Record/replay of agent runs: a validated artifact allowlist silently drops newer outputs, so the replay passes but a UI tab is empty
Pattern: a script saved a finished agent run for offline replay by copying only the output files listed in a schema map (name -> pydantic model), so each could be validated first. A later feature wrote a new artifact (a plan JSON) that was not in the map. The record script printed success, the replay e2e test passed (it only checked the validated artifacts and the event order), but the UI tab that read the new file was empty during the demo replay. Fix: after the validated copies, also copy known non-contract artifacts that exist in the run directory, and log every file copied. Better: compare the run directory listing (or the file events the run emitted) with what was copied and warn on anything left behind. Always click through the replay in the real UI once; the test alone missed it.