JCR
A capability resolver that walks a nested tree of documented operations with Jev choice questions and returns only the instructions an agent needs for the current step.
A capability resolver that walks a nested tree of documented operations with Jev choice questions and returns only the instructions an agent needs for the current step.
The state is the current request and the step being resolved. Every open candidate in the capability tree becomes one choice question keyed choice_N whose options are that node's direct children plus a none option, and each child's returned probability ranks the walk. When the walk narrows to three finalists, one more choice picks between them, and only the selected operations' documentation is returned.
At the capability root, which direct child best provides the exact capability needed for this step?
The state is the agent's request and the current step. Options are the direct children of the node being expanded, each with its name and description, plus a none option; the ids below stand in for one catalog's own paths.
The agent calls the resolver with the task, and the resolver opens the capability tree at its root.
For each open candidate it asks one choice question over that node's direct children, sending the request and step as state.
Child probabilities extend each candidate's score by the geometric mean of its path, and an unsure branch drops out instead of being followed.
Up to three finalists get one more choice, and the resolver returns the selected operations' instructions without executing anything.
It uses Jev as a search policy over a hierarchy rather than as a text generator, and it scores a whole path by the geometric mean of its question probabilities so a confident early choice cannot hide an unsure one deeper in the tree.