@taskflow-corp/cli

@taskflow-corp/cli

Multi-agent orchestration harness for AI coding agents (claude-code, codex, cursor, opencode, pi). Async-await TypeScript API; lifecycle hooks; auto-todos with verify-loop.

Install

npm install @taskflow-corp/cli

Quick start

import { taskflow } from '@taskflow-corp/cli';

await taskflow('hello').run(async ({ phase, session }) => {
  await phase('greet', async () => {
    await session('say-hi', { with: 'claude-code', task: 'Print hello world' });
  });
});

Hooks via .agents/taskflow/config.ts

import { defineConfig } from '@taskflow-corp/cli/config';

export default defineConfig({
  events: {
    afterTaskDone: async (ctx, { spec, result }) => {
      // post-task hook
    },
  },
  todos: { autoExtract: true, maxRetries: 3 },
});

Structured output

Sessions with a zod schema return typed results. claude-code uses native tool-use. codex (on gpt-5 / gpt-5.4 models) uses --output-schema; gpt-5-codex variants fall back to prompt-engineered JSON pending openai/codex#4181. Other adapters (cursor, opencode, pi) use prompt-engineered JSON. Override codex behavior with HARNESS_CODEX_SCHEMA=0|1.

import { z } from 'zod';

const result = await session('summary', {
  with: 'codex:gpt-5.4',
  task: 'Summarize the repo',
  schema: z.object({ title: z.string(), bullets: z.array(z.string()) }),
});
// result is typed: { title: string; bullets: string[] }

Plugins

A plugin contributes hooks, a ctx.plugins.<name> namespace, and optional config fragments:

import type { Plugin } from '@taskflow-corp/cli/core';

export const myPlugin: Plugin = () => ({
  name: 'my-plugin',
  events: {
    afterTaskDone: async (ctx, { spec }) => { /* ... */ },
  },
  ctx: () => ({ hello: () => 'world' }),
});

To get typed access to ctx.plugins.myPlugin.hello() in downstream hooks, module-augment the plugin namespace registry:

declare module '@taskflow-corp/cli/core' {
  interface PluginNamespaces {
    'my-plugin': { hello: () => string };
  }
}

See examples/omai-plugin-starter/ for a fuller scaffold (screen capture + UI-TARS wiring stubs).

Environment variables

Var	Purpose
ANTHROPIC_API_KEY	Required for claude-code sessions and pi:anthropic/*.
OPENAI_API_KEY, GROQ_API_KEY, CEREBRAS_API_KEY, GEMINI_API_KEY	Required per-session for the respective providers.
HARNESS_PI_BIN	Override the pi binary name (default pi). Use omp with @oh-my-pi/pi-coding-agent.
HARNESS_CODEX_SCHEMA	0 forces prompt-engineered JSON for codex; 1 forces native --output-schema.
HARNESS_ADAPTER_OVERRIDE=mock	Swap every agent for the mock adapter — smoke runs with zero token cost.
HARNESS_NO_TTY=1	Force headless JSONL output even when a TTY is attached.
HARNESS_RUNS_DIR=...	Override the runs archive directory (default data/runs).
HARNESS_REAL_TESTS=1	Enable integration tests that make real LLM calls (default-skipped).

Docs

See .claude/skills/taskflow/SKILL.md for the authoring guide.