pi-chat

A pi extension that bridges Discord and Telegram channels to a sandboxed pi session. Each connected channel gets its own Gondolin micro-VM with persistent workspace, shared storage, memory, and skills.

Quick Start

# Install
pi install /path/to/pi-chat
# or
pi -e /path/to/pi-chat

# Configure accounts and channels
/chat-config

# Connect
/chat-connect

Requirements

• QEMU installed (brew install qemu on macOS)
• Gondolin guest image (downloaded automatically on first connect)
• A Discord bot token or Telegram bot token

---

Features

• Discord server channels and Telegram DMs/groups
• Gondolin VM sandbox per connection — tools run inside an isolated Alpine Linux micro-VM
• Persistent workspace and shared storage across sessions
• Streamed preview responses with edit-in-place
• Reply-to-trigger — bot replies are attached to the triggering message
• Durable memory — account-wide and channel-specific memory files
• Skills — agent-created reusable tools, auto-discovered and injected into the prompt
• Encrypted secret exchange — securely pass credentials via browser-based encryption
• Remote control — stop, compact, new session, and status via chat commands
• Chat history tool for searching older messages
• File attachments — send and receive files between chat and the VM

---

Setup

Discord

1. Create a bot at Discord Developer Portal
2. Enable Message Content Intent under Bot settings
3. Run /chat-config → Create account → Discord
4. Enter your bot token
5. Invite the bot to a server (the setup flow provides the invite URL)
6. Select a server and configure channels

Telegram

1. Create a bot via @BotFather
2. Run /chat-config → Create account → Telegram
3. Enter your bot token
4. Add DMs or groups through the guided setup

---

Commands

Command	Description
/chat-config	Configure accounts, channels, and secrets
/chat-connect	Connect to a configured channel
/chat-disconnect	Disconnect the current channel
/chat-status	Show connection status, model, usage, context
/chat-list	List configured channels
/chat-new	Start a new pi session, keeping the chat connection

---

Remote Control

Users in the connected chat can send these commands (with or without mentioning the bot):

Command	Effect
stop	Abort the current turn
status	Show model, usage, context stats
compact	Trigger context compaction
new	Start a new pi session

---

Storage Layout

Everything lives under ~/.pi/agent/chat/:

~/.pi/agent/chat/
├── config.json                          # Accounts, channels, secrets
├── cache/                               # Discovery cache
└── accounts/<account>/
    ├── shared/                          # Mounted as /shared in VM
    │   ├── memory.md                    # Account-wide persistent memory
    │   └── skills/                      # Account-wide skills
    └── channels/<channel>/
        ├── channel.jsonl                # Chat log
        ├── .lock                        # Runtime lock
        ├── workspace/                   # Mounted as /workspace in VM
        │   ├── memory.md                # Channel-specific persistent memory
        │   ├── skills/                  # Channel-specific skills
        │   ├── incoming/                # Downloaded attachments
        │   ├── .secrets/                # Encrypted secrets
        │   └── SYSTEM.md                # Environment modification log
        └── gondolin/                    # VM state
            └── session.json

---

VM Environment

Each connection starts a Gondolin micro-VM with:

• Alpine Linux with bash pre-installed
• /workspace → channel workspace directory
• /shared → account shared directory
• Tools: read, write, edit, bash
• All outbound HTTP/TLS open by default

The agent sees /workspace as its working directory.

---

Memory

Two persistent memory files, injected into the system prompt on every turn:

File	VM Path	Scope
Account memory	/shared/memory.md	Shared across all channels for this account
Channel memory	/workspace/memory.md	Specific to this channel

The agent is instructed to write durable facts and preferences to these files when asked to remember something. Account-wide goes to /shared/memory.md, channel-specific to /workspace/memory.md.

---

Skills

The agent can create reusable tools as skills, following the Agent Skills standard:

• Account-wide: /shared/skills/
• Channel-specific: /workspace/skills/

A skill is either a single .md file (e.g. skills/foo.md) or a directory with SKILL.md plus supporting files (e.g. skills/foo/SKILL.md, skills/foo/run.sh).

Each skill needs YAML frontmatter:

---
name: skill-name
description: Short description of what this skill does
---

Skills are automatically discovered and listed in the system prompt. The agent reads the full skill file before using it.

---

Secrets

Config Secrets (Gondolin HTTP hooks)

Configure secrets at three levels via /chat-config:

• Global — shared across all accounts
• Per account — shared across channels of that account
• Per channel — specific to one channel

Each secret has a value and allowed host patterns. Gondolin replaces placeholder env vars with real values only for outbound HTTP requests to allowed hosts. The agent never sees the real secret value.

Runtime Secrets (encrypted exchange)

For credentials the agent needs at runtime (API keys for skills, OAuth files, etc.):

1. Agent calls the chat_request_secret tool
2. A link to pi.dev/secret is sent to the chat with an embedded public key
3. User clicks, pastes the secret, and gets an encrypted blob
4. User pastes the blob back into chat
5. pi-chat decrypts it (RSA-OAEP + AES-256-GCM) and stores it at /workspace/.secrets/<name>
6. Agent is notified and can use the file

The encrypted blob is useless without the ephemeral private key held in pi-chat's memory.

---

Tools

Tool	Description
read	Read files (routed through Gondolin VM)
write	Create/overwrite files
edit	Precise in-place edits
bash	Execute commands (runs /bin/bash in the VM)
chat_history	Search older messages from the chat log
chat_attach	Queue files to send with the next reply
chat_request_secret	Request a secret from the user via encrypted exchange

---

Credits

pi-chat includes vendored/adapted logic inspired by Vercel Chat SDK (MIT):

• src/render/format.ts
• src/render/streaming-markdown.ts
• src/render/streaming.ts

---

License

MIT