code-review-discipline

# Code review discipline Load this skill at the start of any code review assignment. The lessons here come from real misses that shipped close to deploy and were caught by either another reviewer or a subagent's second pass. ## Subagent confidence is uncorrelated with correctness Running a `code-reviewer` subagent for a "second set of eyes" is cheap and sometimes catches what the primary reviewer missed. It also sometimes produces wrong-premise findings with confident framing. Banked data points from the aweb-aaou federation review chain: - One subagent pass returned **0 valid findings out of 5** — every finding was a wrong-premise about tool names or registration patterns that didn't exist in source. - Another subagent pass returned **2 valid findings** — both were real misses the primary reviewer hadn't caught. The subagent's confidence in BOTH cases sounded the same. Confidence is not a signal of correctness. **Practice**: when a subagent surfaces a finding, treat it as a LEAD, not a conclusion. Open the actual file at the claimed line, verify the claim against code, then decide whether to act. If the subagent claims "X is missing," `grep -n X` first. If the subagent claims "Y is wrong," read Y in source first. Subagent output gets verified-against-code with the same scrutiny as any other review claim — including findings that match your own intuition. This applies symmetrically: the subagent's wrong findings AND its valid findings both need code-level verific...