strip_thinking_tokens crashes on list-shaped streaming delta content (AttributeError: 'list' object has no attribute 'find')

[KSerProject]

Versions: livekit-agents 1.6.3 (the code is unchanged on current main), livekit-plugins-openai, Python 3.12.

Setup: openai.LLM plugin pointed at an OpenAI-compatible provider (Mistral, base_url=https://api.mistral.ai/v1), wrapped in llm.FallbackAdapter.

What happens: the OpenAI-compatible streaming schema allows delta.content to be either a str or a list of typed content parts ([{"type": "text", "text": "…"}]). Mistral has started emitting the list form on some streamed chunks. llm/utils.py::strip_thinking_tokens assumes str | None and calls content.find(...), raising AttributeError: 'list' object has no attribute 'find'. The plugin wraps this as a retryable APIConnectionError; under a FallbackAdapter the failure lands mid-stream ("failed after sending chunk, skip retrying"), so the adapter silently moves the whole session to the fallback leg — a hard-to-diagnose degradation for voice agents.

Repro (no network):

import asyncio
from livekit.agents.llm.utils import strip_thinking_tokens
strip_thinking_tokens([{"type": "text", "text": "Hallo"}], asyncio.Event())
# AttributeError: 'list' object has no attribute 'find'

Suggested fix: flatten list-shaped content to its concatenated text parts before the tag scan — see the linked PR.

[chenghao-mou]

the OpenAI-compatible streaming schema allows delta.content to be either a str or a list of typed content parts ([{"type": "text", "text": "…"}]).

Thanks for raising the issue. Do you know where I can find the source for this claim? I am asking for this because when I look up the doc, it only says "optional string":