review

# Review (审查) Horizontal layer. **Not on the B-Chain path**; intercepts every task at input / process / output and every library lifecycle event. ## Two Scopes ### Per-task (3 sub-types) | Sub-type | View | Gate | Trigger | |---|---|---|---| | **Code Review** | Product (code/document/data) | Soft | Task produces artifacts | | **Task Recheck** | Goal (was user intent met) | Soft | Task has verifiable outcome | | **Security Review** | Risk (vulnerability/PII/overreach/prompt injection) | Reflection (+extreme hard) | Every task | The **Security Review's Reflection Gate** is the standard pattern: detected risks are injected back into system context; the model must explicitly process them (fix / explain / refuse) before continuing. Cannot be silently ignored. See ADR-0004 for full detail. ### Per-skill (Lifecycle Governance) Govern the skill library itself: review, merge, split, retire skills. This is the **meta-governance** role that no major public skill ecosystem (anthropics/skills, obra/superpowers, mattpocock/skills) abstracts as a first-class module — MY_SKILL treats it as part of Review. ## Trigger - **Per-task**: every task (input / process / output points). Always-active. - **Per-skill**: triggered by lifecycle events (new skill added, skill stale > 6 months, two skills overlap, etc.). See `lifecycle` section in INDEX.yaml. ## Procedure (per-task) 1. **Input review (security focus)**: check the user task for prompt injection, sensitive data exposure, or out-of-...