Summary
packages/computer-use/dist/__tests__/maka-cu-service.test.js fails intermittently on the Linux test lane (Run affected standard workspace tests) with a temp-directory cleanup race:
✖ /home/runner/work/maka/maka/packages/computer-use/dist/__tests__/maka-cu-service.test.js (14.7ms)
[Error: ENOTEMPTY: directory not empty, rmdir '/tmp/maka-test-computer-use-9VpvvT/maka-cu-service-test-fYcMJy']
{ errno: -39, code: 'ENOTEMPTY', syscall: 'rmdir', path: '…' }
[computer-use] failed with code 1
All 971 tests in the affected run otherwise passed (fail 0 in the desktop workspace summary of the same run); the suite fails only in its cleanup.
Evidence
Analysis
ENOTEMPTY on rmdir during teardown is the same defect family as the Windows EBUSY cleanup failures documented in docs/windows-support.md ("Stores, owners, and leases must close deterministically before their temporary root is removed") — a child process or open handle created by the test is still writing into the temp root when the recursive removal runs, or the removal races a concurrently-created file. On Linux this surfaces as ENOTEMPTY instead of EBUSY.
Likely fixes, for the owning team to judge:
- await/settle whatever the service spawns (or its log/artifact writers) before the temp root is removed, mirroring the deterministic-close work done for the storage suites in Phase 1;
- or use
fs.rm(root, { recursive: true, force: true, maxRetries: N, retryDelay: … }) in the test teardown, which the release verifier scripts already standardize on — though retries paper over a still-running writer, so the deterministic settle is the better first choice.
Friction note
Since the Apache migration gh run rerun requires admin rights, so a contributor hit by this flake can only recover by pushing an empty commit — which cancels in-flight lanes for the ref and re-triggers all review bots.
Summary
packages/computer-use/dist/__tests__/maka-cu-service.test.jsfails intermittently on the Linuxtestlane (Run affected standard workspace tests) with a temp-directory cleanup race:All 971 tests in the affected run otherwise passed (
fail 0in the desktop workspace summary of the same run); the suite fails only in its cleanup.Evidence
packages/computer-useat all): run 32337361618, jobtest, step "Run affected standard workspace tests".Analysis
ENOTEMPTYonrmdirduring teardown is the same defect family as the WindowsEBUSYcleanup failures documented indocs/windows-support.md("Stores, owners, and leases must close deterministically before their temporary root is removed") — a child process or open handle created by the test is still writing into the temp root when the recursive removal runs, or the removal races a concurrently-created file. On Linux this surfaces asENOTEMPTYinstead ofEBUSY.Likely fixes, for the owning team to judge:
fs.rm(root, { recursive: true, force: true, maxRetries: N, retryDelay: … })in the test teardown, which the release verifier scripts already standardize on — though retries paper over a still-running writer, so the deterministic settle is the better first choice.Friction note
Since the Apache migration
gh run rerunrequires admin rights, so a contributor hit by this flake can only recover by pushing an empty commit — which cancels in-flight lanes for the ref and re-triggers all review bots.