Local-first, zero-prefill multi-agent orchestrator. AMD APU native. No cloud. No NVIDIA.

• Multi-agent pipeline — python -m anvil --pipeline produces structured JSON output from a real LLM, end-to-end, on integrated graphics
• 18 tok/s generation on AMD Radeon Graphics (RADV RENOIR) — no discrete GPU required
• 7.27× KV compression vs fp16 baseline — asymmetric q8_0 keys + 3-bit FWHT Lloyd-Max values
• GBNF constrained decoding — LLM output guaranteed to match your Pydantic/JSON Schema, enforced at the token level via grammar
• Wasmtime sandbox — LLM-generated code executes in WebAssembly isolation, fuel-metered, no Docker

# 1. Bootstrap system + compile llama.cpp with Vulkan
./scripts/anvil_setup.sh

# 2. Download model
./scripts/download_models.sh

# 3. Start inference server
./scripts/start_llama_server.sh

# 4. Run the pipeline
python -m anvil --pipeline

# 5. Run multi-agent chain
python -m anvil --chain

CLI → AgentOrchestrator → InferencePipeline
    ├── LlamaClient      → llama-server (Vulkan)
    ├── GBNFCompiler     → JSON Schema → grammar
    ├── SSDPagePool      → KV blocks → safetensors (SSD)
    └── ContextScheduler → KV context switch between agents

Agents (3 registered + dynamic LLM callbacks):
- CodeAgent    — generates and sandbox-executes code
- ResearchAgent — structured summaries with citations
- DebugAgent   — error analysis and fix suggestions

pytest tests/ -v    # 325 tests, all passing
ruff check anvil/   # production lint clean

Apache 2.0