# bot-keepalive
**Keep your [Grok Bot](https://x.ai/bot) cloud computer awake while long jobs run, for the fewest tokens possible.**
One low-effort subagent, one tool call every ~13 minutes, about **1.8M tokens a day** in our own use. No more jobs dying every time the computer dozes off.
[中文说明](#中文说明)
---
## The problem
Grok Bot puts its cloud computer to sleep after about 5 minutes of "idle". But idle is defined loosely: if no agent is in a conversation and you don't have the remote desktop open, the computer counts as idle, **even if it's running a pile of background jobs**. The whole machine freezes, `nohup` and `tmux` included.
That breaks anything long-running: trading bots, monitors, server-fleet management, ML training and scheduling.
**Why routines don't fix it**
- The shortest routine interval is 5 minutes, the same as the sleep timer.
- Routines don't fire on time. Daily ones can drift 15 to 30 minutes at busy hours.
- So even with a 5-minute routine, the computer keeps falling asleep and freezing for minutes at a time. Jobs die, and restarting them costs more tokens.
- Each routine run is a fresh bot that re-reads the task context and takes several turns before it finishes.
## The fix
A bot turn that is **waiting on a tool call counts as active**, and a single Shell call can block for up to 13 minutes. So:
1. Start your jobs detached (`setsid nohup …`) and register each one in `~/.keepalive/jobs/`.
2. Run **one** global watcher subagent at **low effort**, outside your main chat so it never blocks you.
3. Each turn, the watcher makes one Shell call: a bash loop that checks every 5 seconds whether the jobs are done, for about 12 minutes, then returns.
4. Repeat until every job is done or a hard deadline passes. A lock file makes sure only one watcher runs per account, no matter how many bots start jobs.
5. Grok Bot stops background subagents after 50 minutes, so the watcher runs as a relay: each leg does at most 3 waits (about 39 minutes), exits cleanly, and the main chat starts the next leg.
Jobs on remote compute (Kaggle, Colab) and waits for a quota reset don't need a watcher at all: a one-off routine at the expected finish time, plus an hourly fallback routine that deletes itself, is enough. For jobs over 2 hours, prefer that or remote compute, since every relay leg also wakes the main chat.
The 5-second polling happens inside bash, not in the model. The agent only wakes once per ~13 minutes, for one tool call.
## Cost
| | Routine every 5 min | bot-keepalive |
|---|---|---|
| Keeps the computer awake | No, still freezes between runs | Yes |
| Timing | Drifts, up to 15 to 30 min | Steady, ~13 min per turn |
| Work per wake-up | New bot, re-reads context, several turns | One turn, one tool call |
| Model effort | Whatever the routine uses | Low |
| Measured cost | More | ~1.8M tokens/day |
For a paid plan, that's small change.
## Install
Send your Grok Bot this link and ask it to save it as a skill:
```
https://github.com/realTaki/bot-keepalive/blob/main/skills/bot-keepalive/SKILL.md
```
Every bot on your account can then use it. Next time you start a long job, tell the bot to "keep the computer awake with bot-keepalive until it finishes".
The full method, loop and rules are in [`SKILL.md`](skills/bot-keepalive/SKILL.md).
---
## 中文说明
**Grok Bot 云电脑防休眠，最省 Token 的做法。** 一个 low effort 的 subagent，大约每 13 分钟一次 tool call，我们自己实测一天额外开销约 180 万 Token 额度。
**问题。** Grok Bot 闲置约 5 分钟就休眠，但“闲置”定义得太宽：没有 agent 在对话、你也没开远程桌面，就算闲置，哪怕后台开着一堆任务在跑。整台机器冻住，`nohup`、`tmux` 也一起停。打狗、监控、管服务器集群、ML 任务监控和调度都很难用。
**Routine 为什么不行。** Routine 最短 5 分钟一次，跟休眠时间一样；而且不准时，高峰期误差能到 15 到 30 分钟。所以就算 5 分钟一次，电脑还是反复休眠，任务挂掉，修复重启还要额外花 Token。每次 Routine 都是一个新 bot，要从头读上下文、跑好几个 turn 才结束。
**解法。** 等待 tool call 返回的 turn 算活跃，单次 Shell 调用最长能阻塞 13 分钟。所以长任务脱离会话启动并登记到 `~/.keepalive/jobs/`，再开**一个**全局 low effort 守护 subagent（不阻塞你正常对话）：每个 turn 只执行一条 bash 命令，在命令里每 5 秒检查一次任务，约 12 分钟返回一次，直到全部完成或到截止时间。锁文件保证整个账号只有一个守护。后台 subagent 最多跑 50 分钟就会被平台停掉，所以守护按接力跑：每一棒最多 3 轮（约 39 分钟），干净退出，主会话开下一棒。Kaggle、Colab 这类远端任务和等额度重置不用守护，用预计完成时刻的一次性 Routine 加每小时自删的兜底 Routine 盯就行；超过 2 小时的任务优先这么做。
**安装。** 把上面那个 `SKILL.md` 链接发给你的 Grok Bot，让它存成技能就行。
## License
MIT