Claude Code burns 20 percent of tokens on a useless status line
Claude Code silently wastes up to a fifth of your rate limits on background status updates, but a new script stops the bleed.
Tools · Source: Hacker News
What happened
Claude Code has a massive leak hidden in plain sight. Every thirty seconds, it takes your background agent's entire transcript and sends it to the Opus model. It asks the model to describe the most recent action in three to five words. This answer only updates a tiny status line in the agents panel. It never reaches your actual conversation. Yet, this invisible task burns fifteen to twenty percent of your total tokens.
There is no official setting to turn this off. The feature became the default behavior in August 2026 with version 2.1.232. A group called DRIM caught the anomaly by looking at Claude Code OpenTelemetry exports. They grouped API request events by query source and found a massive red flag. The agent summary was actually costing more than the main conversation itself. Four independent measurements confirmed the burn rate sits between eighteen and twenty-six percent.
Flamerix AI just released a clever fix on GitHub to stop the bleeding. They built a Bun preload script called no-agent-summary.js. It is seventy lines of code that wraps the global timeout function. When it detects the summary timer trying to fire, it returns an inert handle instead. This kills the summary loop without modifying the actual Claude Code binary.
Key facts
- 16.4% — Share of all Claude Code spend wasted on agent summaries before the fix
- 19.7% — Share of all API requests caused by the summary timer
- 9,000:1 — Input tokens per output token ratio for the summary request
- 30 seconds — How often Claude Code resends the full transcript
- 96% — Portion of the summary spend that hits the Opus model
Why it matters
This changes the math for anyone using Claude Code for serious development. You are losing massive rate limit headroom to a purely cosmetic user interface feature. Input tokens might be getting cheaper, but an input to output ratio of nine thousand to one adds up incredibly fast. When you hit a rate limit, your work stops. Stopping this leak instantly reclaims a fifth of your capacity. You get your API limits back for actual coding tasks.
The second order effect is a harsh wake up call about telemetry and local AI developer tools. If you do not monitor your own API traffic, you are completely flying blind. Tool builders are optimizing for slick interfaces and magical user experiences. They are not optimizing for your API bill or your rate limits. You have to audit your own tools because the defaults will drain your resources.
For builders
Audit your local AI tool network traffic
Export your OpenTelemetry data and group your API requests by source to find hidden costs. You pay the bill when tools make invisible background calls. Builders lose rate limit headroom if they do not monitor outbound traffic.
Patch Claude Code with Bun preload
Install the Flamerix AI script globally using the BUN options environment variable. This intercepts the timer and saves your rate limits without breaking background agents. You win back twenty percent of your capacity immediately while keeping your workflow intact.
Watch for silent breakage on updates
The fix relies on matching source text inside the timer callback. If an update renames the log line or moves the timer, the script will fail open and summaries will return. You must verify your telemetry after every Claude Code update.
My take
I build AI products in public, and silent token burns absolutely infuriate me. Hiding a twenty percent tax inside a cosmetic user interface update is terrible product design. We need raw control over our developer tools, not hardcoded money pits that drain our rate limits.
Original reporting: Hacker News. This is my rewrite and opinion.