Mistral tool_calls not parsed - missing 'type' field in API response

[xu-chris]

Bug Description

When using Mistral's API through the OpenAI-compatible format, tool calls are not being parsed because decode_openai_tool_call/1 requires a "type" => "function" field that Mistral does not include in its responses.

Environment

• ReqLLM version: 0.4.0
• Elixir: 1.19.1
• Provider: Mistral (using use ReqLLM.Provider)

Steps to Reproduce

1. Configure a Mistral provider using the OpenAI-compatible format
2. Make a request with tools defined
3. LLM returns tool calls
4. Tool calls are not parsed - response.message.tool_calls is nil despite finish_reason: :tool_calls

Root Cause

In lib/req_llm/provider/defaults.ex, the decode_openai_tool_call/1 function at ~line 877 requires "type" => "function":

defp decode_openai_tool_call(%{
       "id" => id,
       "type" => "function",  # <-- REQUIRED
       "function" => %{"name" => name, "arguments" => args_json}
     }) do
  case Jason.decode(args_json || "{}") do
    {:ok, args} -> ReqLLM.StreamChunk.tool_call(name, args, %{id: id})
    {:error, _} -> nil
  end
end

defp decode_openai_tool_call(_), do: nil  # Falls through here for Mistral

Mistral API returns:

%{
  "function" => %{"arguments" => "{\"city\": \"Paris\"}", "name" => "get_weather"},
  "id" => "lVauww8VE",
  "index" => 0
  # NO "type" field!
}

OpenAI API returns:

%{
  "function" => %{"arguments" => "{\"city\":\"Paris\"}", "name" => "get_weather"},
  "id" => "call_...",
  "type" => "function"  # HAS "type" field
}

Since Mistral omits the "type" field, the pattern match fails and falls through to the catch-all clause returning nil.

Proposed Fix

Add a fallback pattern that doesn't require the "type" field. This is backwards-compatible since it will only match when the stricter pattern fails:

# Existing pattern for OpenAI and others that include "type"
defp decode_openai_tool_call(%{
       "id" => id,
       "type" => "function",
       "function" => %{"name" => name, "arguments" => args_json}
     }) do
  case Jason.decode(args_json || "{}") do
    {:ok, args} -> ReqLLM.StreamChunk.tool_call(name, args, %{id: id})
    {:error, _} -> nil
  end
end

# Fallback for Mistral and others that omit "type" field
defp decode_openai_tool_call(%{
       "id" => id,
       "function" => %{"name" => name, "arguments" => args_json}
     }) do
  case Jason.decode(args_json || "{}") do
    {:ok, args} -> ReqLLM.StreamChunk.tool_call(name, args, %{id: id})
    {:error, _} -> nil
  end
end

defp decode_openai_tool_call(_), do: nil

Why this belongs in ReqLLM defaults

1. This is the generic OpenAI-compatible decoder used by multiple providers
2. Other providers claiming OpenAI compatibility might have the same issue
3. The fix is backwards-compatible - existing OpenAI responses still work
4. Putting this in individual provider overrides would require duplicating the entire decode pipeline

I'm happy to submit a PR with this fix if that would be helpful.