LangSmith

Adds AgentAddress, ExperimentAddress, and EvaluatorAddress types, exposes client.sessions in the JS SDK to resolve an address, and reads LANGSMITH_REVISION_ID for run revision_id. License key and API key salt assignments are now redacted in both the Python and JS SDKs.

Engine usage, spend limits, and pricing now display in LangChain Standard Units at $1 each, and OAuth applications can use a single access token for the LangSmith API and Remote MCP server. Other changes include fetch_runs deprecated in favor of query_runs, cost resolution for OpenAI and Vertex AI-hosted open models, and Claude Sonnet 5.5 and Gemini 3.7/3.8 Flash in the Playground.

Fixed the full OTLP traces endpoint resolving incorrectly across the Python and JavaScript SDKs. Also refactored address handling to use strings.

Adds an address() handle and typed address object for agent addressing, and ends traced runs when a returned ReadableStream errors. Also retries transient WebSocket failures before sandbox command start and exports OTel prompt/completion as JSON strings.

Python gains an ls.address handle for agent addressing, and only the primary write replica now keeps the original run IDs. Also fixed missing OTel exports for tool name, call_id and definitions, preserved responses on wrapped HTTP errors, and resolved relative example attachment URLs.

Restored missing OpenTelemetry exports for tool_call_id and tool_definitions, and fixed tracing of ClaudeSDKClient.receive_messages(). Sandbox now verifies service URL user tokens and proxy callbacks.

Sandboxes are now granted delegated LangSmith access on create, with LangSmith-login service URLs and JS service URL support. Also captured Claude SDK run stop_reason plus tool and response IDs, added agent-scoped addressing for runs and feedback parts in Python, and patched soupsieve ReDoS vulnerabilities.

Evaluators using organization-scoped model configurations no longer attempt per-model OAuth token exchange, and Amazon Bedrock models are no longer rejected; annotation queues can now add selected thread items to a dataset. LangSmith Go and Java SDKs can save Insights report configurations, and the thread traces API accepts trace_filter and tree_filter parameters.

evaluate() and aevaluate() now accept a disable_evaluator_tracing toggle, and sandbox helpers support AWS IAM roles in Python and JavaScript. Sandbox also gains run_config, stdin EOF handling, native file search and range operations, and an SSE exec-api command streaming option behind a feature flag.

Fixed the LiveKit integration dropping system instructions when history is supplied. Release also includes a JS SDK 0.10.4 release and docs for OpenWiki with a weekly refresh workflow.

Runs now report the configured tracing sample rate in both Python and JS. Also fixed sandbox run() waiting about a second on the server's TCP close after exit, and patched vulnerable Vitest, Hono, js-yaml, and sharp test dependencies.

Starting September 14, 2026, traces ingested into LangSmith SaaS with extended retention are kept at most 180 days, existing settings above that are applied as 180 days for new traces, and new settings above 180 days are rejected, while self-hosted and BYOC deployments are unchanged. Organization Operators can no longer grant the Organization Admin role to a service key, and multipart feedback ingest now records a rejection error on the run instead of returning a 400 that dropped the run and all accompanying feedback.

Adds /v2/threads/stats for thread/trace metrics, preserves OAuth/session auth across Gateway routing, and requires download permission for dataset exports alongside stricter validation for malformed IDs, shares, and null artifacts. Trajectory batch reads move to a dedicated gRPC service.

Python client now supports snapshot tags in the sandbox, and rate-limited sandbox WebSocket connections are retried in both Python and JS. Plus dependency and advisory patches.