Plan, result, and events

kazeflow separates execution information by when and why it is used.

Information	When	Responsibility
FlowPlan	Before execution	Review selected targets, dependency structure, partitions, and normalized configuration.
RunResult	After normal completion	Review terminal statuses, timings, in-memory outputs, and failure metadata.
Execution events	During execution	Feed optional progress consumers without coupling presentation to the executor.
Logs	During or after execution	Carry application-configured diagnostic detail.

Planning is metadata-only after loading

Flow.plan() validates the selected dependency closure and configuration without invoking asset bodies. The CLI must first import a Python entry to obtain the flow; that loading boundary is separate from planning.

Partition selection is normalized configuration

For a partitioned closure, a FlowPlan records one explicit selection form: keys, an inclusive bounded range, or deliberate empty work. The involved definitions must share a stable domain, and each definition normalizes the supplied input before attempts exist. An omitted selection, invalid date, reversed range, incompatible domain, duplicate normalized key, or selector on unpartitioned work fails during preflight.

The plan exposes selection kind, domain, and safe counts in portable presentation; it does not expose arbitrary selected key values. See Select partitions deliberately for the command forms and the distinction between empty work and missing configuration.

Results are structured values

Asset failures normally appear in a terminal RunResult rather than being raised to the caller. Task and partition-attempt results remain in deterministic plan and selection order.

RunResult.to_record() creates a new JSON-friendly projection. It omits arbitrary outputs, raw exception objects, and the dedicated raw partition-key field. It does not write a file or promise durable history. Its portable exception messages and tracebacks can still contain application-controlled values; see the portable-record trust boundary.

External cancellation is different from a normally completed result. Cancelling the task awaiting Flow.run_async() propagates asyncio.CancelledError without a synthetic terminal RunResult or complete terminal event sequence.

Events do not own execution

An event consumer observes lifecycle values. It cannot authorize a run or replace the terminal result. A consumer failure propagates as infrastructure failure rather than being reported as an asset failure.