Show HN: Contextrot– Analyze where Claude Code degrades from its own session log (github.com)

Contextrot, a new diagnostic tool for coding agents like Claude Code, reveals how performance deteriorates as the context of a session increases. By analyzing session logs already stored on users' systems, it provides actionable insights on when and why an agent's effectiveness declines, thereby addressing a critical gap in existing tools that focus primarily on benchmark tests rather than real-world usage. The tool can lead to immediate optimizations by identifying specific issues like edit failures, retry loops, and self-corrections, presenting users with clear verdicts ranging from "Context rot detected" to "No measurable rot." This innovation is significant for the AI/ML community because it introduces a practical means to gauge the performance of coding agents under typical working conditions, moving away from synthetic benchmarks. Contextrot utilizes five independent failure signals, maintaining rigorous statistical confidence, allowing developers to understand their agent's degradation points based on their unique workflows. Users can gain insights into the economic impact of inefficiencies by seeing potential cost implications of degraded output, thereby helping them make informed decisions regarding model management and optimization strategies. Additionally, contextrot promises strong privacy protections by operating entirely offline, ensuring that no data leaves the user's device.