Skip to content

README.md

Latest commit

 

History

History
88 lines (61 loc) · 2.46 KB

README.md

File metadata and controls

88 lines (61 loc) · 2.46 KB
Raw

FlowLLM Lite

flowllm.lite is a tiny local CLI flow runner. It is intentionally separate from the full FlowLLM application framework.

It keeps only four concepts:

  • BaseConfig: define input parameters with Pydantic.
  • BaseFlow: hold config and context, then execute steps in order.
  • register: bind an action name to a flow class.
  • fl: read an action and CLI arguments, run the flow, and print JSON output.

Usage

fl --list
fl --demo --x 1 --y 2
DEFAULT_PROXY_HOST_ENV=example.com fl --proxy
fl --proxy --host example.com --port 12345
fl --remote-server --host 0.0.0.0 --port 8765
DEFAULT_REMOTE_HOST_ENV=example.com fl --remote-client --action ping
fl --remote-client --action exec --command "pwd"

Output:

{"result": 3}

The command format is:

fl --action --field value --another-field value

Rules:

  • The first argument is the action and must look like --action.
  • The remaining arguments must appear in pairs: --field value.
  • - in CLI argument names is converted to _ in Python field names.
  • Values are read as strings first, then converted by Pydantic through BaseConfig.
  • The final result only includes fields declared in output_keys.

The built-in proxy flow starts an SSH SOCKS5 proxy. Its host config defaults to the DEFAULT_PROXY_HOST_ENV environment variable, so no server IP is hard-coded.

The built-in remote_server and remote_client flows provide the remote command executor. The client host config defaults to DEFAULT_REMOTE_HOST_ENV.

Writing a Flow

from flowllm.lite import BaseConfig, BaseFlow, register


class AddConfig(BaseConfig):
    x: int
    y: int


@register("add")
class AddFlow(BaseFlow[AddConfig]):
    output_keys = ["result"]

    def build_steps(self):
        return [self.add]

    def add(self):
        self.context["result"] = self.config.x + self.config.y

Run it:

fl --add --x 1 --y 2

Design Constraints

Lite flow is designed to be direct to read and easy to change:

  • One flow class maps to one action.
  • One step is a zero-argument method that reads from self.config.
  • Steps pass intermediate values through self.context.
  • build_steps() declares execution order explicitly.
  • output_keys declares final output explicitly.

This implementation does not handle service startup, remote calls, plugin orchestration, complex dependency injection, or streaming protocols. Use the full FlowLLM framework when those capabilities are needed.