adversarial-review

name	adversarial-review
description	Run an adversarial review of a plan, spec, bead, or PR by dispatching to multiple harnesses with adversarial framing and aggregating findings. Use when you want pressure-testing against a single-model blindspot, high-stakes review, or multi-round critique before committing to a direction.

Adversarial Review

A workflow skill for pressure-testing a target (plan, spec, bead, code change, or PR) by dispatching adversarial review prompts to multiple harnesses and aggregating the findings. Replaces ad-hoc inline quorum invocations with a structured, repeatable workflow.

When to use

• Plans and specs before breaking into beads — catch ambiguity and missing constraints early.
• High-stakes beads — migrations, auth, API contracts, security surface.
• Controversial changes — one model's opinion is not enough; different model families catch different failure modes.
• Multi-round refinement — run 2–3 rounds; use findings to revise; re-run until BLOCKING count drops to zero.

Single-harness bead review (ddx bead review <id>) is fine for routine work. Reserve adversarial-review for the cases that justify the cost.

Workflow

1. Prepare the target

Gather the content to review into a single prompt file. Include:

• The artifact being reviewed (plan text, spec section, bead description + AC, or diff / PR summary).
• The question you want pressure-tested ("Is this design complete?", "What are the failure modes?", "What does this spec leave undefined?").
• Any governing artifacts the reviewer should check against (FEAT-* IDs, CONTRACT-* IDs, ADR links).

Write the prompt to a file, e.g. review-target.md:

## Target

<paste plan, spec excerpt, or diff>

## Governing artifacts

- FEAT-010: docs/helix/01-frame/features/FEAT-010-task-execution.md
- (add others)

## Review question

Find BLOCKING issues: ambiguities, missing constraints, design gaps,
internal contradictions, or anything that would cause implementation
rework. Be adversarial — your job is to find problems, not validate.

2. Dispatch to multiple harnesses

Run each harness in parallel using ddx run --harness <name>. Use --min-power 10 (or --harness for explicit control). Route to harnesses from different model families for maximum coverage.

# Two-harness adversarial dispatch (parallel)
ddx run --harness codex --min-power 10 \
  --prompt review-target.md \
  > findings-codex.md &

ddx run --harness claude --min-power 10 \
  --prompt review-target.md \
  > findings-claude.md &

wait

For three harnesses:

ddx run --harness codex  --min-power 10 --prompt review-target.md > findings-codex.md  &
ddx run --harness claude --min-power 10 --prompt review-target.md > findings-claude.md &
ddx run --harness gemini --min-power 10 --prompt review-target.md > findings-gemini.md &
wait

3. Require structured output

Each harness must produce findings in this format. Include this contract in the prompt:

## Output contract

Produce findings as:

### Findings

| Severity | Area | Finding |
|---|---|---|
| BLOCKING | <area> | <specific issue> |
| WARNING  | <area> | <specific issue> |
| NOTE     | <area> | <observation> |

### Verdict: APPROVE | REQUEST_CHANGES | BLOCK

### Summary
<2–4 sentences>

A finding with no evidence is invalid. Cite the specific line, section, or gap that caused the finding. "Looks fine" is not a finding.

4. Aggregate

Read all findings files and collate:

1. BLOCKING items from any harness → escalate immediately. One BLOCKING is enough to stop.
2. WARNING items with ≥2 harnesses in agreement → treat as BLOCKING.
3. WARNING items unique to one harness → review manually, weigh cost.
4. NOTE items → log; address in a follow-up pass or ignore.

Surface disagreements explicitly. If harness A says BLOCKING and harness B says NOTE for the same area, show both verdicts — do not silently collapse to the weaker one.

# Combine findings for review
cat findings-codex.md findings-claude.md > findings-combined.md

5. Iterate until clean

Revise the target based on BLOCKING findings. Re-run the adversarial review against the revised version. Typical well-scoped plans converge in 2–3 rounds. Stop when:

• No harness returns a BLOCKING verdict, AND
• All WARNINGs have been explicitly accepted or addressed.

Record the round count and final verdict in the bead evidence:

ddx bead evidence add <id> --type review --body findings-combined.md

Adversarial framing

The adversarial framing instruction to include in every review prompt:

You are a critic, not a validator. Your job is to find every way this could fail, every constraint it leaves undefined, every assumption it bakes in without stating, every interface it leaves ambiguous. A BLOCKING finding is anything that would cause implementation rework, a migration hazard, or a spec gap that agents will interpret differently. Do not balance criticism with praise — a useful adversarial review is entirely about what is wrong.

Personas

Bind a specification-enforcer persona for the strictest checks:

ddx run --harness codex \
  --persona specification-enforcer \
  --prompt review-target.md

The specification-enforcer persona refuses drift from governing artifacts and surfaces spec gaps others miss.

Anti-patterns

• Single-harness adversarial review. The point is cross-model coverage. One harness is bead review, not adversarial review.
• No structured output contract. Free-form critique is unaggregateable. Always include the findings table + verdict contract in the prompt.
• Collapsing disagreements. If harnesses disagree, show both verdicts. Hiding disagreement defeats the purpose.
• Running adversarial review on routine beads. Reserve for high-stakes or controversial work. Routine beads use ddx bead review.
• Stopping after one round. One round finds problems. Iteration verifies that the revision actually resolves them.

CLI reference

# Parallel adversarial dispatch
ddx run --harness codex  --min-power 10 --prompt review-target.md > findings-codex.md  &
ddx run --harness claude --min-power 10 --prompt review-target.md > findings-claude.md &
wait

# With persona
ddx run --harness codex --persona specification-enforcer \
  --prompt review-target.md > findings-codex.md

# Store findings as evidence
ddx bead evidence add <id> --type review --body findings-combined.md

# Bead review (routine, single-harness — not this skill)
ddx bead review <id>
ddx bead review <id> --execute --harness claude