review-tasks-plan

/review-tasks-plan - Task planning review

<input_contract> Targets:

• FT-<NNN>: review that product feature;
• no argument: infer the latest decomposed product feature.

Latest-feature inference order:

1. most recently modified .memory-bank/tasks/plans/IMPL-FT-<NNN>.md;
2. most recently modified indexed product task record;
3. highest numeric task ID as a final tie-breaker.

If the target cannot be resolved, stop and ask for FT-<NNN>. FT-000 is not a review target, although its final gate and dependency effect must be reviewed when they constrain product tasks.

A full review reads current:

• Constitution, requirements/RTM, spec backbone/index, task schema, tier policy, task index, and Foundation decision when present;
• the acceptance-closure and task-boundary definitions in .memory-bank/workflows/execute-loop.md#principle-no-task-explosion and #execution-cohesive-task-boundary, plus .memory-bank/workflows/tier-policy.md#task-claim-and-dependency-ownership;
• target feature, implementation plan, all indexed target task records, and their dependency records;
• direct canonical-spec routes and relevant doctor findings. Read full architecture sources in the main context only when another review concern or the architecture review requires it.

Require Global Backbone Planning Revision to be a positive integer. A missing, zero, or invalid revision is a blocking design-readiness gap owned by /spec-design.

An existing report may bound a rerun only when its Planning Revision matches and the exact repair delta is established from the caller handoff, an isolated diff, or direct comparison of cited evidence. Otherwise perform a full review. On a bounded rerun, recheck the delta and every coverage area it can affect; unchanged cited evidence may cover the rest of the feature.

PLANNING_RECONCILIATION_REQUIRED in the target feature blocks review and routes /feature-to-tasks FT-<NNN>.

</input_contract>

<hard_invariants>

• Use a fresh-context reviewer or separate fresh session.
• Before each feature verdict, load and apply this installed skill's references/finding-adjudication.md semantic pack.
• Do not create, edit, close, promote, block, or reconcile reviewed artifacts.
• Task schema, IDs, lifecycle, tier, Foundation, single-card, scope, and status ownership remain governed by their canonical sources.
• touched_files is advisory; a non-empty write_boundary is hard.
• Behavior specs are optional examples, never readiness gates.
• Deterministic doctor findings are evidence, but this semantic review neither reruns nor impersonates /mb-doctor.
• A bounded execution-cohesion probe is review evidence only. Do not require or create a simulation artifact, field, report, status, or replacement queue.
• Rubrics below are coverage criteria, not a mandatory reasoning order or an exhaustive limit on reviewer probes.
• A previous report is non-authoritative cache. Retain evidence only when its cited coverage and inputs are unchanged; the new verdict covers the current whole feature.
• Establish two co-review focuses on a full review. On a bounded rerun, retain each focus only when its target, scope, and governing evidence are unchanged; refresh every other focus through the semantic pack.
• Run /architecture-review only when current accepted evidence leaves a material ownership, dependency, or boundary question that could change the verdict. If delegation fails, perform the same architecture review locally. </hard_invariants>

<operator_decisions> The reviewer never chooses between competing product, design, canonical-path, task-boundary, tier, dependency, or verification interpretations. Route every unresolved feature-related semantic finding to /feature-doctor FT-<NNN>, even when one inspected specification appears to make the defect unambiguous; the doctor validates that finding's governing basis and repair owner.