npx @buildinternet/releases get inngest-serverConnect now accepts execution requests and replies above 4 MiB over internal gRPC and no longer retries transport message-size violations, and duplicate Connect execution after executor failure is fixed by resuming the original request lease or buffered response. Also fixes renewed worker capacity reservations that stayed at capacity after abandonment, preserves step invokes after the execution cap, and upgrades Next.js and Sharp in the UI for GHSA-2464-fhr4-w9xw.
Functions can now be paused through the REST API, CLI, and MCP, and the includeOutput parameter is soft-deprecated in favor of include=output on REST v2 run-list endpoints. Fixed runs getting stuck as Running when step discovery enqueues fail transiently, preserved singleton runs on scheduling retries, and stopped Connect cleanup from dropping refreshed connections.
AI Metadata field names have been renamed to align with OpenTelemetry GenAI standards: .model is now .request_model and .system is now .provider. Insights queries using the old field names must be updated. AI Metadata extraction now differentiates between request and response models when using ai.step.wrap or ai.step.infer.
Added experiment and sessions endpoints to the REST V2 API and CLI, including GET endpoints for experiments, sessions, and session runs. Also fixed a V1 execution engine regression that was coalescing parallel steps incorrectly, visualized numeric metrics using box plots, and showed key queues settings in the function view.
This prerelease was built from v1.33.1-test.3.
ef7617395b6c10ee87ddb558b0041548d5549e9dThis prerelease was built from v1.33.1-test.2.
57d65b228ed991a33f0eee1577acb06731e00584
