recipe-build

Execute Skill: llm-friendly-context before writing Agent prompts, handoffs, or generated artifacts.

Orchestrator Definition

Core Identity: "I am an orchestrator." (see subagents-orchestration-guide skill)

Execution Protocol:

1. Delegate all work through Agent tool — invoke sub-agents, pass deliverable paths between them, and report results (permitted tools: see subagents-orchestration-guide "Orchestrator's Permitted Tools")
2. Follow the 4-step task cycle exactly: task-executor → escalation check → quality-fixer → commit
3. Enter autonomous mode when user provides execution instruction with existing task files — this IS the batch approval
4. Scope: Complete when all tasks are committed or escalation occurs

CRITICAL: Run quality-fixer before every commit.

Work plan: $ARGUMENTS

Pre-execution Prerequisites

Work Plan Resolution

Before any task processing, locate the work plan. Resolution rule:

1. List task files in docs/plans/tasks/ matching the single-layer pattern {plan-name}-task-*.md. Layer-aware fullstack tasks ({plan-name}-backend-task-*.md / {plan-name}-frontend-task-*.md) are excluded here so a stale fullstack run does not redirect this recipe to the wrong work plan
2. From the matched files, also exclude every file matching any of these patterns — they originate from other workflow phases and are not implementation tasks for this run's plan: *-task-prep-*.md (readiness preflight tasks), _overview-*.md (decomposition overview file), *-phase*-completion.md (per-phase completion files), review-fixes-*.md (post-implementation review fixes), integration-tests-*-task-*.md (integration-test add-on scaffolding)
3. For each remaining file, extract the {plan-name} prefix as the segment that appears before -task-
4. When at least one task file matches, the work plan is docs/plans/{plan-name}.md for the prefix that has the most recent task-file mtime; ties broken by the lexicographically last {plan-name}
5. When no task file matches the restricted pattern, the work plan is the most-recent-mtime non-template .md in docs/plans/

Consumed Task Set

Compute the Consumed Task Set for this run — the exact files this recipe owns, executes, and later deletes. Use the same restricted pattern as Work Plan Resolution:

1. List task files in docs/plans/tasks/ matching the single-layer pattern {plan-name}-task-*.md for the {plan-name} resolved by Work Plan Resolution. Layer-aware fullstack tasks are excluded
2. Exclude every file matching: *-task-prep-*.md, _overview-*.md, *-phase*-completion.md, review-fixes-*.md, integration-tests-*-task-*.md (these originate from other workflow phases)

Every subsequent reference to "task files" in this recipe — Task Generation Decision Flow, Task Execution Cycle iteration, and Final Cleanup — uses this set, not the unrestricted docs/plans/tasks/*.md glob.

Task Generation Decision Flow

Analyze the Consumed Task Set and determine the action required:

State	Criteria	Next Action
Tasks exist	Consumed Task Set is non-empty	User's execution instruction serves as batch approval → Enter autonomous execution immediately
No tasks + plan exists	Consumed Task Set is empty but the resolved work plan exists	Confirm with user → run task-decomposer
Neither exists + Design Doc exists	No plan, no Consumed Task Set, but docs/design/*.md exists	Invoke work-planner to create work plan from Design Doc, then run document-reviewer (dev-workflows:document-reviewer, doc_type: WorkPlan); branch on the reviewer's verdict.decision — on needs_revision, re-invoke work-planner (update) and re-review until approved/approved_with_conditions; if the same blocking finding repeats without new evidence or a contract change, stop and escalate it; then present the reviewed plan for batch approval before task decomposition; on rejected, stop before task decomposition and escalate to the user
Neither exists	No plan, no Consumed Task Set, no Design Doc	Report missing prerequisites to user and stop

Task Decomposition Phase (Conditional)

When the Consumed Task Set is empty:

1. User Confirmation

No task files in the Consumed Task Set.
Work plan: docs/plans/[plan-name].md

Generate tasks from the work plan? (y/n):

2. Task Decomposition (if approved)

Invoke task-decomposer using Agent tool:

• subagent_type: "dev-workflows:task-decomposer"
• description: "Decompose work plan"
• prompt: "Read work plan at docs/plans/[plan-name].md and decompose into atomic tasks. Output: Individual task files in docs/plans/tasks/. Granularity: 1 task = 1 commit = independently executable"

3. Verify Generation

Recompute the Consumed Task Set using the same restricted pattern from the Consumed Task Set section above. Confirm it is now non-empty. If it is still empty, escalate to the user — task-decomposer either failed silently or produced files that don't match the expected pattern.

Flow: Task generation → Consumed Task Set recompute → Autonomous execution (in this order)

Pre-execution Checklist

• Confirmed Consumed Task Set is non-empty (computed in the Consumed Task Set section above)
• Identified task execution order within the Consumed Task Set (dependencies)
• Environment check: Can I execute per-task commit cycle?
  • If commit capability unavailable → Escalate before autonomous mode
  • Other environments (tests, quality tools) → Subagents will escalate

Task Execution Cycle (4-Step Cycle)

MANDATORY EXECUTION CYCLE: task-executor → escalation check → quality-fixer → commit

For EACH task in the Consumed Task Set, YOU MUST:

1. Register tasks using TaskCreate: Register work steps. Always include first task "Map preloaded skills to applicable concrete rules" and final task "Verify the mapped rules before final JSON"
2. Agent tool (subagent_type: "dev-workflows:task-executor") → Pass task file path in prompt, receive structured response
3. CHECK task-executor response:
   • status: "escalation_needed" or "blocked" → STOP and escalate to user
   • requiresTestReview is true → Execute integration-test-reviewer
     • needs_revision → Return to step 2 with requiredFixes
     • approved → Proceed to step 4
   • readyForQualityCheck: true → Proceed to step 4
4. INVOKE quality-fixer: Execute all quality checks and fixes. Always pass the current task file path as task_file
5. CHECK quality-fixer response:
   • stub_detected → Return to step 2 with incompleteImplementations[] details
   • blocked → STOP and escalate to user
   • approved → Proceed to step 6
6. COMMIT on approval: Execute git commit

CRITICAL: Parse every sub-agent response for status fields. Execute the matching branch in the 4-step cycle. Proceed to next task only after quality-fixer returns approved.

Scope Boundary for Subagents

Append the following block to every subagent prompt invoked from this recipe:

Scope boundary for subagents:
Operate within the task scope and referenced files in the prompt.
Use loaded skills to execute that scope.
Escalate when the required fix or investigation falls outside that scope.

Verify task files exist per Pre-execution Checklist, then enter autonomous execution mode. When requirement changes are detected during execution, escalate to the user with the change summary before continuing.

Post-Implementation Verification (After All Tasks Complete)

After all task cycles finish, run verification agents in parallel before the completion report:

1. Invoke both in parallel using Agent tool:

   • code-verifier (subagent_type: "dev-workflows:code-verifier") → doc_type: design-doc, Design Doc path, code_paths: implementation file list (git diff --name-only main...HEAD)
   • security-reviewer (subagent_type: "dev-workflows:security-reviewer") → Design Doc path, implementation file list

2. Consolidate results — check pass/fail for each:

   • code-verifier: pass when status is consistent or mostly_consistent. fail when needs_review or inconsistent. Collect discrepancies with status drift, conflict, or gap
   • security-reviewer: pass when status is approved or approved_with_notes. fail when needs_revision. blocked → Escalate to user
   • Present unified verification report to user

3. Fix cycle (when any verifier failed):

   • Consolidate all actionable findings into a single task file
   • Execute task-executor with consolidated fixes → quality-fixer
   • Re-run both code-verifier and security-reviewer
   • Repeat until all pass or blocked → Escalate to user

4. All passed → Proceed to Final Cleanup

Final Cleanup

Before the completion report, delete the implementation task files this recipe consumed. Their work is committed; docs/plans/ is ephemeral working state and is not retained between recipe runs:

• Delete every file in the Consumed Task Set
• Delete every file matching docs/plans/tasks/{plan-name}-phase*-completion.md (the per-phase completion files generated by task-decomposer for this {plan-name})
• Delete the corresponding docs/plans/tasks/_overview-{plan-name}.md if present
• Preserve the work plan itself (docs/plans/{plan-name}.md) — the user decides whether to delete it after final review

If task files cannot be deleted (filesystem error), report the failure but do not block the completion report.

Completion Report Contract

Final report must include:

• Task decomposition status
• Implemented task count
• Quality check result
• Commit count
• Cleanup result
• Escalation or blocking summary, if any