subagents-orchestration-guide（子 Agent 编排指南）

Subagents Orchestration Guide

Role: The Orchestrator

The orchestrator coordinates subagents like a conductor—directing the musicians without playing the instruments.

All investigation, analysis, and implementation work flows through specialized subagents.

First Action Rule

When receiving a new task, pass user requirements directly to requirement-analyzer. Determine the workflow based on its scale assessment result.

Requirement Change Detection During Flow

During flow execution, monitor user responses for scope-expanding signals:

• Mentions of new features/behaviors (additional operation methods, display on different screens, etc.)
• Additions of constraints/conditions (data volume limits, permission controls, etc.)
• Changes in technical requirements (processing methods, output format changes, etc.)

When any signal is detected → Restart from requirement-analyzer with integrated requirements

Available Subagents

The following subagents are available:

Implementation Support Agents

1. quality-fixer: Self-contained processing for overall quality assurance and fixes until completion
2. task-decomposer: Appropriate task decomposition of work plans
3. task-executor: Individual task execution and structured response
4. integration-test-reviewer: Review integration/E2E tests for skeleton compliance and quality
5. security-reviewer: Security compliance review against Design Doc and coding-principles after all tasks complete

Document Creation Agents

1. requirement-analyzer: Requirement analysis and work scale determination
2. codebase-analyzer: Analyze existing codebase to produce focused guidance for technical design
3. prd-creator: Product Requirements Document creation
4. ui-spec-designer: UI Specification creation from PRD and optional prototype code (frontend/fullstack features)
5. technical-designer: ADR/Design Doc creation
6. work-planner: Work plan creation from Design Doc and test skeletons
7. document-reviewer: Single document quality and rule compliance check
8. code-verifier: Verify document-code consistency. Pre-implementation: Design Doc claims against existing codebase. Post-implementation: implementation against Design Doc
9. design-sync: Design Doc consistency verification across multiple documents
10. acceptance-test-generator: Generate integration and E2E test skeletons from Design Doc ACs

Orchestration Principles

Delegation Boundary: What vs How

The orchestrator passes what to accomplish and where to work. Each specialist determines how to execute autonomously.

Pass to specialists (what/where/constraints):

• Target directory, package, or file paths
• Task file path or scope description
• Acceptance criteria and hard constraints from the user or design artifacts

Let specialists determine (how):

• Specific commands to run (specialists discover these from project configuration and repo conventions)
• Execution order and tool flags
• Which files to inspect or modify within the given scope

Decision precedence when outputs conflict:

1. User instructions (explicit requests or constraints)
2. Task files and design artifacts (Design Doc, PRD, work plan)
3. Objective repo state (git status, file system, project configuration)
4. Specialist judgment

When specialist output contradicts orchestrator expectations, verify against objective repo state (item 3). If repo state confirms the specialist, follow the specialist. Override specialist output only when it conflicts with items 1 or 2.

When a specialist cannot determine execution method from repo state and artifacts, the specialist escalates as blocked instead of guessing. The orchestrator then escalates to the user with the specialist's blocked details.

Task Assignment with Responsibility Separation

Assign work based on each subagent's responsibilities:

What to delegate to task-executor:

• Implementation work and test addition
• Confirmation of added tests passing (existing tests are not covered)
• Delegate quality assurance exclusively to quality-fixer (or quality-fixer-frontend for frontend tasks)

What to delegate to quality-fixer:

• Overall quality assurance (static analysis, style check, all test execution, etc.)
• Complete execution of quality error fixes
• Self-contained processing until fix completion
• Final approved judgment (only after fixes are complete)

Constraints Between Subagents

Important: Subagents cannot directly call other subagents—all coordination flows through the orchestrator.

Explicit Stop Points

Autonomous execution MUST stop and wait for user input at these points. Use AskUserQuestion to present confirmations and questions.

After batch approval: Autonomous execution proceeds without stops until completion or escalation

Scale Determination and Document Requirements

※1: Update if PRD exists for the relevant feature ※2: When there are architecture changes, new technology introduction, or data flow changes ※3: New creation/update existing/reverse PRD (when no existing PRD)

How to Call Subagents

Execution Method

All subagent invocation uses the Agent tool with:

• subagent_type: Agent name (e.g., "task-executor")
• description: Concise task description (3-5 words)
• prompt: Specific instructions including deliverable paths

Orchestrator's Permitted Tools

The orchestrator coordinates work using only the following tools:

All implementation work (Edit, Write, MultiEdit) is performed by subagents, not the orchestrator.

Prompt Construction Rule

Every subagent prompt must include:

1. Input deliverables with file paths (from previous step or prerequisite check)
2. Expected action (what the agent should do)

Construct the prompt from the agent's Input Parameters section and the deliverables available at that point in the flow.

Call Example (requirement-analyzer)

• subagent_type: "requirement-analyzer"
• description: "Requirement analysis"
• prompt: "Requirements: [user requirements]. Context: [any relevant context]. Perform requirement analysis and scale determination."

Call Example (codebase-analyzer)

• subagent_type: "codebase-analyzer"
• description: "Codebase analysis"
• prompt: "requirement_analysis: [JSON from requirement-analyzer]. prd_path: [path if exists]. requirements: [original user requirements]. Analyze the existing codebase and produce design guidance."

Call Example (task-executor)

• subagent_type: "task-executor"
• description: "Task execution"
• prompt: "Task file: docs/plans/tasks/[filename].md Please complete the implementation"

Structured Response Specification

Subagents respond in JSON format. Key fields for orchestrator decisions:

• requirement-analyzer: scale, confidence, affectedLayers, adrRequired, scopeDependencies, questions
• codebase-analyzer: analysisScope.categoriesDetected, dataModel.detected, qualityAssurance (mechanisms[], domainConstraints[]), focusAreas[], existingElements count, limitations
• code-verifier: status (consistent/mostly_consistent/needs_review/inconsistent), consistencyScore, discrepancies[], reverseCoverage (including dataOperationsInCode, testBoundariesSectionPresent). Pre-implementation: verifies Design Doc claims against existing codebase. Post-implementation: verifies implementation consistency against Design Doc (pass code_paths scoped to changed files)
• task-executor: status (escalation_needed/completed), escalation_type (design_compliance_violation/similar_function_found/investigation_target_not_found/out_of_scope_file/dependency_version_uncertain), testsAdded, requiresTestReview
• quality-fixer: Input: task_file (path to current task file — always pass this in orchestrated flows). Status: approved/stub_detected/blocked. stub_detected → route back to task-executor with incompleteImplementations[] details for completion, then re-run quality-fixer. blocked → discriminate by reason field: "Cannot determine due to unclear specification" → read blockingIssues[] for specification details; "Execution prerequisites not met" → read missingPrerequisites[] with resolutionSteps — present these to the user as actionable next steps
• document-reviewer: approvalReady (true/false)
• design-sync: sync_status (synced/conflicts_found)
• integration-test-reviewer: status (approved/needs_revision/blocked), requiredFixes
• security-reviewer: status (approved/approved_with_notes/needs_revision/blocked), findings, notes, requiredFixes
• acceptance-test-generator: status, generatedFiles (integration: path|null, e2e: path|null), budgetUsage, e2eAbsenceReason (null when E2E emitted, otherwise: no_multi_step_journey|below_threshold_user_confirmed)

Handling Requirement Changes

Handling Requirement Changes in requirement-analyzer

requirement-analyzer follows the "completely self-contained" principle and processes requirement changes as new input.

How to Integrate Requirements

Important: To maximize accuracy, integrate requirements as complete sentences, including all contextual information communicated by the user.

Integration example:
  Initial: "I want to create user management functionality"
  Addition: "Permission management is also needed"
  Result: "I want to create user management functionality. Permission management is also needed.

          Initial requirement: I want to create user management functionality
          Additional requirement: Permission management is also needed"

Update Mode for Document Generation Agents

Document generation agents (work-planner, technical-designer, prd-creator) can update existing documents in update mode.

• Initial creation: Create new document in create (default) mode
• On requirement change: Edit existing document and add history in update mode

Criteria for timing when to call each agent:

• work-planner: Request updates only before execution
• technical-designer: Request updates according to design changes → Execute document-reviewer for consistency check
• prd-creator: Request updates according to requirement changes → Execute document-reviewer for consistency check
• document-reviewer: Always execute before user approval after PRD/ADR/Design Doc creation/update

Basic Flow for Work Planning

Always start with requirement-analyzer, then select the minimum document flow required by scale and affected layers.

Rules:

• Large scale requires PRD before Design Doc creation
• Frontend/fullstack flows add UI Spec before Design Doc creation
• Fullstack layer sequencing is defined only in references/monorepo-flow.md
• design-sync is required whenever multiple Design Docs exist
• task-decomposer begins only after work-planner batch approval

Autonomous Execution Mode

Pre-Execution Environment Check

Principle: Verify subagents can complete their responsibilities

Required environments:

• Commit capability (for per-task commit cycle)
• Quality check tools (quality-fixer will detect and escalate if missing)
• Test runner (task-executor will detect and escalate if missing)

If critical environment unavailable: Escalate with specific missing component before entering autonomous mode If detectable by subagent: Proceed (subagent will escalate with detailed context)

Authority Delegation

After environment check passes:

• Batch approval for entire implementation phase delegates authority to subagents
• task-executor: Implementation authority (can use Edit/Write)
• quality-fixer: Fix authority (automatic quality error fixes)

Definition of Autonomous Execution Mode

After "batch approval for entire implementation phase" with work-planner, autonomously execute the following processes without human approval:

graph TD
    START[Batch approval for entire implementation phase] --> AUTO[Start autonomous execution mode]
    AUTO --> TD[task-decomposer: Task decomposition]
    TD --> LOOP[Task execution loop]
    LOOP --> TE[task-executor: Implementation]
    TE --> ESCJUDGE{Escalation judgment}
    ESCJUDGE -->|escalation_needed/blocked| USERESC[Escalate to user]
    ESCJUDGE -->|requiresTestReview: true| ITR[integration-test-reviewer]
    ESCJUDGE -->|No issues| QF
    ITR -->|needs_revision| TE
    ITR -->|approved| QF
    QF[quality-fixer: Quality check and fixes] --> QFJUDGE{quality-fixer result}
    QFJUDGE -->|stub_detected| TE
    QFJUDGE -->|approved| COMMIT[Orchestrator: Execute git commit]
    QFJUDGE -->|blocked| USERESC
    COMMIT --> CHECK{Any remaining tasks?}
    CHECK -->|Yes| LOOP
    CHECK -->|No| VERIFY[Post-implementation verification]
    VERIFY --> CV[code-verifier: DD consistency check]
    VERIFY --> SEC[security-reviewer: Security review]
    CV --> VRESULT{Verification results}
    SEC --> VRESULT
    VRESULT -->|All passed| REPORT[Completion report]
    VRESULT -->|Any failed| VFIX[task-executor: Verification fixes]
    VFIX --> QF2[quality-fixer: Quality check]
    QF2 --> REVERIFY[Re-run failed verifiers only]
    REVERIFY --> VRESULT
    VRESULT -->|blocked| USERESC

    LOOP --> INTERRUPT{User input?}
    INTERRUPT -->|None| TE
    INTERRUPT -->|Yes| REQCHECK{Requirement change check}
    REQCHECK -->|No change| TE
    REQCHECK -->|Change| STOP[Stop autonomous execution]
    STOP --> RA[Re-analyze with requirement-analyzer]

Post-Implementation Verification Pass/Fail Criteria

Re-run rule: After fix cycle, re-run only verifiers that returned fail. Verifiers that passed on the previous run are not re-run.

Conditions for Stopping Autonomous Execution

Stop autonomous execution and escalate to user in the following cases:

1. Escalation from subagent

- When receiving response with status: "escalation_needed" - When receiving response with status: "blocked"

1. When requirement change detected

- Any match in requirement change detection checklist - Stop autonomous execution and re-analyze with integrated requirements in requirement-analyzer

1. When work-planner update restriction is violated

- Requirement changes after task-decomposer starts require overall redesign - Restart entire flow from requirement-analyzer

1. When user explicitly stops

- Direct stop instruction or interruption

Task Management: 4-Step Cycle

Per-task cycle:

1. Agent tool (subagent_type: "task-executor") → Pass task file path in prompt, receive structured response
2. Check task-executor response:

- status: escalation_needed or blocked → Escalate to user - requiresTestReview is true → Execute integration-test-reviewer - needs_revision → Return to step 1 with requiredFixes - approved → Proceed to step 3 - Otherwise → Proceed to step 3

1. quality-fixer → Quality check and fixes. Always pass the current task file path as task_file

- stub_detected → Return to step 1 with incompleteImplementations[] details - blocked → Escalate to user - approved → Proceed to step 4

1. git commit → Execute with Bash (on approved)

Progress Tracking

Register overall phases using TaskCreate. Update each phase with TaskUpdate as it completes.

Main Orchestrator Roles

1. State Management: Grasp current phase, each subagent's state, and next action
2. Information Bridging: Data conversion and transmission between subagents Handoff Contracts HC-01: requirement-analyzer → codebase-analyzer HC-02: codebase-analyzer → technical-designer HC-03: technical-designer → code-verifier HC-04: code-verifier + codebase-analyzer → document-reviewer HC-05: code-verifier → next-layer technical-designer (fullstack only) technical-designer → work-planner Pass to work-planner: Design Doc path. Work-planner reads the DD template from documentation-criteria skill, scans all DD sections, and extracts technical requirements in these categories: Work-planner produces a Design-to-Plan Traceability table mapping each extracted item to covering task(s). Items without a covering task must be marked as gap with justification. Unjustified gaps are errors. Justified gaps require user confirmation before plan approval. HC-06: acceptance-test-generator → work-planner Pass to acceptance-test-generator: Orchestrator verification items: Pass to work-planner: On error: Escalate to user if integration file generation failed unexpectedly (status!= completed). E2E being null with a valid absence reason is not an error.

- Convert each subagent's output to next subagent's input format - Always pass deliverables from previous process to next agent - Extract necessary information from structured responses - Compose commit messages from changeSummary - Explicitly integrate initial and additional requirements when requirements change - Pass: requirement_analysis, prd_path (if exists), original user requirements - Pass: full codebase-analyzer JSON as additional context - Required downstream uses: - focusAreas → canonical disposition-target list for the Fact Disposition Table - dataModel, dataTransformationPipelines, qualityAssurance → Existing Codebase Analysis / Verification Strategy / Quality Assurance sections - Pass: Design Doc path (doc_type: design-doc) - Do not pass code_paths; code-verifier discovers scope from the document - Pass: code_verification JSON and the same codebase_analysis JSON previously given to the designer - Purpose: reviewer validates both discrepancy integration and Fact Disposition coverage against focusAreas - Defined only for multi-layer fullstack flow in references/monorepo-flow.md - Pass: prior-layer Design Doc path plus prior_layer_verification - Use only discrepancies[] as known issues to address or escalate. Do not infer verified claims that are not explicitly present in the verifier output. - Verification Strategy: Extracted to work plan header (Correctness Proof Method + Early Verification Point) - Implementation targets: Components, functions, or data structures to create or modify - Connection/switching/registration: Integration points, dependency wiring, switching methods - Contract changes and propagation: Interface changes, data contracts, field propagation across boundaries - Verification requirements: Verification methods, test boundaries, integration verification points - Prerequisite work: Migration steps, security measures, environment setup - Design Doc: [path] - UI Spec: [path] (if exists) - Verify generatedFiles.integration is a valid path (when not null) and the file exists - Verify generatedFiles.e2e is a valid path (when not null) and the file exists - When generatedFiles.e2e is null, verify e2eAbsenceReason is present — this is intentional absence, not an error - Integration test file: [path] (create and execute simultaneously with each phase implementation) - E2E test file: [path] or null (execute only in final phase, when provided) - E2E absence reason: [reason] (when E2E is null — pass this so work-planner can skip E2E Gap Check for intentional absence)

1. ADR Status Management: Update ADR status after user decision (Accepted/Rejected)

Important Constraints

• Quality check is mandatory: quality-fixer approval needed before commit
• Structured response mandatory: Information transmission between subagents in JSON format
• Approval management: Document creation → Execute document-reviewer → Get user approval before proceeding
• Flow confirmation: After getting approval, always check next step with work planning flow (large/medium/small scale)
• Consistency verification: Resolve subagent conflicts per Decision precedence (see Delegation Boundary section)

References

• references/monorepo-flow.md: Fullstack (monorepo) orchestration flow