Tessary call sites

A call site is a place in your code where an agent with a specific function runs. That can be a single independent LLM call, or a whole agent: a prompt plus its tool calls and the loop around them. Either way it's the unit Tessary binds findings, baselines, and regressions to. You tag its spans with tessary.call_site.id, a plain span attribute, no SDK required. From then on, everything Tessary learns about that function accrues to that place in your code.

The binding matters because detection needs history. Classifiers fit a baseline per call site from its own traffic before saying anything moved, and the latency and cost watchers compare each call site against its own past. So a freshly tagged call site is quiet at first while its baseline builds. That's by design: Tessary doesn't report a deviation until it has seen what normal looks like for that call site.

The binding also localizes regressions. When quality shifts, findings arrive already tied to a specific place in the code, which gives triage and root cause analysis a concrete starting point. One practical rule: keep the ids stable across deploys. They're the key your history accrues under, and a changed id means the baseline starts over.