Why the hell would you put 50k lines into a single PR?
replyI mean, why even pretend you’re going to “review” something that large? Just build everything on main.
PR is just an entity to review /do some other LLM processes . Human part check other models review PR's, tests all kind of , security etc. Also human is to checks docs, specs in the pr, db migrations if any , some of the tests related to the PR. We stopped reading the code after opus 4.6. Sometimes for very core parts i skim through files just to make sure if the changes were correct.
replyI would imagine even LLMs would do a better job reviewing smaller PRs than very large ones.
replyBut to reiterate parent's question: why not just do those things continually at that point? Or on a calendar-based basis?
replyYeah exactly, if you have a fully automated SDLC, how do you expect the agent to code review 50k lines properly?
replyIt will take shortcuts and now the entire premise is busted. You now need to build a code review process for large PRs.
This is a common problem. Fully automated SDLC needs to start before the CI/CD.
replyMy suggestion is to have the proper chunking mechanisms and multiple specialised agents. The most important is harness engineering, what we do at dromeas.ai to verify the code that goes to prod is a)have the code mapped before hand for the right agentic context, b)chunks of the right size per model context window c)specialised agents d)deduplication and verification . All before assessing a PR, a commit, a release. Harness engineering is not easy.. Especially when supporting multi model