your first extension

you will write a branches tool, load it, and watch the model call it.

set up a scratch repo

mkdir /tmp/glrs-ext && cd /tmp/glrs-ext && git init
printf 'hello\n' > greeting.txt
git add -A && git commit -m start
git branch fix-login
git branch add-tests
Copy

write the extension

mkdir -p .glrs/extensions
Copy

put this in .glrs/extensions/branches.ts:

import type { Extension } from "@glrs-dev/glrs/extension-api";

const extension: Extension = (g) => {
  g.tool({
    name: "branches",
    description: "List this repository's branches, newest commit first.",
    input: g.z.object({}),
    execute: async () => {
      const shell = await g.exec(
        "git for-each-ref --sort=-committerdate refs/heads --format='%(refname:short) %(committerdate:relative)'",
      );
      return shell.output || "no branches";
    },
  });
};

export default extension;
Copy

three things: a default export taking g, a zod schema from g.z so the extension needs no imports of its own, and g.exec for the shell.

use it

.glrs/extensions is read at startup, so a new run picks it up.

glrs --model anthropic/claude-opus-5 -p "which branch was touched most recently?"
Copy

the model has no git tool, sees branches in its tool list, and calls it:

add-tests, committed 2 minutes ago.
Copy

-p loads extensions exactly as a session does, which makes it the fastest way to check one works.

next: have glrs write the extension, extensions, events