Core infrastructure for a low-latency “co-drawer” system:

• Paper Pro streams stylus strokes (stroke-native input events)
• Desktop server routes those events over WebSocket and triggers an AI worker
• AI emits ghost-layer vector strokes (layer="ai")
• Clients render and animate (server/bridge never render)

This repo is infra-first (not a product demo yet).

• AI never overwrites user ink: AI output is always separate layer="ai".
• No per-point model calls: trigger AI only on stroke_end (micro-pauses later).
• Server routes, clients render: keep payloads incremental and small.
• Rate limit: design for ~50 RPM model caps (throttle + debounce).

• docs/protocol.md (canonical protocol)
• docs/architecture.md
• docs/latency_budget.md
• docs/remarkable_setup.md (connect + install on Paper Pro)

Requirements:

• Python 3.11+
• uv

Install and run:

uv sync
uv pip install -e .
uv run uvicorn codrawer_bridge.server.app:app --reload --host 0.0.0.0 --port 8000

Optional config:

• Copy env.example → .env and edit values (AI throttle knobs, future model keys).

This repo includes a fast local model gateway in model-server/ (OpenAI-compatible).

• Start it:

cd model-server
pnpm install
pnpm dev

• Configure Cerebras (in model-server/.env, not committed):

CEREBRAS_API_KEY=...

• Point the desktop server at it (in .env):

CODRAWER_MODEL_SERVER_URL=http://127.0.0.1:3100
CODRAWER_MODEL_SERVER_MODEL=blazing_fast

If your model supports vision, you can attach a small rendered PNG patch (local area around the stroke) to improve “what’s on the page” awareness:

CODRAWER_MODEL_SERVER_USE_CONTEXT_IMAGE=1
CODRAWER_MODEL_SERVER_CONTEXT_IMAGE_PX=256
CODRAWER_MODEL_SERVER_CONTEXT_IMAGE_WINDOW=0.22

Notes:

• This adds a bit of CPU + payload size (still small at 256×256).
• If the model doesn’t support images, leave it off.

Endpoints:

• Health: GET http://localhost:8000/healthz
• WebSocket: ws://<desktop-ip>:8000/ws/<session_id>
• Viewer: http://<desktop-ip>:8000/viewer/<session_id> (renders user vs AI layers)

See bridge/remarkable/README.md.

Record:

uv run python -m codrawer_bridge.tools.stroke_sim.record_jsonl --ws ws://127.0.0.1:8000/ws/session1 --out out.jsonl

Replay:

uv run python -m codrawer_bridge.tools.stroke_sim.replay_jsonl --ws ws://127.0.0.1:8000/ws/session1 --in out.jsonl --speed 1.0

uv run ruff check .
uv run mypy .
uv run pytest -q

Apache License 2.0. See LICENSE.