gwangee

Qwen 3.5 is a really good local model. I'm using it with personal assistant(https://github.com/daegwang/atombot) every day!

That sounds interesting. Would love to see it when you share it.

Atombot – A tiny but powerful personal AI assistant

github: https://github.com/daegwang/atombot

I wanted a self-hosted personal AI assistant that I could actually understand and enhance end-to-end.

Atombot keeps it simple. Core functionality in just ~500 lines of code (vs 400k in OpenClaw), here's what it does: - Persistent memory with searchable daily history logs - Gateway (support telegram, chat from your phone, allowlist-based access) - Recurring and one-time reminders - Skills system (OpenClaw SKILL.md compatible) - Auto-detects Ollama, LM Studio, Codex on first run. No manual config hunting

Would love feedback!

Give it a description, a plan agent breaks it into tasks, then an execution agent works through them one at a time after the review. - Review and edit the plan before runs - New context for each execution with key information: git diff + a structured progress log - Mix agents: use Claude for planning and Codex for execution, or any combination - Auto-retry on failures

It works with Claude Code and Codex.

Inspired by the Ralph Wiggum pattern, but focused on more structured workflow

Github: https://github.com/daegwang/ralph-code

npm install -g ralph-code cd ~/project ralph-code

Early version. Would love feedback!

I kept having the same experience with AI coding agents: they'd make a mistake, I'd correct them, and later they'd make the exact same mistake again. The instruction files (CLAUDE.md, AGENTS.md) are supposed to solve this, but maintaining them manually is tedious and I'd always forget to update them.

So I built slagent — a CLI that sits in the background, watches your agent sessions, and learns from them.

How it works:

1. It monitors agent log directories and records events — file edits, commands run, test failures, your corrections 2. When you run /review, it sends the session to an AI that identifies patterns like repeated failures, unnecessary retries, or missing project context 3. Each suggestion comes with a target file, rationale, and confidence score. You approve or reject, and approved rules get written into your instruction files automatically (with backups)

Currently supports Claude Code and Codex. Everything is stored locally in ~/.slagent/.

npm install -g self-learning-agent

Still early (v0.1.0) — I'd appreciate feedback on the approach and what you'd want to see next.