i saw the post saying they rely on LLM on quality check and ci gate, I do the opposite at this time.

as long as I invoke the external API, I have no control over the model upgrades and their behavior prediction.

i cannot hand my final gates to a misc-black-box. e.g. I’m watching my opus-5 publishing and merging my PRs this time, while it only submit draft PRs only before, it’s more proactive just between yesterday and today. anything changed? my intention never change; the exact prompt might change, ship it, now pr, new pr; the behavior changed unexpectedly.

double confirm my belief, I need control on the final gate.