Skip to content

Spells

A Spell names an action by what it should accomplish. The implementation might call an Agent, compute a result directly or cause an effect. Consider a Bridge request to answer a question from Archive notes. What answer can the caller expect when recall finds nothing it can use?

In the semantic chat Workflow, RecallSemanticNotes keeps the recall receipt. AnswerSemanticNotes checks which notes can still be used. With an empty selection it prepares an explanation without calling the chat model. With selected notes it requests a local model grant and supplies the source material. SettleSemanticAnswer checks the recall again before saving the reply. Follow the semantic chat tests for the case where permission is withdrawn between recall and settlement. Even an empty result cannot be saved as a reply once access has been revoked.

When adding an executable action, define its inputs, its result and when the caller must be refused. Implement it in Core and bind the owning Workflow through the Workflow registry. The registry checks active revisions and route precedence. Extensions have no Spell registration store today. If an action produces an effect, check permission or the relevant host privilege where the effect is performed.

For an action with several steps, Patterns explains how to describe their order. Weaver executes the score. To find out which steps actually ran, follow the executing owner's observations through Oculus.