Summary
The Mistral integration captures tool_calls in the LLM span's output dictionary but does not create child SpanTypeAttribute.TOOL spans for individual tool calls. This is a tracing depth gap compared to the OpenAI, Anthropic, and Google GenAI integrations, which all decompose tool-use responses into dedicated child tool spans.
Mistral's Chat API and Agents API both support function calling / tool use. When a model response includes tool calls, users currently see them only as entries in the LLM span's output message — they cannot drill into individual tool invocations as separate spans in the Braintrust UI.
What is missing
The Mistral tracing module (py/src/braintrust/integrations/mistral/tracing.py) accumulates tool calls into the output message but never creates child spans:
Non-streaming path: Tool calls from response.choices[0].message.tool_calls are serialized into the output dict.
Streaming path (line ~599): _merge_tool_calls() accumulates streaming tool call deltas into the message dict — stored flat in the LLM span.
Span creation: Only SpanTypeAttribute.LLM spans are ever created. No SpanTypeAttribute.TOOL child spans exist anywhere in the file.
Both _CHAT_METADATA_KEYS and _AGENTS_METADATA_KEYS include parallel_tool_calls as metadata, confirming tool use is a recognized feature — but the span-level decomposition is missing.
Comparison with other integrations in this repo
| Integration |
Tool call handling |
Child TOOL spans? |
| OpenAI (Chat Completions) |
_log_response_tool_spans() in tracing.py |
Yes |
| OpenAI (Responses API) |
_log_response_tool_spans() in tracing.py |
Yes |
| Anthropic |
_log_server_tool_spans() in tracing.py |
Yes |
| Google GenAI |
_finalize_interaction_tool_spans() in tracing.py |
Yes |
| Cohere |
tool_calls stored in LLM span output dict |
No (tracked separately) |
| Mistral |
tool_calls stored in LLM span output dict |
No |
What child TOOL spans should capture
For each tool call in the response:
- Span name: Tool function name (e.g.
tool: web_search)
- Span type:
SpanTypeAttribute.TOOL
- Input: Tool call arguments / parameters
- Output: (empty for client-side tool calls)
- Metadata:
tool_call_id, tool type, tool index
This applies to both client.chat.complete() / client.chat.stream() and client.agents.complete() / client.agents.stream() tool-use responses.
Braintrust docs status
supported (partial) — The Mistral integration page documents chat completions and agents instrumentation. It does not mention tool span decomposition.
Upstream sources
- Mistral Function Calling docs: https://docs.mistral.ai/capabilities/function_calling/
- Mistral supports parallel tool calling via
parallel_tool_calls parameter
- Tool calls in responses follow the OpenAI-compatible format:
id, type: "function", function.name, function.arguments
- Agents API also supports tool use with the same response format
Local files inspected
py/src/braintrust/integrations/mistral/tracing.py — no SpanTypeAttribute.TOOL usage; tool calls stored flat in output message via _merge_tool_calls() (line 599)py/src/braintrust/integrations/mistral/patchers.py — Chat, Embeddings, FIM, Agents, Transcriptions, Speech, OCR, Conversations patchers defined; no tool-span-related logicpy/src/braintrust/integrations/openai/tracing.py — _log_response_tool_spans() creates child TOOL spans (for comparison)py/src/braintrust/integrations/anthropic/tracing.py — _log_server_tool_spans() creates child TOOL spans (for comparison)py/src/braintrust/integrations/google_genai/tracing.py — _finalize_interaction_tool_spans() creates child TOOL spans (for comparison)
Summary
The Mistral integration captures
tool_callsin the LLM span's output dictionary but does not create childSpanTypeAttribute.TOOLspans for individual tool calls. This is a tracing depth gap compared to the OpenAI, Anthropic, and Google GenAI integrations, which all decompose tool-use responses into dedicated child tool spans.Mistral's Chat API and Agents API both support function calling / tool use. When a model response includes tool calls, users currently see them only as entries in the LLM span's output message — they cannot drill into individual tool invocations as separate spans in the Braintrust UI.
What is missing
The Mistral tracing module (
py/src/braintrust/integrations/mistral/tracing.py) accumulates tool calls into the output message but never creates child spans:Non-streaming path: Tool calls from
response.choices[0].message.tool_callsare serialized into the output dict.Streaming path (line ~599):
_merge_tool_calls()accumulates streaming tool call deltas into the message dict — stored flat in the LLM span.Span creation: Only
SpanTypeAttribute.LLMspans are ever created. NoSpanTypeAttribute.TOOLchild spans exist anywhere in the file.Both
_CHAT_METADATA_KEYSand_AGENTS_METADATA_KEYSincludeparallel_tool_callsas metadata, confirming tool use is a recognized feature — but the span-level decomposition is missing.Comparison with other integrations in this repo
_log_response_tool_spans()intracing.py_log_response_tool_spans()intracing.py_log_server_tool_spans()intracing.py_finalize_interaction_tool_spans()intracing.pytool_callsstored in LLM span output dicttool_callsstored in LLM span output dictWhat child TOOL spans should capture
For each tool call in the response:
tool: web_search)SpanTypeAttribute.TOOLtool_call_id, tool type, tool indexThis applies to both
client.chat.complete()/client.chat.stream()andclient.agents.complete()/client.agents.stream()tool-use responses.Braintrust docs status
supported (partial) — The Mistral integration page documents chat completions and agents instrumentation. It does not mention tool span decomposition.
Upstream sources
parallel_tool_callsparameterid,type: "function",function.name,function.argumentsLocal files inspected
py/src/braintrust/integrations/mistral/tracing.py— noSpanTypeAttribute.TOOLusage; tool calls stored flat in output message via_merge_tool_calls()(line 599)py/src/braintrust/integrations/mistral/patchers.py— Chat, Embeddings, FIM, Agents, Transcriptions, Speech, OCR, Conversations patchers defined; no tool-span-related logicpy/src/braintrust/integrations/openai/tracing.py—_log_response_tool_spans()creates child TOOL spans (for comparison)py/src/braintrust/integrations/anthropic/tracing.py—_log_server_tool_spans()creates child TOOL spans (for comparison)py/src/braintrust/integrations/google_genai/tracing.py—_finalize_interaction_tool_spans()creates child TOOL spans (for comparison)