tmux

name	tmux
description	Inspect tmux panes, survey which panes are active, capture pane output, and send text or key presses to a tmux target. Use when the user wants to monitor a tmux session, identify busy panes, read pane content, or message a pane.
triggers	{"intent_patterns":["tmux|pane|session|监控 pane|发送到 pane|抓取 pane","都在干嘛|谁在跑|哪个在跑|哪些在跑|看下 session|看下 pane","盯着|持续看|有变化告诉我|watch 一下|监控一下","给 .* 发消息|给 .* 发 continue|发到 pane|发回车|按 enter|按 ctrl-c|发送按键","抓最近输出|看最近日志|看最近内容|看详细输出|展开看"],"context_signals":{"keywords":["tmux","pane","session","capture-pane","send-keys","watch","continue","Enter","C-c"]},"confidence_threshold":0.8}
priority	8
requires_tools	["bash"]
max_tokens	280
cooldown	30

tmux

Use this skill when the target terminal lives inside tmux. Prefer it over desktop UI automation because tmux exposes pane text directly and supports stable targeting.

Quick Reference

Intent	Command	Key Params
List panes	$PYTHON $SKILL_DIR/run.py list_panes	none
Survey pane activity	$PYTHON $SKILL_DIR/run.py survey --session 7 --lines 8	--session, --lines
Inspect one pane	$PYTHON $SKILL_DIR/run.py inspect --target %5 --lines 20	--target, --lines
Watch for changes	$PYTHON $SKILL_DIR/run.py watch --target %5 --ticks 6 --interval 2 --lines 6 --silence-secs 4	--target or --session, --ticks, --interval, --lines, --silence-secs
Capture pane text	$PYTHON $SKILL_DIR/run.py capture --target 5:1.1 --lines 80	--target, --lines
Send command and read feedback	$PYTHON $SKILL_DIR/run.py send_command --target %5 --text 'cargo test' --lines 12	--target, --text, --keys, --wait_secs, --lines
Send literal text	$PYTHON $SKILL_DIR/run.py send_text --target %5 --text 'cargo test'	--target, --text, --no-enter
Send tmux keys	$PYTHON $SKILL_DIR/run.py send_keys --target %5 --keys C-c,Enter --repeat 2	--target, repeated --keys, --repeat

Workflow

1. Run survey first when the current information is insufficient; it returns summary plus grouped running, idle, and dead panes. Running panes include direct content_lines so you do not need a second inspect just to see what they are doing.
2. Use inspect on the most relevant pane to get real foreground process, recent activity age, and raw preview.
3. Use watch when you need a bounded polling loop that only reports meaningful changes across a pane or session. It now emits compact events with changed and new_lines.
4. Use capture when you need a larger raw scrollback.
5. Use send_command when the user wants to run something in a pane and expects immediate feedback; it types the command, presses launch keys, waits briefly, then returns a compact pane summary.
6. Use send_text only when you intentionally want to type without necessarily executing yet.
7. Use send_keys for pure control sequences such as Enter, C-c, or repeated key presses.

When to Trigger

Trigger this skill as soon as the user is asking about a tmux-hosted terminal, even if they do not say tmux explicitly but the context already points to pane/session work.

• 状态判断: "都在活动吗", "谁在跑", "看下 agent-infer 里哪个 codex 还活着"
• 内容理解: "这些 pane 现在在干嘛", "抓 %7 最近输出", "把 running 的内容总结一下"
• 变化监控: "盯一下 %9", "有变化告诉我", "watch 一下这个 session"
• 发消息/按键: "给 %7 发 continue", "往 pane 里发 cargo test", "按一下 Enter/C-c"
• 运行命令并看结果: "在 %7 里跑 cargo test 然后告诉我结果", "给 pane 发命令并回我反馈"
• 二次确认: 用户先给了 pane 列表或 TTY 信息，接着问 "现在怎样了"、"继续看"

Do not trigger this skill for non-tmux desktop tabs, GUI terminals without tmux, or background daemons that are not writing to a pane.

Response Shape

Default to a conclusion-first answer. Do not narrate pane-discovery mechanics unless the user explicitly asks for them.

1. Start with the answer the user actually wants: how many panes are running, which ones matter, or whether it is safe to message.
2. For each relevant running pane, summarize what it is doing from summary and content_lines, not from raw TTY/process bookkeeping.
3. Mention idle panes only when they are actionable, for example "safe to message".
4. Only surface raw details such as tty, pane_current_command, unmatched background processes, or invisible panes when they change the decision.

Good:

• "agent-infer 里现在明确在跑的是 %7 和 %9。%7 在做 scheduler review，%9 还在跑 benchmark。"
• "%4 现在是 idle shell，可以直接发消息。"

Bad:

• "我看到 ttys000 还有一组 node，%6 之前没了，pane_current_command 是 ..."
• "先列了 pane，再比对了 TTY，再看了 ps ..."

Next-Step Guidance

After using this skill, always end with the single most useful next step instead of a generic "要不要继续看".

• If survey shows running panes, suggest one of:
  • "可以直接抓 %7 最近 20 行详细输出。"
  • "可以只盯 %7 / %9，有变化再汇报。"
• If the user needs more context on one pane, move to inspect.
• If the user needs ongoing observation, move to watch.
• If a pane is idle and actionable, suggest:
  • "可以直接给 %4 发 continue。"
  • "可以发 Enter / C-c 到 %4。"
• If the user wants to run a command and know what happened, prefer:
  • "可以直接发命令到 %7，按 Enter 启动，然后把最新反馈抓回来。"
• If the pane is active, avoid suggesting messages unless the user explicitly wants to interrupt it.

Preferred follow-up prompts:

• "%7 还在跑 scheduler review。可以直接抓它最近 20 行，或者继续 watch 它的变化。"
• "%4 现在空闲，可以直接发命令进去。"
• "%9 还在忙，不建议现在打断；如果你要，我可以只盯它到空闲为止。"

Target Forms

• session:window.pane such as 5:1.1
• %pane_id such as %5
• Positional forms also work for quick use, for example send_keys %5 Enter or send_text %5 cargo test

Notes

• survey uses tmux metadata plus the pane TTY foreground process from ps, which is usually more accurate than pane_current_command alone.
• inspect and survey return a compact pane view by default. Raw pane metadata stays available through list_panes; inspect still includes preview.
• Prefer survey for user-facing answers: it already groups panes by state and gives a top-level summary, so avoid repeating raw process bookkeeping unless the user asked for it.
• watch is intentionally finite. It polls and returns initial, events, final, and summary; it does not run as a background daemon.
• watch --silence-secs N stops early after N seconds with no meaningful change, which is useful when a pane settles back into a prompt or a long-running task goes quiet.
• send_command is the default choice for "run this in the pane and tell me what happened". It sends the text, launches it with Enter by default, then returns feedback from inspect.
• For send_text and send_command, embedded newlines in --text are treated as real Enter key presses. Do not rely on literal \\r or \\n bytes to trigger execution inside tmux.
• send_keys accepts repeated --keys, comma-separated keys, whitespace-separated keys, and --repeat for repeated key presses.

Boundaries

• This skill can only read pane content that exists in tmux.
• It cannot inspect non-tmux terminal tabs or background daemons that do not write to the pane.
• Do not send destructive commands unless the user explicitly asked for them.