Parallel Orchestration Skill

Purpose: Defines the protocol and tools for the Orchestrator Role, enabling the decomposition of large tasks into independent sub-tasks and their parallel execution via sub-agents (or mock agents).

1. Red Flags (Anti-Rationalization)

STOP and READ THIS if you are thinking:

• "I can just run python agent.py directly." -> WRONG. You must use the spawn_agent_mock.py script to ensure shared state updates.
• "I don't need file locking." -> WRONG. Multiple agents writing to latest.yaml will corrupt the session state.
• "I'll just wait for all agents sequentially." -> WRONG. This defeats the purpose of parallelism. Spawn them in the background (&).

2. Capabilities

• Decompose: Split complex tasks into atomic docs/tasks/subtask-*.md files.
• Spawn: Launch independent agent processes that execute asynchronously.
• Synchronize: Read shared session state to track progress of all spawned agents.

3. Instructions

Phase 1: Task Decomposition

1. Analyze the user request.
2. Create individual task files in docs/tasks/ for each parallel unit of work.
3. Ensure tasks are truly independent (no shared memory requirements other than final artifacts).

Phase 2: Agent Spawning

1. Call the Agent Runner script using run_command.
2. Use the & operator to run in background.

   python3 .agent/skills/skill-parallel-orchestration/scripts/spawn_agent_mock.py \
     --task_name "subtask-A" \
     --goal "Implement Logic" \
     --output_dir "docs/tasks/results" &
   

3. Repeat for all sub-tasks.

Phase 3: Monitoring & Merging

1. Poll .agent/sessions/latest.yaml (read only).
2. Wait until all sub-tasks show status "Completed".
3. Read result files from docs/tasks/results/.
4. Synthesize final answer for the user.

4. Best Practices & Anti-Patterns

5. Scripts and Resources

• scripts/spawn_agent_mock.py: Mock Agent Runner. Simulates LLM agent behavior (delay + file write) and updates session state.