Adversarial Review

Adversarial Review

Spawn reviewers on the opposite model to challenge work. Reviewers attack from distinct lenses grounded in brain principles. The deliverable is a synthesized verdict — do NOT make changes.

Hard constraint: Reviewers MUST run via the opposite model's CLI (codex exec or claude -p). Do NOT use subagents, the Agent tool, or any internal delegation mechanism as reviewers — those run on your own model, which defeats the purpose.

Step 1 — Load Principles

Read brain/principles.md. Follow every [[wikilink]] and read each linked principle file. These govern reviewer judgments.

Step 2 — Determine Scope and Intent

Identify what to review from context (recent diffs, referenced plans, user message).

Determine the intent — what the author is trying to achieve. This is critical: reviewers challenge whether the work achieves the intent well, not whether the intent is correct. State the intent explicitly before proceeding.

Assess change size:

Read references/reviewer-lenses.md for lens definitions.

Step 3 — Detect Model and Spawn Reviewers

Create a temp directory for reviewer output:

REVIEW_DIR=$(mktemp -d /tmp/adversarial-review.XXXXXX)

Determine which model you are, then spawn reviewers on the opposite:

If you are Claude — spawn Codex reviewers via codex exec:

codex exec --skip-git-repo-check -o "$REVIEW_DIR/skeptic.md" "prompt" 2>/dev/null

Use --profile edit only if the reviewer needs to run tests. Default to read-only. Run with run_in_background: true, monitor via TaskOutput with block: true, timeout: 600000.

If you are Codex — spawn Claude reviewers via claude CLI:

claude -p "prompt" > "$REVIEW_DIR/skeptic.md" 2>/dev/null

Run with run_in_background: true.

Name each output file after the lens: skeptic.md, architect.md, minimalist.md.

Build each reviewer's prompt using the template in references/reviewer-prompt.md.

Step 4 — Verify and Synthesize Verdict

Before reading reviewer output, log which CLI was used and confirm the output files exist:

echo "reviewer_cli=codex|claude"
ls "$REVIEW_DIR"/*.md

If any output file is missing or empty, note the failure in the verdict — do not silently skip a reviewer.

Read each reviewer's output file from $REVIEW_DIR/. Deduplicate overlapping findings. Produce a single verdict using the format in references/verdict-format.md.

Step 5 — Render Judgment

After synthesizing the reviewers, apply your own judgment. Using the stated intent and brain principles as your frame, state which findings you would accept and which you would reject — and why. Reviewers are adversarial by design; not every finding warrants action. Call out false positives, overreach, and findings that mistake style for substance.

Append the Lead Judgment section to the verdict (see references/verdict-format.md).

Reviewer Lenses

Three distinct adversarial perspectives. Each reviewer adopts one lens exclusively.

Architect

Challenge structural fitness. Ask:

• Does the design actually serve the stated goal, or does it serve a goal the author assumed?
• Where are the coupling points that will hurt when requirements shift?
• What boundary violations exist? Where does responsibility leak between components?
• What implicit assumptions about scale, concurrency, or ordering will break first?

Map findings to: boundary-discipline, foundational-thinking, redesign-from-first-principles.

Skeptic

Challenge correctness and completeness. Ask:

• What inputs, states, or sequences will break this?
• What error paths are unhandled or silently swallowed?
• What race conditions or ordering dependencies exist?
• What does the author believe is true that isn't proven?
• Where is "it works on my machine" masquerading as verification?

Map findings to: prove-it-works, fix-root-causes, serialize-shared-state-mutations.

Minimalist

Challenge necessity and complexity. Ask:

• What can be deleted without losing the stated goal?
• Where is the author solving problems they don't have yet?
• What abstractions exist for a single call site?
• Where is configuration or flexibility added without a concrete second use case?
• Is this the simplest possible path to the outcome, or is it the path that felt most thorough?

Map findings to: subtract-before-you-add, outcome-oriented-execution, cost-aware-delegation.

Reviewer Prompt Template

Each reviewer gets a single prompt containing:

1. The stated intent (from Step 2) 2. Their assigned lens (full text from references/reviewer-lenses.md) 3. The principles relevant to their lens (file contents, not summaries) 4. The code or diff to review 5. Instructions: "You are an adversarial reviewer. Your job is to find real problems, not validate the work. Be specific — cite files, lines, and concrete failure scenarios. Rate each finding: high (blocks ship), medium (should fix), low (worth noting). Write findings as a numbered markdown list to your output file."

Spawn all reviewers in parallel.

Verdict Format

## Intent
<what the author is trying to achieve>

## Verdict: PASS | CONTESTED | REJECT
<one-line summary>

## Findings
<numbered list, ordered by severity (high -> medium -> low)>

For each finding:
- **[severity]** Description with file:line references
- Lens: which reviewer raised it
- Principle: which brain principle it maps to
- Recommendation: concrete action, not vague advice

## What Went Well
<1-3 things the reviewers found no issue with -- acknowledge good work>

## Lead Judgment
<for each finding: accept or reject with a one-line rationale>

Verdict Logic

• PASS — no high-severity findings
• CONTESTED — high-severity findings but reviewers disagree on them
• REJECT — high-severity findings with reviewer consensus