Background subagents burn the weekly quota with no cost visibility, and a rate-limited subagent leaves half-finished work behind

Reporting on behalf of my own usage, with numbers taken from Claude Code's own subagent completion notifications (duration, tokens, tool calls). No estimates.

Environment: Claude Code desktop app, one multi-repo project, one session, 11–13 September 2026.

The actual impact: my work stopped

This is the part that matters most to me. The weekly limit ran out on work I did not ask for, and I could not continue my project until the limit reset. My development stopped — not because of anything I decided, but because the assistant spent my quota on tasks it queued for itself. I was also left with a half-finished repository state that I had to find and repair myself.

What happened

Over two days, eight background subagents ran:

| # | Task | Duration | Tokens | Tool calls | Who asked for it |
|---|---|---|---|---|---|
| 1 | SA firm context | 68.7 min | 589,514 | 296 | me |
| 2 | soft delete | 53.7 min | 544,952 | 198 | me |
| 3 | UI debts | 93.8 min | 669,127 | 298 | me |
| 4 | firm catalogue | 62.6 min | 669,375 | 323 | me |
| 5 | entitlement core | 79.8 min | 484,794 | 198 | assistant expanded scope |
| 6 | legal/consent layer | 52.3 min | 542,646 | 239 | assistant initiated |
| 7 | branch limit | 38.7 min | 447,934 | 160 | assistant expanded scope |
| 8 | service categories | 53.2 min | 477,921 | 223 | me |
| 9 | tooltip pass | died at the limit | never reported | — | me |

Total (1–8): 8h 23m · 4,426,263 tokens · 1,935 tool calls. Average run: 62.9 minutes.

33% of those tokens (1,475,374) went to work I never asked for — the assistant proposed follow-up work, queued it, and ran it. One single self-initiated task cost 542,646 tokens.

Concrete problems

1. No cost visibility before or during a run. Nothing says what a subagent is likely to cost or how much quota is left. I only reconstructed the table above by reading completion notifications one by one, after the fact.

1. No warning as the weekly limit approached. The limit hit while subagent #9 was mid-task.

1. A rate-limited subagent leaves the repo half-finished, silently. #9 had already published a package version (panel-ui v0.18.0) but had not yet committed the dependency update in the two consumer repos. The notification says only "rate limit". Nothing reports what the dead agent left behind. I found it the next day; had it stayed that way, the next package release would have been rejected for a "dirty consumer".

1. A single command taking ~60 minutes is itself the problem. For an hour there is no visibility, no way to steer, and no way to stop — and if it dies at minute 59, that hour is gone.

1. The assistant should not be able to spend my quota on work I did not request. Suggesting follow-up work is fine; running it unasked, for hundreds of thousands of tokens, is not.

What I'm asking for

• Show an estimated cost and remaining quota before a subagent starts.
• Stop and ask past a configurable share of the quota, instead of consuming it silently.
• When a subagent dies on a rate limit, report exactly what it left behind: uncommitted files, the step it was on, the command it was about to run.
• Make quota consumption attributable (which agent, which task) inside the session, not only in post-hoc notifications.
• A refund/credit for the period, since the exhausted quota went largely to unrequested work and cost me working days. I am requesting this through support as well, since a GitHub issue cannot resolve billing.