FluxyChatHow-to Guides
Room Decisions
System One gates on a room. Jev and CLM score options; they do not write chat. Deterministic policy still wins.
Room Decisions
Jev (TypeSafe, Workers AI typesafe/jev) and CLM-8B (Apache-2.0, self-hosted GPU) do not generate messages. They return a typed choice or a noul score. Groq (or your room LLM) still writes text. Thresholds, HITL, and two-key stay in application code.
Author numbers (DeepSWE, Terminal-Bench, BFCL, WikiRacing, “13×”) are theirs. We do not claim them. CLM needs vLLM-class GPUs; that is not on the hosted free plan.
What is on
| Kind | Default | What it does |
|---|---|---|
Floor (shouldRespond) | keyword | Legacy ambient keyword policies. Opt-in: invoke_only, calibrated, silent. |
Tool HITL (approveTool) | shadow | Log System One; extra alignment HITL only in enforce. Two-key remains last word. Auto-handoff stays off unless SYSTEM_ONE_AUTO_HANDOFF. |
Cost router (routeModel) | shadow | Logs skip / small / large. enforce may skip LLM (OK.) or cap tokens / ROOM_LLM_SMALL_MODEL. No published €/room-hour. |
Notify (notifyTriage) | shadow | Reuses existing priority scores. enforce skips push on low, non-mention. Mentions still push. |
| Auto-summon | shadow | Fixed bot roster (max 12) + silence. enforce writes a feed suggestion. Does not add members. Not a 1k-candidate directory. |
| NL policy | shadow | Room nlPolicy or ROOM_NL_POLICY. System One noul → ok / warn / block. enforce rewrites agent text only. Not a medical-device or legal filter. |
| Labels | opt-in | D1 room_system_one_decisions: kind, options, probabilities, model, human outcome (approved / denied / escalated). Hosted needs ROOM_DECISIONS_LABELS=true or decisions.labelsOptIn. Skip rooms linked to no_train connections. |
GET /rooms/:id/system-one-decisions (member JWT). Console: Rooms → Advanced → Room Decisions. Model counts on that page are not a significance test.
fluxy.config
decisions: {
shouldRespond: { mode: "keyword" },
approveTool: { mode: "shadow" },
routeModel: { mode: "shadow" },
notifyTriage: { mode: "shadow" },
autoSummon: { mode: "shadow" },
nlPolicy: { mode: "shadow" },
}Env: ROOM_FLOOR_MODE, ROOM_TOOL_DECISION_MODE, ROOM_ROUTE_MODE, ROOM_NOTIFY_MODE, ROOM_SUMMON_MODE, SYSTEM_ONE_CASCADE=clm_then_jev. ROOM_DECISION_MARGIN_MIN is per-decision-type, not copied from a Jev noul threshold.
Floor modes
- keyword — today’s ambient
message_keywordpolicies (no behavior change). - invoke_only — ambient only if the line contains
@handle(CHI-style: don’t jump into small talk). - calibrated — System One
silencevsspeakon a fixed two-option menu. Low margin → silence. - silent — no ambient dispatch.
@mentions still invoke agents on the mention path. That is not ambient.
Tool gate
Deterministic two-key + tool policy first. System One noul can add HITL. Alignment noul (LlamaFirewall-style “does this serve the user’s goal?”) runs only in enforce. We do not abort the run: extra HITL or a tool error, not a silent kill. Shadow/warn only write the audit row.
Not in this slice
Mid-stream abort. Best-of-N draft pick (Jev lost as a selector on two public benches; don’t sell it). A published cost-per-room-hour. Certified medical/legal compliance.