Windows CI: desktop-compatibility build probe PowerShell child times out at 10 s on hosted runners #7854

Client or integration

Other

Area

Platform (Windows / macOS / Linux)

Summary

tests/clients/desktop-compatibility-build-probe.test.ts › "owned ancestor junctions and persistent retargets fail before proof publication" intermittently fails on hosted Windows runners. The PowerShell child it spawns produces no output and is killed at the test's 10 s deadlineMs (exitCode: null, signal: SIGKILL, timedOut: true, empty stdout and stderr). The same test passes in about 1.5 s on other hosted Windows runs from the same dev base, so the deadline looks too tight for slow PowerShell start-up under runner load, not like a product defect. Separately, the harness reports it as an assertion failure, so the shard does not retry it.

Reproduction

Not deterministic. Observed in:

• Cross-platform CI lane=all run 40540187621, attempt 1, job windows 8/9 (shard 8/9, batch 12/50), head a99745788 (feat(codex): restore LazyCodex role models from a change history #7693). The failing call is at tests/clients/desktop-compatibility-build-probe.test.ts:41 (spawnTestChild([resolveTrustedWindowsPowerShellExe(), "-NoProfile", "-NonInteractive", "-Command", script], { cwd: root, deadlineMs: 10_000 })), with the assertion at :44. Attempt 2 of the same job passed.
• The same test passed (1480 ms) in lane=all run 40530456183, job windows 8/9, head 950062c71 (feat(dashboard): revocable server-side remembered device sessions #7593), on the same dev base d9f594b01.

#7693 does not reach this test. It changes only src/codex/role-model-history.ts and src/server/management/codex-agent-role-routes.ts (imported by management-api.ts and route-registry.ts). The probe test imports src/codex/desktop-compatibility/installed-build.ts (→ lib/test-home-guard, lib/windows-elevation, desktop-app/types), tests/helpers/spawn-test-child.ts and tests/helpers/desktop-package-observation.ts. None of these import the changed files or management routes, and the subprocess is a fixed PowerShell script.

Possible directions: give this test's PowerShell spawn the same start-up allowance other Windows PowerShell tests use, or classify a no-output deadline kill as a retryable infrastructure timeout in the Windows shard runner.

Version

dev at d9f594b01 (2.94.0 line)

Operating system

GitHub-hosted Windows runner (windows 8/9 shard of ci.yml lane=all)

Provider and model

No response

Logs or error output

error: expect(received).toMatchObject(expected)
-   "exitCode": 0,
+   "elapsedMs": 10028.5582,
+   "exitCode": null,
+   "signal": "SIGKILL",
+   "stderr": "",
+   "stdout": "",
+   "timedOut": true,
(fail) owned ancestor junctions and persistent retargets fail before proof publication [10038.62ms]
##[error]Test failure in shard 8/9 batch 12/50 (exit 1); not retrying assertion/test failures.

Screenshots and supporting files

No response

Redacted configuration

No response

Checks

• I searched existing issues and documentation.
• I removed secrets, tokens, account details, request credentials, and personal data.