Project

gi

active

Go coding agent harness with Piclaw-compatible web UI, append-only turn engine, and go-ai inference.

Overview

Gi is a coding agent built on go-ai, shaped by lessons learned from Pi, Piclaw, and Vibes. It combines a Go turn engine, SQLite-backed state, a Piclaw-compatible web UI, and provider-backed inference into a single workspace-centric agent harness.

How it works

The core is an append-only turn engine backed by SQLite: turns, messages, and events are stored in a local database, then streamed to the frontend over SSE. Inference goes through go-ai, with auth loaded from Pi-compatible config and prompts sourced from AGENTS.md. The web UI reuses Piclaw's TypeScript source verbatim, with Gi-specific api.ts and app.ts adapters wiring the existing component tree into Gi's REST and SSE endpoints.

Features
🧠
go-ai inference

Unified provider layer with streaming and GitHub Copilot enterprise token exchange.

🗃
SQLite-backed turn engine

Append-only turns, messages, and events with queueing, cancelation, and replayable state.

🖥
Piclaw-compatible web UI

Piclaw TypeScript source reused verbatim with Gi-specific API/entry adapters.

📡
SSE live updates

Realtime status, draft deltas, thought deltas, and responses streamed to the browser.

📁
Workspace-centric

Reads AGENTS.md, operates against a workspace root, and exposes workspace file APIs.

🧪
Playwright-tested

End-to-end UX tests cover the running web interface against isolated test instances.

Architecture
Piclaw web UI verbatim TS source Gi web/API layer REST + SSE + workspace APIs Turn engine append-only queue / cancel / stream go-ai inference provider auth + streaming SQLite store sessions · messages · events Pi-compatible config auth.json · AGENTS.md Models Copilot / others REST + SSE Go turn engine + Piclaw UI reuse + go-ai inference
Posts