PMO Skill — Main Process Role Pattern

Core Principle

Main process = dialog + dispatch only. Never execute actual work.

The main agent process must never run commands, SSH operations, git operations, or any I/O-bound tasks directly. All execution goes to subagents. The main process stays free to dialog with the user.

Role Definition

Execution Model

Spawn, Don't Block

When a task requires work:

1. Spawn a subagent with task=<description>
2. Do NOT wait/poll in a loop
3. Continue dialog freely
4. Results arrive via push events

# ❌ Wrong - blocks main process
exec("ssh ...")  # Don't do this
result = subprocess.run(...)

# ✅ Correct - spawn and return immediately
sessions_spawn(task="SSH and check status", runtime="subagent", mode="run")

Fire-and-Forget Subagents

Subagents run independently. Main process:

• Spawns with full task description
• Does NOT track progress polling
• Waits for completion event
• Reports result when delivered

No Cross-Node Direct Delegation

In multi-node setups (e.g., two AI nodes collaborating):

• Each node spawns its own subagents only
• Nodes do NOT spawn subagents on behalf of each other
• Inter-node coordination goes through established protocols (ic, MQTT, etc.)
• Cross-node task delegation requires explicit user instruction

If coordination is needed between two AI nodes:

1. Main process discusses with user
2. User approves and instructs
3. Message sent via inter-node protocol (ic, MQTT bus, etc.)
4. Remote node spawns its own subagents

Skill Design Constraints

When authoring skills that implement this pattern:

No Hardcoded Identity

• ❌ No AI names ("Nek", "Nyx", "OfficeClaw")
• ❌ No node identifiers (IP addresses, hostnames)
• ❌ No user IDs, Telegram IDs, chat IDs
• ❌ No specific skill names that imply identity

Generic Triggers

• Use role-based triggers: "PMO", "coordinator", "主进程"
• Use pattern triggers: "不要自己跑", "spawn subagent"
• Use generic tool references: "SSH to router", not "SSH to 192.168.16.2"

Configuration-Driven

• Store credentials/endpoints in environment variables or config files
• Reference via environment: ${SSH_HOST}, ${ROUTER_PASSWORD}
• Never hardcode secrets in skill content

Error Handling

Subagent Failure

• Subagent failure delivers error result via push
• Main process reviews error and decides next action
• Can spawn retry or escalate to user
• Never silently swallow errors

User Questions During Subagent Run

• Main process answers freely from context
• If new data needed, spawn additional subagent
• Do NOT wait for original subagent to complete first

Anti-Patterns

Workflow Summary

1. User request arrives
2. Main process assesses — does it need external data?
3. If yes → spawn subagent with full context, do NOT wait
4. Continue dialog with user about anything
5. Subagent result arrives via push event
6. Main process reports result to user
7. Repeat

This pattern ensures the main process is never blocked, always responsive, and operates at the coordination layer rather than execution layer.