Task Dependencies and Lineage
The dependency graph records declared input references. Parent/child metadata records synchronous invocations separately. These records provide evidence of relationships; callers still execute stages and validate data compatibility.
The Alpha158 example shows input dependencies above and invocation records below. Only declared upstream references become edges in get_task_graph.
Passing upstream identities
All BaseInputParams contain a list of canonical Task IDs:
{
"task_name": "prediction-01",
"upstream_task_ids": ["etl#my-etl#dataset-01", "train#my-train#model-01"]
}Python/HTTP and CLI exec inputs use a JSON array. Names must be canonical and unique; the Task cannot depend on itself. Pass each child Task's upstream IDs explicitly. Declaring an upstream does not run it or validate its business compatibility.
Where the graph comes from
get_task_graph builds edges only from metadata.input_params.upstream_task_ids. It returns the connected input-dependency component, including missing referenced upstreams. Nodes are provisional while their state is not succeeded. parent_task_id and children describe in-process invocations in metadata and do not add graph edges. A parent with no declared input dependencies may therefore appear alone even when it has executed children.
Querying and inspecting

In Studio, open a Task's details and select Dependencies. This complete screenshot uses Playground's synthetic ETL, analysis, training, prediction, and backtest records.
axonx get_task_graph --task-id 'predict#my-predict#prediction-01'
axonx get_task_context --task-id 'predict#my-predict#prediction-01'Inspect actual metadata, start times and artifacts when following an edge. A fixed name can now refer to replaced inputs; pin verified revisions before consumption.
Reading upstream Tasks in plugins
from axonx.task.models.identity import TaskType
train_id = self.input_params.upstream_task_id(TaskType.TRAIN)
train_dir = self.upstream_task_dir(train_id)The helper requires exactly one matching upstream type. Locating a directory neither downloads nor executes it. Use storage.inputs helpers to read metadata, validate paths and pin artifacts. Missing/failed/incompatible data is a consumer error, not resolved by a graph edge.
Naming and deletion
Use distinct instances for reproducible experiments and retain input/model revisions. Deletion does not cascade; downstream records may show missing upstreams.
Sources: identity rules, input models, graph.