fluss-monty

Check and run flows written as Python source in a Monty sandbox. This is the core that the fluss client and the fluss server share. It never talks to a server itself.

from fluss_monty import ManifestAction, arun_python_flow, validate_python_flow

source = '''
async def main(image: MikroImage) -> MikroImage:
    return await segment(image, sigma=1.5)
'''

report = await validate_python_flow(source, [ManifestAction(alias="segment", action=segment_action)])
assert report.ok, report.diagnostics
report.args, report.returns, report.manifest, report.stubs  # what the flow is stored with

returns = await arun_python_flow(source, {"image": "img-1"}, actions, call)

• Validation works in two passes, and neither executes the entrypoint.

  1. The host ast checks the source's shape and reads the entrypoint's signature into ports.
  2. Monty type-checks the source against stubs of the manifest's actions, and nothing else. A call to anything outside the manifest is an unresolved name.

  validate_python_flow_sync is the same check for synchronous callers. Checking runs under tight resource limits (CHECK_LIMITS), because defining the source's functions runs their class bodies, decorators and defaults.

• A run binds only the manifest's actions, as async functions. Each call goes out through the injected call(action, kwargs, call_key).

  • A call key is alias:digest(kwargs):n, the n-th identical call. It doesn't depend on how gathered calls interleave, so a workflow that is run again finds every call it made.
  • A run that ends early cancels the calls it still has in flight.

• Structures are their IDs. @mikro/image is typed MikroImage, a str subclass that only the type checker sees.

uv run pytest