[BUG] Usage accrues with zero client activity — weekly +14 pts across a 9.4h idle window, including Fable I never invoked

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Weekly subscription usage accrues while every client I control is disconnected and idle. The usage dashboard also reports consumption against a Fable-scoped weekly limit, though I have not invoked Fable since 2026-07-17.

I measured a 9.5-hour window in which my only local client issued zero requests, while weekly_all rose 14 percentage points and weekly_scoped/Fable rose 9 percentage points.

Values below are the server's own, read from the cachedUsageUtilization block Claude Code caches from the usage endpoint, cross-checked against /usage and the claude.ai usage page. Times US Eastern (UTC-4).

| Reading | session (5h) | weekly_all | weekly_scoped (Fable) |
|---|---|---|---|
| 2026-08-04 00:04:21 ET | 16% | 3% | 3% |
| 2026-08-04 09:32:52 ET | 6% | 17% | 12% |
| 2026-08-04 10:26:37 ET | 13% | 18% | 12% |

| Interval | Duration | Local activity | Δ weekly_all | Δ Fable |
|---|---|---|---|---|
| 00:04 → 09:32 ET | 9.5 h | ~7 min | +14 pts | +9 pts |
| 09:32 → 10:26 ET | 0.9 h | ~54 min, continuously active | +1 pt | 0 pts |

Continuous heavy local use produces ~1 point/hour. The overnight accrual is an order of magnitude faster and occurred while nothing local was running. Fable-scoped usage did not move at all during the actively-working interval.

I have never reached a weekly limit on this plan before this behavior began. At the observed overnight rate, a 7-day window is exhausted in ~62 hours.

What Should Happen?

Weekly usage should only accrue from requests I initiate. With all clients disconnected and the local client idle for 9.4 continuous hours, the weekly meter should be approximately flat.

Fable-scoped usage should only accrue when I actually run Fable. My last local Fable request was 2026-07-17T18:18:23Z, so the Fable-scoped weekly meter should read 0% for the 2026-08-03 → 2026-08-10 window.

Error Messages/Logs

No errors. Evidence is a full parse of local Claude Code history: 296 transcript files, 474.6 MB, extracting per-request model, token counts, and timestamps.

All-time local usage by model:
  claude-opus-4-8   41,006 calls   79,547,037 output tokens
  claude-fable-5     3,925 calls    8,005,027 output tokens
  claude-sonnet-5    1,611 calls    1,625,780 output tokens
  claude-opus-5        102 calls      143,387 output tokens

- All local Fable usage falls in 2026-07-15 → 2026-07-17. Last local Fable request: 2026-07-17T18:18:23Z. Cause was a model pin in local settings.json, reverted 2026-07-17 14:34 ET. No local Fable usage since.
- Current weekly window: 143 assistant requests total across two sessions, zero Fable.
- Continuous local silence 2026-08-04T03:59:14Z → 2026-08-04T13:25:25Z (9.44 h), covering nearly all of the 14-point increase.
- Local client fully silent 2026-07-25 → 2026-08-03 (10 days, zero requests) while usage continued to climb.

Steps to Reproduce

1. Note weekly usage via /usage or the claude.ai usage page. Record the timestamp.
2. Disconnect all clients: close all Claude Code sessions, disconnect the browser extension and any IDE integrations, confirm no active agents on claude.ai, log out all sessions/revoke devices.
3. Leave the account completely idle for 8+ hours (overnight).
4. Re-check weekly usage.

Observed: weekly_all rose 14 percentage points and weekly_scoped/Fable rose 9 points across a 9.5-hour idle period.
Expected: approximately no change.

Reproduced across multiple weekly windows over several weeks — usage begins climbing again after every reset. Verified via local transcript audit that the local client issued zero requests during the measured window.

Claude Model

Not sure / Multiple models

Is this a regression?

Yes, this worked in a previous version

Last Working Version

Unknown. The behavior occurs while the client is not running, so it is likely not client-version dependent.

Claude Code Version

2.1.221

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

PowerShell

Additional Information

Ruled out locally:

• Scheduled/cron jobs — none; routineFiredWatermark unchanged since 2026-07-05
• Background processes — single interactive claude.exe verified
• MCP servers — none configured, global or per-project
• OS scheduled tasks — audited; the only Anthropic-adjacent candidate uses a local Ollama backend (localhost:11434), never contacts Anthropic, disabled regardless
• Browser extension, IDE integrations, active agents — disconnected
• claude.ai chats / projects / Cowork / Code agents — deleted or confirmed inactive
• API-key usage — no ANTHROPIC_API_KEY in environment; this is subscription usage, not API credits
• Sessions/devices — "log out all sessions" already used once; behavior resumed

Possibly relevant account flags — the only surfaces I can see that would execute server-side and therefore leave no local trace:
hasRemoteEnvironment = true
groveConfigCache = { grove_enabled: true }
hasUsedRemoteControl = true
remoteControlSurfacesSeen = "mobile"

Not asserting causation — I have no visibility into server-side sessions. Flagging because work in a remote environment would produce exactly this pattern: real consumption, zero local footprint.

Requested:

1. Server-side request logs for my account, window 2026-08-04 04:04 UTC → 13:25 UTC — client/surface/user-agent, originating IP, model, and initiating session or agent ID for every request billed in that period.
2. Explanation of Fable-scoped consumption in the 2026-08-03 → 2026-08-10 window given no local Fable requests since 2026-07-17.
3. Confirmation of whether any remote environment, cloud agent, or scheduled routine is provisioned and executing under the account, and a user-facing way to view and terminate them.
4. A way to force-invalidate all issued OAuth tokens. The account uses Google SSO, so there is no password to rotate, and "log out all sessions" did not stop the behavior.
5. Whether any surface defaults to the Fable model without explicit user selection.

Filed here because the claude.ai support messenger has no escalation path — the assistant there confirmed it cannot access server-side logs, audit sessions, issue case numbers, or investigate remote environments. Account identifiers available on request; omitted as this tracker is public.

Plan: Claude Max 20x (default_claude_max_20x), Stripe subscription. Login via Google SSO.