Configuration resolution
How repository and workspace intent become a workload
The normal researcher contract is a checked-in direct run config:
name: train
engine: rayjob
entrypoint: train.py
compute:
workers: 2
gpus_per_worker: 8
storage:
data_pvc: training-data
TauGrid combines:
- Explicit project and target selection.
- Repository or monorepo discovery.
- The project’s checked-in workspace connection descriptor, or its exact machine-local assignment when no descriptor is present.
- Platform-owned workspace defaults.
- Checked-in workload intent.
- Temporary explicit operator overrides.
Ambiguity is an error. Dry-run output must keep default sources visible.
Checked-in descriptors always take precedence. A local assignment is keyed to the physical Git worktree and selected catalog project, so changing one project cannot silently redirect another repository.
tau run validate --config tau/train.yaml
tau run train --dry-run=client
tau run explain-config
See direct run config vs. managed workflow manifest for how Python SDK-generated manifests relate to this default format.