v1.3.1 — The Red Team Depth Release

Redesigned meow:review with a hybrid two-phase adversarial system. Base reviewers stay unchanged; new adversarial personas run as separate focused passes informed by base findings — going deeper, not wider.

What Changed

Hybrid Persona System (Phase A + Phase B)

The review pipeline now has two phases:

• Phase A — existing 3 base reviewers (Blind Hunter, Edge Case Hunter, Criteria Auditor) run in parallel. Unchanged.
• Phase B — 4 adversarial persona subagents run AFTER Phase A, receiving the diff AND a summary of Phase A findings. Each persona has a hostile lens and explicit instructions to find what base reviewers missed.

Persona	Lens	Activates
Security Adversary	Attack surface, injection, auth bypass, supply chain	scope=full
Failure Mode Analyst	Race conditions, cascading failures, resource exhaustion	scope=full
Assumption Destroyer	Implicit assumptions, unvalidated inputs, edge cases	scope=full, domain=high
Scope Complexity Critic	Over-engineering, YAGNI violations, scope creep	scope=full, domain=high

Personas run 2-at-a-time to respect the max-3-agent rule.

Why hybrid? Pure overlays (persona on existing reviewer) dilute attention — the base reviewer's own instructions compete with the persona prompt. Pure separate subagents duplicate infrastructure. The hybrid gives personas undiluted context in their own subagent while reusing the existing review infrastructure.

Scope Gate

Step-01 now classifies diffs as minimal or full:

Signal	Threshold
File count	≤ 3
Line count	≤ 50
Security file touched	Forces full
Domain complexity (via meow:scale-routing)	high → forces full

Minimal scope runs Blind Hunter only — no personas, no artifact verification. Reduces false-positive noise on trivial diffs.

Forced-Finding Protocol

If all reviewers + personas produce zero findings:

1. Re-run the highest-signal reviewer with "look harder" prompt
2. If still zero → accept as clean with note: "Zero-finding review (double-checked)"
3. Maximum 1 re-analysis (no infinite loops)

Prevents rubber-stamp approvals. Adapted from BMAD-METHOD's forced problem-finding.

4-Level Artifact Verification

Added to step-04 verdict (full scope only):

Level	What	Catches
1. Exists	Files compile, exports valid	Build errors
2. Substantive	No stubs, TODOs, empty handlers	Hollow implementations
3. Wired	New exports imported somewhere	Orphaned code
4. Data Flowing	Inputs used, returns consumed	Dead data paths

Levels 1-3 produce WARN/MAJOR findings. Level 4 is informational only (heuristic).

Files Changed

New files (8):

• prompts/personas/security-adversary.md
• prompts/personas/failure-mode-analyst.md
• prompts/personas/assumption-destroyer.md
• prompts/personas/scope-complexity-critic.md
• step-02b-persona-passes.md
• references/artifact-verification.md
• docs/guides/red-team-overview.md

Modified files (9):

• step-01-gather-context.md — scope gate
• step-02-parallel-review.md — conditional dispatch
• step-03-triage.md — forced-finding + Phase A/B dedup
• step-04-verdict.md — artifact verification
• workflow.md — updated flow + variable table
• SKILL.md — v1.2.0 → v1.3.1
• reviewer.md — updated architecture
• AGENTS_INDEX.md — updated capabilities
• memory/patterns.json + memory/security-log.md