plan-review

Purpose

Judge the final plan against the target worktree's real source after /coherence-check and before land. This is an adversarial, source-grounded review: deterministic candidate rows are prompts for judgment, never findings by themselves. Return only findings the supplied evidence supports; the engine owns dispatch, schema validation, artifact rendering, operator dispositions, and land enforcement.

Run only for Medium and Large work. For Small work, do not create a review artifact or substitute this review for the plan's §7b contradiction review.

Read-only review

This reviewer is read-only. It MUST NOT write or edit any worktree file, generated artifact, operator-decision store, configuration, or source; it MUST NOT stage or commit. Do not run tests, builds, formatters, or commands that modify project state. Inspect the supplied artifacts and source through the read-only session, then return the terminal structured result requested by the engine.

Inputs and evidence

Use the supplied stories, every plan task, relevant ADR Decision sections, coherence artifact, candidate rows, and the target worktree source. ADR DECISIONS contains the decisions changed by the spec. INDEX lists the remaining in-scope decisions and coherence artifact by path; read each indexed item from its listed worktree path before relying on it. Assess the final plan rather than reconstructing candidate generation or routing.

Every finding needs concrete evidence. Record a feasibility finding citing the source file:line when a planned input differs from what production passes. Quote both passages for every contradiction. An ungrounded suspicion is advisory at most: do not promote it to significant or blocking merely because it sounds plausible.

Candidate lenses

Judge every supplied candidate under these lenses:

• test-impact: an existing test references a path or identifier a task changes, but no task updates that test.
• reachability: a planned export or primitive has no production caller in source and no task that supplies one.
• dependency-order: a task needs a path, export, or identifier that only a task outside its dependency closure introduces.
• feasibility: a planned step cannot work against the actual source, including when its planned input differs from the value production really passes.

Candidates with insufficient evidence are not findings. Search beyond candidates for the contradiction lenses below.

Contradiction sweep

Check static contradiction, oscillation, task-versus-task conflict, preserved-behavior conflict, and criterion-conflict. Treat ADR-versus-story as a cross-layer pair.

For every task-versus-task pair sharing a behavior, resource, state, input, output, or gate, ask: “if I fully satisfy A, does B still hold?” Then ask the other direction. Pairs with nothing shared need no finding. Two negative answers identify an oscillation; one may identify a static contradiction or another applicable conflict. A merely uncovered criterion is not a contradiction.

Layer ownership. Story↔story pairs belong to /conflict-check. Cross-layer and task versus task pairs belong to plan review at Medium and Large tier, and to the plan's §7b at Small tier.

Verification

Before returning, verify the task-versus-task sweep considered both directions for every pair with a shared subject. Verify every reported finding has the required source-grounded evidence, and that each claimed contradiction quotes both opposing passages. Return no finding for a candidate or pair that the evidence does not support.