spectre-plan-review

name	spectre-plan_review
description	Correct and aggressively simplify plan.md after create_plan. Do not review tasks/code, change Scope, or add speculative completeness.
user-invocable	true

plan_review

Purpose

Produce the smallest correct plan through one evidence wave, correctness review, then subtraction-first simplification.

Inputs

• $ARGUMENTS: feature root/name or descendant plan.md; optional --auto-apply scope-safe, --orchestrated.
• Require {FEATURE_ROOT}/specs/plan.md or route to spectre-create_plan. Authority: concepts/scope.md, specs/prd.md, specs/ux.md, then explicit task_context.md requirements. Reuse existing research.

Working Set

• Resolve one managed FEATURE_ROOT for this work from explicit/current-thread evidence only (physical directory wins; never branch/recency/lifecycle/scans). If none is confirmed, including when the candidate path is occupied, standalone MUST first load and follow Skill(spectre-feature-root) through DONE; orchestrated calls escalate. Keep writes beneath it and pass it unchanged.
• Reports: reviews/plan_correctness.md and reviews/plan_review.md. Resume a sole correctness report only when its post-edit hash matches plan.md; otherwise use absent canonical paths or timestamped siblings.

Outputs + DONE

• plan_correctness.md: route/model metadata and hashes; evidence/unknowns; findings; retained constraints/tests; dispositions and resulting edits.
• plan_review.md: correctness path/hash; deletions, collapses, reuse, deferrals, retained exceptions, test reductions, findings/dispositions, and structural Before → After.
• Reports are deltas; never restate the plan.

DONE when both stages complete; research ran once at most; every mechanism traces to a current requirement, constraint, prerequisite, or verified fact; every new complexity boundary has a valid exception; the plan is smaller in mechanisms, surfaces, process, or tests—or proves no safe reduction; behavior and constraints survive; no correctness Blocker/High remains; scope changes are withheld; findings are disposed; hashes/write bounds pass; and direct-mode Verification is executable.

Method / guardrails

Canonical Scope Invariant: correct and simplify implementation means without narrowing, expanding, or reinterpreting Scope. Mark boundary changes Scope Change Required; never auto-apply them.

1. Evidence. Identify unsupported material claims. If needed, use at most one each @spectre_finder, @spectre_analyst, and @spectre_patterns, in parallel, for citation-first evidence/unknowns only (≤1,000 tokens total).

2. Correctness. Read references/correctness-review.md and send it verbatim to a fresh reviewer with: plan, Scope, task-context, and report paths/hashes; compiled evidence/unknowns; edit mode and write bounds; route metadata.

3. Simplification. After correctness writeback, read references/simplification-review.md and send it verbatim to a second fresh reviewer with: corrected plan, Scope, correctness-report, and output-report paths/hashes; edit mode and write bounds; route metadata. Supply existing evidence; allow one cited anchor spot-check.

4. Writeback. Reviewers write only their report and plan.md, with the report written before plan edits; scope/context/tasks/code remain immutable. --auto-apply scope-safe permits Blocker/High and unambiguous Medium edits; otherwise ask all|blockers|IDs|skip, then continue on the same route. Record addressed|skipped|unresolved|scope-change. Stop on unresolved correctness Blocker/High, scope change, unavailable writeback, or failed schema/hash/scope/Out-of-Bounds checks. The primary may normalize mechanics, never semantics.

5. Route. Run each stage fresh at high effort (20-minute limit): Codex → Claude Code opus; Claude Code → Codex gpt-5.6-sol. Record stage/runtime/model/effort/route. If unusable, use one clean-context native @spectre_reviewer with the same template and context. A usable review is terminal.

Handoff

Return routes, correctness status, deletions/retained exceptions, structural delta, dispositions, reports, withheld scope changes, and updated plan in ≤1,000 tokens. Orchestrated calls return; standalone success recommends spectre-create_tasks or direct spectre-execute.

Escalate-If

Plan missing; a material claim remains unknowable; scope change is required; correctness Blocker/High remains; writeback is unavailable; or schema/hash/scope/Out-of-Bounds checks fail.