Python SDK v0.21.0

Changelog

Features

• (livekit) Add agents integration (#407)

Package: https://pypi.org/project/braintrust/0.21.0/

Full Changelog: py-sdk-v0.20.0...py-sdk-v0.21.0

Python SDK v0.20.0

Changelog

Features

• (autogen) Trace agents inside teams (#419)
• More explicit default git settings (#408)
• (litellm) Trace text completion calls (#406)

Bug Fixes

• (strands) Close spans with no-op OTEL provider (#420)
• (cohere) Support chat stream context managers (#418)
• (openai) Filter omit sentinel from metadata (#426)
• (logger) Do not retry on 413s (#427)
• Update GitPython constraint for security fix (#423)
• Support autoevals inline score typing (#403)
• (cohere) Preserve v2 stream citations (#391)
• (mistral) Add conversation tool spans (#392)
• (langchain) Normalize anthropic cache token metrics (#411)
• (parameters) Improve dict types (#399)

Package: https://pypi.org/project/braintrust/0.20.0/

Full Changelog: py-sdk-v0.19.0...py-sdk-v0.20.0

Python SDK v0.19.0

Changelog

Features

• (openai) Trace streaming audio transcriptions (#375)
• (anthropic) Add prompt cache ttl metrics (#344)
• (temporal) Add temporal integrations auto instrumentation (#367)

Bug Fixes

• (mistral) Add tool spans for completions (#381)
• (cohere) Add tool spans for chat tool calls (#380)
• (devserver) Honor evaluator.project_id when request omits it (#372)

Package: https://pypi.org/project/braintrust/0.19.0/

Full Changelog: py-sdk-v0.18.0...py-sdk-v0.19.0

Python SDK v0.18.0

Changelog

Features

• (openai_agents) Pull cached tokens through into metrics (#364)
• (perf/pydantic_ai) Simplify trace shaping (#357)
• (perf/adk) Reduce tracing serialization overhead (#358)
• (llamaindex) Add native LlamaIndex SDK integration (#355)

Bug Fixes

• Make OpenAI wrapping compatible with ddtrace (#370)
• Handle PEP 649 NameError in inspect.signature on Python 3.14 (#361)
• Use Mapping instead of dict for EvalResult.scores (#360)

Maintenance

• Modernize typing for Python 3.10+ minimum (#356)

Package: https://pypi.org/project/braintrust/0.18.0/

Full Changelog: py-sdk-v0.17.0...py-sdk-v0.18.0

Python SDK v0.17.0

Changelog

Features

• (strands) Add agents integration (#346)
• (crewai) Add new tracing integration (#341)
• (autogen) Add automatic instrumentation (#345)
• (cohere) Trace audio transcription calls (#333)
• (litellm) Instrument rerank and arerank (#323)
• (mistral) Instrument beta conversations api (#324)

Bug Fixes

• (pydantic_ai) Strip unnecessary wrapper tokens (#321)
• (claude_agent_sdk) Preserve tool span output across back-to-back tool_use messages (#322)

Other Changes

• Rm eval status page (#347)

Package: https://pypi.org/project/braintrust/0.17.0/

Full Changelog: py-sdk-v0.16.0...py-sdk-v0.17.0

Python SDK v0.16.0

Changelog

Features

• (cohere) Add integrations api support (#316)
• (openai_agents) Support v0.14.0 TaskSpanData and TurnSpanData (#307)
• (framework) Add classifier support to Eval (#282)

Bug Fixes

• (openai) Re-export NamedWrapper from braintrust.oai for back-compat (#319)
• (types) Re-export public top-level symbols with PEP 484 aliasing (#315)
• (framework) Make classifier labels optional (#311)
• (pydantic) Retype wrapper spans as task to stop cost double-counting (#312)
• (google_genai) Serialize Content objects with nested binary Parts as attachments (#276)
• Use Mapping for dict params, Any for ValidatedParameters (#289)
• Use Sequence instead of list for input parameters (#287)
• Use Mapping for Metadata and ParametersSchema type aliases (#285)
• Use Sequence instead of Iterable for report_failures parameter (#284)
• Use Mapping instead of dict for EvalParameters type alias (#281)

Maintenance

• (nox) Centralize dependency pins in pyproject.toml (#300)
• Replace Optional and Dict with modern syntax in logger.py (#291)

Other Changes

• Widen BraintrustStreamChunk and annotate BraintrustStream.stream (#306)
• Remove @dataclasses.dataclass from TaskProgressEvent (#305)

Package: https://pypi.org/project/braintrust/0.16.0/

Full Changelog: py-sdk-v0.15.0...py-sdk-v0.16.0

Python SDK v0.15.0

Changelog

Features

• Add per-input trial_count support to Eval() (#274)
• (mistral) Trace audio api calls (#275)
• (anthropic) Trace beta managed agents apis (#270)
• (mistral) Trace OCR API calls (#268)
• (openai|litellm) Attach speech outputs (#267)
• (litellm) Trace transcription calls (#264)
• (litellm) Trace image generation calls (#261)
• (openai) Trace responses api tool spans (#253)
• (anthropic) Trace server tool content blocks as tool spans (#255)
• (openai) Preserve streaming refusal output (#258)

Bug Fixes

• CachedSpanFetcher no longer caches empty results permanently (#79)
• (openai) Trace chat streaming audio attachments (#271)
• (openai) Log audio file attachments (#265)
• (adk) Avoid duplicate sync runner invocation spans (#252)
• (integrations) Unify multimodal attachment materialization (#247)

Package: https://pypi.org/project/braintrust/0.15.0/

Full Changelog: py-sdk-v0.14.0...py-sdk-v0.15.0

Python SDK v0.14.0

Changelog

Features

• (openai) Trace images api calls (#245)
• (openai) Instrument Audio API (speech, transcriptions, translations) (#232)

Bug Fixes

• (framework) Split Output TypeVar into Output and Expected (#243)
• (framework) Preserve defaults for unmatched callback params (#241)
• (cli) Reject whitespace in python bundle filenames (#237)
• Resolve reportUnknownMemberType under pyright strict mode (#242)
• (anthropic) Capture structured output metadata (#238)
• (claude_agent_sdk) Log ResultMessage metrices (#225)
• (anthropic) Capture time_to_first_token when using .text_stream (#234)

Maintenance

• Generated SDK types (#248)
• Generated SDK types (#246)

Other Changes

• Include sandbox runtime in list request (#227)

Package: https://pypi.org/project/braintrust/0.14.0/

Full Changelog: py-sdk-v0.13.0...py-sdk-v0.14.0

Python SDK v0.13.0

Changelog

Features

• (google_genai) Trace interactions api methods (#210)
• (openai) Migrate openai wrapper and add claude agent sdk integration (#206)
• (integrations) Add mistral sdk integration (#199)

Bug Fixes

• (openai) Preserve logprobs for streaming chat spans (#215)
• (pydantic_ai) Support moved tool manager module (#224)
• (claude_agent_sdk) Use result message output (#218)
• (anthropic) Convert multimodal base64 inputs to attachments (#212)
• Allow eval filtering with EvalCase objects (#211)
• (claude_agent_sdk) Trace top-level query calls (#216)
• (claude_agent_sdk) Serialize thinking blocks with type (#213)
• (devserver) Tighten cors origin validation (#203)
• (dspy) Add adapter callback tracing (#190)
• (anthropic) Move ephemeral cache creation fields to metadata (#205)
• (anthropic) Capture server-side tool usage metrics (#192)

Maintenance

• Add minimum supported versions for genai and dspy (#220)
• Generated SDK types (#221)

Package: https://pypi.org/project/braintrust/0.13.0/

Full Changelog: py-sdk-v0.12.1...py-sdk-v0.13.0

Python SDK v0.12.1

Changelog

Bug Fixes

• (anthropic) Capture nested usage metrics and metadata (#189)
• Encode request body as UTF-8 bytes to prevent Latin-1 corruption (#194)
• Fix python sandbox parameters (#191)

Package: https://pypi.org/project/braintrust/0.12.1/

Full Changelog: py-sdk-v0.12.0...py-sdk-v0.12.1