cr

# cr — Full Code Review Do a PR's complete review in one pass: **broad static sweep + deep dynamic dig**. This skill is **self-contained** in both methods. Most PR review is static — read the snapshot and you can judge right/wrong, **easy but must not be skipped**. The hard part, the part you can't catch line-by-line, is **deriving dynamic behaviour from static code**: timing / state evolution / concurrency. cr does both tracks. ## Execution mode (choose before you start) **Iron rule: the actual review always runs in a 【clean subagent】; the main session only dispatches + mechanically merges, NEVER reviews in person.** Having just written the code in the main session, carrying the defensive "I know why I did it this way" context, is the worst possible reviewer. When a subagent is spawned via the Agent tool it **does not inherit the main session's history** — it sees only the diff + this skill → that is exactly the source of fresh eyes. The dev session's residual history **must not pollute** triage / modelling / downgrade judgements. Two side-effect red lines: ① the main session **does not make substantive judgements on the subagent's behalf**; ② splitting static out also removes the "static breadth dilutes dynamic modelling" dilution of the main thread. Three tiers by scale (**all enter at least one clean subagent**). **You pick between the first two by scale yourself; hard is enabled ONLY when the user explicitly asks for it — never escalate to it on your own** — howeve...