---

name: operator-hub-and-cli description: "Operate Towhee ops wrappers, local operator registration, Hub revisions, and CLI help validation without unsafe downloads or server startup." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Operator Hub and CLI

Use this sub-skill when a task needs Towhee operator lookup, local operator development boundaries, Hub operator version selection, towhee init guidance, package entry points, or help-only CLI validation.

Fast routing

• For operator names, wrapper behavior, local registration, PyOperator, and NNOperator boundaries, use references/operator-api.md.
• For towhee, towhee init, towhee server --help, python -m towhee, and package entry points, use references/cli-reference.md.
• For failure triage around pkg_resources, Hub downloads/cache, accidental server startup, template writes, registry naming, and NN training prerequisites, use references/troubleshooting.md.
• To validate the installed CLI safely, run scripts/check_cli_help.py. The script checks help output only and must not start servers or initialize templates.

Boundaries and handoff

• Route neural-network training loops, trainer configs, model zoo dependencies, and extended NNOperator.train() usage to training-and-models.
• Route live towhee server startup, HTTP/GRPC clients, Triton, Docker, and port management to serving-and-triton.
• Route pipeline node composition, schemas, batching, flush/debug/profiling, and pipe.input(...).map(...).output(...) design to pipeline-programming.
• Do not use Hub downloads, towhee init, or live servers as routine validation unless the user explicitly requests those side effects and supplies an isolated target.

Minimum safe operating pattern

from towhee import ops, register
from towhee.operator import PyOperator, NNOperator

@register(name='add_operator')
class AddOperator(PyOperator):
    def __init__(self, factor):
        self.factor = factor

    def __call__(self, x):
        return x + self.factor

op = ops.add_operator(10)   # lazy wrapper; loads when called
assert op(2) == 12

Prefer pinned Hub revisions for reproducibility:

op = ops.some_namespace.some_operator().revision('main')

Use .latest() only when intentionally refreshing a cached Hub operator.