Refactor voice control core and robot behavior

CLAUDE.md

@@ -0,0 +1,44 @@
+# CLAUDE.md
+
+## Purpose
+
+Project guidelines for Claude Code working in this repository.
+
+## Project Overview
+
+Robotics project using Dora to control a ULite6 robot arm with Spanish voice commands for children.
+
+## Building and Running
+
+- Build: `dora build <dataflow.yml> --uv`
+- Start: `dora  run <dataflow.yml> --uv`
+
+## Key Packages
+
+- `dora_voice_control`: Voice processing and robot behavior
+- `dora_yolo_object_detector`: YOLO-based object detection
+- `dora_ulite6`: Robot arm control
+- `dora_zed_cpp`: ZED camera (C++)
+- `dora_iobridge`: WebSocket bridge for external clients
+
+## Robot Adapter
+
+Voice control uses a robot adapter pattern to support different robots:
+
+- `ROBOT_TYPE=vacuum`: Vacuum gripper (grab→vacuum_on, release→vacuum_off)
+- `ROBOT_TYPE=gripper`: Parallel gripper (grab→gripper_close, release→gripper_open)
+
+To add a new robot, create a new adapter in `robot/adapter.py`.
+
+## Guidelines
+
+- Keep the event loop responsive. Run slow operations (LLM) in threads.
+- Robot commands are serialized: one command at a time, wait for status before next.
+- State is shared via `SharedState` (thread-safe).
+- Handlers process events, tick callbacks run after each event.
+
+## Safety
+
+- This controls a real robot. Test changes carefully.
+- Workspace bounds are enforced in behaviors.
+- Use `DRY_RUN=true` to test without robot motion.