spectre-plan

name	spectre-plan
description	Orchestrate durable XS–XL planning after confirmed Scope, with adaptive artifacts, review, tasks, telemetry, and explicit pre-code approval. Do not use for scoping, bug diagnosis, read-only work, or execution of approved artifacts.
user-invocable	true
disable-model-invocation	true

plan

Purpose

Turn confirmed Scope into the smallest sufficient approved implementation handoff. spectre-plan-route alone classifies; the primary planning agent owns synthesis, routing, and deterministic finalization. Research agents return evidence only and never write planning artifacts; independent reviewers use only explicit scope-safe reviewer write surfaces.

Inputs

• $ARGUMENTS: confirmed managed Scope/root or descendant artifact. Ordinary repository-changing work without confirmed Scope returns to spectre-scope; reported bugs normally use spectre-fix. Read-only diagnosis/review, release, and execution of approved artifacts remain direct.
• Existing concepts/scope.md, specs/prd.md, ux.md (preferred) or legacy specs/ux.md, task_context.md, plans, reviews, and tasks/index.

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:spectre-feature-root through DONE; orchestrated calls escalate. Keep writes beneath it and pass it unchanged.
• Treat concepts/scope.md as immutable canonical scope; downstream work may not narrow, expand, or reinterpret it without explicit scope-change approval.
• Read references/estimation-guidance.md only at an applicable gate; each estimate excludes user waits and never delays or blocks the gate.

Outputs + DONE

Size	Required durable result
XS	Compact direct specs/plan.md; no review or tasks artifact.
S	Direct plan.
M	Direct plan + one paired plan-review pipeline.
L	Reviewed plan + tasks.json + execute.md.
XL	L outputs + one task-review report before one-time index finalization and pair validation.

DONE when both routing decisions are recorded; every required high-level design gate was explicitly approved and recorded; required children ran in order; declared artifacts and protected inputs validate; reviewer writes stayed stage-owned; any reroute recommendation was resolved; the returned launch command names the routed execution source; every size passed final pre-code approval before an implementation prompt; telemetry is complete|degraded; and canonical Scope is unchanged or a scope-change gate stopped work.

Method / guardrails

1. Check product readiness: unresolved journeys/states/copy/accessibility route to spectre-ux; load-bearing interaction/layout validation routes to spectre-prototype.
2. Make one cheap local scan, then invoke Skill(spectre-plan-route) in initial mode. Run spectre-workflow plan start with the root/scope hash plus the returned semantic record, size/route, reason codes, design flag, probe flags, and protected boundaries; retain PLAN_RUN_ID for the resulting plan.started. Show size/rationale without asking approval of the label.
3. Gather proportional evidence: XS/S stay local except the route's bounded probe; M uses at most two relevant @spectre:finder, @spectre:analyst, or @spectre:patterns; L/XL cover necessary dimensions. @spectre:web-research is only for an external dependency/API/framework decision. Persist accepted evidence in task_context.md.
4. For L/XL, and for XS–M when the decision reports a material user-owned decision, read references/high-level-design-gate.md and follow it through DONE before invoking any child. A current-run explicitly approved ## Selected Design is the only fast path. Show Estimated remaining planning time when valid; after approval emit plan.gate_completed without response prose.
5. Run exactly the chosen route; one plan-review pipeline means one spectre-plan_review invocation with correctness then simplification:
   • XS → Skill(spectre-create_plan) --depth xs --no-review --execution direct.
   • S → Skill(spectre-create_plan) --depth light --no-review --execution direct.
   • M → Skill(spectre-create_plan) --depth standard --no-review --execution direct → Skill(spectre-plan_review) --auto-apply scope-safe --orchestrated.
   • L → Skill(spectre-create_plan) --depth standard --no-review → Skill(spectre-plan_review) --auto-apply scope-safe --orchestrated → Skill(spectre-create_tasks) --depth standard --orchestrated.
   • XL → Skill(spectre-create_plan) --depth comprehensive --no-review → Skill(spectre-plan_review) --auto-apply scope-safe --orchestrated → Skill(spectre-create_tasks) --depth comprehensive --tasks-only --orchestrated → Skill(spectre-task_review) --mode adversarial --auto-apply scope-safe --orchestrated → → .

Handoff

Only after approval: emit plan.completed with artifact hashes (SHA-256 of raw artifact file bytes), counts, planning elapsed time excluding waits, and continuation. XS returns a self-contained spectre-tdd prompt anchored to the compact plan. S–XL return one copy-ready fenced launch command as the primary next step: the host's explicit spectre-execute invocation (Claude Code /spectre: prefix, Codex $ prefix) plus the repo-relative execution source — specs/execute.md for L/XL, the direct specs/plan.md for S/M. Never pass --orchestrated; a user-launched run stays self-owned through final review and proof. Append at most one short runtime-only instruction that cannot live in a durable artifact; durable guidance belongs in the plan, tasks, or task_context.md. Never generate a goal prompt; autonomous goal mode is an explicit opt-in utility outside this route.

Escalate-If

• Scope/root is absent or conflicts, product ambiguity changes behavior, a reviewer gate fails, deterministic artifacts do not validate, or user feedback changes scope.
• Telemetry or a missing estimate never blocks planning; report degraded observation and continue.