What might a task specification and snapshot layout look like for an insurance claim intake task?

A task pack for intake can be delivered as one JSONL file of task specifications plus snapshot files they reference.

A task record might carry these fields:

- **task_id**: TSK-0093 - **snapshot_ref**: snap/claim_lodgement_0093.parquet - **goal**: complete the missing-information request and route the claim - **scripted_events**: a list of objects with trigger, offset_steps and payload, such as a document upload arriving after the agent sets PENDING_INFO - **expected_end_state**: field-level assertions, for example status equals ROUTED and queue_id equals QUEUE-LOSS-ASSESS - **protected_fields**: fields that must not change, such as party.legal_name - **allowed_tools**: the tool names exposed for this task - **difficulty_tags**: duplicate_party, late_attachment, conflicting_dates

Snapshots hold the case, party, document, task and policy_rule tables as Parquet or JSONL, with consistent keys across files. A manifest lists file names, schema versions and a content hash for each file so a loader can verify what it read.

Scripted events are stepped by the environment rather than by wall-clock time, which keeps runs repeatable. The exact event vocabulary, snapshot size and tool set are things to agree in a pilot, since they depend on your harness.