What it does
- Picks wake, tell or log for problems that are not floors.
- Writes what broke, why, and a plan, with a confidence.
- Proposes a fix from the kinds your rules allow.
- Writes the plain-words summary you see on the problem page and in the weekly email.
What it sees
Up to 5 scrubbed samples per problem, the route and status, counts over time and recent deploy markers. Never your stored credentials.
Hard limits
- Cannot demote a floor; floors are decided before it runs.
- Error text is fenced as untrusted data. Instructions inside it are ignored.
- Cannot propose a data change or a kind outside your rules.
- Its words are never read aloud in a phone call.
- Unsure, slow or failed means tell, never log.
Model and allowance
Claude by Anthropic. Data is scrubbed before it is sent and is not used for training. Each account has a monthly AI allowance; when it runs out, problems are still sent to you, without the diagnosis, until the month resets.