mirror of
https://github.com/QwenLM/qwen-code.git
synced 2026-11-07 06:19:43 -10:00
Shaojin Wen
d2e12e404f
* test(managed-agent): crash the approval result-commit barrier A5's last packaged barrier: after the approved tool result commits, hold the post-tool model round and crash the Harness. The next generation must reissue the saved request (deep-equal), retract the published prefix, and finish with one execution, one file effect and one terminal. Part of #14481. * test(managed-agent): gate the result crash on the published prefix Review findings: wait for the public feed to carry the held prefix before SIGKILL and record it in the crash evidence, so the transcript assertion cannot pass on a no-op retraction. Also decouple the source pins from prettier whitespace and the comparison operator. * test(managed-agent): pin the after-result model count Review follow-up: keep the two-request expectation tied to the after-result arm via expectedModelsAfterResults, without reintroducing the indentation coupling that the previous split removed. * test(managed-agent): strengthen the result-commit assertions Review findings: name the two pre-dispatch barriers and pin after-result's exclusion from the held start, assert the stream.reconciled ordering beside the reload text, retain both request bodies in model-request-parity.json, and pin the retraction throw. * ci: record the sdk-java workflow growth The five approval fault arms add steps and comments to sdk-java.yml: 104042 bytes, 4100 over the previous 99942 baseline. Update the ratchet in the same PR, as check-workflow-size.sh prescribes for real growth.