liyuanhao

há 4 mesesdiscuss

the hard part isn't saving state, it's deciding what to forget. yoyo re-reads everything from scratch every session. For yoyo, the repo is his memory. It works at 1,500 lines. won't work at 15,000.

The moment yoyo has to build its own memory management is the most interesting day of the experiment. we're not there yet but it's coming.

liyuanhao

há 4 mesesdiscuss

Hi HN, I'm Yuanhao.

I pointed this agent at its own source code and told it to add a new command. It read itself, edited itself, recompiled, and the command worked. The agent is 260 lines of Rust.

The core idea is that an agent is just a loop:

  Prompt → LLM → Tool calls → Execute → Repeat

yoagent is that loop as a Rust library. You get streaming, built-in tools (bash, file read/write/edit, search), and 20+ LLM providers through 7 API protocols. The CLI example is ~260 lines and I can use it for real work.

The thing I'm most proud of: swap --model and the same 260 lines talk to Claude, GPT, Gemini, Llama via Groq, DeepSeek, or Mistral. One trait (StreamProvider), one implementation for all OpenAI-compatible providers with per-provider quirk flags. Adding a new provider is one file.

No framework to learn. No planning layers, no RAG, no reflection loops. Two traits define everything: StreamProvider (how to talk to an LLM) and AgentTool (how to execute a tool). Read agent_loop.rs and you understand the whole system.

Honest limitations: no permission management, no git integration, tokio is a hard dependency. This is a library, not a product.

MIT licensed. No telemetry. `cargo add yoagent`.

Happy to answer anything about the architecture.

liyuanhao

há 6 mesesdiscuss

I built yolog Desktop because I was tired of losing valuable Claude Code sessions. We've all been there: you solve a complex problem, forget the details, and then hit the same wall two weeks later and have no idea where that conversation went.

If AI coding is how we're all going to work, we need a reliable way to preserve that knowledge. yolog.dev Desktop keeps your sessions searchable and replayable. All local-first, so your data stays on your machine. Currently supporting Claude Code, with more tools on the way.

liyuanhao

há 6 mesesdiscuss

Basically I got tired of losing good Claude Code sessions. You solve something, forget about it, hit the same problem two weeks later and have no idea where that conversation went.

So yolog just keeps everything searchable and replayable. Local-first, no cloud stuff, your data stays on your machine.

The "vibe metrics" came from realizing that raw chat logs are painful to look at. Added some analytics to see patterns (leverage, burn rate, flow state) and a replay mode that makes sessions less boring to review. Actually useful for learning from past sessions instead of just hoarding them.

If AI coding is how we're all going to work, seemed worth having a way to remember what you figured out.

Submissions

[untitled]

[untitled]

Show HN: yoyo - a 200 line baby coding agent evolving in its own Truman Show

Show HN: Yoagent – Build a coding agent CLI in 260 lines of Rust

[untitled]

Show HN: Yolog – Archive and analyze your vibe coding sessions

comments