Documentation
Agent capabilities and limits
What a connected agent can read, what it can change, and what it can never do.
Fourteen tools: nine that read, two that write, and three for the project's image files. Every write goes through the same operations, the same reducer and the same undo a person's edit does.
Reads
The orientation, the document outline, a slide's markup, a slide's stylesheet, the component vocabulary, the design system, the templates and the project list.
Writes
One tool, apply_operations, taking a batch of typed operations: create, duplicate, resize, reorder and delete slides; add, update, group, order and delete elements; insert presets and components; set a slide's markup or its stylesheet; set design tokens, variables and the title; and the whole music vocabulary. One batch is one undo.
Images
An agent can list the project's images, import one from a public https URL, and upload one from your disk through the local file companion. No image ever travels inside a tool call.
What it cannot do
- Send code. The markup and stylesheet operations take Framulus markup and Framulus CSS, both through the same parsers the Code panel uses, and neither can express anything the document model does not already have.
- Reach outside the document. The only outward-facing tool is the https image import, and it is guarded against private addresses on every redirect hop.
- Write as a viewer. Access is re-resolved on every call from the grant and the project's access list.
- Act on more than one document. A call lands in the editor session you have open, and two open editors is a refusal rather than a guess.
Addressing the master
The master is addressed as the slide __master__, so every element operation works on it with no special case.
Standing instructions
Settings holds rules handed to every agent you connect, and a project can add its own from the Agent panel. They are quoted into the orientation and never parsed, matched or executed: they guide, and they cannot cause anything to happen.