kvleak

kvleak

Optimize AI prompt caching and minimize unnecessary spending with kvleak. This powerful tool helps identify where an AI coding agent, such as Claude Code or Codex CLI, is leaking prompt-cache money and offers solutions to improve efficiency.

Efficient Spending Insights

With a typical cache hit rate exceeding 95%, a significant portion of input costs—up to 20%—can still be attributed to cache rebuilds. kvleak effectively highlights the sources of these leaks and provides straightforward solutions, all implemented with a single command.

Key Features

• No Setup Hassles: kvleak requires no API keys, configurations, or network calls. It operates offline, reading session files that AI models save on disk in mere seconds.
• Detailed Leak Reports: Gain insights into session performance with comprehensive audit reports. For example:

  kvleak · prompt-cache leak report
  Cache hit rate: 95.4%  |  Leaked to misses: $537.72  |  Total input spend: $2,602
  

• Optimize Cache Settings: Make adjustments to improve caching performance and reduce unnecessary costs with commands like:

  kvleak tune --apply         # Optimize settings
  

• Real-time Monitoring: Track cache performance live with the kvleak statusline --install command, allowing users to visualize cache countdowns in their status bar.

Why Cache Hits Can Be Misleading

The true cost of a cache hit can be deceptive. A cache miss may appear infrequent, yet it costs significantly more than anticipated—potentially leading to costly rebuilds. kvleak clarifies these dynamics by offering detailed analyses of the underlying costs associated with misses, ensuring every user understands their spending patterns.

Advanced Functions

Several functionalities of kvleak further enhance its value:

• Performance Audits: The kvleak probe command analyzes user history, providing before-and-after comparisons of changes made through adjustments.
• Context Hygiene Management: Methods to refine how context is managed, ensuring optimal performance and lower costs, exemplified by:

  kvleak refine --apply
  

Conclusion

kvleak is an indispensable tool for developers relying on AI coding agents to enhance productivity and efficiency. By providing actionable insights and facilitating reliable optimizations, kvleak ensures that every dollar spent on input is truly justified. For detailed documentation and guidelines, visit the full guide. This tool not only enhances financial efficiency but also contributes to a better understanding of AI cache management.