Jiale Wu

codelord is a production-grade open-source AI coding agent (Claude Code competitor) implemented from scratch with zero agent framework dependency — no LangChain, no LangGraph, no CrewAI. Built as a pnpm monorepo with three stable layers: a reusable agent-core engine, a configuration layer with schema validation and layered loading, and a product shell wrapping the engine into a CLI/REPL with persistence and authentication.

Architecture:
• Tool router with semantic rewriting (e.g., bash `cat ` is auto-routed to typed `file_read` tool, reducing shell-complexity surface at framework level)
• Tool safety policy with four-tier risk classification (safe / write / dangerous / control) plus path-based sensitive-resource protection
• Three-layer trace (Provider → Runtime → Operator) with eventId / seq / turnId, enabling 5-second root-cause analysis of agent behavior
• Multi-provider abstraction (Claude, OpenAI, Anthropic, Bedrock, Google) via the pi-ai library
• Production features: runtime interrupt control, session snapshot and resume, checkpoint and undo on every tool execution, runtime-adjustable reasoning level (off / minimal/low/medium/high/xhigh), context window truncation strategy, operator queue injection

A bridge I built that exposes AI agents as chat participants on enterprise IM platforms like WeCom and Feishu, enabling agent-in-the-loop workflows without forcing users to switch tools. Most LLM tooling assumes a dedicated chat surface (web UI, CLI, and standalone app), but real teams already live inside their existing communication tools, forcing a tool switch breaks adoption.

FEATURES
• A pluggable platform-adapter layer abstracting WeCom and Feishu webhook semantics, message types, and authentication into a uniform interface.
• An agent-orchestration layer that routes incoming messages to the right agent instance, maintains conversation context per thread, and handles streaming responses within IM-platform constraints.
• A deployment story optimized for self-hosted production use with a small footprint, a single binary, and no external state dependency required for basic flows.

It has received 18 stars on GitHub.

A low-code visual editor I built from scratch with a 3-layer architecture (engine, workbench, and material-foundation). The project demonstrates the depth of its framework internals and editor design patterns rarely shown in OSS low-code tools.

KEY CONTRIBUTIONS
• Designed and built a VS Code-style dependency-injection container from the ground up, enabling modular plugin composition and clean separation between core and feature layers.
• Built an invertible mutation model that captures every state change as a reversible operation, making it operational-transform-ready for future real-time collaboration.
• Designed a type-level multi-renderer that targets both Vue and React from a single DSL.
• Wrote a custom DSL compiler with compile-time type safety and runtime extensibility, allowing third-party material packages to plug in without modifying the core.

It is fully open source, single-maintained, and currently has 53 GitHub stars.