inter-agent-chat

Inter-Agent Terminal Chat

Send messages directly into another agent's Claude Code prompt via TIOCSTI terminal injection. Messages enter the receiving agent's context window as user input.

Discovery — PTS Registry

Registration is automatic via SessionStart hook. Do NOT register manually (tty doesn't work inside Claude Code).

Registry: /tmp/agent-pts/<agentName> — one file per agent containing its /dev/pts/N.

Sending

Inject characters into target's terminal input buffer using TIOCSTI ioctl. Only change the target and message:

sudo python3 -c "
import fcntl, termios, os, time
msg = '[@agentB] your message here. reply via inter-agent-chat skill.\r'
fd = os.open('$(cat /tmp/agent-pts/agentB)', os.O_WRONLY)
for c in msg:
    fcntl.ioctl(fd, termios.TIOCSTI, c.encode())
    time.sleep(0) # must yield to avoid paste detection
os.close(fd)
"

On the first message to an agent, end with "reply via inter-agent-chat skill." so they use this skill to reply. After that, just [@agentB] message is enough.

\r submits. \n does NOT submit (just adds a blank line). Always end with \r.

time.sleep(0) is required — without a thread yield between characters, the terminal detects rapid input as a paste and the message won't submit. sleep(0) adds negligible overhead and has been tested reliably up to 32K characters.

For broadcast: loop over /tmp/agent-pts/*, skip yourself.

Message Format

[@agentB] message               # single recipient
[@agentA @agentC] message       # multi-recipient (send to each listed)
[@all] message                  # broadcast to all agents
[@user] I need guidance         # flag for human attention
[@stop]                         # kill switch — stop all inter-agent chatter

Receiving

When you see [@yourName] or [@all] as input:

1. Recognize it as an inter-agent message, not a user command
2. Process the content — answer questions, act on directives
3. Reply if needed — use this skill to inject into the SENDER's pts (not your own)
4. [@stop] — stop immediately. Do not reply. Return to user-directed work.
5. [@user] — for the human, not you (unless you're sending it)