routine-healthcheck · diff
git:20261010.3ba1fbe to git:20261113.8af5195
30 added, 23 removed. Audit A to A.
What this update changes in the checks
No check changed: every fixed check gives the same answer before and after this update.
A tripped check is a label, not a verdict: read the matched text and the lines below. This week's updates that newly trip a check.
---
name: routine-healthcheck
description: >-
Use when auditing Grok Bot routines for token waste, offering a fleet checkup
after template import, deciding cadence, or running the standing
routine-healthcheck wake.
---
# Routine healthcheck
- Scan the fleet's scheduled and event routines for token waste. Report what to change. Do not edit another bot's routines unless the user says to.
-
- ## When to run
+ Audit every bot's scheduled and event routines for wasted wakes and noise, then propose fixes. Report only. Don't change another bot's routines until the user picks which fixes to apply.
- - first run after this template is imported
- - the standing `routine-healthcheck` routine fires
- - the user asks for a routine checkup, token audit, or cheaper wakes
- - before creating a new standing routine
+ ## When
- On first import, offer once: "want a checkup of every bot's routines?" Run only if they say yes. Skip if they decline.
+ - The standing `routine-healthcheck` routine fires (default Monday morning).
+ - First run after import: offer once ("want a checkup of every bot's routines?") and run only on yes.
+ - The user asks for a routine audit, a token check, or cheaper wakes.
+ - Before you create a new standing routine.
- After a checkup (first import or on demand), offer once to create a standing routine named `routine-healthcheck` if one does not already exist. Default cadence is weekly (Monday morning). Ask only if they want a different time. Do not ask again every run.
+ After the first checkup, offer once to create the weekly `routine-healthcheck` routine if it doesn't exist. Don't re-ask on later runs.
- ## How to scan
+ ## Collect
- 1. List every agent folder under `/home/box/agent-data/agents/`.
- 2. For each agent, read `profile.json` (name) and every `automations/*/automation.json`.
- 3. Note schedule vs event trigger, enabled vs paused, and the prompt's quiet-when-nothing rule.
+ 1. List every folder in `/home/box/agent-data/agents/`.
+ 2. For each: the bot name from `profile.json`, then every `automations/*/automation.json`.
+ 3. Per routine, note: cron or event, enabled or paused, runs per day (work it out from the cron), and whether the prompt says to stay silent when nothing changed.
- ## Flags
+ ## Flag
- Call out each hit with bot name, routine name, and one fix line.
+ Each flag names the bot, the routine, the evidence, and a one-line fix.
- - **Too frequent.** Cron denser than hourly (for example every 15 minutes ≈ 96 runs/day). Prefer hourly, a few times a day, or weekday waking hours. Prefer an event listener when the trigger is Slack, GitHub, or similar.
- - **Recurring on a long chat.** A standing digest, sweep, or poll lives on a bot whose transcript is already long. Move that job to a fresh bot with a short chat. Keep talking to the main bot.
- - **Noisy empty runs.** Prompt never says to stay quiet when nothing changed. Add that. Filler like "(no change.)" still costs a wake.
- - **Dead or duplicate.** Paused forever, same job on two bots, or a finite watch that already passed its end condition.
+ - **Too often.** Anything denser than hourly (every 15 minutes is 96 wakes a day). Suggest hourly, a few fixed times, or weekday working hours. If it's watching Slack, GitHub, or mail, suggest an event trigger instead of a poll.
+ - **Talks when it has nothing.** No silence rule, or it sends filler like "(no change)". A message the user can't act on is noise, and noise trains people to ignore the real ones.
+ - **Heavy host.** A frequent routine on a bot whose chat is already long, so every wake drags that history along. Move it to a fresh, small bot.
+ - **Dead weight.** Paused for weeks, finished its job (a watch past its end date), or two bots doing the same job.
+ - **Never acted on.** If the user hasn't responded to or used a routine's output in weeks, propose pausing it.
## Report
- One short list. Bot, routine, flag, suggested fix. No essay. Offer to apply fixes only after they pick which ones.
+ One list, worst first: bot, routine, flag, fix, and estimated wakes saved per week. Add a total. No essay. If everything is fine on a scheduled run, say nothing.
- ## When you create a bot
+ Offer to apply fixes only after the user picks. When creating a new bot for a standing sweep, give it one routine at the slowest schedule that still works.
- If the job is a standing sweep, give it its own bot and one routine at the coarsest useful schedule. Do not hang that routine on a bot whose chat is already long.
+ ## Inspired by
- Credit: [Lauren Tan (@poteto), dr eggbot](https://x.ai/bot/marketplace/bots/dr-eggbot-v2) names this healthcheck and its weekly token/waste audit. See the [source audit](../../../docs/credits.md) for the evidence and license limits.
+ - [Lauren Tan (@poteto), dr eggbot](https://x.ai/bot/marketplace/bots/dr-eggbot-v2) (marketplace template; no reuse license shown, so no text is copied). Does well: names a weekly token and waste audit that stays quiet and waits for the user to pick. Missing: the listing gives the schedule and intent but no flags, cost estimate, or report shape.
+ - [Google SRE Book, Monitoring Distributed Systems](https://sre.google/sre-book/monitoring-distributed-systems/). Does well: every page should be actionable, noisy alerts get ignored, and rarely used rules should be up for removal. Missing: about production alerting, not agent wakes or token cost.
+ - [Google SRE Book, Eliminating Toil](https://sre.google/sre-book/eliminating-toil/). Does well: defines repetitive, automatable work with no lasting value. Missing: no audit procedure for scheduled jobs.
+
+ ## What this version adds
+
+ - Wakes-per-day math and a weekly savings total in the report.
+ - The "actionable or silent" test and a "never acted on" flag, taken from alerting practice.
+ - Event triggers suggested in place of polling.