adversarial-review

v0.9.9: Cross-host adversarial review for coding workflows. Works in Claude Code, Codex, Pi, and Grok from the same source: detects which agent host you're running in and routes the heavy critique to the OTHER agent.

Critique discipline (v0.8.1): break confidence rather than validate; expensive attack surfaces first; material-only findings; mandatory simplicity counterfactuals; steerable focus; terse ship/no-ship summary. Lessons adapted from OpenAI Codex Companion and Matt Pocock's engineering skills.

Routing uses three tiers: T1 Claude and Codex review each other, T2 is the Pi chain (opencode-go DeepSeek V4.1 Flash, GLM 5.3, then Grok 4.7), and T3 is the Gemini Antigravity ladder as last resort. Reviewers get read-only tools. On host Claude, Codex-authored work is reviewed by the Claude main loop itself. Outside host Claude the Codex reviewer defaults to GPT-6 Astra medium; the architect selects high only for auth, sensitive data, migrations, concurrency, or changes spanning multiple modules. The Gemini ladder is gemini-3.8-flash-high, gemini-3.7-flash-high, then gemini-3.1-pro-high.

Simplicity counterfactual

Every plan and code review now asks whether the same required behavior and safety can be delivered with fewer concepts, layers, seams, entry points, or configuration surfaces. It checks YAGNI, pass-through layers, hypothetical adapters, duplicate ownership, and missed reuse of existing in-repo modules.

Effect is conditional, not a framework preference: when the repository already has a canonical Effect implementation for the concern, the reviewer challenges a parallel implementation. It does not recommend introducing Effect merely to reduce line count or enforce framework consistency. Fewer lines are a clue, never the verdict.

The principle

The partner reviews, never the host.

Two models examining the same artifact from different angles catch more issues than either alone. Each has different training biases, blind spots, and reasoning patterns. Disagreements between the two surface the highest-value findings - the ones a single reviewer would miss.

This plugin enforces that principle automatically:

• Running in Claude Code routes by author: Claude-authored work (default) uses Astra high via Pi first; Codex-authored work is reviewed inline by the Claude main loop, and the wrapper only adds T2/T3 opinions, never Codex.
• Running in Codex with a Codex, user, or unknown author uses claude -p first; Claude, Pi, or Grok authors use Codex Astra first. Then T2, then T3.
• Running in Grok uses Codex Astra, then Gemini.
• Codex Astra effort is medium by default. Set ADVERSARIAL_REVIEW_CODEX_EFFORT=high only for auth, sensitive data, migrations, concurrency, or changes spanning multiple modules.
• Antigravity is a one-shot review call with -p, --sandbox, and --mode plan.
• Everything unavailable enters degraded mode with an explicit banner.

How it works

SKILL.md (same file in both hosts)
   │
   ├─ host runs lib/call-external.sh
   │     │
   │     ├─ lib/detect-host.sh  (override -> env -> PPID walk)
   │     ├─ partner = NOT host
   │     ├─ ADVERSARIAL_REVIEW_DEPTH = 1  (anti-recursion guard)
   │     ├─ select ordered partners from host + artifact + author
   │     ├─ try Codex, Claude, direct xAI, or Pi in that route's order
   │     ├─ on fail -> Gemini Antigravity ladder
   │     └─ on fail -> degraded mode (exit 2)
   │
   ├─ host runs its own independent analysis (no peeking at partner output)
   ├─ cross-validate: tag findings [cross-validated] / [external-only] / [host-only]
   └─ return unified output (P0-P3 severity, evidence, recommendation, partner-attribution)

No haiku courier subagent (removed in 0.5.0 - added complexity without value). The main session does the dispatch and synthesis directly.

Skills

Skill	What it does
/adversarial-review:adversarial-review	Single entry point. Classifies input (plan, code, prompt) and runs the matching procedure: plan critique with revised plan, red-team code review with patch, or prompt-engineering analysis (external critique plus host adjudication).

In Codex, after running the install script, the same skills are available as $<skill-name> (Codex prompt-prefix convention).

Install

Claude Code

# Add the marketplace and install the plugin
claude plugin marketplace add ~/repos/skills/plugins/adversarial-review
claude plugin install adversarial-review@adversarial-review

# Reload in current session
/reload-plugins

Codex (additionally)

# Symlinks each skills/<name>/ into ~/.codex/skills/
bash ~/repos/skills/plugins/adversarial-review/adapters/codex-skill/install.sh

Verify both:

claude plugin list                 # should show adversarial-review enabled
ls -la ~/.codex/skills/            # should show the adversarial-review symlink back to this repo

Prerequisites

At least one external partner CLI must be authenticated for cross-host review to work (otherwise you'll get DEGRADED mode):

# When host=claude, partner=codex:
codex login

# When host=codex and author=codex, partner=Pi xAI OAuth Grok 4.7 first:
pi --version

# Host=codex direct fallback after Pi:
grok --version

# Next host=codex fallback:
claude  # interactive once to register OAuth, then `claude -p` works headless

# Optional final fallback:
agy --version  # Antigravity CLI, authenticated through the Antigravity app

For ChatGPT-account Codex users, also add to ~/.codex/config.toml:

forced_login_method = "chatgpt"

(Without this, codex exec returns 404 "Model not found" even though the TUI works. See `references/codex-integration.md