Orchestrate

Coordinate specialized agents to complete complex multi-step tasks efficiently.

Steps

1. Analyze and Decompose

Restate the task in your own words. Break it into discrete subtasks with clear boundaries and explicit dependencies. Classify each subtask by type so you can route to the right agent.

2. Build Execution Plan

Map each subtask to an agent and identify which can run in parallel vs. sequentially.

Agent Routing

Agent	Use For
oh-my-claudecode:executor	Code implementation, file edits
oh-my-claudecode:planner	Feature design, phase breakdown
oh-my-claudecode:architect	System design, interfaces
oh-my-claudecode:debugger	Root cause analysis, bug fixes
oh-my-claudecode:build-fixer	Build errors, type failures
oh-my-claudecode:explore	Codebase discovery, symbol mapping
oh-my-claudecode:verifier	Completion evidence, claim validation
oh-my-claudecode:quality-reviewer	Logic defects, maintainability

Execution Patterns

Sequential (later tasks depend on earlier):

agent-A → agent-B → agent-C

Parallel (independent tasks):

         ┌→ agent-A ─┐
trigger →├→ agent-B ─┼→ synthesize
         └→ agent-C ─┘

3. Execute with Progress Tracking

For each phase, announce it before dispatching:

[Phase 1/N] <Name> — dispatching <agent>
[Phase 2/N] <Name> — parallel dispatch: <agent-A>, <agent-B>
[Phase N/N] Synthesizing results

Dispatch independent tasks in the same message using parallel Task calls. Wait for results before starting dependent phases.

Pass full context to every agent: relevant file paths, function names, code snippets, and the specific subtask scope. Never spawn an agent without context.

4. Synthesize Results

After all phases complete, combine agent outputs into a unified result. Resolve any conflicts between agent outputs. Report what was completed, what changed, and any unresolved issues.

5. Validate

Confirm the work meets the original requirements. If anything is incomplete, dispatch a targeted follow-up agent rather than re-running the full orchestration.

Coordination Rules

1. Each agent owns distinct files — no overlap between parallel agents
2. Pass file paths and function names explicitly in every agent prompt
3. If an agent fails, diagnose before retrying — do not repeat blindly
4. After 3 failures on the same subtask, escalate to oh-my-claudecode:architect
5. Report progress after each phase; do not go silent during long orchestrations

Output Format

## Execution Plan
Phase 1: <Name> [agent]
Phase 2: <Name> [agent-A, agent-B — parallel]
Phase 3: Validation [verifier]

## Progress
[Phase 1/3] <Name> — complete
[Phase 2/3] <Name> — complete
[Phase 3/3] Validation — complete

## Results
<Unified summary of what was built/changed>

## Unresolved
<Any issues that need follow-up, or "None">