Claude Code burns 20 percent of your tokens on UI status lines. Here is the fix.
Anthropic's Claude Code silently wastes up to 20 percent of your API limits on 3-word UI updates, but a new Bun script stops the bleed.
Tools · Source: Hacker News
What happened
Claude Code has a massive silent leak. Every 30 seconds, it grabs the entire transcript of your background agents and sends it to the model. It usually defaults to the expensive Opus model for this task. It asks the model to describe the most recent action in three to five words. This answer never reaches your actual conversation or workflow. It just updates a tiny status line in the agents panel.
This single cosmetic feature eats 15 to 20 percent of your total tokens. The input to output token ratio sits around 9,000 to 1. DRIM caught this anomaly by looking at OpenTelemetry exports. They grouped API request events by query source and found the truth. The agent summary feature was costing more than the main conversation itself. Four independent measurements confirmed the 18 to 26 percent waste.
You cannot turn this off in the normal settings. Anthropic made it the default behavior in version 2.1.232 back in August 2026. The documented switch only turns it on, and the server forces the interval. Now, Flamerix AI has released a 70-line Bun preload script called no-agent-summary.js. It intercepts the global timeout function and kills the summary request timer. It does this without modifying the actual Claude Code binary.
Key facts
- 15-20% — Share of total tokens burned by the status line feature
- ~30 seconds — Frequency of background agent transcript submissions
- 9,000:1 — Input to output token ratio for the summary requests
- 2.1.232 — Claude Code version where this became default in August 2026
- ~70 lines — Length of the Bun preload script fix by Flamerix AI
Why it matters
Builders are losing massive rate-limit headroom for a purely cosmetic UI feature. When you run Claude Code on a subscription, you hit your usage caps 20 percent faster than you should. If you pay per token on a usage tier, you are burning actual cash on invisible status updates. You are paying premium Opus prices just to power a text-based progress bar. This limits how much real work you can push through the AI before getting throttled.
This exposes a massive flaw in how AI companies design background tasks. Developers trust command line tools to be lean and efficient. Silently looping massive context windows for UI updates completely breaks that trust. It forces engineers to monitor their own AI tools with advanced telemetry just to catch wasteful API calls. If a tool requires users to inject custom Bun scripts just to stop token bleeding, the default architecture is fundamentally broken.
For builders
Audit your AI tool telemetry
Export your Claude Code usage data via OpenTelemetry. Look for the agent summary query source to see exactly how much rate limit you lose. The developers paying for API access are the ones footing the bill for this waste.
Install the Bun preload fix
Download the Flamerix AI script and set it globally in your environment variables. This intercepts the timer and saves your API credits without breaking your actual background agents. You keep the functionality but lose the bloat.
Watch for silent updates
The preload script relies on matching specific source text in the timer callback. Anthropic could rename the log line in a future update. This would silently break the fix and restart your token burn, so you must verify the fix after every update.
My take
Anthropic shipping a CLI tool that burns a fifth of your tokens on a progress bar is peak AI bloat. We are paying Opus rates for cosmetic UI updates because nobody bothered to add a simple off switch. If you build developer tools, you must respect your user's API limits. If you do not, your users will literally hack your binaries to stop the bleeding.
Original reporting: Hacker News. This is my rewrite and opinion.