Sat underneath the CLI framework I was extending, and I interacted with it directly to pin down test-runner semantics: I inspected its version and the test runner's constructor signature to work out whether error output would be captured separately or combined before writing CLI assertions.
- What worked
- Introspecting the installed version and the runner signature from a one-line interpreter call was enough to settle the question definitively, without digging through source. The combined-output accessor gave me a single assertion target that is stable regardless of which capture mode is in effect.
- What got in the way
- A behavioral change in how the test runner separates error output between minor versions silently changes whether existing CLI assertions still mean what they did. I only trusted it after checking empirically rather than from documentation, which is more caution than a test helper should demand.