| name | channel |
| description | Connect this agent to another Codex, Claude Code, OpenCode, or Pi agent through the shared /tmp/claude-channels named-channel file protocol. Use when the user asks to open, join, watch, poll, send on, leave, or coordinate with another agent over a channel. |
Channel
Use a shared append-only NDJSON file to talk with another agent in a separate session. This skill is compatible across Codex, Claude Code, OpenCode, and Pi because all participants share /tmp/claude-channels/<channel>.ndjson.
User Interface
The user should invoke this as a skill, for example $channel vrp gpt, or ask in
natural language to join, watch, send on, poll, or leave a channel. Do not ask
the user to run the Python helper. The helper commands below are internal
implementation details for the agent to execute.
Model
Most agent harnesses are turn-based: they only see output returned by the current tool call. Watching a channel in the background needs the harness to wake you on output or exit. In preference order:
-
Codex with exec_command.wake_on_output — stream via output wake. This is the local fork path for https://github.com/openai/codex/issues/22003. The tool may be shown as exec_command or namespaced like functions.exec_command; the wake_on_output parameter is decisive. Launch python3 <HELPER> stream <channel> <agent> with wake_on_output: true, tty: true, and a short yield_time_ms. It keeps running and prints one flushed line per peer message; Codex is re-entered by synthetic background-output user messages. No re-arm is needed. Do not also run foreground poll/listen/wait while the stream is live because they share the same cursor. Keep the returned session_id; stop it with write_stdin Ctrl-C ("\u0003") if the user leaves before the stream exits.
-
Claude Code — wait via Bash(run_in_background: true), or stream via Monitor. Two output/exit-wake paths, both zero-token while idle (kqueue, no inference):
wait (default) blocks until a peer message lands, then exits — Claude Code re-invokes you with it. Handle it, re-launch wait. Shows as a background-job pill. Most robust under heavy traffic.stream via the Monitor tool (Monitor running python3 <HELPER> stream <channel> <agent>, persistent: true) wakes you inline on every message with no re-arm — Claude Code's Monitor is an output-line-wake tool, the same shape as OpenCode's monitor. Lower latency on rapid exchanges and saves the per-message re-arm call. stream self-exits when a peer leaves; stop early with TaskStop. Caveat: Monitor auto-stops a firehose, so prefer wait if the channel is very high-volume.
-
OpenCode or Pi with background tools — monitor / bash_background. monitor running python3 <HELPER> stream <channel> <agent> wakes inline on every message, no re-arm. bash_background running python3 <HELPER> wait <channel> <agent> --timeout 0 wakes once on exit; re-arm after each wake.
-
Foreground bounded listen (portable fallback). If your harness does not support output wake or background-exit wake, use listen --timeout 30: read what returns, act on it, then listen again. This is the stock Codex fallback when wake_on_output is absent. Each listen occupies the current model turn, so only use it when actively waiting with the user present, or when background wake-up is unavailable.
While joined to a live channel, treat listening as active work. With foreground listen, do not stop after one empty poll if the user is waiting for the peer; run another listen --timeout 30 unless the user asks you to stop, the peer leaves, or the channel task is clearly complete. With background wait or stream, end the turn after arming; the harness will bring you back on the next message.
Do not use watch-start for a normal "watch in the background" request. watch-start is log/desktop-notification only and does not put messages into model context. Use it only if the user explicitly asks for desktop/log-only monitoring or accepts that the agent will not be re-entered.
Shell variables often do not persist between tool calls. Resolve the channel name and your agent name once, tell the user which name you adopted, then pass those literal values to every command.
Arguments
Infer these from the user's request:
channel: required. If missing, ask the user for the channel name and stop.agent name: optional. If missing, choose a readable unique name.
Recommended fallback names:
- Codex:
codex-<cwd-basename>-<short-random> (use the helper: python3 <HELPER> name)
- Claude Code: session name if available, else
<cwd-basename>-<CLAUDE_CODE_SESSION_ID[0:4]>
- OpenCode/Pi: WARNING — these harnesses may derive a title from the user's prompt text, so two agents receiving the same prompt (e.g. "watch channel tst") can adopt the same name and silently lose each other's messages. Do NOT use the raw session title as the agent name. Use the helper (
channel.py name) which always generates a unique random suffix, or fall back to a per-session slug.
Both agents must use different names. If a generated name might collide, ask the user for an explicit name.
Fork/collision safety + newer flags. setup stamps a per-session instance-id and, if a different live session already holds the requested name (e.g. a forked session that inherited it), prints a WARNING and auto-adopts a unique name — use the name setup prints. Two instances that share a session id AND the CLAUDE_CODE_CHILD_SESSION marker still need a unique name or a distinct CLAUDE_CHANNEL_IID. To send a message containing shell metacharacters (backticks, parens, globs, $), use send … --stdin and pipe/heredoc the body (printf '%s' "$msg" | … send ch me --stdin) so the caller's shell can't execute them; a lone -/--stdin must be the only token or it errors. On a busy channel where peers come and go, pass --stay to wait/stream so one peer leaving doesn't stop your watch.
Agent-Internal Helper
<HELPER> is the bundled scripts/channel.py shipped alongside this SKILL.md.
Resolve its absolute path once and pass that literal path to every command
below (shell variables do not persist between tool calls, so do not rely on an
exported $HELPER). As a Claude Code plugin the script is at
${CLAUDE_PLUGIN_ROOT}/skills/channel/scripts/channel.py; as a plain skill it is
the scripts/channel.py next to this file.
Use the bundled helper internally when available:
python3 <HELPER> <command> <channel> <agent> [args...]
Commands:
python3 <HELPER> name
python3 <HELPER> setup <channel> <agent>
python3 <HELPER> send <channel> <agent> "hello"
python3 <HELPER> history <channel> <agent>
python3 <HELPER> poll <channel> <agent> --timeout 30
python3 <HELPER> listen <channel> <agent> --timeout 30
python3 <HELPER> wait <channel> <agent>
python3 <HELPER> stream <channel> <agent>
python3 <HELPER> watch-start <channel> <agent>
python3 <HELPER> watch-status <channel> <agent>
python3 <HELPER> watch-log <channel> <agent> --lines 20
python3 <HELPER> watch-stop <channel> <agent>
python3 <HELPER> leave <channel> <agent>
listen defaults to 30 seconds. Re-run it after timeout while still waiting for a peer. The helper prints peer messages as [from] text, skips the current agent's messages, and advances a durable per-agent cursor.
stream is the persistent output-wake path for Codex wake_on_output, Claude Code's Monitor
tool, and OpenCode/Pi monitor. It shares the SAME durable cursor as poll/listen/wait, so
do not run foreground receives while a stream is live. It exits when a peer leaves the channel.
wait is the background-exit wake-up path and shares the SAME durable cursor as
poll/listen (so nothing is seen twice across modes). On macOS it uses
filesystem events rather than fixed-interval polling. Run it as a background
command — in Claude Code, Bash with run_in_background: true. It blocks with
zero model inference until a peer message arrives, prints it, and exits; the
harness then re-invokes you with that output. Handle the message and launch a
fresh wait to keep watching. It exits with a re-arm marker after --timeout
idle seconds (default 1800) as a heartbeat; pass --timeout 0 to block forever.
Add --desktop for a macOS notification too. Because wait advances the shared
cursor, do NOT run a foreground poll/listen while a background wait is
live — they'd steal lines from each other; wait exits the moment it delivers,
so once you're woken there's no lingering process.
watch-start is a DIFFERENT, older daemon: it runs forever, only logs to
/tmp/claude-channels/<channel>.<agent>.watch.log + posts desktop
notifications, and never wakes the agent (you'd have to watch-log/poll later,
which costs a turn). Prefer stream, monitor, or background wait when your
harness supports output or background-exit wake-up. Use watch-start only for
explicit log/desktop-only monitoring. It uses a separate
watch.cursor, so it won't disturb the wait/listen cursor.
Workflow
When joining:
- Confirm the channel name and choose an agent name.
- Run
setup.
- Send
hello.
- Run
history once and summarize existing peer messages.
- Start watching. In Codex, if
exec_command.wake_on_output exists, launch
stream with wake_on_output: true and end the turn. In Claude Code, launch
wait with run_in_background: true (or stream under the Monitor tool for
re-arm-free watching) and end the turn. In OpenCode/Pi, prefer
monitor+stream, then bash_background+wait --timeout 0. Otherwise fall
back to foreground listen; do not substitute watch-start unless explicitly
requested as log-only monitoring.
- Continue turn by turn:
- When the user gives a message,
send it. If a persistent stream is live,
end the turn; otherwise launch a fresh background wait or foreground
listen.
- When Codex background-output wakes from
stream, show the peer messages
from Output: and respond as requested. The stream stays armed; do not
re-arm it.
- When
wait/listen returns peer messages, show them to the user, respond
as requested, then re-arm wait if using background-exit mode.
- With foreground
listen, if it times out and the user is still waiting,
run listen again; before answering "no response", check once more.
- If a peer message says
left the channel (the wait output shows
a peer left the channel), report that the peer left and stop watching —
do not re-arm.
To watch with zero token burn between turns, use the host's real wake mechanism:
Codex stream with wake_on_output, OpenCode/Pi monitor/stream, or background
wait for Claude Code and other background-exit hosts. watch-start is only a
legacy log/desktop notifier and never wakes the model on its own.
Claude Code Notes
Claude Code sessions may use a SessionEnd hook to announce departure. If that hook exists, setup should persist a sidecar with channel and name:
printf '%s\n' "{\"file\":\"/tmp/claude-channels/<channel>.ndjson\",\"me\":\"<agent>\"}" > "/tmp/claude-channels/.session-$CLAUDE_CODE_SESSION_ID"
If running without the helper, preserve the same wire format and cursor semantics described below.
Leaving
Treat these user messages as leave commands: leave, leave the channel, exit, quit, stop watching, /leave, /exit, /quit, disconnect, close the channel, done, bye, goodbye.
On a leave command, run leave, report that you left, and stop polling.
Protocol
The shared transcript is:
/tmp/claude-channels/<channel>.ndjson
Each line is a single JSON object:
{"from":"agent-name","ts":1234567890,"text":"message text"}
Each agent cursor is:
/tmp/claude-channels/<channel>.<agent>.cursor
The cursor stores the last line number processed by that agent. Advance it past all seen lines, including self messages. If the channel file is reset and total lines are less than the cursor, restart from line 0.
Keep channel messages concise and single-purpose. For long code, summaries, or diffs, send a short description and let the user decide whether to relay details.
