An application's assistant is built from three things stored in its dictionary: skills, written as Markdown; tools, written as server scripts with a parameter schema; and agents, which bind a set of each. Until now an external coding agent could read them but not author them through the governed surface. This release adds them as typed families on the development endpoint, with the same verbs, journal and drafts as the rest of the dictionary's code.
Three families
- Skills List, get, write, edit and delete. A skill is prose, so it has no proof; a skill an agent uses cannot be deleted, and the refusal names the agents.
- Tools The body must declare its entry function and pass the platform API check, the parameter schema is proved on every path that sets it, a description and schema are required, and a platform prefix is refused on a new tool.
- Agents Name and description required, the category given by name, an iteration limit, and the agent's tools and skills given as sets that replace its links; each changed set is journalled on its own.
- Attributes A family declares the columns an author sets beside the body, and its verbs publish, accept and return exactly those. An attribute not named keeps its value, an empty one clears it, and widths are measured in bytes as the engine counts them.
Journalled and restorable
- Attributes in the journal A row's attributes are journalled as an item of their own on create, change and delete.
- Back whole A restore plan for a deleted tool, skill or agent recreates it with its body, its attributes and an agent's two sets as they stood. Before, it handed back the body alone.
A procedure, and one loading path
- A prompt to build an agent A development prompt builds an agent in the order the model forces: skills, then tools proved and written, then the agent with its sets. A fresh agent given only the prompt's text built all three.
- The orchestrator reads the dictionary The chat orchestrator now loads the dictionary's assistant tools directly, under their own names, in every world, instead of through a mirror of the application endpoint. The user's AI ceiling is still applied, and a tool named like a platform tool is not loaded.
An assistant is now authored, reviewed and rolled back like any other code in the dictionary, by a person or by an agent working under the same grants.