A GitHub commit d0c31e2 (2026-10-18) landed in weisgnosis/weisgnosis-agents, the project that builds agents and teams that use tools with Scala 3 and ZIO 2. This commit verified, against two real models, request cancellation for a client that uses the Responses API stream, and the flow of using the same client again. The changes consist of only one verification runner and three documents, amounting to 309 added lines and 12 deleted lines.
The verification proceeds by passing real SSE (server-sent events) frames through as they are. The borrowed driver passes actual frames through up to the first reasoning delta, then halts the downstream without closing the upstream reader. The parsed frame event produces a run cancellation signal, and the pending finalizer confirms that the consumer has not yet returned. Continuing the interrupted reasoning context fails before the HTTP driver is invoked again, and a new agent created with the same client completes as READY. The request counts were recorded as one before the new context and two in total. Each run has a 180-second limit, the overall profile has a 420-second watchdog limit, and the abort request is set to 1,024 output tokens while the new-context request is set to 4,096.
The documents record three items — consecutive local tool execution, JSON checkpoint restoration, and controlled HTTP cancellation — all as Passed for the two models qwen3.8-27b and qwen3.8-27b-uncensored-mlx. The four tool and checkpoint results report Completed status, one live_sum(2, 3) call and result pair, a final response of 5, native replay metadata, and positive output usage, and both processes ended with exit code 0. The file header of the added runner states the Apache License 2.0.
However, this check is an ownership inspection with instrumentation in place, and it does not measure whether a controlled halt corresponds to the transmission timing as it is, a physical socket close, or a server-side inference cancellation. Native replay metadata is not evidence that encrypted reasoning was provided either, and native structured output, hosted tools, multimodal data, and other models were not verified. No comparison with Python or performance ranking is derived either. The library source was not changed, and the two runners were compiled with a setting that treats warnings as fatal (fatal warnings).