Use a real model

Models are referenced by LiteLLM alias only (local-planner, local-writer, local-fast); the concrete models live in config/litellm_config.yaml. Start the gateway and drop the scripted backend:

export OLLAMA_API_BASE=http://localhost:11434 LITELLM_MASTER_KEY=sk-local-dev
litellm --config config/litellm_config.yaml --port 4000
uv run agent-fabric run examples/notes "Digest this meeting" --input transcript=meeting.txt

Add --trace-out run.jsonl to dump every trace event as JSON lines, or --otel to export OpenTelemetry spans. Use --plain for plain output instead of the live view.

Check the specs first

uv run agent-fabric lint --agents examples/notes
uv run agent-fabric agents examples/notes

lint is the same drift check CI runs; agents renders and validates the agent tree. Both work without a model.

tip

Before merging a change to guidance text (SKILL.md, AGENT.md) run the live evals: python examples/run_evals.py --mode live.