Hacktoberfest 2026: los issues que los mantenedores marcaron para octubre, abiertos y aptos para principiantes. Explorar issues de Hacktoberfest

Google realtime: on gemini-3.1 a resumed session loses the seeded history, which #6000 no longer replays

Cerrado
  • ¿Algo no está bien?
#7,697 0 comentarios 0 reacciones 0 asignados Ver en GitHub

Los mantenedores suelen responder en 1 día

Nadie ha tomado este issue todavía.

Evaluación

Dificultad
4/5
Tiempo estimado
3-5 días
Aptitud para principiantes
35/100
Tipo de issue
Error
Claridad
Bastante claro
Estado de actividad
Activo
Stack tecnológico
python
Área
ai, audio-video-rtc

Línea de trabajo

Start with the repro script in the issue, which runs google-genai directly with and without history_config set to initial_history_in_client_content, and see which path loses the seeded client content on resume. Then find where the plugin skips chat context replay on resume after #6000 and gate that on whether history_config is set. Done when the 3.1 resume path replays the seeded history and the repro reports kept_seeded_history as true. The issue leaves the replay-versus-diff choice open, so check with maintainers first.

Escrito por el modelo de indexación a partir del texto del issue.

Descripción

What

On gemini-3.1-flash-live-preview, a session resumed with a session resumption handle no longer has the history that was seeded as client content. Since #6000, the plugin skips the chat context replay when it resumes, so after any reconnect or goAway restart the agent has forgotten its initial chat context.

For 3.1, the plugin sets history_config=HistoryConfig(initial_history_in_client_content=True), because the model has no mutable chat context. It then sends the initial chat context once, as client content with turn_complete=False. #6000 assumes "the handle restores the conversation; send only what came after it". With history mode on, the server does not restore that seeded history.

Repro

This uses google-genai 2.29.0 directly, with the same setup the plugin sends for 3.1, and replays nothing on resume, as #6000 does:

import asyncio, os, re
from google import genai
from google.genai import types

MODEL = "gemini-3.1-flash-live-preview"
HISTORY_MODE = os.environ.get("HISTORY", "1") == "1"
client = genai.Client(api_key=os.environ["GOOGLE_API_KEY"])

def config(handle=None):
    return types.LiveConnectConfig(
        response_modalities=["AUDIO"],
        output_audio_transcription=types.AudioTranscriptionConfig(),
        system_instruction="You are a terse test assistant. Answer in as few words as possible.",
        session_resumption=types.SessionResumptionConfig(handle=handle),
        history_config=types.HistoryConfig(initial_history_in_client_content=True) if HISTORY_MODE else None,
    )

async def ask(session, text, state):
    await session.send_realtime_input(text=text)
    said = ""
    async for msg in session.receive():
        u = msg.session_resumption_update
        if u and u.resumable and u.new_handle:
            state["handle"] = u.new_handle
        sc = msg.server_content
        if sc and sc.output_transcription and sc.output_transcription.text:
            said += sc.output_transcription.text
        if sc and sc.turn_complete:
            return said.strip()

async def first_handle(session, state):
    state["handle"] = None
    async for msg in session.receive():
        u = msg.session_resumption_update
        if u and u.resumable and u.new_handle:
            state["handle"] = u.new_handle
            return

async def main():
    state = {}
    async with client.aio.live.connect(model=MODEL, config=config()) as session:
        await session.send_client_content(turns=[
            types.Content(role="user", parts=[types.Part(text="My favourite number is 7351. Please remember it.")]),
            types.Content(role="model", parts=[types.Part(text="Got it, your favourite number is 7351.")]),
        ], turn_complete=False)
        await ask(session, "Remember this code word for later: BLUEFIN. Reply with just: noted.", state)
        await asyncio.wait_for(first_handle(session, state), 10)
    async with client.aio.live.connect(model=MODEL, config=config(state["handle"])) as resumed:
        after = await ask(resumed, "What is my favourite number, and what was the code word? Answer like: NUMBER WORD", state)
    print({"after": after, "kept_seeded_history": bool(re.search(r"7\s?,?351", after))})

asyncio.run(main())

Results, each run resuming with a fresh handle:

history_config Seeded history kept after resume
initial_history_in_client_content=True (what the plugin sends for 3.1) 0 of 11 runs. Typical answers: "Unknown BLUEFIN", "I don't know your number. BLUEFIN was the word."
none 2 of 3

The code word, sent as realtime input, survives the resume. The seeded client content does not.

We hit this in production on a voice agent seeded with each student's prior conversation. After a Gemini drop and reconnect, the tutor forgot everything from before the session. The JS plugin still replays the full context on every reconnect, which is what keeps the seeded history there.

Suggestion

When the session was set up with history_config, i.e. the model has no mutable chat context, keep replaying the full chat context on resume, as before #6000. Alternatively, keep the resumption-diff path only for models whose resumed session is known to hold client content.

lastConsumedClientMessageIndex would be the natural way to tell what a handle covers, but it comes back unset on 3.1 in our runs.

Lenguaje dominante
Python
Estrellas
14.6k
Forks
3.9k
Merge medio
3 d 1 h
PR fusionados (30 d)
166

Preparar el entorno

Este proyecto no incluye contenedor de desarrollo, Dockerfile ni guía de contribución, así que la configuración corre por tu cuenta: empieza por su README y consulta nuestra guía para la primera contribución para los pasos generales.

Primeros pasos

  1. Lee el issue completo y luego la guía de contribución del proyecto.
  2. Comenta en el issue que vas a ocuparte — evita que dos personas hagan lo mismo.
  3. Haz un fork del repositorio y trabaja en una rama.
  4. Abre un pull request que haga referencia al número del issue.

Más de livekit/agents

Todos los issues de livekit/agents

Issues similares

Más issues de Python

Recibe los nuevos issues en tu correo

Un resumen breve de issues de GitHub para principiantes.