Jev Gateway
Jev reads the assistant instructions and recent conversation turns, then answers a typed tool Choice and a Noul question about whether a tool call is needed.
Jev reads the assistant instructions and recent conversation turns, then answers a typed tool Choice and a Noul question about whether a tool call is needed.
Jev reads the assistant instructions and recent conversation turns, then answers a typed tool Choice and a Noul question about whether a tool call is needed. When direct-call mode is enabled, it can also answer typed questions for closed-set arguments. The router requires the tool-choice and Noul answers to agree and applies confidence gates before it steers the language model.
Does the assistant need a tool now rather than reply in plain text?
Assistant instructions and recent conversation turns, including tool-result text when present. Runtime tool names form separate Choice questions which are not charted here.
The gateway keeps recent turns within a state budget. It builds a Choice over the current tool roster and a Noul question about whether a tool is needed; optional closed-set tool arguments add further questions.
The router sends state and questions to the configured Jev provider. Large tool rosters are ranked in shards before the final question.
The router checks answer types, confidence, and agreement. It steers only when the checks pass; errors, low confidence, disagreement, and unsupported tools pass through to the language model.
It places a fast typed decision step in front of coding-agent tool selection and fails open to the existing language model if the Jev request fails.