contextrot provides personal context-rot analytics to help identify when coding agents start underperforming. It runs locally without configuration, analyzing session transcripts to deliver actionable insights on context fill and failure rates. Optimize performance with precise reports free from external dependencies.
contextrot
Overview
contextrot is a powerful tool designed to assess the performance degradation of coding agents due to context saturation during sessions. It provides personal context-rot analytics that operate 100% locally and require zero configuration. By analyzing session transcripts stored on disk, contextrot identifies when coding agents begin to fail, what causes this decline, and how it impacts productivity.
Key Features
- Local Analysis: contextrot leverages session transcripts to deliver insights without the need for cloud integration, ensuring user privacy and data security.
- Simple Usage: Launch contextrot easily with just a command:
uvx contextrot - Detailed Reports: Obtain comprehensive reports that outline degradation thresholds, confidence intervals, and recommendations for improvement.
Performance Insight
As coding agents fill their context with information, their performance can suffer, even before reaching known limits. contextrot investigates individual sessions to reveal unique degradation points based on one’s specific projects, models, and interaction styles. This diagnostic tool informs users whether their systems are performing optimally or require adjustments.
Verdicts Provided
Each report from contextrot includes a straightforward verdict:
| Verdict | Description |
|---|---|
| ✗ Context rot detected | Indicates a significant increase in failure rate as context fills. |
| ! Edge rot | Remains flat until near context limit, then increases — suggests compacting information before reaching limits. |
| ✓ No measurable rot | The system operates effectively with a consistent failure rate. |
| ? Not enough data | Encourages further usage for additional data collection and analysis. |
Comparison with Other Tools
contextrot differs fundamentally from benchmarking tools by focusing on personal usage data rather than synthetic benchmarks. Here’s how it compares against other tools:
| Tool | Focus | Limitations |
|---|---|---|
| ccusage | Reports usage costs | Does not assess output quality |
Claude Code /context | Displays current context content | Lacks historical performance analysis |
| Langfuse / Phoenix | Analyzes built applications | Requires extensive setup and instrumentation |
| Chroma's research | Benchmarks degradation | Doesn't reflect personal workload conditions |
Reliability and Accuracy
contextrot’s methodology relies on five independent failure signals extracted from coding sessions, including edit failures and retry loops, to provide accurate assessments. Statistics are calculated with conservative confidence intervals to ensure reliable results.
Supported Agents
contextrot currently supports:
- Claude Code (✅ today)
- Future plans for Codex CLI, OpenCode, and others.
Future Developments
Upcoming features include interactive fixes based on reports and support for more agent adapters, contributing to enhanced community insights regarding context rot.
In summary, contextrot serves as an essential tool for understanding and improving the functionality of coding agents, empowered by local data analysis and clear performance insights.
No comments yet.
Sign in to be the first to comment.