​

Supervisor assist copilot for text and chat threads

​

Two copilots in the inbox, and when each fires

​

Enablement, roles, and scope

• The panel and the stream require an elevated role, owner, admin, superAdmin, or supervisor, plus an inbox read or write scope on API keys.
• Every request is scoped to your workspace. A conversation id from another workspace returns a plain 404 before the stream opens, and missing or insufficient credentials return 401/403 as JSON, never as a half-opened stream.
• The panel hides itself on finished threads (closed, resolved, archived). If the thread is already closed when a stream connects, you get event: status with terminal: true followed by event: end { "reason": "terminal" }.

​

The rescue loop: read the card, whisper or take over

• recommended_action — intervene, escalate, offer, or monitor.
• headline — a terse imperative summarising the ask.
• rationale — one short paragraph grounded in what the customer and agent actually wrote.
• suggested_message — ready-to-relay text: a private coaching whisper on intervene, a ready reply on offer; omitted on monitor and escalate.
• urgency — low, medium, or high; the panel’s colour cue.
• sentiment_score — a rolling customer-sentiment estimate on [-1, +1].

1. Whisper the agent (intervene). The suggested tip goes to the assigned agent as a private coaching note; the customer never sees it. Use it when the agent is one fact or one tone shift away from recovering the thread.
2. Hand over a reply (offer). The suggested reply is inserted into the thread’s composer through the same insert path as AI-draft suggestions, so it is reviewed and sent by a person. Nothing reaches the customer without a click.
3. Take over (escalate). The panel points you at the takeover surface. The handoff itself, including what the customer sees during the transition, is covered in supervisor takeover and warm handoff. Reserve it for genuine risk: churn threats, compliance breaches, abuse, or an agent clearly out of their depth.
4. Keep watching (monitor). The calibrated default. No action needed.

​

The card on the wire

curl -N "https://orbit.devotel.io/api/v1/inbox/conversations/conv_8f2k/supervisor-assist/stream" \
  -H "X-API-Key: dv_live_sk_your_key_here" \
  -H "Accept: text/event-stream"

const es = new EventSource(
  "https://orbit.devotel.io/api/v1/inbox/conversations/conv_8f2k/supervisor-assist/stream",
  { headers: { "X-API-Key": "dv_live_sk_your_key_here" } },
);
es.addEventListener("assist", (ev) => renderCard(JSON.parse(ev.data)));
es.addEventListener("end", (ev) => {
  if (JSON.parse(ev.data).reason === "max-duration") es.close(); // reconnect to resume
});

{
  "id": 12,
  "generated_at": "2026-08-28T09:41:22Z",
  "model": "claude-sonnet-4-6",
  "recommended_action": "offer",
  "headline": "Offer the 30-day price lock",
  "rationale": "The customer has asked twice whether their current rate survives the plan change; a concrete lock answer removes the last blocker before they upgrade.",
  "suggested_message": "You can keep your current rate for 30 days after the plan change. I can apply the lock today if you confirm.",
  "urgency": "medium",
  "sentiment_score": -0.2
}

​

The dialer variant: compliance prompts on outbound calls

• Configure the disclosure content itself in AI disclosure setup.
• Watch disclosure and do-not-call handling across the floor in live compliance monitoring.

​

Tuning and telemetry

• Cadence is fixed, as on voice. Two new turns, a 12-second minimum gap, and the trailing 6,000 characters. There is no tenant-tunable threshold on the supervisor side; if your floor wants a tighter cue, render or discard cards client-side.
• The inbox_copilot threshold is a different dial. auto_approve_confidence_threshold (default 0.9) marks operator drafts as eligible for one-click approval in the composer. It never sends anything on its own, and it does not gate supervisor cards.
• Reading usage. Team-level handling reads live in inbox team performance. For the operator co-pilot, the accept/edit/discard rollups are described in Inbox AI co-pilot.
• The contract is advisory. Every suggestion on both surfaces carries require_human_review: true. A person clicks send on every customer-visible message; the copilot recommends, it never dispatches.

​

Failure modes

• No cards on a live thread — either the thread is healthy and resting on monitor, or the model layer is degraded while the stream stays open. A stream that holds connected and keeps pinging for over a minute with no assist frames is the degraded-model signature; the panel keeps showing the last card it received.
• Thread already closed at connect — event: status with terminal: true, then event: end { "reason": "terminal" }. Reconnecting is pointless; the coaching queue covers retrospection.
• The 30-minute cap elapses — event: end { "reason": "max-duration" }; reconnect with the same URL and continue watching.
• 401 / 403 / 404 before the stream — a role, scope, or workspace mismatch. Errors arrive as clean JSON because authorization completes before the SSE handshake opens.

​

Related

• Inbox AI co-pilot — the operator-facing drafting drawer, kill switch, and thresholds
• Real-time agent-assist whisper coaching — the voice sibling of this loop
• Supervisor live monitoring for voice calls — listen, whisper, and barge on live calls
• Supervisor takeover and warm handoff — the takeover path an escalate card points you to
• Supervisor live monitoring for digital queues — the digital floor view this copilot sits beside
• AI disclosure setup and live compliance monitoring — the dialer compliance pair
• Inbox team performance — reading handling after the rescue