chore: refactor and add tests (#130)

* chore: add tests for end call

* Update pipecat module

* fix: allow interruptions from deepgram flux

* Add VadUserTurnStrategy

* chore: add test for voicemail detection

api/services/pipecat/event_handlers.py

@@ -10,16 +10,13 @@ from api.services.pipecat.in_memory_buffers import (
     InMemoryTranscriptBuffer,
 )
 from api.services.pipecat.pipeline_metrics_aggregator import PipelineMetricsAggregator
-from api.services.workflow.disposition_mapper import (
-    apply_disposition_mapping,
-    get_organization_id_from_workflow_run,
-)
 from api.services.workflow.pipecat_engine import PipecatEngine
 from api.tasks.arq import enqueue_job
 from api.tasks.function_names import FunctionNames
 from pipecat.frames.frames import Frame, LLMContextFrame
 from pipecat.pipeline.task import PipelineTask
 from pipecat.processors.audio.audio_buffer_processor import AudioBufferProcessor
+from pipecat.utils.enums import EndTaskReason
 
 
 def register_event_handlers(
@@ -83,18 +80,17 @@ def register_event_handlers(
     @transport.event_handler("on_client_disconnected")
     async def on_client_disconnected(_transport, _participant):
         call_disposed = engine.is_call_disposed()
-
         logger.debug(
             f"In on_client_disconnected callback handler. Call disposed: {call_disposed}"
         )
-        engine.handle_client_disconnected()
 
         # Stop recordings
         await audio_buffer.stop_recording()
 
-        # Only cancel the task if the call is not already disposed by the engine
-        if not call_disposed:
-            await task.cancel()
+        # End the call
+        await engine.end_call_with_reason(
+            EndTaskReason.USER_HANGUP.value, abort_immediately=True
+        )
 
     @task.event_handler("on_pipeline_started")
     async def on_pipeline_started(_task: PipelineTask, _frame: Frame):
@@ -114,9 +110,6 @@ def register_event_handlers(
         # Stop recordings
         await audio_buffer.stop_recording()
 
-        call_disposition = await engine.get_call_disposition()
-        logger.debug(f"call disposition in on_pipeline_finished: {call_disposition}")
-
         gathered_context = await engine.get_gathered_context()
 
         # Add trace URL if available (must be done before conversation tracing ends)
@@ -129,13 +122,6 @@ def register_event_handlers(
         # also consider existing gathered context in workflow_run
         gathered_context = {**gathered_context, **workflow_run.gathered_context}
 
-        organization_id = await get_organization_id_from_workflow_run(workflow_run_id)
-        mapped_call_disposition = await apply_disposition_mapping(
-            call_disposition, organization_id
-        )
-
-        gathered_context.update({"mapped_call_disposition": mapped_call_disposition})
-
         # Set user_speech call tag
         if in_memory_transcript_buffer:
             call_tags = gathered_context.get("call_tags", [])

api/services/pipecat/pipeline_builder.py

@@ -57,6 +57,10 @@ def build_pipeline(
     # Insert voicemail detector after STT if enabled
     # Note: We intentionally do NOT use voicemail_detector.gate() to allow TTS
     # frames to continue flowing during classification (non-blocking detection)
+
+    # Note: We must keep user_context_aggregator after voicemail_detector
+    # or else, LLMContextFrames generated from user_context_aggregator will
+    # start generating LLM Completion from Voicemail Classifier
     if voicemail_detector:
         logger.info("Adding native voicemail detector to pipeline")
         processors.append(voicemail_detector.detector())

api/services/pipecat/run_pipeline.py

@@ -52,9 +52,11 @@ from pipecat.processors.aggregators.llm_response_universal import (
     LLMUserAggregatorParams,
 )
 from pipecat.transports.smallwebrtc.connection import SmallWebRTCConnection
-from pipecat.turns.user_mute import MuteUntilFirstBotCompleteUserMuteStrategy
+from pipecat.turns.user_mute import (
+    CallbackUserMuteStrategy,
+    MuteUntilFirstBotCompleteUserMuteStrategy,
+)
 from pipecat.turns.user_start import (
-    ExternalUserTurnStartStrategy,
     TranscriptionUserTurnStartStrategy,
 )
 from pipecat.turns.user_start.vad_user_turn_start_strategy import (
@@ -547,7 +549,7 @@ async def _run_pipeline(
 
     if is_deepgram_flux:
         user_turn_strategies = UserTurnStrategies(
-            start=[VADUserTurnStartStrategy(), ExternalUserTurnStartStrategy()],
+            start=[VADUserTurnStartStrategy(), TranscriptionUserTurnStartStrategy()],
             stop=[ExternalUserTurnStopStrategy()],
         )
     else:
@@ -556,9 +558,16 @@ async def _run_pipeline(
             stop=[TranscriptionUserTurnStopStrategy()],
         )
 
+    # Create user mute strategies
+    # - CallbackUserMuteStrategy: mutes based on engine's _mute_pipeline state
+    user_mute_strategies = [
+        MuteUntilFirstBotCompleteUserMuteStrategy(),
+        CallbackUserMuteStrategy(should_mute_callback=engine.should_mute_user),
+    ]
+
     user_params = LLMUserAggregatorParams(
         user_turn_strategies=user_turn_strategies,
-        user_mute_strategies=[MuteUntilFirstBotCompleteUserMuteStrategy()],
+        user_mute_strategies=user_mute_strategies,
         user_idle_timeout=max_user_idle_timeout,
     )
     context_aggregator = LLMContextAggregatorPair(
@@ -606,7 +615,7 @@ async def _run_pipeline(
         @voicemail_detector.event_handler("on_voicemail_detected")
         async def _on_voicemail_detected(_processor):
             logger.info(f"Voicemail detected for workflow run {workflow_run_id}")
-            await engine.send_end_task_frame(
+            await engine.end_call_with_reason(
                 reason=EndTaskReason.VOICEMAIL_DETECTED.value,
                 abort_immediately=True,
             )