braintrust

Changelog

Features

• (transformers) Add HuggingFace Transformers integration (#676)

Bug Fixes

• (claude_agent_sdk) Stop assigning query usage to the final call (#670)
• (claude_agent_sdk) Prevent orphan spans for racing MCP tools (#689)
• (harbor) Default to shared Harbor project (#688)
• (mistral) Trace cached prompt tokens (#684)
• (framework) Pass eval case fields to scorers (#679)

Maintenance

• (api) Generate REST models from pinned OpenAPI spec (#690)
• (api) Add client bootstrap and service facade (#655)

Package: https://pypi.org/project/braintrust/0.34.0/

Full Changelog: py-sdk-v0.33.0...py-sdk-v0.34.0

Changelog

Features

• (harbor) Add native Harbor evaluation plugin (#631)
• (cursor_sdk) Instrument Cursor agent runs (#665)
• (ai-sdk) Add Vercel AI SDK instrumentation (#664)

Other Changes

• Replace GitPython with direct git commands (#669)

Package: https://pypi.org/project/braintrust/0.33.0/

Full Changelog: py-sdk-v0.32.0...py-sdk-v0.33.0

Changelog

Features

• (functions) Add async function invocation (#659)
• Add pipecat integration (#543)
• (api) Add policy-aware HTTP transport (#653)

Bug Fixes

• (framework) Preserve remote eval row origins (#660)
• (logger) Honor internal BTQL dataset limits (#658)
• (livekit) Disable audio attachments by default (#656)
• (otel) Filter spans before adding origin attributes (#652)

Package: https://pypi.org/project/braintrust/0.32.0/

Full Changelog: py-sdk-v0.31.1...py-sdk-v0.32.0

Changelog

Bug Fixes

• (adk) Align Google usage token totals (#649)
• (google_genai) Preserve tool use tokens as metadata (#637)
• (google_genai) Align usage metadata metrics (#630)

Other Changes

• (api) Establish transport boundary for API client (#641)

Package: https://pypi.org/project/braintrust/0.31.1/

Full Changelog: py-sdk-v0.31.0...py-sdk-v0.31.1

Changelog

Features

• (devserver) Support remote dataset version and environment selection (#622)

Bug Fixes

• (adk) Avoid logging stream cleanup as errors (#626)
• (langchain) Prevent async runs from inheriting ended spans (#624)
• Eval parameters omit null description (#613)
• (pydantic_ai) Allowlist kwargs pass-through, drop eager message reshape (#610)
• (strands) Allowlist metrics, add span_origin, fix agent metadata reads (#612)
• (temporal) Pass Exception directly to span.log(error=...) (#611)

Package: https://pypi.org/project/braintrust/0.31.0/

Full Changelog: py-sdk-v0.30.1...py-sdk-v0.31.0

Changelog

Bug Fixes

• (openai) Allowlist metadata pass-through, drop eager response serialization (#606)
• (openai) Preserve async parsing for streaming responses (#607)
• (openrouter) Allowlist metadata pass-through, drop eager serialization (#608)

Package: https://pypi.org/project/braintrust/0.30.1/

Full Changelog: py-sdk-v0.30.0...py-sdk-v0.30.1

Changelog

Features

• (perf/huggingface_hub) Optimize how dict attributes are set (#591)

Bug Fixes

• Normalize list score dicts in remote evals (#595)
• (mistral) Drop eager response serialization, use allowlist metadata pass-through (#604)
• (llamaindex) Stop emitting llm spans over delegated provider calls, add provider metadata (#602)
• (livekit_agents) Allowlist metadata fields, drop eager serialization (#603)
• (litellm) Use underlying provider for metadata, drop eager serialization (#594)
• (langchain) Add provider metadata, route tools to metadata, allowlist span fields (#593)
• (instructor) Add model/provider metadata, drop __future__ import (#592)
• (google_genai) Add provider metadata, route tools to metadata, drop eager serialization (#590)

Package: https://pypi.org/project/braintrust/0.30.0/

Full Changelog: py-sdk-v0.29.0...py-sdk-v0.30.0

Changelog

Features

• Per-integration instrumentation.name on span origin (#576)

Bug Fixes

• (crewai) Add provider metadata, route tools to metadata, drop eager serialization (#587)
• (dspy) Align span shape with spec, drop over-serialization, avoid nested llm spans (#588)
• (cohere) Drop redundant _try_to_dict pass in tool-call iteration (#586)
• (claude_agent_sdk) Add provider metadata, drop deep-copy serialization, allowlist hook fields (#585)
• (bedrock_runtime) Improve metrics, tools, and attachments on spans (#584)
• (anthropic) Drop redundant bt_safe_deep_copy passes in tracing (#583)
• (agno) Route tools to metadata, materialize media attachments, propagate exception objects (#582)
• (agentscope) Improve data attached to spans (#581)
• (adk) Fix llm span metadata (provider, tools, errors) (#580)

Package: https://pypi.org/project/braintrust/0.29.0/

Full Changelog: py-sdk-v0.28.0...py-sdk-v0.29.0

Changelog

Features

• Add span origin provenance (#570)

Bug Fixes

• (litellm) Use allowlist for span metadata (#578)
• Preserve custom PYTHON_ENV span origin (#577)

Package: https://pypi.org/project/braintrust/0.28.0/

Full Changelog: py-sdk-v0.27.0...py-sdk-v0.28.0

Changelog

Features

• (anthropic) Trace messages.parse outputs (#562)

Bug Fixes

• Function invocation 405s on EU/self-hosted (strip /v1/proxy from proxy_conn) (#541)
• Avoid duplicate spans when a full fetch follows a typed fetch (#546)
• Fix Temporal replay span parent propagation (#560)

Other Changes

• Expose realtime flag for span fetcher (#542)

Package: https://pypi.org/project/braintrust/0.27.0/

Full Changelog: py-sdk-v0.26.0...py-sdk-v0.27.0

• Improved LangChain integrations with simplified parsing for both TypeScript and Python

• Added JSON attachment SDK support

​

• Add OpenTelemetry compatibility mode for TypeScript. This allows OTel spans to work with Evals

​

• Added Google GenAI wrapper support

• Updated Mastra wrapper methods from generateVNext / streamVNext to generate / stream

• Moved langchain-js braintrust dependency to peer dependencies

• Fixed handling of attachments for Anthropic to avoid large base64 strings in UI

• Fixed preservation of result object when returning from wrappedStreamObject in AI SDK

• Fixed LanguageModelV1#supportsUrl being a function, not a property

​

• Added wrapClaudeAgentSDK for automatic tracing of Claude Agent SDK applications

• Improved Anthropic wrapper to log consistent input/output format

• Fixed AI SDK model detection in wrapGenerate callback

• Added SpanComponentsV4 support

​