macOS Codex weekly quota switches 26% → 95% → 26% used with inconsistent future reset timestamps

Environment and impact

Date: 2026-09-09. macOS desktop Codex, ChatGPT Pro account. Exact desktop build not verified. Account identity was unchanged between the two direct quota reads. All observations concern limitId codex, primary windowDurationMins 10080.

The UI showed approximately 5% weekly allowance remaining after unattended scheduled work. It subsequently showed 74% remaining. Direct quota tool responses also changed, so this is not established as a UI-only rendering issue. The user paused useful scheduled inspections because the earlier reading suggested unexpectedly severe quota depletion.

Observed evidence

Times are ET (UTC−04:00):

Reset timestamp 1789435346 means 2026-09-14 21:22:26 ET; 1789205174 means 2026-09-12 05:26:14 ET. Neither reset time had arrived during these observations. The read operations did not redeem any reset credits. Available reset-credit count remained 2. No account switching was performed by the assistant.

Minimal response projections:

{"limitId":"codex","primary":{"usedPercent":95,"windowDurationMins":10080,"resetsAt":1789205174},"planType":"pro","availableResetCredits":2}
{"limitId":"codex","primary":{"usedPercent":26,"windowDurationMins":10080,"resetsAt":1789435346},"planType":"pro","availableResetCredits":2}

Expected / actual

Expected: the same account and quota category should return a coherent current quota window, or expose a clearly identifiable entitlement/window change. Reading usage should not be mistaken for redemption.

Actual: reported used percentage and future reset timestamp changed together to a different window, then returned to the previous window. A repeatable trigger has not been established. Refresh/query preceded the visible correction but causality is unproven.

Please investigate inconsistent cached quota snapshots, window selection, entitlement propagation or another account-side cause. These are hypotheses, not established root causes. We are not claiming a confirmed billing defect or that scheduled work actually consumed 69 percentage points.

Privacy and diagnostic scope

This report intentionally excludes private project transcripts, trading data, API keys, auth tokens, account identifiers and reset-credit IDs. Detailed local runtime audit exists separately, including timestamps and task grouping. Do not attach the entire conversation automatically; request a targeted, reviewed excerpt if additional diagnostics are necessary.

Official feedback instructions: https://learn.chatgpt.com/zh-Hans/docs/reference/troubleshooting (type / in the app composer to access feedback; session sharing is optional).

Related reports

Possibly related symptom reports: #44205, #44254 and tracking issue #41220. This report adds a different observed sequence: the same codex weekly counter returned 26% used, then 95% used with a different future reset timestamp, then returned to 26% with the earlier timestamp, without reset-credit redemption. These reports are not assumed to share one root cause.

No repository files or tests are named. Start by comparing repeated local runtime token_count snapshots with authenticated get_usage_limits reads for limitId codex, keeping the account and window parameters fixed. Done means identifying why the used percentage and resetsAt switch between windows, or documenting a reproducible trigger and the evidence needed for further investigation.