A human example
The user says to hide the left panel, but the command is named Toggle sidebar.
What the caller supplies
The caller supplies available command IDs, their descriptions and the current interface state.
What happens next
Jev selects the command ID. The application checks permissions and asks for any required confirmation.
Illustrative example, not a recorded result.
Potential value: medium
Useful for natural-language command menus and overlapping tools. It adds little when a rule or explicit user instruction already chooses the tool.
Evidence confidence: moderate
A small command demo and tool prediction study report useful selections. End-to-end task success is still uncertain.
The rating describes support for this claim. It is separate from Jev's returned probability. How we assign ratings.
Evidence, including disagreement
- Choosing tools from real MCP inventories. reported. Better tool prediction can still take longer.
- A personal app routes two recipe requests. anecdotal. Descriptions can distinguish two plausible routes; reliability remains untested.
- Select an editor command from an informal request. reported. An authored demo maps descriptions to commands better than its name matcher.
The next test
50 user phrasings over fixed tool inventories, including unavailable actions, ambiguous requests and no suitable tool.
Compare against
- Command-name matching
- Direct agent selection
Measure
- Correct next action
- Completed tasks
- Extra host turns
- Unnecessary calls
Decision after the test
Require at least baseline task completion and a measured reduction in total time or host work.
The report will retain inputs, question versions, every attempt and failure examples. We will update the confidence rating after reviewing the result.
Use a related Engine recipe
Recipes are implementation starting points. Their presence does not mean the protocol above has passed.