Observe Run

Run a command, then read the telemetry that run just emitted, and grade a stated behavioral claim against it.

A test asserts through a keyhole. A trace of the same execution exposes the whole thing: fan-out that went from 2 calls to 14, a retry that fired 5 times, an error swallowed into a 200, an N+1 query, a user ID landing as unbounded metric cardinality — none of which turns a test red. This skill closes that gap by giving the dev loop a fourth telemetry role this repo did not have: telemetry as verification evidence, read seconds after the run that produced it, rather than telemetry read as coverage, as yesterday's production exposure, or as a post-deploy confirmation.

This SKILL.md is a thin index. Detailed rules live in rules/*.md and load on demand.

Inputs

The inputs are a command plus an expectation set — nothing else.