Fix 1 -- Opening-Prompt: processSessionOpening in serviceCommcoach.py prüft jetzt ob es die erste Session ist (isFirstSession) und gibt der AI einen expliziten Prompt, der das Erfinden von Kontext verbietet.
Fix 2 -- Stabiler Transcript: onresult in CommcoachCoachingView.tsx nutzt jetzt processedResultIndexRef um nur neue Results zu verarbeiten. Finalisierte Teile werden stabil akkumuliert, kein Flackern mehr.
Fix 3 -- Hintergrundgeräusche-Timeout: Neuer silenceTimerRef mit 5s Timeout. Wenn nach onspeechstart kein Text kommt, wird isUserSpeaking automatisch zurückgesetzt.
Fix 4 -- Stop-Button: "Stop" Button erscheint im Session-Header wenn TTS läuft (via isTtsPlaying State, synchronisiert per 200ms Interval mit isTtsPlayingRef).
Fix 5 -- Weitersprechen-Button: lastTtsAudioRef speichert das zuletzt gespielte Audio. stopTts setzt wasInterrupted = true. "Weitersprechen" Button erscheint nach Unterbrechung und spielt das Audio erneut ab.
Fix 6 -- Paralleles TTS: Neue _generateAndEmitTts() Hilfsfunktion. In processMessage und processSessionOpening wird TTS als asyncio.create_task parallel zu _emitChunkedResponse gestartet.
Fix 7 -- JSON-Response: Die AI antwortet jetzt als JSON mit text, speech, documents. Neuer Prompt-Block wird in buildCoachingSystemPrompt angehängt. _parseAiJsonResponse() und _saveGeneratedDocument() im Backend. processMessage und processSessionOpening nutzen die neue Struktur.
Fix 8 -- Loading-States: Neuer actionLoading State in useCommcoach. Alle async Funktionen setzen setActionLoading('key') vor dem Await und null im finally. Buttons zeigen Loading-Text und werden disabled.
Fix 9 -- Umlaute: Alle deutschen Strings in allen CommCoach-Dateien (Backend + Frontend) korrigiert: ae->ä, oe->ö, ue->ü.
Extend TrusteePosition with documentType, payeeIban, payeeName, payeeBic, paymentReference and dueDate fields. Overhaul AI extraction prompts to capture full document details in desc (line items, addresses, conditions), extract QR/ESR payment references and IBAN from invoices, and tag each record with its documentType. Sanitise and normalise all new fields on write.
Made-with: Cursor
Normalize AI-derived date/time and numeric fields before writes, tighten extraction prompts for unix timestamps, and prevent /positions crashes by normalizing or deleting irreparably invalid legacy TrusteePosition records during reads.
Made-with: Cursor
Keep aggressive hybrid persistence rules, but use final speaker hints as transient context to react to direct bot addressing even when audio capture is silent.
Made-with: Cursor
Read and log per-chunk capture diagnostics (trackId, readyState, rms, nativeSampleRate) from bot audioChunk messages to support root-cause analysis of silent audio streams.
Co-authored-by: Cursor <cursoragent@cursor.com>
- New config fields: botAccountEmail, botAccountPassword for dedicated MSFT account
- BrowserBotConnector passes credentials + backgroundImageUrl to bot service
- Service passes config credentials to connector in joinMeeting
- Enables: full language settings, virtual background, no lobby wait
- Fallback: anonymous join when no bot account configured
Co-authored-by: Cursor <cursoragent@cursor.com>
- SPEECH_TEAMS prompt: Rule 1 (bot name mentioned) now has highest priority
and ALWAYS triggers a response, even for greetings/smalltalk
- Bot-name detection in transcript overrides cooldown timer (immediate trigger)
- Added multilingual instruction: AI responds in the language it's addressed in
- Voice interface: clean camelCase mapping (audio_content -> audioContent)
- Config save: added "config" to allowed update fields in interfaceFeatures
- Default TTS voices for common languages in connectorVoiceGoogle
Co-authored-by: Cursor <cursoragent@cursor.com>
- HTTP POST endpoints: /bot/transcript/{sessionId} and /bot/status/{sessionId}
- Fallback for when Azure App Service blocks WebSocket upgrade
- Debug logging [WS-DEBUG] in WebSocket handler for message tracking
- Handle websocket=None in AI response pipeline (HTTP mode)
Co-authored-by: Cursor <cursoragent@cursor.com>
- Voice test endpoint generates sample text dynamically via AI in selected language
- Fixed config save: added "config" to allowed update fields in interfaceFeatures
- Clean camelCase mapping in interfaceVoiceObjects (audio_content -> audioContent)
- Default TTS voices for common languages in connectorVoiceGoogle
- Fixed updateFeatureInstanceConfig -> updateFeatureInstance with config field
Co-authored-by: Cursor <cursoragent@cursor.com>
- URL extraction: _extractTeamsMeetingUrl() handles SafeLinks, invitation text, and clean URLs
- Status mapping: added in_meeting, launching, navigating to bot status map
- Voice test endpoint: POST /{instanceId}/voice/test for TTS preview
- Improved error logging in browserBotConnector (shows exception type + repr)
- Improved AI trigger logging in service (shows buffer size, cooldown state, trigger reasons)
Co-authored-by: Cursor <cursoragent@cursor.com>
- Add getIsoTimestamp() to timeUtils for JS-compatible ISO 8601 strings
- Replace getUtcTimestamp() (epoch float) with getIsoTimestamp() for all
teamsbot session/transcript/response timestamp fields (startedAt, endedAt,
creationDate, lastModified, SSE event timestamps)
- Pass config.language to browser bot in join request for captions spoken language
Co-authored-by: Cursor <cursoragent@cursor.com>
