jj-mailbox: File-Based Agent Messaging

You have access to a jj-mailbox — a shared file-based messaging system that lets you communicate with other agents. Messages are JSON files in a jj (Jujutsu) version-controlled repo.

Scope

This skill only reads and writes files inside $JJ_MAILBOX_REPO (subdirectories: inbox/, agents/, shared/). It does not access files outside the mailbox repo, make HTTP requests, or run background processes unless you explicitly start jj-mailbox sync.

Prerequisites

• jj-mailbox CLI — included as jj-mailbox.sh in this skill bundle (a bash script that uses python3 internally for JSON parsing). Copy it to your PATH or run directly.
• jj and git — Jujutsu version control with git backend.
• Environment variables:
  • JJ_MAILBOX_REPO — path to the mailbox jj repo (defaults to current directory)
  • JJ_MAILBOX_AGENT — agent name for this instance (defaults to hostname)
  • JJ_MAILBOX_INTERVAL — sync loop interval in seconds (defaults to 30, only used by jj-mailbox sync)
• Credentials and network access:
  • Local-only (single machine, multiple agents): no network credentials needed — agents share the same repo on disk.
  • Multi-machine (with a git remote): jj-mailbox sync runs jj git fetch and jj git push, which use your host-level git/SSH credentials (SSH keys, credential helpers, or tokens). Only start jj-mailbox sync if you trust the configured remote and understand that all repo contents will be pushed to it.

How It Works

• Each agent has an inbox directory: inbox/{agent-name}/new/
• To send a message, write a JSON file to the recipient's inbox
• To receive messages, read files from your own inbox
• The jj-mailbox sync command handles jj git fetch/push in a loop — this is opt-in and only needed for multi-machine setups; it never starts automatically

Your Identity

Your agent name is set by the environment variable JJ_MAILBOX_AGENT. Your mailbox repo is at the path set by JJ_MAILBOX_REPO.

Sending a Message

Use the jj-mailbox CLI:

jj-mailbox send <recipient> "<subject>" "<body>"

Or write the file directly:

cat > inbox/<recipient>/new/$(date -u +%Y-%m-%dT%H-%M-%SZ)_${JJ_MAILBOX_AGENT}_msg-$(head -c4 /dev/urandom | xxd -p).json <<EOF
{
  "version": "0.1",
  "id": "msg-$(head -c4 /dev/urandom | xxd -p)",
  "timestamp": "$(date -u +%Y-%m-%dT%H:%M:%SZ)",
  "from": "${JJ_MAILBOX_AGENT}",
  "to": "<recipient>",
  "type": "message",
  "subject": "<subject>",
  "body": "<body>",
  "refs": [],
  "metadata": {}
}
EOF

Checking Your Inbox

jj-mailbox inbox

Or read files directly from inbox/${JJ_MAILBOX_AGENT}/new/ (sorted by filename = sorted by time).

Processing a Message

After reading a message, move it to processed/:

mv inbox/${JJ_MAILBOX_AGENT}/new/<filename>.json inbox/${JJ_MAILBOX_AGENT}/processed/

Seeing Other Agents

jj-mailbox status

Or check agents/ directory — each subdirectory is an agent with profile.json and status.json.

Shared Space

Write to shared/ for content all agents can access:

• shared/tasks/ — shared task board
• shared/knowledge/ — shared knowledge base
• shared/artifacts/ — shared outputs (files, reports, etc.)

Rules

1. Never modify another agent's processed messages — they are immutable history
2. Always include from, to, subject, body in messages
3. Use refs to link replies to original messages for threading
4. Keep messages small — for large content, write to shared/artifacts/ and reference the path
5. Check your inbox regularly — other agents may be waiting for your reply
6. Update your status in agents/{name}/status.json when starting/finishing tasks
7. Never put secrets, credentials, or sensitive data in the mailbox repo — if a git remote is configured, all repo contents may be pushed to that remote