Claude Code Hidden Problem Analysis
A measured analysis of Claude Code cache, rate-limit and context bugs that inflate token usage on Max plans.
At a Glance
Read the independent research on GitHub at no charge. Claude Code access is separate.
Engagement
Available On
Alternatives
Listed Nov 2026
About Claude Code Hidden Problem Analysis
Claude Code Hidden Problem Analysis is a GitHub repository by ArkNill that documents measured client-side bugs and server-side quota behavior in Anthropic's Claude Code. It combines proxy-captured request data, JSONL session analysis, rate limit header analysis and CHANGELOG cross-references through v2.1.140. The README states it is based on community research and personal measurement and is not endorsed by Anthropic.
What It Is
The repository is a research and documentation project rather than a software tool. It tracks 11 confirmed client-side bugs (B1-B5, B8, B8a, B9, B10, B11, B2a) and 3 preliminary findings. It records which bugs are fixed, possibly fixed or unfixed across Claude Code releases. The author reports the analysis started when a Max 20 plan hit 100% usage in about 70 minutes, and later grew with community contributions.
Evidence and Method
The author used a transparent monitoring proxy (cc-relay) to capture per-request data and rate limit headers. The README reports 45,884 requests across 320 sessions between April 1 and 22. These were cross-checked against independent community datasets and against Anthropic's April 23 postmortem. Chapters cover the dual 5-hour and 7-day quota windows, thinking-token accounting, microcompact and tool-result budget behavior, and per-environment cache ratios.
Opus 4.7 Advisory
A dedicated advisory recommends staying on v2.1.109 with Opus 4.6. It cites a 2.4x averaged five-hour quota burn on Opus 4.7, model pin bypass, cache metering anomalies and an autocompact threshold of about 195K tokens. It also describes sycophancy and narrow-scope behavior patterns, rated by evidence strength.
Practical Guidance
A quick-fix guide covers version pinning with DISABLE_AUTOUPDATER=1, avoiding --resume and --continue, starting fresh sessions periodically, and checking whether the npm or standalone build is in use. Companion documents provide a usage guide, an advanced guide on hooks, subagents and monitoring, an issue index and a Korean translation.
Community Discussions
Be the first to start a conversation about Claude Code Hidden Problem Analysis
Share your experience with Claude Code Hidden Problem Analysis, ask questions, or help others learn from your insights.
Pricing
Public Research
Read the independent research on GitHub at no charge. Claude Code access is separate.
- Analysis of prompt caching, context usage and rate limits
- Benchmark results and a timeline of reported findings
Capabilities
Key Features
- Catalog of 11 confirmed Claude Code client-side bugs with fixed/unfixed status
- CHANGELOG cross-reference through v2.1.140
- Proxy dataset of 45,884 requests across 320 sessions
- Rate limit header analysis of 5-hour and 7-day quota windows
- Opus 4.7 upgrade advisory
- Postmortem cross-check against Anthropic's April 23 postmortem
- Quick-fix and usage guides
- Korean translation
