What changed
Chamoji's site reports a benchmark run on the project's own repository. According to the site, 30 real Claude Code sessions answered 5 questions about chamoji's own rules, with half the sessions connected to chamoji and half using git alone. The comparison is about how much an agent has to read before it can answer.1
Chamoji reports 476,616 input tokens in total for the sessions with git alone, against 266,578 with chamoji, which it presents as 1.8 times the input tokens without the tool. The agent answered in 1 turn at the median with chamoji, against 4 with git alone, and made 0 tool calls at the median, against 3. The site adds that the saving was not uniform: by question it ranged from 0.6 times to 4.0 times.1
How it works
According to the site, the session figures are what Claude Code reported for claude-sonnet-5-5 on 19 November, with 3 runs a question, and both sets of figures were measured on chamoji's own repository at commit 61f7810. The figures in its comparison table are estimates at four characters a token rather than exact tokenizer counts. The site says the session brief those sessions used ships in chamoji 0.17.0.1
The site also points to limits in its own result. The two questions with the smallest saving were the two whose rule the session brief had left out, according to Chamoji. For builders who want to try it, the site says the command-line tool is free and open source, installs with npm i -g @chamoji/cli and needs Node.js 22 or newer.1