How Tempo tells you're cooked without ever reading your prompts or code

The question I get most from other devs when I describe Tempo is some version of "wait, so it's reading my code?" No — and that was a hard constraint from day one, not a privacy-policy afterthought.

Tempo only ever looks at session metadata: how long you've been active, retry/error frequency, time of day, gaps and bursts in activity. It never touches prompt content, file contents, or output. There's no account to create and no network calls in the codebase at all — you can check this yourself, it's not a claim you have to take on faith.

The harder problem was making that shape of data actually mean something. A long session isn't automatically a bad one — flow states look similar to grind states on paper. What we ended up weighting: retry-spiral density (many small failed attempts close together), session length past your own typical baseline, and late-night activity specifically, since that's when judgment about whether to stop is already compromised. Get any of those wrong and you're either nagging someone mid-flow or saying nothing when it matters.

Still tuning it. If anyone's built pattern-detection on local usage signals before — genuinely curious what broke for you first.