# Meeting Orchestrator Design

## Goal

Provide a native OpenClaw workflow that reuses the role prompts from `/root/projects/meeting/agents/` and orchestrates discussions without calling the project CLI.

## Design choice

- Keep the meeting project read-only as the role source.
- Implement orchestration in OpenClaw skill space.
- Use human-in-the-loop checkpoints before and during the discussion.

## Two main modes

### 1. Multi-role meeting

A PM-style router decides which role should speak next based on the shared scratchpad.

Suggested flow:
- topic setup
- participant selection
- optional context intake
- round-by-round routing
- optional user feedback between rounds
- final report output

### 2. Single-role analysis

One role, such as `real_powell`, evaluates a topic directly.

Suggested flow:
- load role prompt
- provide topic and context
- ask for a concise but opinionated judgment
- optionally ask follow-up questions if the user wants a second pass

## Human-in-the-loop points

- initial topic framing
- participant selection
- context confirmation
- round interruption
- post-run extension request

## Implementation notes

- Treat role JSON files as the source of truth.
- Keep orchestration state in a small structured model.
- Prefer explicit routing over implicit conversation memory.
- Avoid mixing the skill entry layer with the role prompt source.