8 sessions on node:test with a coding agent: the agent failed at least once in 3, and 2 finished without a human stepping in.
Takeaway: Next time I'd start by telling the agent explicitly which ROMs I actually care about and that I want tests tied to real gameplay symptoms (freezes, missing sprites) rather than emulator-accuracy test suites like blargg/nestest.
Takeaway: Next time I'd front-load the dev-mode vs production proxy-probing distinction in the first prompt, since that's exactly the false-positive Codex caught later.
Takeaway: I'd start the next prompt by explicitly scoping whether I want just the fix or the fix-plus-refactor loop in one go, since the refactor pass doubled the session length.
Takeaway: Next time I'd front-load the exact list of window.confirm call sites in the first prompt so the agent doesn't have to rediscover them twice across two PRs.
Takeaway: Next time I'd open with the constraint I ended up stating later: reuse the existing Files UI/tool, no separate assistant-only flow — that would've saved a full exploration-and-pitch round trip.
Takeaway: Next time I'd ask upfront for the fix to include a manual verification step (actually reload and check tabs) before calling it done, since the first commit passed all checks but never actually wired the restore into the boot path.
Takeaway: I'd open with the regolamento PDF next time instead of finding it halfway through — it settled several 'is this a bug or a feature' questions I burned turns guessing at from code alone.
Takeaway: I'd start the next session the same way — running Bugbot on the branch diff before merging — since it caught a real high-severity bug I'd have missed otherwise.