isair/jarvis

Task 1d — Make low-confidence threshold and phrase configurable Part of Task 1: Low Confidence Voice Feedback. Depends on #374 (Task 1b) and #375 (Task 1c). What to do Once TTS and visual feedback are working (Tasks 1b and 1c), expose the key values as user-configurable settings. Add low_confidence_threshold (float, default matching current hardcoded value) to config.json / settings schema. Add low_confidence_phrase (string, default "I didn't quite catch that") to config. Read both values at runtime in the relevant handlers so users can tweak sensitivity and the spoken phrase without touching code. Files to touch Config schema / config.json defaults Handler introduced in Task 1b (read phrase from config) Listener threshold check in Task 1a (read threshold from config) Acceptance criteria low_confidence_threshold is read from config (falls back to current default if absent) low_confidence_phrase is read from config (falls back to "I didn't quite catch that" if absent) Changing

Task 1c — Visual indicator for low-confidence rejection Part of Task 1: Low Confidence Voice Feedback. Depends on #373 (Task 1a). What to do Using the low-confidence event from Task 1a, update the desktop widget UI to show a visual indicator when confidence is too low. Find where the desktop widget receives listener state updates. Add a new transient state (e.g. LOW_CONFIDENCE) that displays a visual cue — for example a subtitle line like "Didn't catch that...", a colour change on the listening indicator, or a brief shake/pulse animation. The indicator should auto-dismiss after ~2 seconds and return to normal idle/listening state. Do not change any TTS logic — this is UI only. Files to touch Desktop widget / tray UI layer (wherever listener state is rendered) src/jarvis/listening/state_manager.py (if a new state constant is needed) Acceptance criteria Widget shows a visible low-confidence cue when transcript is rejected Cue disappears automatically after ~2 seconds No vis

Task 1b — TTS response for low-confidence transcript Part of Task 1: Low Confidence Voice Feedback. Depends on #373 (Task 1a). What to do Using the low-confidence event surfaced in Task 1a, trigger a TTS response when Whisper confidence falls below the rejection threshold. In src/jarvis/reply/ or wherever the listener dispatches results, handle the low-confidence event. Call the TTS engine (in src/jarvis/output/) to say a short phrase such as "I didn't quite catch that" or "Sorry, could you repeat that?". The phrase should ideally be a constant or config value, not hardcoded inline. Do not speak this if Jarvis is already speaking or in cooldown (respect existing TTS guard logic). Files to touch src/jarvis/reply/ (dispatch/handler) src/jarvis/output/tts.py (if a new helper method is needed) Acceptance criteria Jarvis speaks an audible response when confidence is too low Response phrase is a named constant (not a magic string inline) Does not double-speak if TTS is already acti

Task 1a — Expose confidence score on low-confidence rejection Part of Task 1: Low Confidence Voice Feedback. What to do In src/jarvis/listening/listener.py, find where Whisper transcripts are filtered out due to low confidence (currently logs 🔇 Low confidence (x.xx): ...). Extract the confidence value and the reason for rejection into a structured object or named return value (e.g. a LowConfidenceEvent dataclass or a dict {"confidence": float, "transcript": str}). Emit or return this event so downstream code (reply engine, output layer) can react to it. Do not change any TTS or UI behaviour yet — this subtask is purely about surfacing the data. Files to touch src/jarvis/listening/listener.py Acceptance criteria Low-confidence rejection produces a structured event/value (not just a log line) The confidence score and raw transcript are both accessible from outside the listener Existing behaviour is unchanged (nothing new is spoken or displayed yet)