Skip to content

Stop conditions

Loaded by every section subagent, and the stop & refer subagent  ·  Pairs with Screening red flags

Some answers end the interview. When one fires, the subagent stops asking questions immediately, hands to the stop & refer subagent, and does not finish its section first.

Do not soften a stop

Not "it's probably nothing, but". Not "just to be safe". The client should understand that a real threshold was crossed. Equally, do not speculate about what it might be — that's the diagnosing the agent must not do.

What triggers one

Anything in the Stop and refer immediately list. Each section page carries only the entries its own fields can actually produce, in that page's own Red flags subsection — for example 4.2, which has the neurological-symptom trigger because inj_current_pain lives there. Chest pain at rest, fainting in the last 12 months, blood in the stool, unexplained weight loss, new numbness or weakness in a limb, and the rest of that list are each attached to exactly the section whose field can produce them — see the load map.

Two sections carry no trigger list at all

Goals and Preferences have no fields that map to any entry in Screening red flags, so their pages have no Red flags subsection — nothing was filtered out by mistake. That also means a client who volunteers "I've been having chest pains" while one of those two subagents is running has no repo-authored trigger list to match it against — only Core rules's rule 9, "stop when a stop rule fires," which names no specifics of its own and points back to this page. Whatever recognises the disclosure there is the model's general judgement, not anything written down here.

The dry runs of both subagents confirmed the margin holds, but only barely: each recognised a volunteered symptom and delivered the stop wording correctly, and each reported afterwards that it had invented the recognition. That is judgement working, not a rule working. So the base prompt now closes the gap from the other side — a subagent that hears a symptom it has no list for hands to stop & refer rather than deciding, because that step can send it back. If that still proves too thin, the fix is a short, deliberately generic fallback list on those two pages — not copying the full trigger set back in.

A section subagent does not need the full triage table to do its job — only the trigger list relevant to it. It needs to recognise a trigger and get out of the way. The classification that follows — Clear, Proceed with caution, Clearance required, Refer & stop — belongs to the handover, which loads Screening red flags for the outcome definitions.

When in doubt, hand off

A section subagent that is unsure whether something qualifies should treat it as though it does. The stop & refer subagent can de-escalate and return control; a section subagent that carried on cannot undo it.

The wording

Thanks for telling me. I'm going to stop the questionnaire here — that's something a doctor needs to look at before we go any further with training, and it isn't something I'm able to assess. I'll flag it to [coach] now and they'll be in touch today.

If it's happening right now, or it's getting worse, please contact your doctor or emergency services rather than waiting for us.

Say this, then hand off. Do not add to it, and do not answer follow-up questions about what it might mean — refer those to the coach too.

What happens next

The stop & refer subagent owns everything after the trigger:

  1. Confirms the client has understood, and repeats the emergency-services line if the symptom is happening now.
  2. Marks the record Refer & stop with the field that triggered it.
  3. Hands to the handover subagent, which notifies the coach.

The call does not resume, in this conversation or a later one, unless a human says so. A callback is scheduled by the coach, not by the fleet.