@pstdio/tiny-ai-tasks

Composable AI workflows and tool-using agents for TypeScript/JavaScript. Stream LLM outputs, call tools mid-stream, summarize history, and iterate.

Install

npm i @pstdio/tiny-ai-tasks @pstdio/tiny-tasks

Provide your API key via OPENAI_API_KEY or pass apiKey to the LLM task.

Quick Start

import { createLLMTask } from "@pstdio/tiny-ai-tasks";
import { MemorySaver } from "@pstdio/tiny-tasks";

const llm = createLLMTask({ model: "gpt-5-mini" });

const opts = { runId: "demo", checkpointer: new MemorySaver() };
const messages = [{ role: "user" as const, content: "Write a haiku about databases." }];

for await (const [assistant] of llm(messages, opts)) {
  if (assistant) console.log(assistant.content);
}

Add a Tool

import { createLLMTask, Tool } from "@pstdio/tiny-ai-tasks";

const weather = Tool(async (location: string) => ({ location, temperature: 72, condition: "Sunny" }), {
  name: "weather",
  description: "Get current weather for a location",
  parameters: {
    type: "object",
    properties: { location: { type: "string" } },
    required: ["location"],
  },
});

const llm = createLLMTask({ model: "gpt-5-mini" });

for await (const [assistant] of llm({
  messages: [{ role: "user" as const, content: "What’s the weather in Miami?" }],
  tools: [weather],
})) {
  if (assistant) console.log(assistant.tool_calls);
}

A Tiny Agent

import { createAgent, createLLMTask, Tool } from "@pstdio/tiny-ai-tasks";

const news = Tool(async (topic: string = "general") => ({ topic, articles: [{ title: "Demo" }] }), {
  name: "news",
  description: "Fetch the latest news for a topic.",
  parameters: { type: "object", properties: { topic: { type: "string" } } },
});

const agent = createAgent({
  template: [{ role: "system", content: "You are a concise analyst." }],
  llm: createLLMTask({ model: "gpt-5-mini" }),
  tools: [news],
});

for await (const [msgs] of agent([{ role: "user", content: "Summarize today’s tech news." }])) {
  if (msgs) console.log(msgs);
}

Summarize & Trim History

import { createLLMTask, createSummarizer, truncateToBudget } from "@pstdio/tiny-ai-tasks";

const callLLM = createLLMTask({ model: "gpt-5-mini" });
const summarize = createSummarizer(callLLM);

const history = [
  { role: "system" as const, content: "Rules" },
  { role: "user" as const, content: "Long conversation…" },
];

// Deterministic truncation (no LLM)
const compact = truncateToBudget(history, { budget: 200 });

// Or LLM-based compression of the middle slice
for await (const [msgs] of summarize({ history, opts: { budget: 200, markSummary: true } })) {
  if (msgs) console.log(msgs);
}

Optional Scratchpad Tool

import { createScratchpad, createScratchpadTool, createLLMTask } from "@pstdio/tiny-ai-tasks";

const scratch = createScratchpad();
const scratchTool = createScratchpadTool(scratch);

const llm = createLLMTask({ model: "gpt-5-mini" });
const input = { messages: [{ role: "user" as const, content: "Plan a trip" }], tools: [scratchTool] };

for await (const _ of llm(input)) {
}
console.log("scratch:", scratch.get());

API Overview

• createLLMTask(options): Stream assistant deltas with accumulated tool_calls.
• createAgent({ template?, llm, tools?, maxTurns? }): Minimal loop that plans → runs tools → continues.
• Tool(run, definition): Define a JSON‑schema validated tool; returns { messages, data? } or a plain value.
• createToolTask(tools): Route a ToolCall to the matching tool and package a ToolResult.
• toOpenAITools(tools): Convert tools to OpenAI’s tools shape.
• messages: filterHistory, mergeHistory, toBaseMessages.
• summarize: truncateToBudget, createSummarizer, roughCounter, TokenCounter.

Notes

• Pass credentials via env or apiKey; use baseUrl for proxies/compatible endpoints.
• The last stream chunk may include usage metrics; the final assistant message includes them.
• Import MemorySaver from @pstdio/tiny-tasks for demo checkpointing; swap for persistence in production.