Why Claude Code burns half its weekly quota in a day – 7 reverse‑engineered bugs

A developer reverse‑engineered Claude Code and identified seven inter‑related bugs—most notably an Extra Usage mode that silently reduces cache TTL to five minutes—causing token‑quota exhaustion up to 1.8× more expensive and triggering a costly “death‑spiral” for CLI users.

Machine Heart
Machine Heart
Machine Heart
Why Claude Code burns half its weekly quota in a day – 7 reverse‑engineered bugs

Claude Code’s recent quality decline sparked concerns when its reasoning depth dropped from about 2,200 characters in late January to 720 characters in late February and further to 560 characters in early March, leading users to distrust its ability to handle complex engineering tasks.

Root causes identified

The author reverse‑engineered the CLI source and uncovered seven bugs that compound token consumption. The most severe bug is that the Extra Usage mode silently disables the one‑hour cache and forces a five‑minute cache TTL without any user notification. This downgrade raises the cost per 220 K‑token round from $0.22 to $0.61, a 1.8× increase, turning a $30 Extra Usage budget from roughly 135 rounds to only 48 rounds.

When the cache TTL is reduced, any pause longer than five minutes (e.g., a bathroom break) forces a full context rebuild, immediately charging the Extra Usage balance. Combined with other cache‑related bugs, this creates a “death spiral”: accelerated quota depletion triggers Extra Usage, which then forces more frequent cache resets, rapidly draining the balance and locking the user out for hours.

Other six bugs

Native installer cache corruption : The bundled Bun runtime corrupts cache prefixes on each request. Switching to an npm installation and verifying the claude binary is a symlink resolves the issue.

Attachment loss in versions 2.1.69–2.1.90 : Session restoration drops key attachment types, causing full cache misses. Fixed in v2.1.91.

Unbounded auto‑compression retries : Failed compression loops indefinitely; 1,279 sessions logged >50 consecutive failures. Fixed in v2.1.89.

Tool result truncation : Bash tool limited to 30 KB and Grep to 20 KB; truncated output breaks cache prefixes. Limits are visible in ~/.claude.json under cachedGrowthBookFeatures.

Core cache‑downgrade bug (described above).

Fake rate‑limit errors : Large conversation logs generate synthetic model entries with zero tokens, falsely indicating rate‑limit hits. Still unfixed.

Server‑side compression deletion : The server silently removes tool results during compression, breaking cache without client‑side patches. Still unfixed.

The author emphasizes that these bugs interact multiplicatively; triggering bugs 1, 3, and 5 together can exhaust a week’s quota in under two hours.

Mitigation steps

Prefer the npm installation over the native installer.

Upgrade to v2.1.91 or later.

Patch the cache‑TTL function in cli.js to always return true (forcing a one‑hour cache). Note that the patch is overwritten on each update.

Apply the npm‑install fix and verify the binary is a symlink.

Community feedback confirms that users who switched to the npm install saw immediate normalization of token consumption, while users of the VS Code plugin, desktop app, or web UI did not encounter these bugs, indicating the issue is specific to the CLI native installer.

Recent improvements

Claude Code v2.1.92 adds detailed cost breakdowns via the /cost command and proactive cache‑expiry warnings in the status bar for Pro users, providing better transparency and reducing silent over‑charging.

Reference: https://www.reddit.com/r/ClaudeAI/comments/1sbqalg/i_reverseengineered_why_claude_code_burns_through/

Original Source

Signed-in readers can open the original source through BestHub's protected redirect.

Sign in to view source
Republication Notice

This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactadmin@besthub.devand we will review it promptly.

CLIAItoken costBug AnalysisClaude Code
Machine Heart
Written by

Machine Heart

Professional AI media and industry service platform

0 followers
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.