JevAutoModeMiddleware

Uses Jev to decide, before execution, whether a guarded tool call is safe enough to run automatically.

For every tool call whose name appears in guardedTools, this middleware asks Jev a yes/no risk question built from the conversation history and the tool input. Calls whose P(safe) is at least safetyThreshold pass through; the rest get a synthetic DENIED tool result written to the conversation state and are never dispatched to the toolkit. Tool calls that were already explicitly allowed by the user (state ALLOWED) skip the check entirely, so this guardrail cannot override a human confirmation.

The check runs in #onActing, after any outer middleware that filters or rewrites tool calls and immediately before the core permission/execution pipeline. This makes it complementary to the deterministic PermissionEngine: a call that passes the Jev assessment is still subject to permission deny rules inside the core acting pipeline, so this middleware adds a model-driven pre-check for the specific tools it guards without replacing the rule-based engine.