contextrot measures Claude Code degradation using session logs
The contextrot utility analyzes the exact context fill level at which an agent begins to fail. It reads local JSONL transcripts that Claude Code already writes to disk, requiring no API keys or network requests.
The tool extracts five failure signals at each step: failed code edits, repeated calls after errors, re-reading already processed files, self-corrections, and tool errors. It then correlates these with context fill levels, calculating statistics using 95% Wilson score intervals.
The report provides one of four verdicts: degradation detected, edge degradation at the window limit, no degradation, or insufficient data. You can run it via uvx contextrot or pip3 install contextrot. Support for Codex CLI, OpenCode, and Gemini CLI is planned. The project is under the MIT license.
Related stories
- 57% of public Claude Code subagents inherit Bash
- Toolog keeps a forensic log of Claude Code tool calls
- Stemma compiles agent rules into CLAUDE.md and AGENTS.md
- Spotify's shunt blocks Claude Code reads over 350 lines
- Context Engineering Kit adds judge agents to Claude Code
- ccswitch juggles Claude Code accounts from one config
Comments
No comments yet. Be the first.
Join the conversation
Sign in with Google to leave a comment. Your name and avatar come from your Google profile, and the comment appears after moderation.