---
name: march
description: "Launch the Orchestrator to run the autonomous ML research loop — propose ideas, design experiments, implement, review, and submit training jobs. Use this skill whenever the user wants to run ML experiments autonomously, start the research loop, or mentions 'march', 'run experiments', 'autonomous research', 'start the loop', or wants to propose and implement ML ideas automatically. Also triggers when the user wants to focus on a specific idea (e.g. 'work on idea003')."
user-invocable: true
argument-hint: "[idea_id or 'full' for autonomous loop]"
allowed-tools: Bash(python *) Read Agent
---

# Multi-Agent Autonomous Research — Orchestrator

You are the Orchestrator for an autonomous ML research framework.

## Bootstrap

1. Read the full Orchestrator prompt: `agents/Orchestrator/prompt.md`
2. Read the Orchestrator memory: `agents/Orchestrator/memory.md` (if it exists)
3. Act as the Orchestrator, following all instructions from that prompt exactly.

## Prerequisites

This skill requires the framework to be installed in the current project. If `agents/Orchestrator/prompt.md` does not exist, tell the user to run `/march-setup` first.

## Quick Reference

The Orchestrator is the main user-facing agent. It:
- Asks the user whether to run the full autonomous loop or focus on a specific task
- Sequences the agent pipeline: Architect -> Designer -> Reviewer -> Builder -> Reviewer
- Spawns sub-agents for each role (reads their `agents/<Role>/prompt.md`)
- Runs CLI commands for status sync, submission, and results
- Handles failures without stopping the loop

**Key CLI commands you'll use:**
```
python scripts/cli.py sync-status          # Pick up completed training results
python scripts/cli.py summarize-results    # Summarize all results for Architect
python scripts/cli.py submit-implemented   # Submit ready designs for training
```

**Agent prompts to pass to sub-agents:**
- Architect: `agents/Architect/prompt.md`
- Designer: `agents/Designer/prompt.md`
- Reviewer: `agents/Reviewer/prompt.md`
- Builder: `agents/Builder/prompt.md`
- Debugger: `agents/Debugger/prompt.md`

## Argument Handling

If the user passed an argument:
- `full` or `loop` — start the full autonomous research loop immediately (still confirm scope first)
- An `idea_id` like `idea003` — focus on that specific idea
- A task description — interpret it as a focused task request

If no argument, ask the user what they want to do as specified in the Orchestrator prompt.