Skip to content

workflows.md

Latest commit

 

History

History
87 lines (65 loc) · 2.37 KB

workflows.md

File metadata and controls

87 lines (65 loc) · 2.37 KB
Raw

Workflows

The repository now includes a first-class persisted workflow surface for multi-step scientific runs:

  • manage_session(action="create", ...)
  • manage_session(action="register_dataset", ...)
  • run_workflow(...)
  • resume_workflow(...)
  • get_status(kind="workflow", ...)
  • get_status(kind="operation", ...)

Internally, the default workflow template is a deterministic sequence that runs:

  1. validate_hpc_setup
  2. probe_path_access
  3. inspect_mesh
  4. inspect_variable
  5. validate_dataset
  6. calculate_area
  7. calculate_zonal_mean when a valid face-centered variable is available

Each workflow stores:

  • workflow status and per-step state
  • progress events
  • a final result handle with a JSON artifact
  • session-visible result and operation references when a session_id is used

Recommended Usage

For repeated work across multiple datasets, create a session first:

from uxarray_mcp.tools import create_session, register_dataset, run_workflow

session = create_session("baseline-analysis")
dataset = register_dataset(
    session["session_id"],
    grid_path="/path/to/grid.nc",
    data_path="/path/to/data.nc",
    name="baseline",
)

workflow = run_workflow(
    session_id=session["session_id"],
    dataset_handle=dataset["dataset_handle"],
    variable_name="temperature",
)

Then inspect it later:

from uxarray_mcp.tools import get_workflow_status, get_result_handle

status = get_workflow_status(workflow["workflow_id"])
summary = get_result_handle(status["result_handle"])

Relationship to the Example Script

The repository still includes:

uv run python scripts/agentic_hpc_loop.py \
  --grid-path /gpfs/fs1/home/<username>/path/to/grid.nc \
  --data-path /gpfs/fs1/home/<username>/path/to/data.nc \
  --poll-seconds 5 \
  --timeout-seconds 300

That script remains useful as a lower-level example of explicit remote polling and branching. The new workflow tools are the supported persisted runtime for the common probe → inspect → validate → analyze sequence.

Current Scope

The first workflow implementation is intentionally narrow:

  • one canonical workflow template
  • JSON-backed local state
  • explicit resume support
  • stage-based progress events instead of percentage completion

This keeps the workflow layer predictable and makes it compose well with the new session, comparison, remapping, and export tools.