glrs is given its own documentation and told to write an extension when asked for something it cannot do. you will ask for a capability it lacks and watch it build one.
mkdir /tmp/glrs-authored && cd /tmp/glrs-authored && git init
printf 'hello\n' > greeting.txt
git add -A && git commit -m start
glrs --model anthropic/claude-opus-5
type this:
add a /branches command that lists this repo's branches with their last commit date
there is no /branches command and no tool that lists branches.
it reads its own documentation before writing anything:
read 9-reference/11-extensions.md 142 lines
read 1-tutorials/2-first-extension.md 70 lines
write .glrs/extensions/branches.ts
those paths are under glrs's install directory, not your project.
cat .glrs/extensions/branches.ts
a g.command() call, a g.exec() for the git work, and a default export.
/reload
/branches
/reload re-reads extensions without restarting.
the system prompt lists the reference pages by path and tells the model to read them
when asked about glrs itself. it is pointed at documentation rather than at
packages/, so what it can read about glrs is what it can rely on.
next: design, extensions