TypeScript SDK

​

OpenAI Agents trace data

• Updated the OpenAI Agents adapter for current ResponseSpanData, which exposes model input and response data through _input and _response. Legacy input and response fields remain supported.
• Fixed empty OpenAI Agents spans by restoring normalized input, output, model, token usage, model parameters, and reasoning data.
• Added task and turn span handling, including usage and turn metadata.
• Nested LLM input and output now propagate through intermediate spans to the nearest parent agent, keeping agent and root trace data populated.

​

Decorator compatibility

• Added support for current Stage 3 TypeScript method decorators while retaining compatibility with legacy experimentalDecorators emit.
• Both bare and configured @trace / @span usage work with synchronous and asynchronous class methods.

​

Examples and release compatibility

• Updated the DeepAgents example and install target for the current API and required LangChain runtime peers.
• Refreshed provider examples with current models, including Claude Sonnet 5, Gemini 3.8 Flash, Mistral Small, and gpt-5.6-terra.
• Added AWS credential guidance and updated the Bedrock example to the Claude Sonnet 5 global inference profile.
• Kept provider and framework SDKs on demand so the base development install remains compatible with the Node 18 CI matrix.
• Pinned npm 11.5.1 in the release workflow for Node 20 and npm trusted-publishing compatibility.

​

Release metadata update

• Advanced the published SDK version to v0.4.1.
• No runtime or public API changes from v0.4.0.

​

Initial public release of the TypeScript SDK

• Tracing core: init, Client, trace and span contexts, withTrace / withSpan, method decorators, buffered transport with retry, shutdown flushing, and programmatic scoring.
• Context propagation: concurrency-safe session, user, trace, and span state built on AsyncLocalStorage.
• Provider wrappers: OpenAI, Anthropic, Google Gemini, Mistral, and AWS Bedrock, including streaming, time-to-first-token, normalized input/output, and token usage.
• Agent integrations: LangChain, LangGraph, DeepAgents, Claude Agent SDK, OpenAI Agents SDK, and Vercel AI SDK.
• Universal trace shape: normalized CHAIN → AGENT → LLM / TOOL trees compatible with traces produced by the Python SDK.
• Package support: Node.js 18 or newer, dual ESM and CommonJS builds, TypeScript declarations, and provider-specific package exports.