Device Reliability Agent

Task

Build a Device Reliability Agent that investigates barcode-scanner telemetry and reports emerging failures before they disrupt customers.

Requirements

1. Return a new state with the same seven fields: status, steps, pending, observations, answer, seen. Never mutate inputs. Statuses are ready, tool_pending, done, or stopped. Terminal states are unchanged. max_steps is an integer from 1 to 20, not a boolean; invalid values raise ValueError containing max_steps.
2. A decision event contains action. A tool action is {type: "tool", id, name, arguments}. Allow inspect_fleet, read_device_logs, read_releases, and search_runbooks only. IDs must be nonblank, previously unused strings; arguments must be a dictionary. Invalid actions raise ValueError containing action. Keep only id, name, arguments in pending, append the ID to seen, add one step, and set tool_pending. Only ready states accept decisions; otherwise raise ValueError containing pending.
3. A tool_result event contains call_id and a dictionary result. Match it to pending, append {call_id, name, result} to observations, clear pending, and return to ready. A result whose call_id is already in observations is an unchanged replay. An unmatched result raises ValueError containing call_id. A result can contain an error; record it so the model can decide what to do next.
4. A final action is {type: "final", text, citations}. Require nonblank text and a nonempty list of string citations. Each citation must match a document ID in a previous observation’s result.documents list. Otherwise raise ValueError containing citation. Set done, add one step, and store {text, citations} in answer.
5. Before accepting any decision, if steps >= max_steps, set stopped and clear pending instead. Tool results do not consume a model step. Unknown event types raise ValueError containing event. The HTTP layer validates size and structure; this function enforces ordering, tool policy, and evidence.