contributor examples

Your first TokenPrint change

Use Python 3.11/3.12 and Node 20.9+. Run python3 scripts/start.py from the repository root, or reopen the repository in its development container. The container installs dependencies without downloading weights; launch the model only when needed. For frontend-only work, run npm ci && npm run dev in frontend/, then choose Try a recorded demo.

Add a visualization panel

1. Copy examples/contributors/TokenProbabilityPanel.tsx to frontend/components/ui/TokenProbabilityPanel.tsx. It reads actual recorded top-k probabilities, preserves their original normalization, and provides an empty state.
2. Add { id: "token_probability", label: "Token probabilities", purpose: "Inspect captured next-token probabilities." } to the ANALYSIS tools in frontend/lib/debuggerTools.ts.
3. Import the component in frontend/components/ui/DebuggerPane.tsx and add token_probability: <TokenProbabilityPanel /> to views.
4. Open the recorded demo, then Debugger → Token probabilities. Confirm values match frontend/public/demo/hello-world.json, that empty state works before loading, and that navigating away unmounts the panel.
5. Run npm run test:unit and npm run build in frontend/.

Keep inference outside React components. Declare missing measurements rather than deriving fake attention, timing, or causal evidence from unrelated values. Shared scene labels use SceneHtml; passive labels are collision-filtered and remain below dialogs.

Add a model capability adapter

Run the small example:

PYTHONPATH=backend backend/.venv/bin/python examples/contributors/model_adapter.py

Copy its class to backend/app/inference/adapters/<family>.py, give it an exact model_type match, and add an instance to SPECIFIC_ADAPTERS in that directory's __init__.py. Register before the generic fallback. Add positive and negative config fixtures to backend/tests/test_capabilities.py.

A capability adapter describes support; it does not implement a new inference architecture. Before enabling a capability, implement and verify the corresponding runtime module layout in the engine/instrumentation adapters. Test against a real small checkpoint: compare attention to an independent forward pass, check layer/head shapes, ensure hooks are removed after errors, and demonstrate that an intervention changes the measured output. Keep unverified capabilities disabled with a reason, as the example does.

Checks and a useful PR

npm run test:unit covers pure logic; npm run test:store covers streaming state. backend/.venv/bin/python -m pytest backend/tests runs backend tests (install pytest first). The browser tests require a frontend on port 3000; first-use tests deliberately block the backend. A PR should state the user-visible behavior, data source, and actual checks run. See beginner tasks for current starting points.