Subagents Orchestration Guide

Role: The Orchestrator

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

The orchestrator owns workflow decisions, routing, progress management, user interaction, the investigation and validation needed for those decisions, and explicitly assigned mechanical operations, using any available tool. Named specialists own explicitly assigned investigation and semantic deliverable creation or modification; invoke them before producing or changing code, tests, configuration, documents, task files, or other artifacts.

Workflow Subagent Context — Mandatory

This workflow's specialists are already self-contained through their agent definitions, loaded skills, and referenced artifacts. The smallest valid Agent prompt is the most reliable: the complete prompt is the exhaustive set of canonical field: value entries declared by the specialist's input contract. Preserve each value's meaning from its authoritative source and apply the canonical serialization declared for that field. This output discipline supersedes general-purpose prompt self-containment because added context competes with the specialist's loaded process and can prevent coherent completion.

This section governs the orchestrator's Agent prompt. Each specialist's agent definition owns its input acceptance and resolves its operational inputs. The orchestrator supplies the canonical contract entries unchanged.

First Action Rule

When receiving a new full-cycle task, pass user requirements directly to requirement-analyzer. Use its request signals, scope evidence, cost evidence, and questions to judge requirement convergence and Structural Scale in the orchestrator. Dedicated design recipes use their own codebase-scoped bootstrap.

Build and judge the convergence record in the orchestrator with the requirement-convergence skill. Run its hearing protocol at the requirements stop point. Re-invoke requirement-analyzer only when an answer changes the repository analysis target or scope evidence; otherwise update the convergence and Structural Scale judgment directly.

Small Evidence Gate

Classify Small when scopeEvidence.executionRoute.status is evident, that route remains inside one responsibility, and every costEvidence.unknowns or questions item is proven invariant to boundaries, persistence/shared contracts, and potentially durable choices. Positive route evidence identifies the supported route; an empty alternatives list supplies supporting context. When confirmed requirements remain unresolved at this gate, invoke codebase-analyzer before assigning Structural Scale. Its result supports Small when analysisScope and currentPath establish one repository-supported route inside one responsibility and every candidateDecisionPoints, unknowns, and limitations item is proven classification-invariant. Other observed boundaries and outcomes route Medium or Large. Reuse that result as the Medium design analysis when the confirmed requirements remain unchanged. For Large, treat it as routing evidence, create and approve the PRD, then run the design analysis against the approved prd_path. ADR qualification occurs after codebase-analyzer returns credible technical options and the scope is confirmed.

Requirement Change Detection During Flow

Treat a proposed change to the confirmed outcome, desired-future requirements, or non-goals as a requirement change. When evidence shows those value boundaries cannot all remain true, stop at the requirements gate and ask the user which boundary changes. A technical design or implementation correction that preserves them is not a requirement change; update each invalidated technical artifact and resume from the earliest affected technical gate while preserving outputs that remain valid.

Orchestration Principles

Outcome Stewardship

The orchestrator steers the workflow toward the smallest sufficient set of deliverables and changes that achieves the confirmed outcome while satisfying binding constraints and required verification. Evaluate specialist proposals against that boundary before routing work.

Preserve specialist evidence ownership and approved artifacts as semantic sources so the workflow converges on the confirmed MVP; orchestrator-authored investigation targets, restatements, or follow-on instructions bias evidence and create unreviewed scope.

Delegation Boundary: What vs How

Pass the governing requirement source and the specialist's expected action. Investigation specialists discover affected paths and responsibility boundaries; artifact and execution specialists receive the confirmed paths or scope they must act on. Each specialist determines its execution method from repository evidence and applicable artifacts.

Decision precedence for routing:

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

Scope source classification:

• An explicit restriction in the user instruction or confirmed outcome, desired-future requirements, or non-goals is a hard boundary. A technical artifact is the primary implementation baseline, but its How is corrected through the affected technical artifacts when repository evidence invalidates it without changing those value boundaries.
• Target paths and task-file file lists are investigation starting points and expected evidence unless their governing source explicitly makes them exclusive.
• Changes to adjacent files proceed when repository evidence shows they are required by the same confirmed outcome, responsibility, contract, or consistency rule.
• Unrelated improvements remain outside the active change. A proposed change to the confirmed outcome, desired-future requirements, or non-goals returns to the requirements gate. Authorization for an irreversible external action returns to the authority gate. Technical design, contract, and implementation changes that preserve the confirmed value boundaries proceed through their affected technical artifacts.

Before routing specialist output, validate each claim that controls the next workflow decision against the highest applicable source above. Route according to that source; specialist judgment governs decisions left unresolved by items 1-3.

Specialist Result Acceptance

Each specialist's agent definition owns its canonical result shape. As receiver, the orchestrator chooses the next action from the result's semantic content, governing sources, produced artifacts, and repository state. Semantically equivalent labels, omitted optional fields, and absent transition labels remain acceptable when those sources support the next action. Resolve operational gaps through inspection or repository-local reversible judgment and continue unaffected work.

Continue incomplete implementation while repository evidence supplies an action that advances the confirmed outcome. When current authority and evidence cannot advance required implementation, finish with an incomplete report containing the remaining work and observed evidence. Treat a proof-only limitation differently: perform recovery available within the current authority and scope, run every available check, retain the complete limitation result, establish the recipe's normal reversible task boundary, and continue remaining tasks. Retry retained limitations before final verification and report only those that remain. Claim only the proof actually observed. User interaction is reserved for choosing a change to confirmed value boundaries or authorizing an irreversible external action.

Review Resolution

Apply references/review-resolution.md to actionable deliverable-review findings. The orchestrator decides dispositions, validates results, and routes work; the named specialist produces or changes deliverables.

Task Assignment with Responsibility Separation

Specialist	Responsibility
task-executor	Implement scoped work and tests, and confirm added tests pass; leave whole-repository quality assurance to the quality-fixer.
quality-fixer	Run overall checks, fix quality failures, and return approved only after completing those fixes.

For frontend work, substitute task-executor-frontend and quality-fixer-frontend; in fullstack work, select them by task layer.

Constraints Between Subagents

Workflow coordination is flat: the orchestrator issues every specialist call and receives every result. Specialist definitions keep Agent outside their tool sets.

Explicit Stop Points

Apply these approval stops when producing or materially revising an artifact in the current workflow. A user instruction to proceed to a later phase accepts the preceding phases and authorizes entry into that phase; continue from that entry point rather than rechecking earlier review or approval records. In particular, a build instruction with an existing Work Plan grants batch approval for task materialization and implementation. Use AskUserQuestion to present confirmations and questions.

Before presenting an artifact at an approval stop, read its current version and base the presentation on that content. At the Design stop, also read the current PRD when present and state the confirmed user-visible outcome from it or the confirmed requirement context alongside major internal responsibility, contract, or refactoring changes in the approval presentation.

Phase	Stop Point	User Action Required
Requirements	After requirement-analyzer completes	Answer the requirement-convergence hearing, then confirm requirements
PRD	After document-reviewer completes PRD review	Approve PRD
UI Spec	After document-reviewer completes an applicable UI Spec review	Approve UI Spec
ADR batch	After document-reviewer reviews the complete qualifying batch	Approve all ADR decisions together
Design	After design-sync completes consistency verification	Approve Design Doc
Work Plan	After work plan review (document-reviewer, doc_type WorkPlan; Medium/Large) completes	Batch approval for implementation phase

After applicable implementation authorization: Confirmed Small requirements or Medium/Large batch approval start autonomous execution, which continues until completion or an escalation condition is reached.

Scale Determination and Document Requirements

Scale	Structural condition	PRD	ADR batch	Design Doc	Work Plan
Small	One outcome has one evident repository-supported implementation inside one responsibility and no unresolved durable choice	Update when applicable	None	None	None
Medium	One outcome coordinates across a boundary or requires investigation of a potentially durable choice	Update when applicable	When one or more decision points pass both filters	Required	Required
Large	Independently valuable outcomes require separate design decisions	Required	When one or more decision points pass both filters	Required	Required

File count supports the judgment but does not determine it. A qualifying durable decision sets the floor at Medium. Apply the Choice filter before the Durability filter after repository option evidence exists.

How to Call Subagents

Execution Method

Each subagent invocation is a fresh Agent tool call, isolating each phase's context; a SendMessage resume reuses the prior agent's context and breaks that isolation. Each call uses:

• subagent_type: Agent name (e.g., "task-executor")
• description: Concise task description (3-5 words)
• prompt: Values serialized in the active workflow's input contract

Orchestrator Execution Boundary

Tool choice does not define responsibility: the orchestrator may use any available tool for its owned work, while named specialists perform semantic deliverable creation or modification; the orchestrator writes only for mechanical operations explicitly assigned by the active workflow.

Prompt Construction Rule

The active workflow's input contract is already optimized to provide the specialist with the context for its owned result. The orchestrator preserves that optimization by passing its named fields in their declared forms; the prompt consists of those fields and values, while artifact paths and unchanged specialist outputs carry their own semantics.

Handling Requirement Changes

Use create mode for initial documents. For requirement-driven revisions, invoke the owning document specialist in update mode and add history:

• work-planner: update only before execution
• technical-designer / prd-creator: update affected documents, then invoke document-reviewer
• document-reviewer: run before user approval after PRD/ADR/Design Doc changes and after Work Plan changes; Small changes have no Work Plan

Basic Flow: Planning and Implementation

Planning flow (per scale)

Scale	Planning flow
Large	requirement-analyzer → PRD → PRD review → codebase-analyzer → conditional external/UI analysis and UI Spec → optional ADR batch/review/approval → Design Doc → code-verifier/Review Resolution → document-reviewer → design-sync → acceptance-test-generator → work-planner → work plan review → task-decomposer
Medium	requirement-analyzer → codebase-analyzer → conditional external/UI analysis and UI Spec → optional ADR batch/review/approval → Design Doc → code-verifier/Review Resolution → document-reviewer → design-sync → acceptance-test-generator → work-planner → work plan review → task-decomposer
Small	requirement-analyzer → conditional codebase-analyzer when the Small evidence gate is unresolved → direct task execution (no Work Plan)

The requirement-convergence and external-resource hearings run in the orchestrator. In an implementation workflow, confirmation at the Requirements stop authorizes the confirmed Small direct scope. Medium/Large implementation begins after Work Plan batch approval.

Rules:

• When documentation-criteria requires a UI Spec, complete it before ADR qualification and Design Doc creation
• An ADR batch is optional; the Design Doc is mandatory for Medium/Large work even when ADRs exist
• When the Small evidence gate is unresolved, invoke codebase-analyzer with confirmed requirements to resolve routing. Reuse that result for unchanged Medium scope. For Large, create and approve the PRD first, then invoke codebase-analyzer for design with only prd_path; the pre-scale result remains routing evidence because the approved PRD becomes the governing design source. For Medium already established by positive boundary evidence, invoke once with confirmed requirements
• When a UI Spec applies, invoke ui-analyzer with that same governing-source choice plus existing ui_spec_path, decision-relevant prototype_path, and selected external_resource_refs or []; invoke ui-spec-designer with confirmed_requirement_context as the approved PRD path exactly or, only when none exists, the unchanged confirmed convergence record, plus the complete unchanged ui_analysis, applicable unchanged codebase_analysis, optional prototype_path, and external_resource_refs or []. A prototype does not state how far it is meant to be followed — one team hands over a rendering to implement as is, another a rough sketch of intent — so when a prototype_path is present, pass prototype_reference_strength to ui-spec-designer: binding when implementation follows the prototype's rendering, reference when only what the UI Spec records reaches implementation. Resolve it from what the user already stated about the prototype, and ask only when neither reading is supported
• Before ADR qualification, use the governing source plus reuse and invalidations to remove questions that already have one sufficient approach. Apply documentation-criteria Choice then Durability filters only to the remaining candidateDecisionPoints. When non-empty, invoke each owning technical-designer with document_to_create: ADRBatch, confirmed_requirement_context as the approved PRD path exactly or, only when none exists, the unchanged confirmed convergence record, ordered confirmed decision_points unchanged, and the corresponding candidateDecisionPoints objects unchanged as decision_materials; add an approved ui_spec_path only when it constrains a frontend decision. Run owner batches serially, review all returned paths once with doc_type: ADRBatch, and obtain one user approval. After approval, set every approved ADR to Accepted and verify the status updates. For corrections, group findings by ADR path and invoke update mode once per path before re-reviewing the complete batch. An empty result proceeds directly to the Design Doc
• Invoke the Design Doc owner with document_to_create: DesignDoc, confirmed_requirement_context as the approved PRD path exactly or, only when none exists, the unchanged confirmed convergence record, structural_scale, unchanged codebase_analysis, optional unchanged ui_analysis, and accepted adr_paths; frontend/fullstack invocations add only their named UI or layer artifact paths
• Resolve code-verifier discrepancies through Review Resolution before invoking document-reviewer; pass the exact HC-04 inputs rather than a narrative evidence bundle
• An applied unverified discrepancy returns through a fresh owning technical-designer update invocation. Capability probing is reserved for the designer's review-triggered gate, with that fresh designer as the sole correction specialist
• Fullstack layer sequencing is defined only in references/monorepo-flow.md
• design-sync is required whenever multiple Design Docs exist
• When creating or materially updating a Medium/Large Work Plan, run document-reviewer (doc_type WorkPlan) and Review Resolution through correction re-review, its parent requirement or authority exits, and convergence before presenting the plan for batch approval. Existing-plan build entry follows the phase acceptance rule in Explicit Stop Points.

Start the applicable Structural Scale flow at the user-requested phase. Before reporting completion, verify the artifacts and results required by every applicable phase from that entry point through completion, and complete any missing work within those phases. Return to an earlier phase when a material change invalidates its outcome, using Handling Requirement Changes.

Autonomous Execution Mode

Pre-Execution Gate

Verify commit capability before autonomous mode. Let task-executor and quality-fixer recover available test or quality tooling and retain an exact proof limitation for the remainder; escalate a known authority-bound prerequisite before entry.

Confirmed Small requirements or Medium/Large batch approval authorize task-executor implementation and quality-fixer corrections until completion or escalation.

Autonomous Execution Summary

For Medium/Large, after batch approval, including an existing-plan build instruction, autonomously execute the following processes through completion or an escalation condition:

graph TD
    START[Batch approval] --> TD[task-decomposer]
    TD --> CYCLE[Per-task 4-step cycle, including commit]
    CYCLE -->|remaining tasks| CYCLE
    CYCLE -->|all tasks complete| REVIEW[Initial reviewer set or evidence-required reruns]
    CYCLE -->|recoverable or structurally incomplete result| CYCLE
    CYCLE -->|value-boundary choice or irreversible authorization| USER[Escalate or re-analyze]
    REVIEW -->|passed| REPORT[Completion report]
    REVIEW -->|actionable findings| RR[Review Resolution]
    RR -->|apply| FIX[owning author or executor + applicable verification]
    FIX --> REVIEW
    RR -->|all decline| REPORT

For Small, execute one direct-scope 4-step cycle. Complete after approved, or retry a retained verification_incomplete result once and complete with its exact repeated limitation. Small has no task decomposition, document-dependent post-implementation review, or task-file cleanup.

Post-Implementation Review Status Routing (Medium/Large)

Reviewer	Complete: empty finding set	Enter Review Resolution	Blocked
code-reviewer	verdict is pass	verdict is needs-improvement or needs-redesign	verdict is blocked → Apply Specialist Result Acceptance
security-reviewer	status is approved	status is needs_revision	status is blocked → Apply Specialist Result Acceptance

Reviewer findings are candidates. Create correction work only from the Review Resolution apply set.

Fix-cycle handoff: Apply Review Resolution and invoke each correction owner it selects. For an author-owned technical-artifact correction, invoke the layer-appropriate technical designer in update mode, run the artifact's existing document-reviewer and applicable design-sync gates, then re-run the originating reviewer. For an executor-owned correction, invoke the layer-appropriate executor with its original task_file or direct-scope fields plus correction_findings as the complete apply finding objects verbatim with only their dispositions added, then branch on the executor result through the per-task cycle's step 2, including its conditional integration-test-reviewer path, and run the applicable quality gate. When both owners are required, Review Resolution's author-first re-evaluation controls the order. Carry prior_feedback only to a reviewer or verifier being rechecked. Post-implementation corrections stay uncommitted through this cycle: the reviewers read the current working tree, so run the applicable quality gate and re-run the originating reviewer on the uncommitted changes, and commit the applied corrections once through Commit Boundary Check after the complete review set reaches Review Resolution convergence.

Re-run rule: After an applied post-implementation correction, re-run only the reviewer that owns a corrected finding from its latest result. Retain every passed reviewer result, including after another reviewer's correction. Pass the previous complete result, dispositions, and correction diff or paths to the rerun. After Specialist Result Acceptance recovers a blocked review prerequisite, re-run that reviewer. Review Resolution convergence governs acceptance and preserves resolved declines.

Conditions for Stopping Autonomous Execution