Repository navigation
Retain bounded cold-worker failure diagnostics - #270
Conversation
|
|
Frozen source 8b006e1, test SHA256
The actual runtime child101 from #269 is still unclassified until this diagnostic source yields its bounded stderr. No unchanged runtime repeat, tag or upload is selected. Exact-source review and normal current-head CI remain gates; a new compiled parent/build receipt is being returned separately for a bounded diagnostic observation. |
|
Security source-only PASS for exact I independently fetched the frozen head, rehashed the sole changed file, and reviewed the complete one-file delta. The correction retains at most 65,536 bytes plus a one-byte refusal sentinel from each captured stream, includes lossy-decoded stderr only when the owned child exits unsuccessfully, keeps the 30-second owned-child deadline and executable-image checks, and leaves successful structured output unchanged. The synthetic exit-23 path runs before client construction or credential reads; the predecessor omission, oversized-reader refusal, and private-executable identity control are appropriately distinct. No library, package, workflow definition, or production runtime bytes change. This is approval only for producing a new exact build/parent identity and one bounded diagnostic observation under the already reviewed #269 runner safeguards. The prior r4 result remains NONPASS and cannot be relabeled. The new build/parent hashes, complete controller identity, retained private diagnostics, exact-ID cleanup/absence, and remaining current-head checks still gate any attempt. This is not a package, logging behavior, #980, tag, upload, or reusable arbitrary-output supervisor approval. |
|
Diagnostics source accepted on main 188db8b, tree 884ae0ef358603d9c971dee9661a0507e0a198d1, exactly equal to reviewed8b006e1. All19 PR checks and all four actual-main workflows pass: CI, Public Boundary, CodeQL, SonarQube. Remote topic branch is deleted and independently absent. The one reviewed exact-package diagnostic observation retained the actual child error HTTP409 worker_not_registered. The harness registered the setup worker but did not register each fresh child ID. This is a fixture setup failure; source fix #271 registers the child before polling, preserving the workflow/log assertions. No logger or package success is inferred. Full candidate outcome, private bounded diagnostics and exact-ID cleanup remain on #269. |
The installed logging parent hides the error when a cold child exits unsuccessfully. Read at most64KiB plus an oversize sentinel from each child stream, refuse oversized output, and include bounded stderr with a failed status. Successful structured observations, held executable identity, workflow scenario and30-second owned-child deadline are unchanged. This is acceptance-test diagnostics only; the library/package bytes do not change. Relates to #269 and durable-workflow/.github#227.
Validation: restoring only the old child-runner body makes the synthetic pre-client diagnostic control fail. Corrected source passes all ordinary controls (3 PASS/2 SDK flows ignored), including oversized-stderr refusal and the existing private-copy executable-identity control, against the same checksum-bound4.17 candidate with Cargo1.86/offline build. The first all-control setup refusal on noexec temp is preserved; the successful fixture uses a dedicated private64MiB exec scratch mount. Exact validation, source review. Exact-head CI/main checks and the separately reviewed diagnostic runtime/cleanup remain required; no product, package or publication result is inferred.