feat(mistral): migrate LLM to Conversations API with provider tools support

[jeanprbt]

Summary

• Migrates the Mistral AI LLM plugin from the Chat Completions API to the Conversations API,
  enabling support for Mistral's built-in provider tools: Web Search, Code Interpreter, and Document Library
• Adds MistralTool base class with WebSearch, CodeInterpreter, and DocumentLibrary provider tool definitions
• Extends LLM constructor with full completion parameter support (top_p, presence_penalty, frequency_penalty, random_seed, tool_choice)
• Switches default TTS response format from pcm to mp3

Details

Conversations API migration

The Conversations API is stateful — the server retains conversation history across calls. This enables:

• First call: start_stream_async() sends the full context, model, instructions, and tools
• Subsequent calls: append_stream_async() sends only new client-originated entries (function.result, message.input), avoiding redundant re-transmission of the full history

Context diffing uses ChatContext.is_equivalent() (following the OpenAI Responses API pattern) to detect whether the new context extends the previous one. If it diverges, the conversation is reset with a fresh start_stream_async().

A pending tool call verification safeguard ensures all function calls from the previous response have results before attempting to append.

Streaming event handling

The Conversations API uses typed SSE events instead of choice deltas.

Event	Handling
ResponseStartedEvent	Captures conversation_id for stateful tracking
MessageOutputEvent	Emits ChatChunk with text content
FunctionCallEvent	Accumulated in _PendingFunctionCall buffer, flushed as complete FunctionToolCall on next non-function event (since arguments are sent delta-wise incrementally)
ToolExecution{Started,Delta,Done}Event	Provider tool execution (server-side) — arguments accumulated silently, logged once on completion
ResponseDoneEvent	Extracts token usage
ResponseErrorEvent	Raises APIStatusError

Provider format converter

Replaces the old to_chat_ctx() (which delegated to OpenAI format) with to_conversations_ctx() that produces Mistral's flat entry format:

• System/developer messages → extracted as instructions string (returned via MistralFormatData)
• User messages → {"type": "message.input", ...}
• Assistant messages → {"type": "message.output", ...}
• Function calls → {"type": "function.call", ...}
• Function results → {"type": "function.result", ...}

LLM completion parameters

The LLM constructor and update_options() now accept the full set of CompletionArgs parameters supported by the Conversations API: temperature, top_p, max_completion_tokens, presence_penalty, frequency_penalty, random_seed, and tool_choice.

TTS default response format

Switches the default TTS response format from pcm back to mp3, since it is the most reliable default for most setups.

Usage

from livekit.plugins import mistralai

llm = mistralai.LLM(model="mistral-medium-latest")

# With provider tools
agent = Agent(
    llm=llm,
    tools=[
        mistralai.tools.WebSearch(),
        mistralai.tools.CodeInterpreter(),
        mistralai.tools.DocumentLibrary(library_ids=["<your-library-id>"]),
    ],
)

[devin-ai-integration]

Devin Review found 3 new potential issues.

View 13 additional findings in Devin Review.