Skip to content

Quick Start

After installing the plugin, here's how to run your first review in under a minute.

Basic code review

In a Claude Code session, point it at a file or directory:

/adversarial-reviewing src/auth/

This runs all 5 specialists (Security, Performance, Quality, Correctness, Architecture) with default settings: 2 self-refinement iterations, 350K token budget.

Quick mode

For a fast check with fewer specialists:

/adversarial-reviewing src/ --quick

Quick mode uses 2 specialists (Security + Correctness), 2 iterations, and a 150K budget.

Thorough mode

For a deep review:

/adversarial-reviewing src/ --thorough

Thorough mode uses all 5 specialists, 3 iterations, and an 800K budget.

Save the report

/adversarial-reviewing src/ --save

Writes the report to docs/reviews/YYYY-MM-DD-<topic>-review.md.

Strategy document review

/adversarial-reviewing artifacts/strat-tasks/ --profile strat

Runs 6 strategy specialists and produces per-document verdicts (Approve/Revise/Reject).

What happens during a review

  1. Phase 1: Each specialist analyzes the code independently in isolation, self-refining through 2-3 iterations until findings converge
  2. Phase 2: Specialists debate each other's findings through mediated challenge rounds
  3. Phase 3: The orchestrator resolves disagreements, deduplicates, and classifies agreement levels
  4. Phase 4: A structured report is generated with findings by severity and remediation roadmap
  5. Phase 5 (with --fix): Findings are classified, Jira tickets drafted, and fixes implemented via worktree branches

Each phase has programmatic validation: structure checks, injection detection, budget tracking, and convergence detection.