The first failure is often not a red test. It is a closed door the agent never checked: forks blocked, LLM policy forbids autonomous PRs, or the template tripwire that marks the submission as a bot.
Running those checks after an afternoon of patch generation is the expensive order. The cheap order is grep the contributing docs, check cross-fork PR age, and stop before writing if the wall is already there.
At Ascendra Ventures we bake that preflight into agent delivery work. A correct diff that cannot be filed is not a deliverable. It is unfinished research.
Do you wire those policy checks as hard exits, or only as warnings in a log?