Openai-runner NPM

openai-runner

A simple openai task runner driven by openai-like models' Function Calling API.

You can use this tool to quickly write automated unit tests, E2E tests (e.g. Playwright) and much more.

Support Typescript out of box.

Simple Use Case

import { setup, SetupOptions, Action, CreateActions, Action } from 'openai-runner';

// demo config talking to local model llama 3.2
const options: SetupOptions = {
  model: "hermes-3-llama-3.2-3b",
  baseURL: "http://localhost:1234/v1/",
  apiKey: "",
};

/**
 * actions map providing function definition and actual function call
 * */
const createActions: CreateActions = (context: any) => {
    return {
        walk: {
            // actions can be async 
            fn: async ({ secs }: { secs: number }) => {
                console.log(
                    `Robot walked for ${secs} seconds in a ${context.weather} weather`
                );
            },
            // [OPTIONAL] pre validate suggested args by AI to fail fast
            parse: (args: string) =>{
                return z.object({
                    secs: z.number()
                }).parse(JSON.parse(args));
            },
            name: "walk",
            description: "instruct robot to walk for given seconds",
            parameters: {
                type: "object",
                properties: {
                    secs: {
                    type: "number",
                    description: "number of seconds the robot should walk",
                    },
                },
            },
        } as Action,
        // more actions ...
    }
}

// prompt builder can also be async
const buildPrompt = (task: string, context: { weather: string }) => `
    This is your task: ${task}.

    * use provided tools to finish the task.
    * Upon finishing the task, chat to user "Nice robot! And the weather is ${context.weather}" as the final message without saying anything else.
`;

const ai = setup(options, createActions, buildPrompt);

await ai('Ask the robot to walk for 2 minutes and sit for 1 hour.', {
    weather: "sunny",
});

 # expected output

Function:  walk
Params:  {"secs":120}
# console.log
Robot walked for 120 seconds in a sunny weather
Call result:  [empty]

Function:  sit
Params:  {"secs":3600}
# console.log
Robot sat for 3600 seconds in a sunny weather
Call result:  [empty]

Step finished. Final message from AI assistant:
Nice robot! The weather is sunny.

$$$ Token Usage
┌───────────────────┬────────┐
│ (index)           │ Values │
├───────────────────┼────────┤
│ prompt_tokens     │ 79     │
│ completion_tokens │ 157    │
│ total_tokens      │ 236    │
└───────────────────┴────────┘

You can also provide a system prompt so you don't need to send detailed instructions everytime. Optionally you can leave action creater as undefined to skip the function calling completely.

const ai = setup(
        { ...options, systemPrompt: "You are a code generator. Only output code block without extra info." },
        undefined,
        (task) => task
      );

await ai(
    "Write a dummy js code to instruct a robot to walk for 5 mins.",
    { weather: "rainy" }
);

 # expected output

    ```javascript
        function startRobotWalk() {
          console.log("Starting the robot's walking routine...");
          
          let walkingDuration = 300000; // Duration in milliseconds (5 minutes)
          setTimeout(endWalkingRoutine, walkingDuration);
        }

        function endWalkingRoutine() {
          console.log("The robot has finished its 5-minute walk.");
        }

        startRobotWalk();
    ```

$$$ Token Usage
┌───────────────────┬────────┐
│ (index)           │ Values │
├───────────────────┼────────┤
│ prompt_tokens     │ 44     │
│ completion_tokens │ 131    │
│ total_tokens      │ 175    │
└───────────────────┴────────┘