Summary
The Mistral Beta Conversations API (/v1/conversations) is not instrumented. Calls to client.beta.conversations.start(), client.beta.conversations.append(), and their streaming variants produce zero Braintrust tracing. This is a documented beta API on the Mistral platform for multi-turn agentic conversations where the model autonomously executes server-side tools.
The Braintrust Mistral integration instruments client.agents.complete() (agent-configured chat completions), but the Conversations API is a distinct, higher-level agentic execution surface where Mistral manages conversation state server-side and autonomously executes tools including code interpreter, web search, image generation, and document library.
This is the Mistral equivalent of the Anthropic Managed Agents API tracked in #259 — both are beta, server-side agentic execution surfaces from major providers that produce zero tracing today.
What is missing
| Mistral Resource |
Method |
Instrumented? |
client.chat |
complete(), stream() |
Yes |
client.agents |
complete(), stream() |
Yes |
client.beta.conversations |
start(), start_stream() |
No |
client.beta.conversations |
append(), append_stream() |
No |
client.beta.conversations |
restart(), restart_stream() |
No |
client.beta.conversations |
get(), list(), delete(), get_history(), get_messages() |
No (CRUD — lower priority) |
Why this matters
Conversations is an agentic execution surface, not a simple chat API. When a conversation runs:
- The model autonomously decides which tools to invoke (code interpreter, web search, image generation, document library, custom functions)
- Tool execution happens server-side — results are returned as part of the conversation outputs
- Responses include
outputs arrays with message entries, tool execution entries, function calls, and agent handoffs
- Token usage is reported via
ConversationUsageInfo
- Streaming is supported for real-time output delivery
This is functionally equivalent to the agentic frameworks already instrumented in this repo (Anthropic Managed Agents, Pydantic AI agents, Agno agents, Google ADK, OpenAI Agents SDK, Claude Agent SDK).
Minimum instrumentation
At minimum, start() and append() (and their streaming variants) should create spans capturing:
- Conversation-level span (type
task): conversation ID, model, total duration, aggregate token usage
- Output detail: message content, tool execution results, function call results
- Tool spans: child spans for server-side tool executions (code interpreter, web search, image generation, document library)
- Metrics: token usage from
ConversationUsageInfo, time-to-first-token for streaming
- Metadata: model, tools configuration, conversation ID, guardrails config
Braintrust docs status
not_found — The Mistral integration page documents chat completions, embeddings, FIM, and agents. No mention of the Conversations API.
Upstream sources
- Mistral Conversations API endpoints: https://docs.mistral.ai/api/endpoint/beta/conversations
- Mistral Agents & Conversations guide: https://docs.mistral.ai/agents/agents
- Supported tools: web search, web search premium, code interpreter, image generation, document library, custom function tools
- Python SDK:
client.beta.conversations.start(), .append(), .restart(), and streaming variants
- Beta status: accessible to all API users, documented in the official API reference under the Beta section
Local files inspected
py/src/braintrust/integrations/mistral/patchers.py — defines patchers for Chat, Embeddings, Fim, Agents; zero references to conversation, conversations, or betapy/src/braintrust/integrations/mistral/tracing.py — wrapper functions for chat, embeddings, FIM, agents only; no conversation wrapperspy/src/braintrust/integrations/mistral/integration.py — integration class registers 4 composite patchers; no ConversationsPatcherpy/src/braintrust/integrations/mistral/test_mistral.py — no conversation test casespy/noxfile.py — test_mistral session tests against LATEST and 1.12.4; no conversations coverage
Relationship to existing issues
Summary
The Mistral Beta Conversations API (
/v1/conversations) is not instrumented. Calls toclient.beta.conversations.start(),client.beta.conversations.append(), and their streaming variants produce zero Braintrust tracing. This is a documented beta API on the Mistral platform for multi-turn agentic conversations where the model autonomously executes server-side tools.The Braintrust Mistral integration instruments
client.agents.complete()(agent-configured chat completions), but the Conversations API is a distinct, higher-level agentic execution surface where Mistral manages conversation state server-side and autonomously executes tools including code interpreter, web search, image generation, and document library.This is the Mistral equivalent of the Anthropic Managed Agents API tracked in #259 — both are beta, server-side agentic execution surfaces from major providers that produce zero tracing today.
What is missing
client.chatcomplete(),stream()client.agentscomplete(),stream()client.beta.conversationsstart(),start_stream()client.beta.conversationsappend(),append_stream()client.beta.conversationsrestart(),restart_stream()client.beta.conversationsget(),list(),delete(),get_history(),get_messages()Why this matters
Conversations is an agentic execution surface, not a simple chat API. When a conversation runs:
outputsarrays with message entries, tool execution entries, function calls, and agent handoffsConversationUsageInfoThis is functionally equivalent to the agentic frameworks already instrumented in this repo (Anthropic Managed Agents, Pydantic AI agents, Agno agents, Google ADK, OpenAI Agents SDK, Claude Agent SDK).
Minimum instrumentation
At minimum,
start()andappend()(and their streaming variants) should create spans capturing:task): conversation ID, model, total duration, aggregate token usageConversationUsageInfo, time-to-first-token for streamingBraintrust docs status
not_found — The Mistral integration page documents chat completions, embeddings, FIM, and agents. No mention of the Conversations API.
Upstream sources
client.beta.conversations.start(),.append(),.restart(), and streaming variantsLocal files inspected
py/src/braintrust/integrations/mistral/patchers.py— defines patchers forChat,Embeddings,Fim,Agents; zero references toconversation,conversations, orbetapy/src/braintrust/integrations/mistral/tracing.py— wrapper functions for chat, embeddings, FIM, agents only; no conversation wrapperspy/src/braintrust/integrations/mistral/integration.py— integration class registers 4 composite patchers; no ConversationsPatcherpy/src/braintrust/integrations/mistral/test_mistral.py— no conversation test casespy/noxfile.py—test_mistralsession tests againstLATESTand1.12.4; no conversations coverageRelationship to existing issues
client.beta.agents,client.beta.sessions) not instrumented #259 tracks Anthropic Managed Agents API (analogous beta agentic execution surface from another provider)client.ocr.process()) not instrumented #222, [BOT ISSUE] Mistral: Audio APIs (client.audio.speech,client.audio.transcriptions) not instrumented #223 track other Mistral endpoint gaps (OCR, Audio) — those are standalone execution APIs, while Conversations is a higher-level agentic orchestration surfaceclient.agentsinstrumentation covers agent-configured chat completions, NOT the multi-turn conversation execution API