Create schedule

The action is derived from the shape: agent_id + message creates an agent_run task that fires as a durable agent run carrying the guardrail-screened message (endpoint_id/payload are rejected); agent_id without message is the pre-46a compatibility shape, stored as agent_run with the exact message scheduled trigger and its legacy delivery fields retained but never executed; message + model (+ optional tools) with NO agent_id creates an agent_run task for which the engine makes a hidden personal agent — persona framed around the message, the named tools at their default policy — and binds it (the response carries its agent_id; the agent is deleted with the schedule; model is required because the engine has no organization default; model/tools are rejected on every other shape); no agent_id and no message creates the original webhook action, which requires endpoint_id. Responses expose action_type, the stored message (agent_run only), the engine-computed owned_by_caller, and reads_from — dataset or project, which of the task's two data facts the runs it fires READ when a dataset.query call names none of its own (2026-09-22). reads_from is derived from the same resolution the fired run obeys, not restated from the body: dataset whenever the task carries a dataset_id, project when it carries only a project_id. It is ABSENT when the task carries neither — there is then no place to read, and saying project would assert one that does not exist — and absent on every webhook schedule, which fires no run. It appears identically on this create, on GET /v1/schedules/{id}, on PATCH /v1/schedules/{id} and on each row of GET /v1/schedules.