contextrot

Does your coding agent get worse as its context fills?
Find out on the sessions already on your disk — and what actually moves its failure rate.

See the story → priyanshu-byte-coder.github.io/contextrot

Everyone says coding agents get worse as their context fills — /clear often, keep it small, use subagents. That advice comes from lab benchmarks. contextrot checks it against your own real sessions. Sometimes it's true. Often it isn't, and something else is what's hurting you.

No setup, no API keys. It reads the transcripts your agent CLI already keeps, runs entirely on your machine, and makes zero network calls.

uvx contextrot

or pip install contextrot then contextrot (Python 3.9+).

You get one of four honest answers:

A tool that can say "you're fine" is a tool you can trust when it says you're not.

contextrot factors

Context fill is one suspect. This checks the others with the same statistics — time of day, how long the agent runs without you, mistakes piling up, which model, which agent — and ranks what genuinely separates your good steps from your bad ones.

contextrot install statusline --apply    # Claude Code: a live meter in your status bar
contextrot status --setup tmux           # any agent: one line for tmux, Starship, your prompt

A report you run once is forgotten. A meter you see every turn isn't — it shows how full the window is, how many turns are left, and goes red where your curve says it should, not at a generic 70%.

contextrot share --copy

Nobody knows what context rot looks like on real work across many people, because nobody has had the data. This prints your curve as anonymized numbers — no project names, paths, prompts or code — and sends nothing. You read it, then paste it into a Share your curve issue. Clean curves count as much as rotten ones. What's in it, exactly →

Every command animates when you're watching and prints plain output when you're not. The guide covers every command, the status line segment by segment, troubleshooting and the FAQ. The showcase shows each screen.

Adding an agent is one small file with a fixture and a test — the paved first contribution.

Agent CLIs log every step to local transcripts, with token counts and what happened. For each step, contextrot records how full the context window was and whether one of five failure signals fired — a failed edit, a retried call, a re-read of a file already read, a tool error, or an "I apologize, let me fix that". Then it measures how the rate of those changes as context fills.

The statistics are conservative on purpose: Wilson 95% confidence intervals, visible sample sizes, and a threshold declared only when the evidence clears the baseline — one noisy bucket can't scare you. It's observational, and it says so: deeper context also means later in the task. The full method is in methodology.md.

Local files in; terminal, or a local HTML file, out. Zero network calls — there's no HTTP client in the codebase. share prints and copies to your clipboard; whether anything leaves your machine is entirely your decision.

The most valuable first PR is an adapter for the agent CLI you use — see CONTRIBUTING.md. Bugs and ideas are welcome in issues and discussions. If it told you something useful about your setup, a ⭐ helps other people find it.

MIT — see LICENSE. Release notes are in the changelog.