The SDK continued strengthening integrations with LLM providers and shipping sandbox infrastructure. The Python client added WebSocket transport for sandbox commands with retry logic and TTL support, while the JavaScript side gained support for Claude Agent SDK tracing with refined generator input tracking and streaming input handling. Across both languages, the team shipped deeper Anthropic integration fixes—cache token accounting for costs, system message visibility in traces, and improved tool call parsing—plus expanded project introspection APIs like list_project_issues and insights report queries. Runtime reliability improved through memory leak fixes in closure tracing, exception propagation in async experiment results, and compatibility updates for OpenTelemetry 1.40.0.
The JavaScript SDK absorbed most of the month's work, with two major themes emerging. The team shipped Anthropic Claude Agents support with proper streaming aggregation and fixed environment compatibility, released as JS versions 0.5.11 through 0.5.14. For Python, work centered on sandbox infrastructure—TTL support, TCP tunnel reconnection, and yamux multiplexing landed alongside experiment result reliability fixes that eliminated busy-waits and handled concurrent operations in pytest-xdist. Earlier in the month, trace ID propagation to child runs and insights report APIs shipped.