review

Review — Intent-Grounded Review of the Current Session

You are invoked via /lens:review [target]. Run the engine, then render. You are read-only: never commit, edit, or block — produce the artifact; the human decides.

Step 0: Create the task list (standalone only)

Create the in-session progress task list per references/task-tracking.md — one TaskCreate per stage, all status:"pending": scope, intent, analyze, verify, reconcile, render, report, plus setup only when .claude/lens/settings.md is absent. Keep each returned taskId. As the pipeline runs, mark each review-owned task (setup, reconcile, render, report) in_progress when its step begins and completed when it ends; the four engine-owned tasks (scope, intent, analyze, verify) are transitioned by the engine via the taskIds you hand it in Step 2.

Step 1: First-run setup (guard)

If .claude/lens/settings.md is absent, mark the setup task in_progress, run references/setup.md: ask via AskUserQuestion (1) gitignore the .claude/lens/ dir? (offer "artifacts only — track the registry" so teams can share finders) and (2) the default output path; persist to settings.md, then mark setup completed. On later runs, just read settings (no setup task exists — see Step 0).

Step 2: Run the engine

Invoke lens:engine (Skill tool), passing the target + the project finders registered in settings + taskIds = { scope, intent, analyze, verify } so the engine transitions those four stages as it runs. It returns a review-findings object in context. The full engine input/return contract is declared in ../engine/references/engine-api.md. Empty scope (empty diff / no repo): key the empty-scope branch on result.emptyScope === true — the engine's discriminator — not on an empty findings[]. When result.emptyScope is true, mark the review-owned reconcile and render tasks deleted, mark report completed with a one-line "nothing to review", and skip Steps 3–5 — there is nothing to reconcile, render, or write. (The engine already marked scope completed and intent/analyze/verify deleted, so every stage is accounted for once, with no double-transition and no stranded stage.) Clean review (real diff, zero findings): when result.emptyScope is falsy but findings[] is empty, a review did run and simply found nothing — fall through to Steps 3–5 and render a normal ship / no-findings artifact (verdict ship, adherence still shown). Do not take the empty-scope branch.