adversarial-code-review

Multi-perspective adversarial code review with git-isolated worktrees. Two independent reviewers (Architect + Inspector) each produce JSON findings, two cross-review passes pressure-test them, and a synthesis rapporteur collapses everything into a single ranked report.

For Hermes Agent, Claude Code, Codex, or any LLM CLI.

How it works

ARCHITECT ──→ reviews code (architecture, security, concurrency)
INSPECTOR ──→ reviews code (bugs, edge cases, error handling)
CROSS_1 ────→ challenges INSPECTOR findings with ARCHITECT perspective
CROSS_2 ────→ challenges ARCHITECT findings with INSPECTOR perspective
SYNTHESIS ──→ ranks, cross-validates, produces final report

Comparison

Feature	adversarial-code-review	adverse (addyosmani)	alecnielsen/adversarial-review	agent-review-panel
Cross-model debate	✅ Architect↔Inspector	❌ Single-reviewer	❌ Single-round	✅ 4-6 panel
Git worktree isolation	✅	❌	❌	❌
Cross-review rounds	✅ 2 rounds of devil's advocate	❌	❌	❌
JSON findings with schema	✅	❌	❌	❌
--project-dir mode	✅ Review existing codebase	❌	❌	❌

Quick start

# Review changes on a branch
python3 scripts/adversarial_review.py --diff-git

# Review a whole project directory
python3 scripts/adversarial_review.py --project-dir /path/to/project \
  --a-cmd "claude-tmux --model best" \
  --b-cmd "codex exec -C /path/to/project"

Output

Artifacts land in --out (default .adversarial-review):

• final.json — machine-readable verdict (APPROVE|REQUEST_CHANGES|REJECT)
• review.md — ranked report with per-finding evidence
• 01_architect.txt … 05_synthesis.txt — per-phase raw output

Dependencies

• Python ≥ 3.11
• Git ≥ 2.5
• Two LLM CLIs (one for architect, one for inspector)

Uses adversarial-common as the shared engine.

License

0BSD — see LICENSE.