🧩 fix: Keep Text and Tool Calls From Mistral Typed Delta Chunks

[jordanmarques]

Summary

When a Mistral model reasons (Mistral Large 4, or GLM 5.3 served by Mistral), its Chat Completions
stream sends delta.content as a list of typed chunks instead of a string:

{"delta": {"content": [
  {"type": "thinking", "thinking": [{"type": "text", "text": " done."}], "closed": true},
  {"type": "text", "text": "A freelancer"}
]}}

The chunk that closes the reasoning also carries the first words of the answer, and sometimes the
first tool-call delta, the only one with the call's id and name. _streamResponseChunks skips
every chunk whose content is not a string ([WARNING]: Received non-string content from OpenAI),
so:

• answers start mid-word ("ancer is a self-employed…");
• tool calls lose their name and id, so no tool runs and the agent ends with an empty response.

This format is Mistral's documented streaming behavior for reasoning models; only
reasoning_effort: "none" returns plain strings, and some models (GLM 5.3) refuse that value.

Change

LibreChatOpenAICompletions._convertCompletionsDeltaToBaseMessageChunk flattens a list
delta.content before the LangChain conversion:

• text blocks become the string content;
• thinking blocks become delta.reasoning, which attachLibreChatDeltaFields already maps to
  additional_kwargs.reasoning_content.

String content and other providers are unchanged.

Testing

• New src/llm/openai/mistralContentChunks.test.ts: mixed reasoning + text chunk, reasoning +
  tool-call chunk, reasoning-only chunk.
• src/llm/openai suite: 211 tests pass; tsc --noEmit and ESLint clean.
• End to end against a Mistral-backed OpenAI-compatible gateway, 10 tool-calling runs per model:

	Tool calls lost (GLM 5.3)	Answer start
Before	2 / 10	truncated
After	0 / 10	complete, GLM 5.3 and Mistral Large 4

🤖 Generated with Claude Code