jev is a decision model, not a chat model. It is served behind a "System One"-style endpoint that takes POST { "model", "state", "questions" } and returns typed judgments — an answers object plus token usage. That means it cannot be added to ~/.kimi-code/config.toml under [models] with protocol = "openai": the agent loop would send chat-completions payloads (/chat/completions, messages, tools) that the decisions endpoint rejects.
The right way to consume it in Kimi Code is the @jkudish/jev-mcp MCP server in compatible-endpoint mode. The server speaks the decisions protocol for you and exposes twelve judgment tools to the agent: jev_verify, jev_screen, jev_noul, jev_find, jev_rerank, jev_classify, jev_decide, jev_compare, jev_extract, jev_audit, jev_review, jev_gate.
This works with any gateway that implements the same