caspar-execute

name	caspar-execute
description	Build the planned feature from a compact execute.md index and sliceable tasks.json detail, dispatching parallel task waves through deterministic checks, lightweight sentinel review, and final adversarial review/validation. Trigger after execute.md/tasks.json exist or to resume a partially-built wave plan. Do NOT trigger for scoping/planning, unplanned bug-fixing, or dead-code cleanup (caspar-prune).
user-invocable	true

execute

Execute tasks in parallel waves without loading the full task graph into primary-agent context. Read execute.md whole, slice tasks.json only for selected parent tasks, verify each wave before advancing, adapt for spec compliance, run final adversarial review/validation over the cumulative diff, and emit a manual test guide.

Inputs

• $ARGUMENTS — optional path to an execute index or wave/scope hints.
• Required default artifact: {OUT_DIR}/specs/execute.md with Task Detail Source pointing to tasks.json. If absent → stop, route to caspar-create_tasks.

Working Set

• branch = git rev-parse --abbrev-ref HEAD (fallback unknown); OUT_DIR = docs/tasks/{branch}.
• EXECUTE_INDEX = arg path || {OUT_DIR}/specs/execute.md.
• Resolve TASKS_JSON:
  1. Use ## Task Detail Source line Tasks JSON: <path> when present.
  2. Else if index basename is execute.md, use adjacent tasks.json.
  3. Else if basename ends .execute.md, use sibling .tasks.json.
  4. Otherwise stop and ask for the matching task detail JSON path.
• SCOPE_DOCS = existing paths listed in the execute index ## Document Manifest; read each before execution and pass paths to subagents.
• Wave diff per gate: git diff <parent-of-first-wave-commit>..HEAD; files-touched manifest; verbatim ACs from this wave's selected parent-task slice.

Method / guardrails

Resolve + load index.

• Read EXECUTE_INDEX whole. It is the token-efficient orchestration index: manifest, task source, summary, wave plan, parent-task index, slicing rules.
• Do not read TASKS_JSON whole. Use targeted parsing only: status projection, selected parent-task slices, reviewer criteria/context slices, and status updates.
• After any JSON write, re-parse TASKS_JSON before planning the next wave.

Adaptive wave loop — until all parent tasks are done or skipped:

1. Batch. Use the execute index's Wave Plan + Parent Task Index to choose pending parent ids. Assign <=3 sequential parent tasks per @dev; batches may span phases when wave guidance allows. End a batch before any dependency boundary.
2. Dispatch wave. Launch parallel @dev (one per batch).
   • Before dispatch, extract only the selected parent task ids from TASKS_JSON using jq, node -e, or targeted read/edit mechanics. Include minimal phase labels plus assigned parents, subtasks, ACs, context, and status fields.
   • Inline that slice under <task_assignment>. Self-check: the prompt contains selected parent ids and no unrelated parent ids.
   • @dev receives: <task_assignment>, SCOPE_DOCS paths, and waves 2+ Prior-Wave Context. It MUST read scope docs first, use <task_assignment> as the only task source, and not read any tasks file.
   • Then it loads Skill(caspar-tdd), executes tasks sequentially with red/green TDD, commits after each parent task, uses native related-test commands, and returns compressed Implementation Insights + E2E Completeness Check.
   • E2E signal: ⚪ Complete · 🟡 Gap [missing functionality] · 🔴 Blocker [needs other-task changes].
3. Per-wave verification gate (order is load-bearing):
   • 3a — Deterministic pre-gate: detect and run lint/typecheck/build from package.json/pyproject.toml/Cargo.toml/Makefile; fix until green. Never invoke reviewers while deterministic checks fail.
   • 3b — Sentinel selector: classify the wave before reviewer dispatch:
     • skip: deterministic checks passed; isolated implementation; no downstream dependency; no shared contract, replaced path, or E2E gap. Do not dispatch a reviewer; record sentinel_review: skipped.
     • wiring: default when the wave adds the first vertical slice, touches UI/data/control-flow integration, produces output consumed by later tasks, replaces an old path, exposes a feature, or reports 🟡/🔴 E2E status.
     • risk: use only when touched code involves auth, permissions, payments, PII, migrations, public APIs, secrets, network input, destructive actions, or data correctness.
     • If unsure, choose wiring rather than dual review.
   • 3c — Lightweight sentinel review: dispatch at most one @reviewer only for wiring or risk. Build the prompt only from wave diff, verbatim ACs/context from the selected slice, relevant scope docs, and files manifest. Forbidden: dev reports, implementer rationale, orchestrator paraphrase.
     • wiring lens: Defined -> Connected -> Reachable; grep usage, trace UI/API render or call path backward, flag dead computations, orphaned outputs, duplicate data sources, and old active paths.
     • risk lens: security + correctness, including scope adherence for the risk surface.
     • Reviewer output must be either CLEAN or CRITICAL/HIGH findings only. No Medium/Low/nits/style/speculative architecture. Every finding must include file:line, reproducible failure/exploit scenario, concrete evidence chain, smallest scope-safe fix shape, and sha256(file_path + line + finding_category). No evidence chain means CLEAN.
   • 3d — Bounded fix loop: CRITICAL/HIGH findings get <=3 fix waves. Reappearing hash = reviewer disagreement; escalate and do not re-queue. Halt on test-file changes >0.5x implementation-file changes or cumulative fix diff growth >25% per iteration. Re-run 3a, then re-run 3b/3c only when the selector still requires sentinel review after fixes.
   • 3e — Exit: deterministic checks pass and no required sentinel CRITICAL/HIGH remains, or cap reached and user notified.
4. Mark complete. Edit TASKS_JSON status fields for assigned subtasks/parents to done; preserve indented valid JSON; re-parse immediately.
5. Reflect. Read completion reports for scope signals and E2E gaps. All ⚪ → next wave. Otherwise adapt.
6. Adapt only for spec compliance. Edit TASKS_JSON directly: append gap tasks/subtasks with status: "pending", mark obsolete work skipped, and add learned context to future tasks. If parent ids, titles, dependencies, or wave membership change, update only affected Wave Plan / Parent Task Index rows in EXECUTE_INDEX, then re-read the index. No nice-to-haves; no scope expansion.
7. Next wave. Recompute pending status from TASKS_JSON projections, gather prior completion reports into ## Prior-Wave Context, repeat.

Final adversarial code review + validate. After all parent tasks are done/skipped and deterministic checks pass, run the expensive review once over the cumulative feature diff:

• Run Skill(caspar-code_review) with {OUT_DIR} --orchestrated over the cumulative diff. That skill owns the pinned high-effort opposing-runtime reviewer, same-contract native fallback, adversarial lenses, evidence rules, and saved report.
• Reviewer inputs are limited to cumulative diff, files-touched manifest, SCOPE_DOCS, and relevant TASKS_JSON slices. Do not use dev reports or implementer rationale as evidence.
• Read the saved report. CRITICAL/HIGH findings enter the bounded fix loop; Medium/Low findings are summarized but do not block completion unless the user asks.
• Then @analyst runs Skill(caspar-validate) (caspar-validate) narrowed to cross-wave integration audit, scope-creep audit, and dead-computation sweep over the cumulative diff. Pass SCOPE_DOCS plus TASKS_JSON; do not use execute.md as the validation source.
• High-priority review or validation gaps → dispatch @dev to fix, rerun deterministic checks, then rerun Skill(caspar-code_review) or only the affected validation check.

Outputs + DONE

• Complete implementation, committed per parent task.
• TASKS_JSON statuses reflect completed/skipped/adapted work and parse after final write.
• {OUT_DIR}/test_guide.md or {OUT_DIR}/testing/{branch}_test_guide.md from Skill(caspar-create_test_guide).
• Completion summary: tasks done · waves · sentinel review counts (skip/wiring/risk) · per-wave fix-loop counts · final review status · validation status · test-guide path · Task Evolution Summary · E2E Gaps Addressed · Unresolved Findings.
• DONE when: every wave passed deterministic checks plus any required sentinel review; all planned tasks are done/skipped; final adversarial review and cross-wave validate are clean or gaps fixed/accepted; test guide written; summary returned.

Handoff

Report the summary inline (counts, fix-loop iterations, unresolved findings, test-guide path), then suggest next:

• caspar-clean — run prune + risk-based tests + sweep/commit
• caspar-test — strengthen automated tests
• caspar-rebase — tidy history before merge

Escalate-If

• execute.md missing, missing required sections, or cannot resolve/parse tasks.json → stop; route to caspar-create_tasks.
• A subagent prompt slice contains unrelated parent task ids → fix the slice before dispatch.
• Fix loop hits cap, a hash recurs, or a circuit breaker trips → halt and surface; do not force past it.
• Deterministic pre-gate cannot pass → fix before review; never advance a red wave.