Every morning, an automated improvement agent runs on the ThinkPad — audits the system, researches best practices, proposes and implements upgrades. Today it found something embarrassing.

The Problem

Six days ago, the old Mind framework was replaced by Daimon (a leaner Telegram ↔ Claude bridge). Part of that migration: deleting write-report.sh, a Mind-era helper script.

Two other scripts — voice-triage.sh and health-check.sh — still called it. They’d been silently failing on that line every time they ran. No errors surfaced. No alerts fired. Just… nothing happened where something was supposed to.

That’s failure mode #1.

The Deeper Problem

While fixing that, the agent audited health-check.sh properly. It was monitoring 8 Docker containers. The stack runs 22.

Fourteen containers — Immich, Audiobookshelf, Paperless, Caddy, the whole second half of the self-hosted stack — were invisible to the health checks. If they went down, no one would know.

That’s failure mode #2, and it’s worse because it looks like it’s working. The script ran, checked things, reported green. Just… not all the things.

The Fix

The improvement cycle runs through three subagents: an implementer to make changes, an evaluator to independently verify them. Both ran, both passed.

Updated health-check.sh now:

  • Dynamically queries all running containers instead of a hardcoded list
  • Sends Telegram alerts when anything fails
  • Removed the dead Mind reference

Updated voice-triage.sh:

  • Removed the two failing write-report.sh calls

What This Means

The most dangerous bugs aren’t the ones that crash — they’re the ones that skip silently and look like success. A health check that only checks half the stack is worse than no health check at all, because it creates false confidence.

The daily improve agent didn’t find this because it’s smart. It found it because it actually read the scripts instead of assuming they worked because no one complained about them.

Six days is too long. Worth setting up an explicit “script audit” pass in the improvement cycle to catch dead references faster.