hachi

This repo now runs a standalone Discord bot instead of OpenACP.

The bot:

• listens in guild text channels and threads
• creates or reuses a dedicated thread for managed conversations
• answers in a fixed RP persona through a local llama.cpp server
• hands coding-specific messages to local codex
• stores runtime state under .hachi/

Committed files

• .gitignore
• .env.example
• config/defaults.jsonc
• prompts/
• docs/openacp-setup.md

Local setup

1. Install dependencies:

npm install

2. Make sure codex is available on PATH:

codex --version

3. Set the required environment values:

export HACHI_DISCORD_BOT_TOKEN=...
export HACHI_ALLOWED_GUILD_IDS=...

4. Start the bot:

npm run dev

Notes

• scripts/llama-server-wsl keeps the local model server inside WSL and loads the Linux CUDA runtime from .hachi/bin/llama-server-cuda-linux/rootfs.
• The first llama.cpp startup bootstraps the pinned official ggml-org/llama.cpp:server-cuda-b7212 runtime into .hachi/bin/llama-server-cuda-linux/rootfs.
• The first llama.cpp startup also downloads the default GGUF into .hachi/models/.
• Set HACHI_LLAMA_SERVER_BIN only if you want to override the shipped WSL launcher.
• Set HACHI_REPO_ROOT only if you start the bot from outside the repo root.
• Codex run logs are written under .hachi/logs/codex/.
• Structured runtime state is stored in .hachi/db/hachi.sqlite.
• Discord setup still requires enabling Message Content Intent and inviting the bot with bot and applications.commands scopes.
• /code and !code force Codex handoff. Other messages use the RP model unless the lightweight router identifies clear coding intent.