Establish a single concurrency standard: in async code, never block the event loop; in multi-thread/state-machine code, make ownership and cancellation atomic; and in cleanup/shutdown, use bounded, interrupt-friendly paths. 1) Keep async nonblocking - Any `subprocess.run`, CPU-heavy work (ffprobe/ffmpeg, compressors, summary generation), synchronous SDK calls, or blocking I/O must be offloaded. - Pattern: ```py # BAD: blocks event loop result = subprocess.run([...], timeout=60) # GOOD result = await asyncio.to_thread(subprocess.run, [...], timeout=60) # or result = await adapter._run_blocking(func, *args) ``` 2) Re-check cancellation/flags after waits - If a worker can be paused in `Event.wait()` (or similar), it must re-check the controlling mode/flag at every boundary where it could restart capture/playback/processing. - Pattern: ```py await voice_done.wait() if cli_ref._voice_continuous is False: return # cancel boundary ``` 3) Lock/claim correctly; snapshot atomically - When multiple pieces of state must be consistent (e.g., status + payload), take the same lock for a single snapshot. - When there is a single owner (wake-word listener, shared callback transport, adapter connection-local services), implement a real lease/claim/release so late callers can’t steal delivery. 4) Shutdown/cleanup must be bounded and interrupt-friendly - Never perform blocking cleanup (e.g., `shutdown(wait=True)`) before the code path that needs to deliver an interrupt/watchdog action. - Prefer bounded waits or non-blocking cancellation first: ```py # BAD if interrupt must run promptly pool.shutdown(wait=True, cancel_futures=True) # BETTER: cancel first, then bounded wait, then allow interrupt to proceed pool.shutdown(wait=False, cancel_futures=True) # or wait with a timeout in your own logic ``` 5) Add regression coverage for the real concurrency boundary - For cross-thread queueing (e.g., `put_threadsafe()`), tests must involve a real worker thread, not only same-loop calls. - For ownership/leases and cross-process locks, add OS-process regressions (where supported) to validate the intended guard behavior.