration and tools:

adversarial-review doctor

Remove Version 2 Hooks

Version 3 removes the automatic background hooks from version 2. If you used version 2, remove the legacy hooks from your settings:

adversarial-review uninstall --v2-hooks

If you installed version 2 globally, include the global flag:

adversarial-review uninstall --v2-hooks --global

How to Run

1. Select a Route

The roundtable provides two execution routes:

• spawn: Runs sub-agents with your current host CLI.
• swarm: Runs parallel lanes across headless coding agents.

To receive an automatic recommendation, run:

adversarial-review recommend

The command evaluates your quota and diff size, then recommends spawn or swarm.

2. Execute the Review

To run the roundtable on your current repository diff:

adversarial-review run --route auto

You can choose a specific route directly:

adversarial-review run --route spawn

To review a specific specification or plan file, pass the target path:

adversarial-review run --target docs/spec.md

3. Read the Results

The run produces a result.json file in the run state directory. Inspect these fields:

• gateVerdict: Displays PASS or BLOCKED.
• closingList: Lists all defects that you must fix before merge.
• gaps: Lists files and areas that seats did not inspect.

Never fix defects before the ruling. Wait until the judge issues the final verdict before you write code changes.

Review Stages

The roundtable advances through seven structured stages:

1. FIND: Seats inspect the material independently. Seats do not share initial findings.
2. TABLE: The lead compiles all findings and shares them with the full table.
3. DISPUTE: Paired seats debate contested claims.
4. LAST CALL: Seats state any final concerns before the scope freezes.
5. RULING: A fresh judge reviews surviving claims and issues the gate verdict.
6. PATCH REVIEW: Seats evaluate proposed code patches against their conditions.
7. VERIFY: Seats inspect the final diff lines to make sure that the fixes resolve the defects.

License

MIT