Regression watch

4.1 nerf-watch

Where the evidence comes from

nerf-watch reads the session logs that Claude Code and Codex already write on your machine and compares your own history across CLI versions and over time. It flags changes you did not make:

• a different model answered than the one you picked;
• reasoning effort dropped;
• the context window shrank;
• cache writes per turn jumped, or the cache hit rate collapsed, after a CLI update;
• tool calls started failing more often.

Install command and a terminal recording: nerf-watch on the home page.

Everything runs locally. nerf-watch share builds the report from structured fields only: detector, severity, agent, model ids, CLI versions, dates, sample counts and before and after values. It contains no prompts, responses, tool output, file paths, project names, session ids, or user or host names, and it makes no network request. It prints exactly what would be shared before you share it.

4.2 Report

How to report

1. Run npx nerf-watch share (Node.js 20 or newer). It prints the anonymized JSON and a link to a prefilled report on this repository. Add --open to open the link in your browser.
2. Read the JSON. The issue is public.
3. Open the link, say what you observed, and submit. If the JSON was too long for the link, paste it into the report field.

A scheduled job reads open issues labeled regression-report once a day and whenever one is opened, edited, closed or relabeled. It validates the JSON against a strict schema, rejects anything that is not valid, too large, or contains path-like or email-like text, and aggregates the findings per agent, CLI version, model and signal. Closing an issue removes it from the table at the next run.

4.3 Reports

Reported regressions