XAI :livekit-plugins-xai: RealtimeSession treats in-progress input transcriptions (status="in_progress") as final, duplicating user conversation items

[areebkhan-tech]

Bug Description

xAI's realtime API streams partial input transcriptions through the
conversation.item.input_audio_transcription.completed channel with
status: "in_progress", and sends the final with status: "completed".
(OpenAI, by contrast, streams partials via ...transcription.delta and sends
...completed only once.)

The xAI RealtimeSession overrides this handler but still delegates to the OpenAI
base implementation, which ignores status and always emits
InputTranscriptionCompleted(is_final=True):

# livekit/plugins/xai/realtime/realtime_model.py
def _handle_conversion_item_input_audio_transcription_completed(self, event):
    if remote_item := self._remote_chat_ctx.get(event.item_id):
        if remote_item.item.type == "message" and remote_item.item.text_content == event.transcript:
            remote_item.item.content.clear()
    super()._handle_conversion_item_input_audio_transcription_completed(event)  # emits is_final=True

So every in-progress partial is surfaced as a final transcription. Downstream,
AgentActivity finalizes a user ChatMessage and emits conversation_item_added
on each is_final=True, so one user utterance produces many
conversation_item_added (and user_input_transcribed(is_final=True)) events
with the same item_id and a growing transcript.

Observed (grok-voice realtime), single utterance:
completed status=in_progress "What is"
completed status=in_progress "What is my name? What is your"
completed status=in_progress "What is my name? What is your Wi-Fi?"
completed status=completed "What is my name? What is your Wi-Fi?"
-> 4 user conversation items emitted instead of 1.

This is xAI-specific (xAI's status-based streaming), so the fix belongs in the
xAI plugin's RealtimeSession, not in the shared OpenAI base handler.

Expected Behavior

The xAI RealtimeSession should honor xAI's status field: an "in_progress"
transcription is an interim and should be emitted as is_final=False (like a
delta), so it does not finalize a conversation item; only status="completed"
should yield is_final=True (and the single final conversation item).

Suggested handling in the xAI override (its own implementation for the realtime
session, without changing OpenAI base behavior):

def _handle_conversion_item_input_audio_transcription_completed(self, event):
    # xAI sends interim transcripts as "completed" with status="in_progress"
    if getattr(event, "status", None) == "in_progress":
        self.emit(
            "input_audio_transcription_completed",
            llm.InputTranscriptionCompleted(
                item_id=event.item_id,
                transcript=event.transcript,
                is_final=False,
            ),
        )
        return
    # status == "completed" (or absent): treat as final via base handling
    if remote_item := self._remote_chat_ctx.get(event.item_id):
        if remote_item.item.type == "message" and remote_item.item.text_content == event.transcript:
            remote_item.item.content.clear()
    super()._handle_conversion_item_input_audio_transcription_completed(event)

Result:

• in_progress -> is_final=False -> interim only, no duplicate conversation item.
• completed -> is_final=True -> exactly one final conversation item.
  Interim user_input_transcribed events are preserved for UI placeholders.

Reproduction Steps

1. Start an AgentSession with livekit-plugins-xai RealtimeModel
   (e.g. grok-voice-think-fast-1.0), input audio transcription enabled.
2. Subscribe to conversation_item_added (and/or input_audio_transcription_completed).
3. Speak one sentence.
4. Observe multiple is_final=True / conversation_item_added events for the same
   item_id with a growing transcript, instead of a single final one.

Operating System

macOS

Models Used

No response

Package Versions

livekit-agents === 1.5.17
livekit-plugins-xai == 1.5.17
livekit-plugins-openai == 1.5.17

Session/Room/Call IDs

No response

Proposed Solution

Additional Context

No response

Screenshots and Recordings

No response

[ByteMaster-1]

i'd like to work on this bug, will raise PR shortly.