The agent asks the shell for recent history and the working tree with ordinary git commands.
Before each bash call runs, rtk rewrites the command into its token-saving form; the transcript still shows the command the agent typed.
The output is rtk's compact form, for example * main clean — nothing to commit instead of git's paragraph, and the agent answers from it.
Protagonist: rtk.
Point: common commands get cheaper for the model without the model knowing or caring.
# rtk demo
1. The agent asks the shell for recent history and the working tree with ordinary `git` commands.
2. Before each `bash` call runs, rtk rewrites the command into its token-saving form; the transcript still shows the command the agent typed.
3. The output is rtk's compact form, for example `* main clean — nothing to commit` instead of git's paragraph, and the agent answers from it.
Protagonist: rtk.
Point: common commands get cheaper for the model without the model knowing or caring.