openclaw-model-orchestrator

# OpenClaw Model Orchestrator

Dispatch tasks across multiple LLMs from chat. Uses AAHP v3 structured handoffs for minimal token overhead.

## Orchestration Modes

### Fan-Out
Split a task into parallel subtasks, each executed by a different model.
A planner model decomposes the task, workers execute in parallel, a reviewer merges results.

```
/orchestrate --mode fan-out --task "Build a REST API with auth" --planner copilot-opus --workers copilot52c,grokfast --reviewer copilot-sonnet46
```

### Pipeline
Chain models sequentially. Each model refines the previous model's output.
Ideal for plan -> implement -> review -> polish workflows.

```
/orchestrate --mode pipeline --task "Design and implement a caching layer" --planner copilot-opus --workers copilot52c,copilot-sonnet46 --reviewer copilot-opus
```

### Consensus
Send the same question to multiple models, then synthesize the best answer.
Identifies agreement, disagreement, and unique insights across models.

```
/orchestrate --mode consensus --task "What are the security risks of this API design?" --workers copilot-opus,gemini25,sonnet --reviewer copilot-opus
```

## Smart Recommendations

The orchestrator auto-classifies tasks and recommends optimal model combinations:

```
/orchestrate recommend "Build a REST API with JWT auth"
```

Returns: task classification, recommended planner/workers/reviewer, reasoning, and a ready-to-run command.

Use `help` as any flag value for context-aware recommendations:
```
/orchestrate --task "Audit security" --planner help
```

## Task Profiles

Pre-configured model combinations optimized for common task types:

| Profile | Planner | Workers | Reviewer | Use Case |
|---------|---------|---------|----------|----------|
| coding | copilot-opus | copilot52c, grokfast | copilot-sonnet46 | Feature development |
| research | gemini25 | gemini-flash, copilot-flash | copilot-opus | Analysis, investigation |
| security | copilot-opus | copilot-sonnet46, gemini25 | sonnet | Security audi...

# openclaw-model-orchestrator

> Multi-LLM orchestration for OpenClaw. Fan-out, pipeline, and consensus patterns with AAHP v3 handoffs.

## Install

```bash
clawhub install openclaw-model-orchestrator
```

## Quick Start

```bash
# Get help
/orchestrate help

# See available models
/orchestrate models

# Get recommendations for a task
/orchestrate recommend "Build a REST API with JWT auth"

# Fan-out: parallel subtasks across models
/orchestrate --mode fan-out --task "Build REST API" --planner copilot-opus --workers copilot52c,grokfast --reviewer copilot-sonnet46

# Pipeline: sequential refinement
/orchestrate --mode pipeline --task "Design caching layer"

# Consensus: compare model answers
/orchestrate --mode consensus --task "Security risks of this design?" --workers copilot-opus,gemini25,sonnet
```

## How It Works

1. **Classify** - Auto-detects task type (coding, research, security, review, bulk)
2. **Recommend** - Suggests optimal model combination based on task + available models
3. **Execute** - Runs orchestration with AAHP v3 structured handoffs between models
4. **Merge** - Reviewer model synthesizes final output

All inter-model communication uses AAHP v3 handoff objects (structured JSON, no raw chat history) for up to 98% token reduction.

## License

MIT