Repository navigation
Stop a running thread when usage passes a limit you set, per provider and window #33541
Replies: 1 comment 1 reply
|
+1, with one twist: pause until back on pace instead of a hard stop. A thread that is ahead of pace doesn't need a human decision, it needs to wait until the even-pace line catches up. That also gives a concrete answer to open question 3. When to wake. Window start Example: weekly window, day 3 of 7 (even pace 43%), used 60%, so 17% ahead. With Hysteresis. Pause at Pace is per account, threads are many. So this is scheduling, not a per-thread rule:
Where it pauses (Q1). Default: at the end of the current turn, since that is clean and resumable. Optional hard cap (for example used ≥ 90% of the window, or ahead by more than 25%) that interrupts mid-turn, for the single-long-turn case. How it looks. Reuse the Snoozed state and wake machinery from #31072: "Paused · weekly +12% ahead of pace · wakes ~3:40 PM", with Run anyway (one-time override) and Wake now. On wake, the continuation is the same as usage-limit auto-resume. Freshness (Q4). Running Claude threads already receive rate-limit windows on the stream (see #30310 / #24114), so threads that are spending report fresh readings at turn boundaries for free. Polling is only needed for the wake check, and that is once per computed wake time, not periodically. Settings sketch, per provider instance and window, off by default: |
Uh oh!
There was an error while loading. Please reload this page.
Problem
I want to use an expensive model like Fable for some tasks, but I can't leave it running unattended. One long run can use up a whole usage window before I look back at it.
Today a thread only stops when the provider rejects it at the limit. By then the usage is already gone, and the rest of my work on that provider has to wait for the reset.
Proposal
An optional stop rule that you configure per provider instance and per limit window, including model-specific windows where the provider reports them:
It would be off by default, with no built-in thresholds.
When a rule trips, the thread stops in the same recoverable state as a real usage limit:
Why this fits T3 Code
T3 Code already has the data across providers: each window reports
usedPercentandresetsAt, and Usage → Limits already shows pace. Nothing acts on those readings yet, so pace is information you only benefit from if you happen to be watching.Related
Open questions
Written by Claude Opus 5.5 via T3 Code.
All reactions