Quick Start

Commands

Command	Description
/research init	Initialize a new research project. Creates .omc/research/ directory structure and pipeline.yaml.
/status	Show current pipeline stage, agent health, and pending tasks.
/gate [stage] [type]	Set the gate type for a stage. Types: human, auto-judge, auto.
autopilot	Run the pipeline automatically between stages. Pauses at gates.
ralph	Enter a tight agent loop within the current stage (e.g., code-test-fix).
ultrawork	Execute multiple independent tasks in parallel across agents.

Commands run in the Orchestrator session

All commands are issued to the Orchestrator (the main Claude Code session). The Orchestrator then dispatches work to the appropriate agents.

Workflows

Morning: Hands-on Ideation

You have a vague idea and want to explore it interactively.

You:  /research init "efficient attention mechanisms"

      # Orchestrator creates project, invokes Scout for literature
      # Scout returns recent papers on efficient attention

You:  The linear attention papers look promising. 
      What if we combined RetNet's recurrent form with flash attention?

      # Orchestrator refines the idea, invokes Judge for evaluation
      # Judge returns structured verdict: novelty 7/10, feasibility 8/10

You:  Good. Let's move to design. I want three baselines.
      /gate design human

      # Orchestrator advances to Design stage
      # Planner creates experiment plan
      # You review and approve before Implementation begins

Morning sessions are collaborative

During ideation, keep gates on human so you stay in the loop. Your domain expertise + the system's breadth = better ideas.

Evening: Autopilot Overnight

Your experiment design is approved and implementation is ready. You want training to run overnight without intervention.

You:  /gate implementation auto-judge
      /gate training auto
      /gate analysis auto-judge
      autopilot

      # Orchestrator drives the pipeline:
      # 1. Coder implements the design
      # 2. Judge reviews code → auto-approves if clean
      # 3. Training launches, monitoring activates
      # 4. Phase 1: active watch for first 1000 steps
      # 5. Phase 2: CronCreate patrols every 30 min
      # 6. Training completes → Analysis stage
      # 7. Judge evaluates results → reports in morning

Check in the morning

Autopilot pauses at human gates and when errors exceed retry limits. Check /status when you wake up to see where things stand.

Intervene Anytime

You can always step in, even during autopilot.

      # Training is running on autopilot...

You:  /status
      # Shows: Stage 4 (Training), step 45000/100000, loss 2.31

You:  The loss plateaued. Reduce learning rate by 10x and continue.

      # Orchestrator instructs Coder to modify config
      # Training resumes with new LR
      # Autopilot continues

      # Autopilot is writing the paper...

You:  Stop. The related work section needs to mention the RetNet paper.

      # Orchestrator pauses Writer
      # Scout fetches RetNet paper details
      # Writer resumes with additional context

You are always the highest authority

Human input always takes priority over any automated decision. The system pauses, incorporates your input, and continues. No work is lost.

Gate Configuration

Gates control how much human involvement each stage requires.

Gate Type	Behavior	Best For
human	Pauses and waits for your explicit approval	Ideation, critical design decisions
auto-judge	Judge evaluates and auto-approves if criteria pass	Implementation, analysis
auto	Proceeds automatically without review	Training, routine tasks

Change gates at any time:

/gate ideation human        # Always want human input here
/gate implementation auto-judge   # Let Judge review code
/gate training auto          # Training just runs

Default gate configuration

By default, all gates are set to human. This is the safest starting point. As you gain confidence in the system, gradually open gates to auto-judge and then auto.

The recommended progression:

1. First project: all human
2. Second project: auto-judge for implementation and analysis
3. Established workflow: auto for training, auto-judge for implementation/analysis
4. Never set ideation to auto — your ideas are the input

Execution Modes

The three modes are orthogonal — they control different axes and can be combined.

Mode	Controls	Scope
autopilot	Stage transitions	Between stages
ralph	Retry/fix loops	Within a stage
ultrawork	Parallelism	Across agents

Combine freely:

• autopilot + ralph: auto-advance stages, auto-fix errors within stages
• autopilot + ultrawork: auto-advance with parallel agent execution
• All three: maximum automation

Next Steps

• Research State — understand what gets persisted
• Monitoring — how training is watched
• Pipeline Overview — the seven stages in detail