Review Implementation

Review Implementation (post-implementation)

Review a completed task's diff through three lenses, each a read-only critic subagent. The counterpart to ../critique-plan/SKILL.md: that guards the plan before coding; this guards the result after.

Silence is the default output. A clean diff produces no findings — that is the normal case. A finding survives only if a senior engineer, looking at this diff, would actually make the change. State each as a concrete cost, not a preference. Never invent refactors to look thorough.

ALWAYS check .ab-method/structure/index.yaml FIRST for paths (task location, domain model, review.md). Review only what this task changed — the cohesive diff across its missions (git diff for the commit range). This is not a whole-codebase audit; that's /improve-codebase-architecture.

Look for it under the project root (the current working directory) first — a project's own copy is how it customises its paths, so it always wins. Only if the project has none (AB Method installed as a plugin rather than with npx ab-method), read the bundled default: ../../../.ab-method/structure/index.yaml relative to this SKILL.md. Either way, every path the index names is relative to the project root, never to the folder the bundled file lives in.

Process

1. Gather diff + context

The changed files and their git diff, plus (read only what exists): UBIQUITOUS_LANGUAGE.md / CONTEXT.md (canonical terms, spotting reinvented concepts), docs/architecture/* (the documented way things are built here), docs/adr/ (don't propose what an ADR settled), and the task's unresolved-questions.md if it has one.

Black boxes are deliberate — seed every critic with the file. A TODO(UQ-n) seam whose entry is recorded there is a decision the user signed off on: ship a placeholder rather than guess. No lens may flag it as slop, a shallow module, a speculative abstraction, or dead code, and none may propose "just implement it properly" — the answer isn't theirs to pick. Two things are fair game and should be reported: a TODO(UQ-n) marker with no matching entry (an orphan black box nobody recorded), and a placeholder that leaked past its named seam into several call sites — the seam was supposed to contain it, and containing it again is a safe fix.

2. Spin up THREE read-only critics — in parallel

Spawn three subagents in one batch, named exactly:

Each is read-only — analyses the diff, returns a findings list (often empty), edits nothing. Seed each with the diff + context + its lens file — including the task's unresolved-questions.md when it exists, since every lens would otherwise read its placeholders as defects (each lens file says so too). On Codex (spawn_agent is one level deep) spawn them at the orchestrator's own level — flat; on Claude they may nest. Flat works on both.