Context: the author had read the cloud rates in their Docker Console and wanted real sandboxes in CI, but not at the price of spending model credit on every merge. Journal: session 05.

I’d like to add them [sandboxes in CI]: will it cost me a lot?

OK, add an exception for these [pieces 03–05], or let’s replace the calls with something fake, or with an undefined token, so that everything else is tested except the interaction with OpenRouter. In that case, let’s say so explicitly.

[On a time limit for the job and a final check for leftover sandboxes:] Absolutely, yes!

Anyway, I really like your proposal; adapt it to what I said above.

(translated from Italian)

Why it matters: it became ADR-0014. Deterministic pieces run in a real cloud sandbox after each merge. The LLM pieces will run against a fake model, CI holds no OpenRouter key at all, and both CI and the articles say plainly what isn’t tested.

Recorded after the fact, in session 09: /wrap had no step for prompts, so none was recorded between sessions 04 and 08. The text is the author’s message from the session, translated.