recipe-build

Explicit User Instruction: The user explicitly instructs and authorizes every subagent call named in this recipe. Execute each applicable call when its prerequisites are met.

Execute Skill: llm-friendly-context before writing Agent prompts, handoffs, or generated artifacts. Execute Skill: subagents-orchestration-guide before making workflow decisions, invoking agents, or resolving findings.

Orchestrator Definition

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

Local authority gate: Make this recipe's workflow decisions and validate each returned result directly; delegate semantic deliverable production to the named specialist.

Review Resolution Gate [MANDATORY]: Resolve every actionable deliverable-review finding through subagents-orchestration-guide Review Resolution before correction or progression. Before the first finding disposition, read references/review-resolution.md from the loaded subagents-orchestration-guide skill.

Execution Protocol:

1. Invoke named specialists for deliverable production — pass deliverable paths between them and validate their results (see subagents-orchestration-guide "Orchestrator Execution Boundary")
2. Follow the 4-step task cycle exactly for each task in the Consumed Task Set: execute → branch on executor result → quality-fix → commit. Corrections produced outside that cycle reuse its executor-result branching and quality gate, and defer only the commit to their own phase's rule
3. Enter autonomous mode when the user provides execution instruction with an existing Work Plan or task files — this IS the batch approval
4. Scope: Complete consumed task-set execution, post-implementation verification, consumed-task cleanup, and completion reporting in order, or stop autonomous execution at the current phase for a valid user-owned escalation. Advance only when the current phase's stated transition condition is satisfied.

CRITICAL: Commit only after quality-fixer returns approved or verification_incomplete. A quality-fixer pass authorizes a commit at a defined commit point; it does not create one.

Work plan: $ARGUMENTS

Pre-execution Prerequisites

Work Plan Resolution

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

1. Use the work plan explicitly supplied in $ARGUMENTS when present.
2. Otherwise group single-layer task files by the existing {plan-name}-task-*.md naming contract and map each group to docs/plans/{plan-name}.md. Exclude layer-aware fullstack task sets.
3. When task groups produce no candidate, use the only Work Plan under docs/plans/ when exactly one exists.
4. Select the sole candidate. When multiple candidates remain, present them for selection. When none exists, continue through the missing-prerequisite branch below.

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

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:

Task Materialization Phase (Conditional)

When the Consumed Task Set is empty:

1. Task Materialization

Invoke task-decomposer using Agent tool:

• subagent_type: "dev-workflows-fullstack:task-decomposer"
• description: "Materialize work plan tasks"
• prompt: "Read work plan at docs/plans/[plan-name].md and output individual single-commit task files in docs/plans/tasks/."

2. Verify Generation

Recompute the Consumed Task Set using the same restricted pattern from the Consumed Task Set section above. When it remains empty, apply Specialist Result Acceptance: validate the invocation and returned artifacts, correct recoverable input or naming errors, and rerun.

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 is unavailable → Apply Specialist Result Acceptance before autonomous mode
  • Other environments (tests, quality tools) → Quality agents retain proof limitations while the task cycle continues

Task Execution Cycle (4-Step Cycle)

MANDATORY EXECUTION CYCLE (per task in the Consumed Task Set): execute → branch on executor result → quality-fix → commit

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

1. EXECUTE: invoke Agent tool (subagent_type: "dev-workflows-fullstack:task-executor") → Record the current HEAD as diffBase, pass task_file: [path], and receive the structured response
2. BRANCH ON EXECUTOR RESULT:
   • status: "escalation_needed" or "blocked" → Apply subagents-orchestration-guide Specialist Result Acceptance
   • requiresTestReview is true → Identify the changed integration/E2E test files in the current changes and invoke integration-test-reviewer with them as changedTestFiles, plus diffBase, taskFile, prompt-only claims, and mutationEvidence
     • approved → Proceed to step 3
     • blocked → Apply Specialist Result Acceptance
     • needs_revision → Pass qualityIssues unchanged into the Review Resolution Gate; return to step 1 for rerouted corrections and derive convergence from correction re-review prior_feedback_reconciliation
   • status: completed → Proceed to step 3
3. QUALITY-FIX: Invoke quality-fixer with task_file, upstream mutationEvidence, and qualityCommand when available (caller first, otherwise current task)
   • stub_detected → Return to step 1 with quality-fixer's incompleteImplementations array unchanged as the canonical incompleteImplementations field
   • blocked → Apply Specialist Result Acceptance
   • verification_incomplete → Retain the complete result for final retry and proceed to step 4
   • approved → Proceed to step 4
4. COMMIT: Apply subagents-orchestration-guide Commit Boundary Check, then execute git commit after quality-fixer returns approved or verification_incomplete; append its verification trailers for the latter

Use each subagent's semantic result and repository evidence through Specialist Result Acceptance; canonical status fields provide the normal routing shortcut. Proceed to the next task after step 4 and retain any verification limitation with its status kept proof-limited.

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 Review (After All Tasks Complete)

Before invoking post-implementation reviewers, apply subagents-orchestration-guide's retained verification limitation retry. Continue with the reviewers after clearing or retaining each result; include only repeated limitations in the completion report.

Resolve the Work Plan's readable Design Doc; missing input blocks review.

Emit these Agent calls in one assistant message, then await both:

• code-reviewer (subagent_type: "dev-workflows-fullstack:code-reviewer") → review the completed implementation with the resolved typed governingDocuments, the actual files changed by completed tasks as implementationFiles, and the Work Plan path
• security-reviewer (subagent_type: "dev-workflows-fullstack:security-reviewer") → review the completed implementation against the same typed governingDocuments

Apply subagents-orchestration-guide's Post-Implementation Review status-routing and fix/re-run rules. Present the unified report; proceed to Final Cleanup after the complete review set reaches Review Resolution convergence.

Final Cleanup

Before the completion report, commit the post-review corrections applied at Review Resolution convergence when any remain uncommitted, applying subagents-orchestration-guide Commit Boundary Check, then delete the implementation task files this recipe consumed. Their work is then committed; docs/plans/ is ephemeral working state and is not retained between recipe runs:

• Delete every file in the Consumed Task Set
• Preserve the work plan itself (docs/plans/{plan-name}.md) — the user decides whether to delete it after final review

If task-file deletion fails with a filesystem error, report the failure and continue to the completion report.

Completion Report Contract

Final report must include:

• Task materialization status
• Implemented task count
• Quality check result
• Verification limitations that remained after final retry
• Commit count
• Cleanup result
• Declined actionable findings with ID, governing reason, and evidence, when any occurred
• Escalation or blocking summary, if any