ygg

Problem

I started ygg because I wanted to use local models on real repositories. The agents I tried often sent large hidden prompts, assumed reliable cloud-grade tool calls, and made it difficult to see the model's context and permissions.

Approach

ygg uses a small, explicit agent loop. Model traffic goes directly to the selected endpoint, and users choose which tools are available. Project instructions load only from trusted workspaces. Conversations are stored as append-only, branchable JSONL on disk.

What I built

I built four Rust crates covering provider-independent inference, the agent runtime, the coding product, and terminal rendering. Version 0.3.0-alpha supports OpenAI Responses, OpenAI Chat Completions, and Anthropic Messages, along with streaming tool use, reasoning controls, image and audio input, session repair and export, prompt templates, skills, extensions, and ten bundled themes.

Engineering focus

The main work covers typed provider requests and streams, capability validation, bounded file and process tools, crash-safe sessions, deterministic configuration, terminal sanitization, differential rendering, and release testing. A regression test holds the reviewed stable prompt scaffold at 1,895 bytes plus dynamic paths. Tool schemas and optional project context are outside that count.

I'll keep using the pre-1.0 release on repository work, improve local-model support, publish reproducible measurements, and fix rough edges before stabilizing the interfaces.