Show HN: Ctxdiff – Git diff for your LLM agent's context window (github.com)

A new tool called **ctxdiff** has been announced, designed to serve as a local-first debugger for the context windows of large language model (LLM) agents. By integrating seamlessly with various LLM providers like OpenAI, Anthropic, and Google Gemini, ctxdiff captures and records the context for each call made by the agent. This information is stored in a single-file SQLite database, allowing developers to track changes turn by turn, highlighting modifications that may disrupt an agent's performance. The significance of ctxdiff lies in its ability to pinpoint exactly what the model processed during each interaction, addressing challenges associated with prompt cache issues and context engineering—parameters that greatly influence model behavior. With features such as content-hashed block storage, token budgeting, and an intuitive HTML dashboard for inspecting traces, ctxdiff introduces a new level of transparency and debugging capability for AI developers. Unlike traditional tracing tools that aggregate data across multiple runs, ctxdiff's focused approach allows for a granular analysis of order and cost, making it essential for optimizing LLM performance. Its compatibility with both Python and JavaScript further enhances its accessibility for developers working across different environments.