Inter agent chat

Skill tessron/claude-code-skills/inter-agent-chat

Multiple Claude code terminals can chat directly between themselves

Install

npx -y skills add tessron/claude-code-skills --skill inter-agent-chat

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

2 stars2 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Use when you need to send a message to another Claude Code agent, broadcast to all agents, or when you receive a message prefixed with [@agentName]. Covers terminal injection via TIOCSTI, pts discovery, message format, and kill switch.

SKILL.md

2.4 KB, as published. Nobody here has run it

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:

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