subagent-review ยท diff

git:20260702.67ef2cf to v1.0.0

9 added, 10 removed. Audit A to A.

---
name: subagent-review
license: MIT
+ metadata:
+ version: "1.0.0"
description: "USE FOR: Guardian/critic reviews using guardian-specified native reviewer perspectives. DO NOT USE FOR: implementation, dispatcher fan-out, model/tier overrides, or public posting."
---
# Subagent Review
- Runtime-native reviewer perspectives. No implementation, approval, fan-out,
- or model/tier overrides.
-
- Defaults: security, architecture, historian, code, QA.
+ Runtime-native reviewer perspectives (defaults: security, architecture,
+ historian, code, QA). No implementation, approval, fan-out, or overrides.
## 1. Workflow
1. Run the cheapest relevant verifier first.
2. Each role: five subagent perspectives + one self-review (six inputs).
Guardian uses five Codex; critic uses guardian-specified (default: all
five). Guardian narrows; critic never self-selects.
3. Self-review accompanies the wave, never replaces it. Direct-only applies
to guardian-labeled trivials; self-review still required.
4. Subagent waves run in parallel; close before role self-review and
second-wave validators.
- 5. Add the data reviewer only for specialized questions; give each subagent
- bounded read-only paths, context, and output shape.
- 6. Deduplicate across six inputs per role; produce critic packet (six) or
- guardian verdict (twelve: six guardian + six critic).
+ 5. Add the data reviewer only for specialized questions; give each
+ subagent bounded read-only paths, context, and output shape.
+ 6. Deduplicate per role; produce the critic packet (six) or guardian
+ verdict (twelve).
## 2. Role Rules
- Guardian: five Codex subagents + one self-review (six); aggregates all six
into the intermediary result. MUST request critic for substantive reviews.
- Critic: guardian-specified subagents (default: all five) + one self-review
(six); sends all six to guardian for twelve-input aggregation.
- Guardian aggregates twelve (six + six) into the final verdict.
- Postman request to critic authorizes launches. On block: `BLOCKED:
perspective launch not permitted`; no direct-only fallback.
- Trivials may omit subagents when guardian so labels. Self-review still
required. Critic states the criterion.
- Critic sends result to guardian, not orchestrator.
- Reviewer subagents must not edit, commit, push, or approve.
- If <five subagents complete, return `BLOCKED: fewer than required subagent
perspectives completed` unless pre-authorized. Critic replies include
`Required perspectives`, `Perspectives launched`, `Self-review: complete`.
## 3. Output
- Perspective, paths, evidence, severity, confidence, gaps. Guardian aggregates
- final verdicts.
+ Perspective, paths, evidence, severity, confidence, gaps.