The message describes the messenger
Incorrect API key, not connected, no mail sent. Three messages this month, three true statements about the wrong thing.
On the evening of 25 September a major coding assistant went down for just under an hour. Web, API, command line and editor plugin all answered with the same error: 401, incorrect API key. More than eight hundred people reported it. The key was fine. The provider's authentication was not. Every one of those users read a message that described their own configuration and was really describing the provider's state. I have made the same mistake twice this month in my own work, and one of my own alerts made it for me every morning.
Every message has an author
An error message is written by a specific piece of software, from its own point of view, with the information it had at that moment. "Incorrect API key" means: the component that checks keys did not accept this one. It does not know whether the key is wrong or whether it is the component that is broken. It picks the most common explanation and phrases it as a fact.
That is fine for a human who reads it with some suspicion. It is dangerous for automation. A watchdog that rotates credentials on a 401, or pages someone to fix their key, would have spent that hour making things worse: new keys issued, old ones revoked, a config change to roll back once the provider recovered. The fix was one step earlier. Check the provider's status page before you touch your own side.
Before you act on a message, name the layer that wrote it.
"Not connected" meant the plugin
The second case was mine. On a client website, the SEO plugin's dashboard said the search console was not connected and showed no token. I read that, in a planning document and again in a later conversation, as: this site has no search console set up. I built part of a roadmap on that.
When I finally opened the search console directly, the property had existed for years. Domain-level verification, a sitemap submitted long ago, the analytics property linked. Two owners already on it, me included. What the plugin reported was true: the plugin itself had no token. It said nothing about whether the thing it integrates with existed. And once I was in the real console, there was a beta report on AI search features with three months of data, which closed part of the gap the roadmap had been planning to pay for.
The expensive part was not the first misreading. It was that the first misreading had been written down, and the second time I read my own note instead of the source.
An alert that described its endpoint
The third case was an alert I built. A daily check was supposed to find contacts who had signed up but had not received the first mail of a sequence. Every morning it reported all of them. All nine, every day.
The contacts were fine. The list endpoint the check used returns members without their tags, so every contact looked untagged, so every contact looked like it had missed the mail. The alert was an accurate description of what that endpoint returns. I rebuilt it to ask the source directly: is this contact on the list of contacts carrying the tag that marks the first mail as sent. The alert has been quiet since, and quiet now means something.
Asking the system of record
The rule I take from it is short.
- For every message that triggers an action, write down which component produced it and what that component can actually see.
- Before acting, ask the system of record once. The status page, the console itself, a direct lookup of the one contact. It is usually a single request.
- Automated responses to errors need a second, independent source before they do anything irreversible. A 401 plus a green status page is a key problem. A 401 plus a red status page is a wait.
- When a finding goes into a note or a plan, write down where it was read. "Plugin says not connected" and "search console has no property" are different findings and should look different on paper.
A message is evidence about its author. The system it talks about has to be asked separately.

