joenandez/spectre/spectre-plan

plan

Purpose

Turn confirmed Scope into the smallest draft and handoff. spectre-plan-route classifies; primary owns plan. Scope remains immutable.

Inputs

• $ARGUMENTS: confirmed Scope—thread or managed root/descendant. Repository changes without Scope return to spectre-scope; bugs use spectre-fix. Read-only diagnosis/review, release, and approved-artifact execution remain direct.
• Existing Scope/PRD, ux.md (preferred) or legacy specs/ux.md, task_context.md, and prior plans.

Working Set

• Reuse a managed FEATURE_ROOT only when explicit/current-thread evidence ties it to this work (physical directory wins; never branch/recency/lifecycle/scans); distinct work ignores ambient roots. Otherwise, including on collision, standalone MUST first load and follow @skill-spectre:spectre-feature-root through DONE; orchestrated calls escalate. Keep writes beneath it and pass it unchanged.
• Immutable canonical Scope: concepts/scope.md when present, else confirmed thread; never change it without explicit scope-change approval.
• Read references/estimation-guidance.md when applicable; each estimate excludes waits and never delays or blocks the gate.
• Reuse current-request knowledge results/loads; otherwise follow Project knowledge routing. Search actual task across mixed knowledge/work previews; assess, then exact-load applicable records. Standalone #tag: exact search --tag, assess previews, then exact-load applicable matches; never guessed loads. Refine only for an unresolved question or new subject. Keep IDs/revisions in task_context.md; previews/unloaded candidates are not evidence.

Outputs + DONE

Every route writes specs/plan.md and task_context.md with Scope/authority-bound ## Minimum Solution Selection; no review/task/index/task-review artifacts.

DONE when task_context.md binds initial/observed routing to draft/authority hashes; Scope is unchanged; plan.completed records draft; handoff gives alignment/Trade-offs; telemetry is complete|degraded, or authority/scope stops it.

Method / guardrails

Child skill DONE is internal: resume the next Plan step in this turn. Only Plan's own handoff or escalation ends it.

1. Unresolved journeys/states/copy/accessibility route to spectre-ux; load-bearing interaction/layout validation routes to spectre-prototype.
2. Scan once; invoke Skill(spectre-plan-route) in initial mode. Never print its record; its DONE and limits end at the record. In the same response, run spectre-workflow plan start with root/scope hash, record, size/route, reasons, design/probe flags, boundaries; retain PLAN_RUN_ID for plan.started. Report size/rationale in one progress line beside that call; continue to step 3.
3. Gather proportional evidence: XS/S local except probe; M ≤2 relevant @spectre:finder, @spectre:analyst, or @spectre:patterns; L/XL necessary dimensions. Live external captures belong to Execute, never Plan's evidence dispatch. Dispatch no dimension cited by path, knowledge ID/revision, or thread decision; evidence: SUFFICIENT with uncertainty below HIGH and full citation gathers locally at any size. The fresh evidence-only challenger is wave-independent: M triggers it for durable state, identity, public contract, migration, dependency, or workflow/lifecycle; L/XL always, with the wave or alone. Primary persists accepted evidence/IDs/revisions in task_context.md. @spectre:web-research only decides external API/framework. Refine knowledge search after affected files are known; load a work body only for a stated question, never reload an unchanged revision, and pass workers compact findings with provenance.
4. Read references/minimum-solution.md after initial classification. From accepted evidence, select incumbent-first and persist ## Minimum Solution Selection in task_context.md, bound to Scope/authority/accepted evidence. XS/S decide locally; primary applies the canonical simpler-wins result.
5. Invoke Skill(spectre-plan-route) in observed mode with completed selection before drafting; never print its record, and continue to step 6 in the same response. Plan automatically uses the observed route with no paid rerun or user tier gate unless Scope, explicit design, or authority conflicts.
6. Draft once by loading ${CLAUDE_PLUGIN_ROOT}/skills/spectre-plan/references/create-plan.md with the observed route-mapped depth: XS → light --no-review --execution structured; S inherits it; M/L → standard --no-review --execution structured; XL → comprehensive --no-review --execution structured. Never review/task here. XS uses light, not xs, so no Execution Mode: direct.
7. Validate Routing Observations and owned concepts conform to selected record; emit plan.reclassified with plan hash, records, regret. Append observed record to task_context.md with draft raw-byte hash plus Scope/authority hash; emit unchanged plan.completed artifact hashes/counts/planning elapsed. Telemetry failure is degraded and never blocks handoff.
8. From specs/plan.md derive requested outcome, approach, material decisions, Scope/anti-scope boundaries, credible risks, verification intent, and ordering/shared-contract constraints. Immediately before Execute show Trade-offs verbatim (None allowed): Execute accepts them; feedback revises the draft. On Scope change, selected-design conflict, missing irreversible decision, or unavailable authority, withhold the handoff. Otherwise return exactly one copy-ready fenced command: /spectre:spectre-execute <repo-relative plan.md> --origin plan --preflight-plan <xs|light|standard|comprehensive>. observed XS → xs; S → light; M/L → standard; XL → comprehensive. Launching that command is the user's alignment signal. Never pass --orchestrated or wait for delivery insurance.

Handoff

Execute owns assessment/review/tasks/gates/resume. Never generate a goal prompt. Render Execute: resolved absolute plan path, --origin plan, resolved preflight depth.

Escalate-If

• Scope/root conflict; Scope/design contradiction, missing irreversible decision, unavailable authority, or invalid artifacts.
• Telemetry/estimate failure never blocks; report degradation and continue.