Assignment

Your steps

1. Register your team and download agent-config.json.
2. Fill in the ordinary chat model configuration; never put the team credential in the model prompt.
3. Implement parsing, allowlisted dispatch, and the observe/decide/act loop.
4. Run practice first, then initialize one ranked run when your lecturer opens it.

Plain-text protocol

The model may return a whole JSON object or one complete JSON code fence. Use ordinary system, user, and assistant messages; do not use native function calling.

{"type":"tool","name":"get_inventory","arguments":{"room_id":"r_OBSERVED_ID"}}

{"type":"final","answer":{"person_id":"p_ID","room_id":"r_ID","weapon_id":"w_ID"}}

Tool catalog

• get_briefing — receive statement/report IDs and the case rules
• get_witness_statement — read the observed floor and jacket color
• get_forensics_report — read the incident wing and weapon material
• list_people — list people and jacket colors
• list_rooms — list room IDs, floors, and wings
• list_weapons — list weapon IDs and materials
• get_access_log — check people present in one room
• get_inventory — check objects recorded in one room

Rules that matter

• One active run per team.
• Malformed final JSON does not consume a run.
• One valid final answer consumes the run, correct or incorrect.
• Retrying a lost network response is not the same as making another guess.