You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Preserve exit status 130 after a headless submit stops #1331
On Linux with Bun 1.4.0, the headless submit subprocess intermittently exits with status 1 after the first SIGINT. The test holds entry cleanup open, verifies the socket and subprocess stay open, and releases cleanup. It expects the interrupted submission to join and the subprocess to exit with status 130.
Evidence
Command bun run test:bun 3/10 on 7429ad262ca8dba7c3c0833a69a64829acdad55e.
packages/cli/tests/repl-headless-submit-process.test.ts:213, G: retains the socket through exact Stop and held cleanup, then exits with status 130. Expected { code: 130, signal: null }. Observed { code: 1, signal: null }. The sibling blocked-stdin case passed. First failing job
The failed-job retry passed the same test and shard on the identical commit and runtime. Passing retry
Capture the child’s output and finished receipt when the exit-status assertion fails. Identify the cause of the exit and keep the assertions that the socket stays open through cleanup and that the joined interruption exits with 130.
Recurrence during final corpus measurement at exact main 6207d248231be67c9093689abade6609769d05d4: Bun measurement job 134663622215, Bun 1.4.0 on Linux x64. repl-headless-submit-process.test.ts:217, the same G held-cleanup case, expected {code:130,signal:null} and received {code:1,signal:null}. The blocked-stdin sibling passed. The measurement stopped after 107 of 467 applicable files and uploaded an explicitly failed runtime record.
The complete first log and failed record were preserved before any retry. The log reports that a failure journal was saved under workspace .xmd/test-artifacts/.../root-1, but the measurement workflow uploaded only test-weights-bun.json; the saved journal is absent from its artifact. The assertion also omits the child's stdout/stderr/finished receipt, so this occurrence still does not identify which exit-1 path was selected.
The original #1331 diagnostic hold and its instrumented x64 work are retained under /private/tmp/xmd-1331-delivery.AxA8DB/. The retained Planner is being reconciled for a diagnostic-first continuation using current source and this new evidence. No production exit-status change, weights retry or assembly/splicing has been performed. Other runtime measurements continue independently.
The diagnostic slice is PR #1414, exact signed abc64a127b51bbd286698a550fa2723c8b836f2b, with the original Planner and independent Codex Astra both PASS. It preserves joined child output/receipts on an exit mismatch and uploads measurement failure journals separately. The deliberate mismatch control confirms the diagnostic includes the actual status and matching admitted/finished identities; Deno and Node pass, as does the single Linux x64 Bun 1.4.0 probe under emulation.
No exit-1 owner was identified, no production exit logic changed, and no second focused configuration was justified. This issue remains open. The one fresh full measurement required by #1408 is 44279296345, attempt 1 at that reviewed source. Any recurrence should supply the first child diagnostics and saved journals back to this same repair; it will not trigger a blind retry.
Activity
taras commentedon Nov 18, 2026
Recurrence during final corpus measurement at exact main
6207d248231be67c9093689abade6609769d05d4: Bun measurement job 134663622215, Bun 1.4.0 on Linux x64.repl-headless-submit-process.test.ts:217, the same G held-cleanup case, expected{code:130,signal:null}and received{code:1,signal:null}. The blocked-stdin sibling passed. The measurement stopped after 107 of 467 applicable files and uploaded an explicitly failed runtime record.The complete first log and failed record were preserved before any retry. The log reports that a failure journal was saved under workspace
.xmd/test-artifacts/.../root-1, but the measurement workflow uploaded onlytest-weights-bun.json; the saved journal is absent from its artifact. The assertion also omits the child's stdout/stderr/finished receipt, so this occurrence still does not identify which exit-1 path was selected.The original #1331 diagnostic hold and its instrumented x64 work are retained under
/private/tmp/xmd-1331-delivery.AxA8DB/. The retained Planner is being reconciled for a diagnostic-first continuation using current source and this new evidence. No production exit-status change, weights retry or assembly/splicing has been performed. Other runtime measurements continue independently.taras commentedon Nov 18, 2026
The diagnostic slice is PR #1414, exact signed
abc64a127b51bbd286698a550fa2723c8b836f2b, with the original Planner and independent Codex Astra both PASS. It preserves joined child output/receipts on an exit mismatch and uploads measurement failure journals separately. The deliberate mismatch control confirms the diagnostic includes the actual status and matching admitted/finished identities; Deno and Node pass, as does the single Linux x64 Bun 1.4.0 probe under emulation.No exit-1 owner was identified, no production exit logic changed, and no second focused configuration was justified. This issue remains open. The one fresh full measurement required by #1408 is 44279296345, attempt 1 at that reviewed source. Any recurrence should supply the first child diagnostics and saved journals back to this same repair; it will not trigger a blind retry.