adversarial-review（对抗性审查）

Adversarial Review

Structured multi-agent review loop. Catches what a single agent misses.

Session store: ~/.openclaw/workspace/reviews/ Process: Init session → spawn Opus reviewers → collect redlines → position on each → produce v2 → deliver

Complexity Self-Assessment

Run this check whenever you produce a substantial document. Score 1 point per signal present. If score ≥ 3, offer the review loop without being asked.

Score 0–2 → skip. Simple doc, don't add noise.

Score 3–6 → offer. *"This scores [N]/10 on complexity. Want me to run the review team on it before we act?"*

Score 7–10 → strongly recommend. Don't just offer — make the case. *"This scores [N]/10 on complexity — multiple interdependent systems, production consequences, hard to reverse. I'd strongly recommend running the review team before we act on this. Today's taxonomy strategy was a 10/10 and the review caught 14 issues including multiple production-breaking bugs."*

Quick Reference

Session Directory Structure

~/.openclaw/workspace/reviews/{YYYY-MM-DD}-{slug}/
├── input/
│   └── {original-filename}       ← copy of doc being reviewed
├── redlines/
│   ├── reviewer-{role}.md        ← raw output per reviewer
│   └── combined.md               ← synthesize.sh output (sorted by severity)
├── positions.md                  ← farsight agree/disagree log
└── output/
    └── {slug}-v2.md              ← final document

Review Types

For full persona prompt templates → read references/reviewer-personas.md For pre-configured bundles → read references/review-types.md

Spawning Reviewers

Spawn ALL reviewers simultaneously — parallel, not sequential. Independent reviewers find different issues.

Model Selection

A/B testing note: If Sonnet misses a CRITICAL issue that Opus would have caught on a 3–6 doc, upgrade that doc type to Opus permanently. Track findings in references/model-notes.md as patterns emerge.

Key parameters for every reviewer spawn:

model: anthropic/claude-opus-4-6   ← or sonnet for 3-6 scored docs
mode: run
runTimeoutSeconds: 300
label: reviewer-{role}

The task field contains the full reviewer prompt from references/reviewer-personas.md plus the document content to review.

Positioning Rules

For EVERY redline, take an explicit position. No skipping.

All CRITICAL redlines default to AGREE unless strongly defensible. At least 1 DISAGREE expected — if zero, you may be rubber-stamping.

Write positions to positions.md in the session directory.

v2 Requirements

• Revision table at the top (what changed and why)
• All AGREE + MODIFY changes incorporated
• Rejected redlines documented in an appendix ("considered and rejected")
• Version bumped, date updated
• Saved to output/{slug}-v2.md

Quality Bar

A good review session produces:

• ≥2 CRITICAL issues (if zero, reviewers weren't adversarial enough — re-spawn with harder prompt)
• ≥1 DISAGREE from farsight (if zero, consider whether the doc was genuinely perfect or just unchallenged)
• A v2 meaningfully different from v1

Redline Format

**[REDLINE-{TYPE}-{NNN}]** {Section reference}
**Claim:** What the document says
**Challenge:** The specific objection or gap
**Severity:** CRITICAL | MAJOR | MINOR
**Suggested Resolution:** What should change

Full spec → read references/redline-format.md