Jev Chat Gate

With mention-only routing, an agent waits to be tagged, so an open question to the group or an opportunity to add useful information can pass by without a direct invitation. The gate adds selective participation without a tag: it evaluates the recent conversation and admits an unsolicited answer only when there is a useful reason to join.

Whether there is an unanswered request or something concrete to contribute, whether someone has already answered, and whether joining would intrude on a conversation between people. The README's workflow image names the checks as unresolved request, concrete value, redundancy and intrusion.

Direct requests keep their usual answering path, with no relevance scoring, while still respecting participation controls. The README's behavior table shows a tag or a direct reply to a confirmed message going straight to the direct path.

The send permit is denied. Before sending an ambient draft, the gate also rechecks its reply branch, freshness and current policy, because the opportunity to help may have passed while the model was drafting. The offline demo shows a group question admitted, another person replying, and the draft stopping.

Quiet mode restores direct-request-only participation, so the agent only answers when tagged or replied to. Off mode blocks new text participation. Duplicate and bot-budget controls still apply in both modes, and the demo shows a direct request still admitted in quiet mode.

No. The demo uses synthetic scores, so per the README it verifies routing behaviors — an untagged question entering the answering path, a human reply stopping the ambient draft, a direct request working in quiet mode — rather than Jev's conversational judgment. The tests also cover duplicate suppression, mode changes, bot-turn budgets and evaluator failures across restarts.

No. The gate changes when the agent may participate; answer quality, personality and factual accuracy still depend on your answering model and its context. The README calls usefulness a judgment, not a guarantee that every admitted answer will help, and your main model can still choose silence.