Brings the python fork in line with the new feature/pc-tools work that
landed in rust. Same UX, same trigger phrases, same env vars.
config.py:
+ LLM_AUTO_FALLBACK = True
+ LLM_AUTO_FALLBACK_MIN_CHARS = 4
Mirrors the rust config knobs. The old "скажи X" trigger still works
(VA_TBR strips it before intent matching, so what remains hits this
fallback anyway) — but now ANY unrecognized query goes to Groq, not
just "скажи" ones.
main.py:
va_respond: replaced the fuzz-on-"скажи" gate with a unified auto-
fallback that fires whenever no command matches, GROQ_TOKEN is set,
and the utterance is at least LLM_AUTO_FALLBACK_MIN_CHARS long.
+ _speak(text): recorder.stop() → tts.va_speak → start, used by the
new action handlers so the mic doesn't pick up the assistant.
+ _set_clipboard(text): subprocess to PowerShell Set-Clipboard, no
new python deps.
+ _strip_trigger(voice, triggers): shared helper for codegen/file_search
where we need the text AFTER the trigger phrase.
New action handlers (registered in run_action dispatch):
+ groq_codegen — extracts query, calls Groq with strict "code only"
system prompt at temp 0.2, strips ```lang ... ``` fences, puts the
result in the clipboard, speaks "Код в буфере, сэр. Строк: N".
+ ocr_screen — PowerShell System.Drawing screenshot to a temp PNG,
shells to tesseract.exe (-l rus+eng), speaks a 120-char preview,
clipboard. Looks for tesseract on PATH first then standard install
dirs; gives a clear "Tesseract не установлен" if missing.
+ sysinfo — CIM-based battery/cpu/ram/disk via PS one-liners, time
via python datetime (gets us localized weekday for free). Topics:
time / battery / cpu / ram / disk / all.
+ file_search — strips trigger, recursive Get-ChildItem in Desktop/
Documents/Downloads (depth 3, top 5), opens explorer /select on
the first hit, speaks count and basename.
+ audio_devices_panel — opens mmsys.cpl. Proper COM-based device
switching (like the rust IPolicyConfig version) is deferred — TODO
in backlog.
commands.yaml:
Total 33 commands (was 19). New:
+ open_notepad / open_calculator / open_explorer / open_terminal
/ open_settings / open_task_manager / lock_screen / screenshot
— all via shell action (no AHK exes needed).
+ volume_up / volume_down — five chained keys=volumeup/volumedown
presses each, like rust's _volume_helper.ps1 with Times=5.
+ media_stop — keys=stop.
+ sysinfo_time / sysinfo_battery / sysinfo_cpu / sysinfo_ram /
sysinfo_disk / sysinfo_all — sysinfo action.
+ file_search, codegen, read_screen, switch_audio_device.
Replaced (were AHK exes that Priler ships and we don't have):
~ music → keys playpause
~ music_off → keys playpause (toggle)
~ music_next → keys nexttrack
~ music_prev → keys prevtrack
Removed (Priler hardware/setup specific, can't sensibly port):
- gaming_mode_on / gaming_mode_off (display config switch)
- music_save (depends on the specific player AHK was driving)
- switch_to_headphones / switch_to_dynamics (replaced with the
generic switch_audio_device that opens mmsys.cpl)
Kept as-is (worked already, no AHK dependency):
open_browser / open_youtube / open_google (url action)
sound_off / sound_on (pycaw via system.volume_mute/unmute)
thanks / stupid / off
Smoke-tested: yaml.safe_load → 33 commands, ast.parse main.py → OK.
Portability: USERPROFILE for screenshot path, $env:SystemDrive for disk,
PATH lookup for tesseract before falling back to standard install dirs.
No hardcoded user-specific paths.
Adds optional 'browser' field to url-action; runtime resolves it via
config.BROWSER_PATHS (subprocess.Popen with the chosen exe). When
omitted, falls back to webbrowser.open (system default). Constructor
GUI gains a browser dropdown on URL action and on multi-step URL.
Existing open_browser/youtube/google updated to use yandex by default
since user prefers Yandex Browser.
Replace the fixed 10s post-wake-word listen window with a VAD-driven
loop. Once at least 500ms of speech have been seen, 1.2s of silence
closes the window. Hard cap at 15s as a safety net.
Tunable in config.py via VAD_AGGRESSIVENESS, COMMAND_END_SILENCE_MS,
COMMAND_MIN_SPEECH_MS, COMMAND_MIN_LISTEN_MS, COMMAND_MAX_LISTEN_MS.
uses a second KaldiRecognizer with a JSON grammar of WAKE_WORDS + [unk] for
the wake phase, then hands off to the full-vocab recognizer for the 10s
command window. drops the pvporcupine dependency entirely.
