A fresh GitHub issue in the Claude Code repository (number 77018) asks Anthropic to expose rate-limit utilization in headless contexts — through stream-json, --output-format json, or a dedicated usage subcommand. The ask is narrow but the pain is broad: when you run Claude Code non-interactively, you can’t easily see how much of your quota you’ve burned, so you can’t throttle agent loops before they stall.
The problem with invisible limits
In an interactive session, you get subtle cues — slower responses, a warning, a soft block. In a headless run — a cron job, a CI pipeline, a long background agent — those cues disappear. The agent keeps issuing requests until it hits a hard limit and stops, often mid-task, with no clean signal about how close it was to the edge.
Exposing utilization (the issue proposes a used_percentage style field) would let automation make decisions: back off, queue work, or hand a task to a different model when one is saturated. That’s exactly the kind of silent quota burn that bites teams running agents all day.
Why headless matters more now
Headless and background agents are no longer edge cases. The same workflows that promote heavier weekly limits also push more work into unattended runs — and unattended runs are precisely where you can’t watch the meter. The feature request is really about observability for autonomous use, not a nicety for interactive users.
What you can do today
Until Anthropic ships it, the workarounds are crude: wrap calls with your own token/request counting, add artificial pacing between steps, and design agent loops to fail gracefully when a request errors. None of that replaces a real utilization signal, but it keeps unattended runs from silently falling over.
The issue is open and unlabeled beyond the request itself at time of writing — worth tracking if you run Claude Code anywhere it isn’t watched.
FAQ
Q: What does Claude Code issue 77018 ask for? It requests that Claude Code expose rate-limit utilization in headless contexts, via stream-json, –output-format json, or a usage subcommand, so automation can see how much quota remains.
Q: Why does this matter more for headless runs than interactive ones? Interactive sessions show soft warnings as you approach limits. Headless runs — cron jobs, CI, background agents — have no visible meter, so they can stall mid-task with no clean signal about how close they were to the limit.
Q: How can I avoid silent rate-limit stalls today? Wrap agent calls with your own request counting, pace steps artificially, and make loops handle rate-limit errors gracefully. These are stopgaps until a real utilization signal ships.
Tired of deciding which AI subscription to keep? aiFiesta bundles GPT, Claude, Gemini, Grok, DeepSeek, Perplexity and more for $12/mo — less than half of a single premium chat sub.