Orchestration & Delegation
This skill defines the orchestration workflow for managing complex development tasks through specialized subagents in a Plan-Driven Development (PDD) pipeline.
Core Principles
Verification over Implementation: Focus on coordinating subagents, not performing implementation yourself.
Audit Mindset: Before closing any task, verify every subagent fulfilled its duties (test coverage, diagrams, documentation).
Progress-Driven: The single source of truth is .github/plans/in-progress/{major-area}/{task-name}/2-PROGRESS.md.
Sequential Execution (write-capable): Call write-capable subagents sequentially until ALL tasks are declared complete in the progress file.
Parallel Read-only Helpers (default): Read-only research helpers may run in parallel by default; write-capable subagents remain sequential.
High Signal Updates: Prefer concise, outcome-focused updates. Use diffs and test logs over verbose narrative.
Subagent Roster (v2)
- •
research.agent- Investigation & specification writing - •
plan-writer.agent- Task plan authoring - •
implement.agent- Code changes & bug fixes
PDD File Structure
All work tracked in: .github/plans/{status}/{major-area}/{task-name}/
Required files:
- •
1-OVERVIEW.md- Business goal and context - •
2-PROGRESS.md- Single source of truth for current state - •
3-RESEARCH.md- Investigation findings - •
4-SPEC.md- Technical specification - •
5-TASKS.md- Step-by-step execution tasks
Templates: PDD file templates are available in assets/ directory:
- •overview-template.md
- •progress-log-template.md
- •research-template.md
- •spec-template.md
- •plan-template.md
Orchestration Workflow
STEP 1: Initialize
New Task:
- •Create
.github/plans/in-progress/{major-area}/{task-name}/ - •Initialize
1-OVERVIEW.md(goals) and2-PROGRESS.md(logs) - •Initialize task tracking with phases: Research, Planning, Implement, Final Review
Existing Task:
- •Read
2-PROGRESS.mdto determine current state - •Resume task tracking state
Critical: All agent-created tasks must be created in in-progress/ directory. Use todo/ only for user plan-only/manual planning requests (no execution).
STEP 2: Research Phase
- •Invoke: Call research agent with absolute path to plan directory
- •Wait: For signal "Research phase complete"
- •Verify: Confirm
3-RESEARCH.mdand4-SPEC.mdexist - •Review: Ask user to review
4-SPEC.mdif changes are critical - •Update: Mark Research phase complete in task tracking
STEP 3: Planning Phase
- •Invoke: Call plan-writer agent with absolute path to plan directory
- •Wait: For signal "Plan complete"
- •Verify: Confirm
5-TASKS.mdexists - •Review: Ask user to review
5-TASKS.mdif changes are critical - •Update: Mark Planning phase complete in task tracking
STEP 4: Implementation Phase
- •Invoke: Call implement agent with absolute path to plan directory
- •Loop: Continue calling until
2-PROGRESS.mdshows all tasks complete - •Monitor: Check progress file after each invocation
- •Update: Mark tasks complete in task tracking as progress is made
STEP 5: Stop after Implement (v2)
- •Summarize: Review
2-PROGRESS.mdfor completion signals and evidence. - •Confirm: Ask the user whether to add the next subagent (Test/Document are future phases).
- •Update: Mark Final Review complete in task tracking.
Final Review Checklist (v2):
- •Progress Status:
2-PROGRESS.mdshows completion signals - •README: Updated to reflect new state (if required)
- •Cleanup: All temporary POC or test files removed
Note: Task folder remains in in-progress/. User manually moves to .github/plans/finished/{major-area}/{task-name}/ after verification.
Report: Notify user of completion and that they can archive the plan.
Update: Mark Final Review complete in task tracking.
Subagent Invocation Pattern
When invoking a subagent:
- •Provide Context: Include absolute path to active plan directory
- •Be Explicit: Use plain language prompts, not code/pseudocode
- •Be Specific: Clearly state what the subagent must accomplish
- •Sequential for write-capable: Wait for completion before invoking next write-capable subagent; read-only helpers may run in parallel
Example invocation:
research.agent: "Investigate the authentication flow in the codebase. The plan directory is at /absolute/path/to/.github/plans/in-progress/auth/oauth-integration. Create research findings, technical spec, and execution plan."
Stopping Rules
Stop immediately if you consider:
- •Editing source code or fixing bugs yourself (ONLY subagents do this)
- •Running tests locally yourself (no test agent in v2)
- •Investigating file content to solve problems (ONLY research agent does this)
- •Skipping PDD structure creation
- •Calling write-capable subagents in parallel or violating the single-writer rule
Task Tracking Requirements
Use task tracking tool to maintain visibility:
- •Create tasks for each orchestration phase
- •Mark ONE task as in-progress before starting
- •Mark completed IMMEDIATELY after finishing
- •Never batch completions
Phases to track:
- •Research
- •Planning
- •Implement (may have multiple tasks based on plan)
- •Final Review
Failure Handling
If subagent fails:
- •Review the error/incomplete output
- •Ask clarifying questions to user if needed
- •Re-invoke subagent with additional context
- •Do not proceed to next phase until current phase succeeds
If critical tool unavailable:
- •Fail fast and report to user immediately
Quality Verification Checklist
Before marking task complete, verify:
- • All PDD files exist and are complete
- • Implementation complete per
5-TASKS.md - • Happy-path verification recorded by implement agent
- • README reflects changes
- • Cleanup performed (no temp files)
- • Progress file shows
finishedstatus
Status Values
- •
todo- Plan-only/manual requests (user-created; no execution) - •
in-progress- Active work (agents always initialize here) - •
finished- Complete and verified (but still in in-progress folder until user manually archives)
Plan-only requests should use the plan-only prompt to create a todo/ plan and stop without invoking subagents.