adversarial-review

name	adversarial-review
description	Get an independent second opinion from a separate AI model on code, specs, diffs, PRs, or GitHub issues. Use when the user asks for "second opinion", "adversarial review", "independent review", or "codex review". Supports subagent (fresh-context sub-Claude, default), interactive (tmux + codex), codex exec, claude -p, and AWS Bedrock backends. Arguments: [target] [--backend subagent|interactive|codex|claude|bedrock] [--model name] [--effort low|medium|high] [--quick]
license	MIT
compatibility	`subagent` needs nothing beyond Claude Code itself. `interactive` needs tmux and codex CLI. `codex` / `claude` / `bedrock` need their respective CLIs or AWS credentials. Requires git.
allowed-tools	Bash(codex:*) Bash(claude:*) Bash(gh:*) Bash(git:*) Bash(uv:*) Bash(tmux:*) Read Grep Glob Agent
metadata	{"author":"galsapir","version":"1.2.0"}

Independent second opinion from a separate AI model.

Workflow

1. Detect Target

Parse $ARGUMENTS:

• File path(s) → code review
• diff / changes → git diff review
• #N or issue URL → gh issue view review
• PR URL or PR #N → gh pr view + gh pr diff review
• No args → ask user via AskUserQuestion

Extract flags: --backend, --model, --effort (codex only, default medium), --quick.

2. Reconnaissance

Gather context before building the review prompt:

1. Read the target (files, diff, issue, or PR)
2. Identify related files (imports, tests, config) and project context (README, CLAUDE.md)
3. Assemble a 2-5 sentence context summary: what the project does, what the target accomplishes, key related files

3. Select Backend

Priority: explicit --backend flag > auto-detect.

Auto-detect order: subagent (always available inside Claude Code — default) → codex (if which codex) → bedrock (if scripts/bedrock-review.py present and AWS creds) → claude (fallback).

subagent is the cheapest path — zero API keys, fresh context, no subprocess. Pick interactive when you want to drive the reviewer yourself (asking follow-ups, iterating). Pick codex / claude / bedrock when you need a specific external model.

Default models: codex=gpt-5.4, claude=sonnet, bedrock=eu.anthropic.claude-sonnet-4-6-v1:0. subagent runs on whatever model Claude Code is using. interactive uses codex defaults unless --model is passed. Override with --model.

4. Build Review Prompt

Read references/review-prompt.md and substitute placeholders:

• {{CONTENT_TYPE}} → code, spec, diff, issue, pr
• {{REVIEW_TARGET}} → file paths or inline content
• {{CONTEXT}} → project/architectural context from step 2
• {{MODE}} → full (default) or quick (if --quick flag or user said "delegate"/"just review")
• {{GUIDANCE}} → focus areas, related test files, key dependencies

Write assembled prompt to /tmp/ar-prompt-$$.md.

5. Execute Review

Capture git state before (git diff --stat), then run the chosen backend.

subagent (default)

Invoke the Agent tool directly:

• subagent_type: general-purpose
• description: Adversarial review (<content_type>)
• prompt: the full contents of /tmp/ar-prompt-$$.md

Use the subagent's returned text as the raw reviewer output for step 6. The subagent has its own tools (Read, Grep, Glob, Bash) and can explore the repo, but runs in a fresh context with no prior conversation.

interactive (tmux + codex)

Pre-reqs: must be running inside tmux (tmux list-sessions succeeds) and codex must be on PATH. If either fails, print a clear error and fall back to asking the user to pick another backend.

1. Launch a new window with codex reading the prompt via stdin:

   tmux new-window -n adv-review "codex < /tmp/ar-prompt-$$.md"
   

2. Tell the user: "Opened tmux window adv-review. Drive the review to convergence, then in that window type: output the final review JSON only. When done, come back here and say done (or cancel)."
3. Wait for the user to signal completion.
4. On done: capture the full pane buffer:

   tmux capture-pane -t adv-review -p -S - > /tmp/ar-capture-$$.txt
   

   Extract the last valid JSON object: pipe through jq -R -s 'split("\n") | map(try fromjson // empty) | last'. If that yields nothing, try extracting the last {...} block and validating with jq .. If extraction still fails, ask the user to paste the JSON directly.
5. Kill the window: tmux kill-window -t adv-review.
6. Feed the extracted JSON to step 6 as the raw reviewer output.

codex (non-interactive)

codex exec "$(cat /tmp/ar-prompt-$$.md)" --sandbox read-only --model <model> -c model_reasoning_effort=<effort> (use stdin for prompts >100KB; effort defaults to medium)

claude -p

claude -p "$(cat /tmp/ar-prompt-$$.md)" --model <model> --output-format text

bedrock

uv run scripts/bedrock-review.py /tmp/ar-prompt-$$.md --model <model> --region eu-west-1

---

After execution, run git diff --stat && git status --short. If any new changes appear, flag immediately.

6. Parse and Present Results

The reviewer returns a single JSON object (see prompt template). Strip any surrounding markdown code fences, then parse.

On parse success, render as markdown:

## Independent Review (via [backend] / [model])

**Verdict**: [SHIP | ITERATE | RETHINK]

[summary]

### Understanding
[understanding — omit this section entirely in quick mode or if null]

### Findings
[Each finding as its own block with every certificate field — severity, confidence, what, premises (each with evidence), trace, conclusion, alternative_hypothesis, suggestion, location. Do not drop premises, trace, or alternative_hypothesis for brevity.]

### Strengths
[omit section entirely if empty]

### Questions for the Author
[omit section entirely if empty]

Preserve every certificate field verbatim. Do NOT editorialize or soften. If you disagree with a finding, add a labeled ## Note from Claude section after the review — never mutate the reviewer's output.

On parse failure, print the raw reviewer output under a warning:

## Independent Review (via [backend] / [model]) — raw output

> ⚠️ Reviewer did not return valid JSON. Output preserved verbatim below.

[raw text]

Do not attempt to synthesize findings from malformed output.

Clean up temp file.