A command-line agent workspace for designing, running, and comparing agent behavior.

终端智能体Harness 工程LangGraphAgents SDK技能驱动流式交互实时成本本地优先

Lilac-CLI turns Markdown skill files into swappable agent identities, while the added harness project gives LangGraph and OpenAI Agents SDK workflows a local place to be exercised, inspected, and refined.

RuntimeBun

InterfaceInk

Skill format.md

Agent layerLangGraph / Agents SDK

Repository

/home/lry/Projects/TsRepo/lilac

Technical Stack

BunTypeScriptReactInkLangGraphOpenAI Agents SDKgray-matterink-text-inputink-spinnerOpenAI-compatible APIs

01 / Product premise

Agent identity becomes a local, versionable design surface.

Instead of baking behavior into code, Lilac loads persona, model, temperature, and constraints from Markdown frontmatter. A new assistant mode can be created by adding a skill file, while harness runs keep those behaviors testable outside a single chat session.

02 / Interface system

React patterns brought into the terminal without losing terminal speed.

The UI is composed with Ink components for header, message list, input, spinner states, and streaming text. The result feels closer to a professional developer tool than a plain prompt loop.

03 / Agent harness

LangGraph and OpenAI Agents SDK workflows can be exercised as engineering artifacts.

The harness project adds a separate layer for structured agent experiments, making graph-based flows, SDK-driven tool calls, and repeatable behavior checks easier to run without disturbing the terminal interface.

04 / Operating feedback

Live token estimation makes model usage visible during the session.

A lightweight token utility feeds the header cost monitor, so the interface keeps both conversation state and resource pressure visible while responses stream.

What I Built

Markdown skills define agent persona, model, temperature, and instruction constraints.

A harness project separates agent workflow experiments from the core terminal interface.

LangGraph supports graph-shaped orchestration for multi-step and multi-agent behavior.

OpenAI Agents SDK integration gives the project a path toward typed tools, handoffs, and traceable agent runs.

Provider-agnostic API client works with OpenAI-style services including GPT, DeepSeek, and local Ollama-compatible endpoints.

Gradient terminal typography, side-line message layout, and loading motion give the CLI a designed, premium feel.

Bun keeps startup and TypeScript execution tight enough for daily command-line use.