joenandez/spectre/spectre-execute

execute

Preserve parallelism, authority, gates, repair.

Inputs

• $ARGUMENTS: path; --origin plan|fix; --preflight-plan <xs|light|standard|comprehensive>; wave hints; --orchestrated; orchestrated-only --finalization-owner parent. Default owner: self.
• --review-profile final-only, valid only with --orchestrated --finalization-owner parent; it defers intermediate review and requires the orchestrating caller to invoke Skill(spectre-code_review) exactly once; caller owns sequencing, never semantic review.
• structured: execute index + resolvable tasks.json; malformed structured input escalates, never becomes a plan.
• fix-source: bug-report path/root or identifying content—not --origin—selects fix; load references/fix-source.md, not Plan preparation.
• plan-direct: explicit readable plan, including paths, Plan handoff markers, and legacy Execution Mode: direct headers. The explicit supplied plan wins over ambient task artifacts and enters this preparation contract.
• No path: use existing same-run source evidence first, then a plan at the confirmed root, then structured-only fallback. No-path selected plans receive this preparation before dispatch. Never infer from branch/recency.
• --preflight-plan <depth> remains a preparation-depth hint, not authorization, classification, or task-graph switch; a depth hint must not create authority pause. Execution Mode: direct is a legacy coordination hint. No selected readable plan needs a completeness/header ceremony.

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.
• Keep the invocation checkout.
• Structured detail precedence: declared Tasks JSON → adjacent tasks.json → sibling .tasks.json; else escalate. Read whole; detail targeted slices.
• Structured/plan-direct read references/telemetry.md, start/resume local event run with origin, follow primary/worker authority. Local workflow store is the sole lifecycle/progress authority; plans/tasks immutable. Summaries retain origin/shape/category/elapsed/aggregate tokens; Plan-direct emits events. Telemetry failure degrades observation, never delivery authority.
• SCOPE_DOCS: manifest paths or scope/UX/research cited by the plan.
• On first selected readable-plan use/resume, read references/plan-direct.md for preparation state. On the first verification failure, review finding, E2E gap, or proof failure, read references/repair-policy.md.
• Maintain compact verification ledger by HEAD, check id, changed/dependency surface, result, attribution, disposition; controls reruns. Never persist raw output or per-wave evidence/checkpoint/report files.

Outputs + DONE

• Commit implementation batches/canonical artifacts. Task/plan immutable; Execute evidence/state local/uncommitted.
• self: one final comprehensive review; proof ends PASS.
• parent: IMPLEMENTATION_READY + ACCEPTANCE_PENDING manifest; no final review, proof, or acceptance claim. With --review-profile final-only, every phase records final-only and the parent receives FINAL_REVIEW_PENDING.
• DONE: selected work/adaptations done|skipped; current affected verification; each phase final-only or reviewed once for recorded compounding risk; every finding dispositioned; no safe authorized work or stale/uncovered final surface; self final review + end-only proof when applicable; no primary-authored planned work; structured telemetry complete or honestly degraded.

Method / guardrails

• Plan preparation. For selected readable plans, resolve root, plan, Scope, depth/state before run creation. Reuse applicable task_context.md assessment by plan/authority hashes; after scope-safe byte-only review edits, rebind if topology/uncertainty is unchanged; reroute only if absent/materially invalidated. Missing assessment dispatches Skill(spectre-plan-route) to a fresh child agent for one classification. Hash-bound observed size, never depth, decides Plan Review: XS/S record review:not-required:<size> and skip; M+ reuse a closed correctness+simplification chain only with external-attempt/recorded-failure+fallback provenance, completed route/reports, finalized plan hash/disposition, and protected files unchanged; reject route-only/interrupted/NOT DONE; else resume hash-valid or user-decided partial correctness or dispatch ${CLAUDE_PLUGIN_ROOT}/skills/spectre-execute/references/plan-review.md --auto-apply scope-safe --orchestrated once to fresh child. Review-worthy risk reclassifies M+; no hidden XS/S exception. Mechanical repairs stay local; semantic corrections use child. Scope/explicit-design changes remain withheld; scope-safe result proceeds without a second user gate. ATOMIC/DIRECT use the bounded local workstream/Active Wave pattern. For STRUCTURED, pass finalized plan path/hash, closed review evidence, and assessment depth; STRUCTURED dispatches a fresh child with ${CLAUDE_PLUGIN_ROOT}/skills/spectre-execute/references/create-tasks.md --orchestrated (L → standard, XL → comprehensive). No automatic task review; when a user explicitly requests a task-graph audit, load ${CLAUDE_PLUGIN_ROOT}/skills/spectre-execute/references/task-review.md. Select only a validated pair bound to finalized plan and closed review chain. For all preparation children, consume child DONE inside the same Execute run before proceeding.

1. Resolve and resume. Persistent execution authority: on initial entry and every continuation/resume, including after compaction, reload this contract with original source arguments and continue the same durable run; never implement from memory or substitute another workflow. Rebuild the frontier from the store: run start with original --source returns the same runId; run status --run-id <id> --json reports it. Compare selected source/preparation hashes before trusting mappings; references/plan-direct.md owns hash-bound pair reuse and source-mode continuity. Fresh structured selection persists its resolved JSON event source so resubmitting the plan resumes the same run. Trust only completed/skipped; every other value is redo-or-verify. INVALID_TASK_TRANSITION means the log is ahead: re-read, never force. Source order alone does not block concurrent dispatch; detail only the ready frontier.
2. Batch and dispatch. Read references/parallel-scheduling.md before first dispatch; apply it to every source and frontier. Waves guide; phases bound review. Batch at most three sequential parents or bounded plan-direct workstreams. Assign plan-required real-input probes to @spectre:dev before dependent implementation; independent ready work may proceed. Require a sanitized real capture/fixture and observed result against the plan's stop condition; static inference cannot replace them. Every source-owned workstream, including single/sequential work, goes to @spectre:dev; the primary only orchestrates, verifies, accepts, and records state; it never implements planned work. For risk, use Claude claude-opus-5-5 or Codex gpt-6-sol; otherwise defaults. Give exact assignment, context, threatened invariant, and shared contract; omit unrelated ids. Structured dispatch includes exact <workflow_telemetry>. Require the worker to load ${CLAUDE_PLUGIN_ROOT}/skills/spectre-execute/references/tdd.md, run focused checks, commit, return compressed insights, and report E2E Complete|Gap|Adaptation.
3. Gate accepted batches. Before any reviewer, the primary independently runs affected lint/typecheck/build and native focused non-harness tests selected from the returned diff and demonstrated dependency/consumer boundaries; gate batches while other assignments remain in flight, letting early parents ride the next covering gate. Never run a repository baseline/root suite, full app harness, benchmark, or broad qualification here. Record stable check ids/result against current HEAD without raw output or new evidence files. Attribute failures and follow references/repair-policy.md; implementation reports are leads only.
4. Capture boundaries. After exact run/work identity, invoke Skill(spectre-work-record) once at start with truthful intent/scope. Workers return compact findings/evidence; after each accepted batch primary invokes Skill(spectre-capture) only for qualifying reusable knowledge. For meaningful blocked handoff/resolved resumption, invoke Skill(spectre-work-record); never task, batch, check, review, commit, or routine-progress update. At terminal completion run finish first, then invoke Skill(spectre-work-record) once with truthful seven-section account and --source-run-id <exact-run>; capturing earlier leaves a start-only delivery receipt Ship can never select. Authorized successful self sends execution: finalized with remainingWork: "None.", independent of Ship, any PR, and later review/CI/merge/closure; parent stays IMPLEMENTATION_READY/ACCEPTANCE_PENDING; blocked/failed/incomplete runs stay non-final with truthful residual work. Capture failure does not block verification, review, proof, lifecycle authority, or delivery; return recovery input.
5. Record and adapt. Record accepted completion only in local state; emit primary completion against passing gate; worker submission never acceptance. Never mutate source task/plan artifacts for lifecycle state. Add/split/skip/reorder only source-required derivative work; amend the canonical graph only when its definition was wrong. New derivatives join when assigned.
6. Route intermediate review by compounding risk or final-only profile. With --review-profile final-only, record each verified completed phase as final-only without loading review routing or dispatching a reviewer. Otherwise a phase may be reviewed only after all source-owned tasks/workstreams and adaptations are done|skipped under current affected verification; completion alone is not a trigger. On first phase completion, read references/review-routing.md.

• Inputs: phase diff/commits/files, risk, scope, verbatim requirement/AC slices—not dev rationale, pending phases, unrelated files.
• Check named risk, Defined → Connected → Reachable, E2E completeness, security/correctness, dead/orphaned outputs, duplicate sources, old active paths.
• Return CLEAN or evidence-backed CRITICAL/HIGH with file:line, invariant, observable scenario, evidence chain, smallest in-thread fix. Persist only local route/dispositions, never an intermediate review report. Follow repair policy: one consolidated pass, affected verification, no repair re-review.

1. Repeat until the selected completion projection is satisfied.

Finalization

Compute coverage from the ledger. Run only stale or uncovered checks; never blanket-rerun cumulative verification or a root suite. Expensive harness/performance/full qualification runs only for the final relevant candidate unless a source task produces a prerequisite or product-consumed qualification artifact blocking downstream work. Cache by candidate definition hash = relevant inputs + scenario/config + command; permit one run per hash, rerun only after changed inputs or diagnosed infrastructure invalidation.