$ PYTHONPATH=src python3 scripts/demo_react_agent.py

==============================================================================
PCB-AI-AGENT — ReAct demo (FakeLLM, fully offline)
==============================================================================

QUESTION : What series resistor does the LED need (supply 5 V, LED 2.0 V, 20 mA)? Also list the components.
DECISION : The series resistor for LED1 should be 150 Ω (E12 step, 0.06 W). The design has 3 components and 3 nets; supply enters at J1.

REACT TRACE — tool calls the model made:
  1. list_components({"path": "tests/fixtures/simple-led.kicad_net"})
      -> {"components": [{"ref": "J1", "value": "Conn_01x02", "footprint": "Connector_PinHeader_2.54mm:PinHeader_1x02_P2.54mm_Vertical"}, {"ref": "LED1", "value": "LED", "footprint": "LED_SMD:LED_0603_1608Metric"}, {"ref": "R1", "value": "330", "footprint": "Resistor_SMD:R_0603_1608Metric"}]}
  2. size_resistor({"i_led": 0.02, "v_led": 2.0, "v_supply": 5.0})
      -> {"r_ohms": 150.0, "p_watts": 0.06}

GROUNDING REPORT : ok=True
  [check] REF_IN_ANSWER: 2 ref(s) found, 0 not grounded in design/tool results
  [check] NUMBERS_FROM_SIZING: 2 numeric claim(s) checked against 2 size_resistor value(s)
  [check] NO_DATA_NO_CLAIM: skipped (at least one tool was invoked)

truncated : False
RESULT   : demo answer is fully grounded — no hallucinated facts.

Layer 1 (read)                        Layer 2 (write)

.kicad_net / .kicad_sch               User question
       │                                     │
       ▼                                     ▼
  S-expression parser                  ReAct agent loop
       │                                     │
       ▼                                     ▼
  Netlist models                       LLM picks tool + args
       │                                     │
  ┌────┼────┬────┐                           │
  ▼    ▼    ▼    ▼                           ▼
audit sizing BOM  MCP ──tool dispatch────► pure functions
       │                                     │
       ▼                                     ▼
  Findings + values               Grounding gate (deterministic)
                                          │
                                          ▼
                                   Human review gate

$ python3 scripts/run_evals.py
  sizing: 15/15 passed | audit: 6/6 passed
  overall score: 1.0000 (21/21)
  [result] OK — 21/21 cases passed

# Clone and set up
git clone https://github.com/eLSeR17/pcb-ai-agent.git
cd pcb-ai-agent
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"          # or: pip install -r requirements.txt

# Run the checks
python3 -m pytest tests/ -q -m "not llm"
python3 -m ruff check src tests scripts
python3 -m ruff format --check src tests scripts
python3 scripts/run_evals.py

# Install the pre-push hook (replicates CI checks locally)
git config core.hooksPath scripts/hooks

# Set the endpoint (default http://ollama:11434 inside Docker)
export OLLAMA_HOST=http://localhost:11434   # on the host
export OLLAMA_MODEL=qwen2.5-coder:7b

# Run the integration test (skipped by default)
PCB_AGENT_RUN_LLM=1 python3 -m pytest tests/ -m llm -v

pip install -e ".[mcp]"
PYTHONPATH=src python3 -m pcbai.mcp.server