tmux-command

---

name: tmux-command description: This skill should be used when the user asks to "run a slash command in another Claude session", "trigger /clear (or /compact, /model, /context, /rewind, …) in a tmux pane", "drive built-in commands in a running agent session over tmux", "send a slash command to the Claude/codex/opencode instance in session X", or otherwise wants to execute a harness command that the agent running in that session cannot invoke itself. Distinct from tmux-talk: this sends the raw command with no message envelope.

tmux-command

Inject a harness's built-in slash command (or the keystrokes that drive its pickers) into another agent session running in a tmux pane. Built-in commands like /clear, /compact, /model, /context, /rewind, /resume are interpreted by the CLI front-end, not the model — so the agent running in that session cannot trigger them itself. An outside driver must type them into the pane. That is this skill.

This is not tmux-talk. tmux-talk wraps a message in a From:/To-reply: envelope for a human/agent reader. Here the receiver is the CLI parser, so the bytes sent must be exactly the command — there is no wrapper.

TMUXCMD=~/.claude/skills/tmux-command/bin/tmux-command

Subcommands

tmux-command list                          # discover panes: address, running cmd, title
tmux-command context <target> [N]          # gather before acting: the pane's id line + N screen lines (default 200)
tmux-command send  <target> <command...>   # type a literal command line, then submit it
tmux-command keys  <target> <key...>       # send raw keys (Enter Down Up Escape BSpace) to drive a picker
tmux-command read-screen <target> [N]      # capture N lines (default 200) — re-check / verify the effect

target is tmux's session:window.pane, or a shorthand the shared resolver expands (same grammar in /tmux-talk):

form	means
dotfiles:2.1	full address
dotfiles:2	window 2, pane defaults to 1
dotfiles:orca	window named orca (looked up; ambiguous/unknown names error)
:2.1 / :2	same session
.2	same session and window, pane 2
:self	the pane you (the driver) are running in

Every verb resolves its target through tmux-shared/bin/tmux-resolve first, so a bad, unknown, or ambiguous target fails loudly instead of tmux silently retargeting the active pane.

Flow: context → decide → send → verify

Start by running context — it returns everything you need to reply, so you never have to assemble the gather by hand:

TMUXCMD=~/.claude/skills/tmux-command/bin/tmux-command
TARGET="work:0.0"

# 1. Gather the context you need to reply — one call.
$TMUXCMD context "$TARGET"

# 2. Decide the command from what came back, then send it.
$TMUXCMD send "$TARGET" "/compact focus on the auth refactor"

# 3. Verify — a TUI returns no exit code, the pane is the only signal.
sleep 2
$TMUXCMD read-screen "$TARGET" 40

context prints the pane's cmd=/title= line and its live screen. That one read tells you the harness (→ which references/<harness>.md to open), whether it's idle or mid-generation, and which screen is up — everything that decides what to send and whether it's safe to send now. Claude Code panes report cmd=<version-number> (e.g. 2.1.175), not cmd=claude, so identify them by title= (✳ Claude Code) or the UI.

Commands that open a picker need keys

Some commands open an interactive list instead of acting immediately (/model, /permissions, /resume, /agents). send opens the dialog; drive the selection with raw keys:

$TMUXCMD send "$TARGET" "/model"        # opens the model picker
sleep 1
$TMUXCMD keys "$TARGET" Down Down Enter # move to the choice and confirm
$TMUXCMD read-screen "$TARGET" 20       # confirm the new model

Gotchas

• Verify, never assume. The target is another process with no exit code back to you. After every send, read-screen and confirm the command actually ran — especially for destructive ones (/clear, /compact, /rewind).
• Home screen swallows the first Enter. When context showed Claude Code's "describe a task" home screen, send may leave the command sitting in the input (the first Enter is rebound to "collapse"/"create"). If the post-send read-screen confirms it never submitted, send one more keys "$TARGET" Enter.
• One command per send. send appends exactly one Enter. To chain commands, call send again after verifying the first landed.

References

• references/claude-code.md — which Claude Code commands need this skill (and which the in-session agent can already invoke itself), picker commands, and safe-to-test commands
• references/codex.md, references/opencode.md — per-harness command sets