ghost-orchestrator 0.1.2a1

Experimental smart routing and LLM-powered orchestration for Ghostclaw analysis

Ghost Orchestrator

Experimental adaptive plugin routing for Ghostclaw.

"Let the LLM conduct the plugin orchestra — or keep it simple with smart heuristics."

---

⚠️ Prerequisites

• Ghostclaw ≥ 0.2.1b0 (v0.2+ with QMD support). Install from GitHub:
  pip install git+https://github.com/openclaw/ghostclaw@develop (or a released version with QMD)
• QMD enabled for full functionality:
  ghostclaw config set use_qmd true

Without QMD history, the orchestrator falls back to heuristic scoring only.

---

🎯 What Is It?

Ghost Orchestrator is an experimental plugin router for Ghostclaw that intelligently selects which analysis plugins to run on a repository. Instead of running every plugin, it:

1. Discovers available plugins via entry points
2. Examines the repository (languages, size, file patterns)
3. Consults past analysis history via QMD vector similarity (if enabled)
4. Optionally asks an LLM to create a plan
5. Executes selected plugins in parallel (configurable concurrency)
6. Deduplicates issues and aggregates results into a single report

Goal: Reduce noise, improve relevance, and lower compute cost by avoiding unnecessary plugin runs.

Compatibility: Requires Ghostclaw v0.2.1-beta or later (v0.2+). Depends on QMD storage, plugin auto-discovery, and async adapter infrastructure.

---

🏗️ Architecture

┌─────────────┐
│ Repository  │
└──────┬──────┘
       │
       ▼
┌──────────────────────────────────────────────┐
│          Ghost Orchestrator                   │
│  ┌────────────────────────────────────────┐  │
│  │ 1. Discover Plugins                    │  │
│  │    - Scan entry points                 │  │
│  │    - Load capabilities                 │  │
│  └────────────────────────────────────────┘  │
│  ┌────────────────────────────────────────┐  │
│  │ 2. Generate Analysis Plan              │  │
│  │    ├─ VectorAdvisor (QMD similarity)  │  │
│  │    ├─ HeuristicFallback               │  │
│  │    └─ LLMPlanner (optional)           │  │
│  └────────────────────────────────────────┘  │
│  ┌────────────────────────────────────────┐  │
│  │ 3. Execute Selected Plugins            │  │
│  │    - Parallel or sequential            │  │
│  │    - Merge results                     │  │
│  └────────────────────────────────────────┘  │
└──────────────────────────────────────────────┘
       │
       ▼
┌─────────────┐
│   Report    │
└─────────────┘

---

🚀 Quick Start

Installation

# From PyPI (once published)
pip install ghost-orchestrator

# Or from local build
pip install dist/ghost_orchestrator-0.1.0a1-py3-none-any.whl

# Or editable install for development
cd ghostclaw-experimental/orchestrator
pip install -e .

This registers orchestrator as a Ghostclaw plugin via entry points. No further setup needed.

Usage

Basic: Enable QMD and Run

# 1. Enable QMD in Ghostclaw (for vector similarity)
ghostclaw config set use_qmd true

# 2. Run analysis — orchestrator auto-loads and selects plugins
ghostclaw analyze /path/to/repo --no-ai --no-cache

# 3. Check the report: you'll see which plugins were selected
# Look in .ghostclaw/storage/reports/ for "Orchestrator selected: ..."

Orchestrator runs automatically on every analysis when it's installed. You can adjust its behavior via Ghostclaw config.

---

⚙️ Configuration

Add to your Ghostclaw config (~/.ghostclaw/ghostclaw.json):

{
  "orchestrator": {
    "use_llm": false,
    "llm_model": "openrouter/anthropic/claude-3-sonnet",
    "vector_weight": 0.7,
    "heuristics_weight": 0.3,
    "max_plugins": 8,
    "max_concurrent_plugins": 4,
    "plugin_history_lookback": 50,
    "enable_plan_cache": true,
    "plan_cache_ttl_hours": 24
  }
}

Options:

Field	Type	Default	Description
use_llm	bool	false	Use LLM for planning instead of vector+heuristics
llm_model	string	openrouter/anthropic/claude-3-sonnet	Model to use when use_llm=true
vector_weight	float	0.7	Weight given to QMD vector similarity (0-1)
heuristics_weight	float	0.3	Weight given to rule-based heuristics (0-1)
max_plugins	int	8	Maximum plugins to execute per analysis
max_concurrent_plugins	int	4	How many plugins to run concurrently (async)
plugin_history_lookback	int	50	How many past runs to consider for vector scoring
enable_plan_cache	bool	true	Cache analysis plans to avoid re-planning
plan_cache_ttl_hours	int	24	How long to keep a cached plan (TTL)

Note: Plugin capabilities are auto-discovered via entry points. Built-in plugins have known descriptors; custom plugins can provide PluginCapability metadata. Orchestrator is always enabled when installed — no separate enabled flag needed.

---

📦 Package Distribution

The package is published to PyPI as ghost-orchestrator. Requires Ghostclaw ≥0.2.1b0 (install separately if not already).

---

🔧 Development

Project Structure

orchestrator/
├── src/ghost_orchestrator/
│   ├── __init__.py
│   ├── models.py           # Data classes
│   ├── vector_advisor.py   # QMD-based plugin ranking
│   ├── llm_planner.py      # LLM-based planning
│   ├── orchestrator.py     # Main engine
│   └── plugin.py           # Ghostclaw plugin wrapper
├── tests/
│   └── test_orchestrator.py
├── pyproject.toml
└── README.md

Running Tests

cd orchestrator
pytest tests/ -v
# 7 tests: config validation, models, orchestrator creation, plan generation, fallback

Manual Testing

# Quick smoke test
python3 -c "
from ghost_orchestrator.plugin import OrchestratorPlugin
import asyncio

plugin = OrchestratorPlugin(config={})
result = asyncio.run(plugin.analyze('/path/to/repo', []))
print('Selected:', result.get('red_flags'))
print('Plan:', result.get('metadata', {}).get('orchestrator_plan'))
"

---

📊 Current Status

Phase: Pre-PyPI Release (2026-03-21)

Feature	Status
Plugin discovery via entry points	✅ Done
VectorAdvisor with QMD + heuristic fallback	✅ Done
LLMPlanner (OpenRouter/OpenAI)	✅ Done
Orchestrator engine (plan → execute)	✅ Done
Ghostclaw plugin wrapper	✅ Done
QMD store integration	✅ Done
Repository profiling (languages, frameworks, git)	✅ Done
Plugin capability descriptors	✅ Done
Result deduplication	✅ Done
Parallel plugin execution	✅ Done
Plan caching	✅ Done
Detailed metrics (latency, cache hit, qmd_hit)	✅ Done
Benchmark framework	✅ Done (scripts ready)
Packaging & publishing	✅ Done (built, validated)

Note: Orchestrator loads and runs selected plugins via entry points. Results are merged, deduplicated, and annotated with metrics.

---

🔮 Roadmap

Already Implemented (v0.1.0a1)

• See status table above — all core features are complete.

Future Enhancements

• Benchmark on diverse repos to quantify impact (plugin reduction, coverage)
• Tune LLM prompts and add cost tracking
• Consider collaborative filtering or reinforcement learning
• Possibly integrate into Ghostclaw core (if evaluation positive)

---

🧪 Research Questions

---

🧪 Research Questions

1. Effectiveness: Does orchestration reduce plugin count without missing important issues?
2. Speed: What's the latency improvement? (Planning cost vs execution savings)
3. QMD value: Is vector similarity actually predictive, or are heuristics enough?
4. LLM vs heuristic: Is the added LLM cost worth it for plan quality?
5. Cold start: How does it behave on repos with no QMD history?

We'll measure by:

• Running baseline (all plugins) vs orchestrated (selected subset) on same repos
• Comparing issue sets (precision/recall)
• Timing each plugin execution
• Tracking QMD hit rates

---

🤝 Contributing

This is an experimental module. If you'd like to help:

1. Try it on your own repos and report findings
2. Implement missing pieces (see Roadmap)
3. Benchmark and compare strategies
4. Suggest better heuristics or advisor algorithms

Guidelines:

• Keep changes isolated to orchestrator/
• Write tests for new features
• Update this README and CHANGELOG.md (to be created)
• Follow Ghostclaw's coding style (black, ruff, mypy)

---

📚 See Also

• Ghostclaw Plugin System: ../plugins/README.md
• QMD Documentation: (coming soon)
• Ghostclaw Core: https://github.com/your-org/ghostclaw

---

📝 License

MIT — same as Ghostclaw core.

---

Status: Experimental — may change or be removed without notice.