---
name: muster-fire-drill
description: Operate Muster's fictional tabletop fire-drill page through its bounded WebMCP tools, including plan inspection, scripted injects, facilitator-confirmed records, coverage checks, and human-approved review. Use only for training scenarios, never live emergencies or building control.
---

# Muster operator

Operate the same visible exercise state as the human facilitator. Keep every claim tied to a returned tool result. Before an unfamiliar or multi-step run, read [docs/OPERATOR_GUIDE.md](docs/OPERATOR_GUIDE.md).

## Non-negotiable boundaries

- Treat every site, person, count, dimension, date, route, hazard, equipment record, and lesson as a fictional training fixture.
- Do not present Muster as a live sensor, fire-spread model, evacuation authority, compliance check, or emergency-service integration.
- Never infer floor clearance, equipment adequacy, human intent, emotion, personality, or competence.
- Do not record an action or participant signal until the human facilitator confirms what was observed.
- The agent may stage a report; only the human Fire Safety Manager may approve it in the page. Approval is not a WebMCP tool.
- Stop using Muster during a real emergency and follow the official local emergency process.
- A normal browser may expose only the manual interface. Do not claim native WebMCP discovery unless `document.modelContext` registration is observed in a compatible browser.

## Tool surface

Muster defines one manager and nineteen specialist tools (20 total):

| Tool | Input | Operator use |
|---|---|---|
| `run_drill_manager` | `intent`: `orient`, `find_gaps`, `inspect_equipment`, `read_history`, `read_status`, `inspect_zone`, `rehearse`, or `prepare_review`; optional `zone_id` | Route a high-level request. `rehearse` opens the guided sequence; every state change still waits for a human step and report approval is never exposed as a tool. |
| `read_plan` | none | Read the selected fictional building, floor, plan revision, aggregate occupants, exits, and missing role. |
| `start_drill` | none | Start the fictional scenario and add the initial smoke inject. No alarm, call, dispatch, or building control occurs. |
| `send_inject` | `inject_id`: `stair` or `roster` | Add one authored complication after the exercise has started. |
| `record_action` | `action_id`: `reroute`, `account`, or `assist` | Preserve a facilitator-confirmed team action and named owner. |
| `check_coverage` | none | Find active injects without the required recorded response; this does not judge a real response. |
| `stage_report` | none | Create an after-action draft from the exercise log. Human approval remains required. |
| `inspect_zone` | `zone_id`: `west`, `east`, `meeting`, `studio`, `lobby`, or `electrical` | Focus a zone and return fictional occupancy, assistance, nearest exit, and plan distance. |
| `compare_routes` | `zone_id`: `west`, `east`, `meeting`, `studio`, or `lobby` | Compare fictional plan distances and scripted route availability, never direct a live evacuation. |
| `analyze_route_sketch` | `zone_id`: any declared zone; `points`: 2–80 `{x, y}` points inside the fictional 900×610 plan | Measure a facilitator-drawn path, identify its nearest endpoint exit, and flag scripted availability. A qualified human must approve any real route. |
| `read_drill_guide` | `phase`: `before`, `during`, or `after` | Read an exercise checklist, not emergency instructions. |
| `read_hazard` | none | Read the current authored scenario phase; it is not a fire-spread prediction or sensor feed. |
| `read_floor_register` | none | Read aggregate fictional zone counts and assistance ownership; no personal data is returned. |
| `read_status_board` | none | Read facilitator-recorded exercise status without claiming floor clearance. |
| `prepare_team_handoff` | `person_id`: `responder-a-rahman`, `responder-mei-lin`, `responder-d-kumar`, `responder-s-tan`; `room_id`: `west`, `east`, `meeting`, `studio`, `lobby`; `task`: `room_check`, `assistance_brief`, `equipment_check` | Preview a named assignment and cyan room-to-room path. The agent cannot confirm it. Ask the human to inspect and press Confirm assignment. Stale proposals require a new preview. |
| `read_site_context` | none | Read fictional setting, assembly areas, access notes, and explicit live-data limits. |
| `read_room_profile` | `room_id`: `west`, `east`, `meeting`, `studio`, `lobby`, or `electrical` | Focus a room and read its fictional use, protection fixtures, and operating context without inferring cause. |
| `read_equipment` | none | Read equipment shown on the plan; presence, serviceability, adequacy, and compliance are not certified. |
| `read_lessons` | none | Read dated fictional exercise findings and their recorded changes. |
| `record_human_signal` | `role_id`: `fsm`, `security`, `warden-east`, `warden-west`, or `mobility`; `signal`: `confirms`, `uncertain`, `disagrees`, or `delayed` | Record only a facilitator-observed signal, with no psychological or competence inference. |

## Safe operating loop

1. Orient with `read_plan`, `read_site_context`, and `read_floor_register`.
2. Ask the facilitator before changing exercise state.
3. Use `start_drill`, then add at most the requested scripted inject with `send_inject`.
4. Inspect relevant zones, rooms, hazards, routes, or equipment before proposing a training response. Use `analyze_route_sketch` only for a facilitator-drawn path on the fictional plan.
5. Ask what the team actually did; only then call `record_action` or `record_human_signal`.
6. Call `check_coverage`. If gaps remain, name them and return control to the facilitator.
7. Call `stage_report` only when asked or after the facilitator accepts the recorded actions. State that approval is still human-only.

Keep summaries short: observation, source tool, unresolved gap, and next human decision. Never turn fictional proximity or history into a cause claim.

## Room-scale 3D walkthrough

### Assignment loop

Read `read_floor_register` for `team_positions`. Prepare `prepare_team_handoff({person_id:"responder-s-tan",room_id:"studio",task:"assistance_brief"})`. Inspect the returned source, destination, diagram distance, checks and `requires_human_confirmation`. Ask the human to use **Confirm assignment** on the page. Do not try to bypass this control or claim arrival. Then read `read_status_board` to verify `team_assignments` and `team_positions`. A stale preview must be regenerated after state changes. This assignment record does not satisfy a separate `record_action` or prove an assistance task happened.

Three existing tools also accept bounded visual options:

- `read_room_profile({room_id: "studio", view: "3d"})` opens the F07 cutaway and returns room bounds, horizontal dimensions, fixture people and located equipment. `view` may be `current`, `plan` or `3d`. The 3 m elevation, door openings and furniture are inferred.
- `read_equipment({item_id: "MCP-07-L1", view: "3d"})` opens and focuses the manual call-point marker. Other IDs: `EX-07-W1`, `HR-07-E1`, `PWD-07-S1`, `signal-7e`. Both inputs are optional. No device is operated.
- `compare_routes({zone_id: "studio", preview_exit: "A", checkpoint: 2, view: "3d"})` displays one route checkpoint. Checkpoints 0–4 cover room, doorway, corridor, stair approach and accounting handoff. A blocked Stair B caps playback at checkpoint 3. This previews geometry; it records neither evacuation nor a team action.

All three mutate only view/focus or inspection context; their annotations are therefore not read-only. Confirm intended view changes with the user when appropriate. After a route preview, summarize the returned `walkthrough.logic`: observed fixture → scripted constraint → option comparison → human check. It is an explicit rule explanation, not hidden model reasoning or a safety recommendation.

The human may select **Print pack** for two A3 pages, resolution-independent SVG, a 5500 × 4250 PNG, or browser Print/Save PDF. Exports are a timestamped training snapshot, not emergency signage. They preserve missing owners, unverified assembly counts, and separate authored versus measured diagram distances. Exporting does not approve the report.