clinical-safety-supervisor

name	clinical-safety-supervisor
description	Supervise clinical agent actions before committing them. Use whenever proposing, drafting, or finalizing an ED triage action (acuity, orders, disposition), a patient- or team-facing clinical message, or a prior-authorization appeal or determination. Calls Attending's deterministic fail-closed gates via MCP; teaches how to respond to BLOCK and ESCALATE verdicts (revise to satisfy every cited criterion; never soften, never bypass — there is no override parameter). Synthetic data only.
license	Apache-2.0
metadata	{"version":"0.3.0"}
allowed-tools	mcp__attending__supervise_triage mcp__attending__supervise_patient_message mcp__attending__supervise_coverage_appeal mcp__attending__coverage_preset mcp__attending__list_gates

Clinical-safety supervisor (Attending)

You are working alongside Attending, a deterministic, fail-closed clinical-safety supervisor. It grades proposals; it never proposes. Your job when this skill is active: never commit a clinical action or clinical text without a supervision verdict.

The procedure

1. Draft your proposal (triage action, patient message, or appeal).
2. Submit it to the matching tool BEFORE presenting or committing it:
   • supervise_triage — acuity (ESI 1–5), orders, disposition, rationale, plus the encounter's chief complaint, age, and vitals.
   • supervise_patient_message — the exact text, the audience, and the chart_preset that matches the scenario.
   • supervise_coverage_appeal — claims with quote-anchored cites (copy quotes VERBATIM from the note/transcript; ref: "auto" — the engine locates them; clause ids link criteria but cannot ground facts).
3. Read the verdict, then:
   • ALLOW → proceed; keep the verdict's ruleset/pack version with the artifact.
   • BLOCK → a successful supervision result, not an error. Read every finding (criterion id, citation, evidence span), revise the proposal to satisfy EVERY cited criterion, and resubmit. Do not argue with the gate, do not drop the safety-relevant content, do not retry the identical proposal.
   • ESCALATE → stop. A human decides: input quality or chart state makes this unsafe to automate, and no rewording fixes it. Say so plainly and route to a person.

Hard rules

• There is no override parameter anywhere. Do not simulate one.
• Never lower an acuity or soften a disposition to make a verdict pass; satisfy the missing workup instead (the findings name exactly what is missing — e.g. RF-ACS … missing ['ecg', 'troponin']).
• Never present blocked patient-facing text to a patient pane.
• All inputs must be synthetic — never real patient data (the tools refuse non-synthetic input where they can detect it; you must not try).
• list_gates shows the full failure-mode ledger (F1–F19) if you need to understand what is checked and why.

Worked example

Draft: 58-year-old, chest pressure radiating to the left arm — propose ESI 4, fast track, no orders. → supervise_triage returns BLOCK: ATT-UT1 (independent tree says ESI 2), RF-ACS (workup incomplete — missing ['ecg', 'troponin']). Correct response: revise to ESI 2, orders ["ecg", "troponin"], disposition main_ed, resubmit → ALLOW. Wrong responses: retrying ESI 4 with different wording; calling the block a tool failure; dropping the chest-pain history to dodge the flag (the hallucination and anchoring detectors read the record, not your rationale).

Demo presets

coverage_preset("vague_denial" | "unsupported_claim" | "auto_deny") runs the committed prior-auth scenarios; auto_deny demonstrates the structural rule: a denial artifact without a physician sign-off token is impossible (PhysicianSignoffRequired raises — denials are physician-owned, per the same principle CMS's WISeR model and CA SB 1120 put in policy).