windows-call-detection

name

windows-call-detection

description

Use when detecting whether the user is in a call / on the microphone on Windows, or gating SpeedReader MCP agent speech on call state (Core/call_detection.py, mcp.pause_when_mic_in_use). Covers the CapabilityAccessManager registry mic-in-use signal, the recursive packaged/NonPackaged walk, fail-open-to-False behavior, and how the speak tool suppresses (not queues) speech during a call. Use to avoid blocking speech on detection errors or coupling detection to tkinter.

Windows call / microphone detection

GUI-free "is the user in a call?" detection, used to optionally pause MCP agent speech. See Core/call_detection.py.

When to Use

• Detecting microphone-in-use / "in a call" state on Windows.
• Changing how the MCP speak tool gates on call state (mcp.pause_when_mic_in_use).
• Adding another detection signal or platform.

The signal (mic-in-use via the registry)

• Windows records per-app microphone consent under HKEY_CURRENT_USER\SOFTWARE\Microsoft\Windows\CurrentVersion\CapabilityAccessManager\ConsentStore\microphone.
• Each app key has a LastUsedTimeStop QWORD (FILETIME). == 0 means the mic is currently in use; a non-zero value is the time it stopped. We treat "any app holding the mic" as a proxy for "in a call".
• Packaged (Store) apps store the value directly on their key; desktop apps live one level down under the NonPackaged container. REPEAT: you MUST recurse into subkeys — checking only the top level misses desktop apps like Zoom/Teams classic.

Implementation rules (HIGH-RISK — be repetitive)

• microphone_in_use(scan=None) is the public entry. scan is injectable for tests; default walks the registry.
• REPEAT: fail open to False — any error (non-Windows, missing key, access denied) returns False so a detection failure NEVER blocks/suppresses speech. Wrap the scan in a broad except Exception.
• REPEAT: non-Windows returns False early (sys.platform). Import winreg lazily inside the scan, never at module top (keeps Core importable/testable cross-platform).
• _key_active(winreg, key) is the recursive unit: check LastUsedTimeStop == 0 on the key, then recurse into every subkey. Both winreg and the key are passed in so tests drive it with a fake winreg + nested-dict tree — NO real registry needed (see tests/test_call_detection.py).

Gating MCP speech (suppress, not queue)

• mcp.pause_when_mic_in_use (bool, default false) in config.json turns this on. It is threaded config.pause_when_mic_in_use → start_http_in_thread → McpHost → build_mcp(pause_when_mic_in_use, call_active); call_active defaults to microphone_in_use and is injectable.
• In the speak tool: if pause_when_mic_in_use and call_active(): return "Skipped..." — REPEAT: it suppresses and returns a message, it does NOT queue/defer/wait for the call to end (waiting would block the agent's tool call).
• REPEAT: gating lives ONLY in the MCP speak tool — it must NOT touch SpeakService or the GUI, so the user's own reading is never paused.
• Off by default; only affects agent/MCP speech.

Testing

• Unit-test _key_active with a fake winreg backed by a nested dict (values/subkeys); cover packaged active, NonPackaged active, and all-stopped. Unit-test microphone_in_use with injected scan returning True/False and one that raises (asserts fail-open). See tests/test_call_detection.py.
• Test the gate via the MCP tool with a fake service + call_active=lambda: True/False (see tests/test_speak_gating.py); assert the service is/ isn't called. REPEAT: never touch the real registry or pyttsx3 in tests.