go-test-utc-midnight-budget-window
Solution written to ~/SOLUTION-go-test-utc-midnight-budget-window.md.
Note on verification scope: the scheduler source repo isn't checked out in this environment (only Go toolchain files exist), so I couldn't run go test against the real tree. To keep the solution genuinely verified rather than theoretical, I reconstructed the exact described pieces (LoadBudgetSpends in budget.go, scheduler.UTCDayStart, and the julianday-based predicate) into a minimal Go module and proved both the bug and the fix:
now.Add(-2h) / now.Add(-1h)): at 00:30Z → 0.00s, at 01:30Z → 2.25s (the exact reported 2.25 vs 6.25), at 12:00Z → 6.25s. This explains both the flake and why CI (runs at 18:56Z/22:51Z) stayed green.scheduler.UTCDayStart(now).Add(+1h/+2h)): spent_daily = 6.25 at every swept (hour, minute) of the UTC day, including the 00:00–02:00 window that was failing. All tests pass.The key insight captured in the fix: anchor tick timestamps to the start of the current UTC day rather than negative offsets from now, so spawned_at always stays inside the active window. This is safe because the predicate only has a lower bound (julianday(spawned_at) >= julianday(UTCDayStart(now))), so future spawned_at values are fine.
The markdown is self-contained: title, root cause (with a boundary table), the exact code fix, verification commands (go test -count=1 -run TestSchedulerGap066_ProjectsPayloadBudget ./internal/api/), and a reusable detection heuristic.
# Evidence - Problem class: go-test-utc-midnight-budget-window - Model: openrouter/deepseek/deepseek-v4-flash-0731 - Solved: 2026-08-25T01:13:22.804Z - Verification: solution produced by pi in sandbox; see signatures.json
{"description": "TestSchedGap066_ProjectsPayloadBudget failed fresh runs 00:00-02:00 UTC (spent_daily 2.25 vs 6.25): test inserted ticks at now.Add(-2h)/now.Add(-1h) while the daily budget window resets at UTC midnight (julianday(spawned_at) >= julianday(UTCDayStart(now)) in LoadBudgetSpends, budget.go) \u2014 the -2h tick crossed the boundary and dropped out. CI stayed green because runs (18:56Z/22:51Z) never hit the window; gitreins guard also passed (diff-scope ran no tests with a clean tree). Fix: anchor tick times to scheduler.UTCDayStart(now)+1h/+2h (always inside the current UTC day, future spawned_at is fine \u2014 no upper bound on the predicate). Verify with go test -count=1 -run TestSchedGap066_ProjectsPayloadBudget ./internal/api/ inside the window. Detection heuristic: budget/date-window assertions using time.Now()-relative offsets + a UTC-midnight window boundary = run-at-any-hour flaky.", "environment": "", "language": "", "model": "openrouter/deepseek/deepseek-v4-flash-0731", "problem_class": "go-test-utc-midnight-budget-window", "provider": "openrouter", "solved_at": "2026-08-25T01:13:22.805Z", "version": ""}Solution written to ~/SOLUTION-go-test-utc-midnight-budget-window.md.
Note on verification scope: the scheduler source repo isn't checked out in this environment (only Go toolchain files exist), so I couldn't run go test against the real tree. To keep the solution genuinely verified rather than theoretical, I reconstructed the exact described pieces (LoadBudgetSpends in budget.go, scheduler.UTCDayStart, and the julianday-based predicate) into a minimal Go module and proved both the bug and the fix:
now.Add(-2h) / now.Add(-1h)): at 00:30Z → 0.00s, at 01:30Z → 2.25s (the exact reported 2.25 vs 6.25), at 12:00Z → 6.25s. This explains both the flake and why CI (runs at 18:56Z/22:51Z) stayed green.scheduler.UTCDayStart(now).Add(+1h/+2h)): spent_daily = 6.25 at every swept (hour, minute) of the UTC day, including the 00:00–02:00 window that was failing. All tests pass.The key insight captured in the fix: anchor tick timestamps to the start of the current UTC day rather than negative offsets from now, so spawned_at always stays inside the active window. This is safe because the predicate only has a lower bound (julianday(spawned_at) >= julianday(UTCDayStart(now))), so future spawned_at values are fine.
The markdown is self-contained: title, root cause (with a boundary table), the exact code fix, verification commands (go test -count=1 -run TestSchedulerGap066_ProjectsPayloadBudget ./internal/api/), and a reusable detection heuristic.
# Evidence - Problem class: go-test-utc-midnight-budget-window - Model: openrouter/deepseek/deepseek-v4-flash-0731 - Solved: 2026-08-25T01:13:22.804Z - Verification: solution produced by pi in sandbox; see signatures.json
{"description": "TestSchedGap066_ProjectsPayloadBudget failed fresh runs 00:00-02:00 UTC (spent_daily 2.25 vs 6.25): test inserted ticks at now.Add(-2h)/now.Add(-1h) while the daily budget window resets at UTC midnight (julianday(spawned_at) >= julianday(UTCDayStart(now)) in LoadBudgetSpends, budget.go) \u2014 the -2h tick crossed the boundary and dropped out. CI stayed green because runs (18:56Z/22:51Z) never hit the window; gitreins guard also passed (diff-scope ran no tests with a clean tree). Fix: anchor tick times to scheduler.UTCDayStart(now)+1h/+2h (always inside the current UTC day, future spawned_at is fine \u2014 no upper bound on the predicate). Verify with go test -count=1 -run TestSchedGap066_ProjectsPayloadBudget ./internal/api/ inside the window. Detection heuristic: budget/date-window assertions using time.Now()-relative offsets + a UTC-midnight window boundary = run-at-any-hour flaky.", "environment": "", "language": "", "model": "openrouter/deepseek/deepseek-v4-flash-0731", "problem_class": "go-test-utc-midnight-budget-window", "provider": "openrouter", "solved_at": "2026-08-25T01:13:22.805Z", "version": ""}