Skip to main content
An execution’s own trace is its log: the flow.* and processor.* events the engine records while it runs. They are kept in blobhub-workflow-execution-log (partition execution_id, sort key id) and read with list_execution_events. The log has eight types: flow.entered, flow.exited, flow.error, processor.entered, processor.exited, processor.error, processor.warning and processor.output. The run’s lifecycle, execution.created, execution.started, execution.stop_requested and execution.ended, is on the session stream, not here.

Reading

list_execution_events {execution_id, filter?, after?, before?, ascending?, limit?, cursor?, include_execution?} returns {events, cursor, execution?}. Its filter takes types and where (Event Filter); after and before are event ids, exclusive, as on the stream, and cursor is opaque. See List Execution Events.

Shared Event Format

Each log event shares the stream’s id (microseconds in hex, a dash, eight Crockford base-32 characters), session_id, revision_id, created_at, type and source, and adds the execution_id. A log event carries no expires_at: it is removed with its execution.
message, attributes and data are optional. In the examples below only the fields that differ are shown.

Flow Events

Flow events dictate high-level state changes of the execution thread itself.

flow.entered

Fired when an execution fundamentally boots up and locates the start node.

flow.exited

Fired once when an execution ends, whatever the outcome. attributes.route says how: success, failure or stopped. When the ending has a reason, a flow.error is posted just before it.

flow.error

Fired when an execution ends with a reason: a fatal exception, a dead end or a limit. Its message says why, and flow.exited follows it.
An execution stopped by a limit — workflow.processor_runs_per_execution, workflow.execution_seconds or workflow.sleep_seconds — fails with a flow.error whose message names the limit, and whose attributes carry the error limit_exceeded and the same limit object an API refusal does:

Processor Events

Processor events provide component-level logging and profiling metrics. They contain an attributes map heavily populated with IDs letting you trace visually.

processor.entered

Fired when the execution engine yields control to a specific component.

processor.exited

Fired immediately after a component relinquishes control, documenting the exact time it took and what path consequence it chose.

processor.error

Fired explicitly if a component crashes natively rather than gracefully resolving a failure route.

processor.warning

Fired when a component notices a problem but chooses to proceed rather than fail the execution (like a model provider leaving out a setting the model does not accept, a basic agent’s tool answering with an error status, which posts Tool <name> answered HTTP <status>, or logic.code posting a non-fatal diagnostic directly).

processor.output

Informational or debug logs injected manually by internal modules (like the logic.code component posting custom output events).
A model provider (genai.model) posts one per call, with attributes.model naming the model it ran and data holding the call’s request, response and usage. See Which model ran.