Some AI-assisted changes look reasonable locally but miss deployment configuration, pipeline stages, or dependencies outside the repository. How do you catch those gaps without making infrastructure review a constant bottleneck?
How do you catch infrastructure gaps in AI-generated pull requests?
Answers
Put the production rules somewhere developers and their tools can actually find them. A short document in the repository could describe the application topology, supported services, networking constraints, deployment process, and what a new module needs in its pipeline.
That gives both the developer and an AI tool some context before review. It won't guarantee a correct change, so someone still needs to check the diff and verify that the proposed dependencies fit the environment.
Reply to this answer
Move the repeatable checks into the pipeline. If every module needs unit and integration tests, check that the relevant stages exist. Tests for a new module should also be wired to the right pipeline, not just added under an unrelated one.
Use path-based reviewer rules for sensitive configuration files too. An automated review tool can flag possible infrastructure changes, but it shouldn't be the only safeguard for a new service or a networking change.
Reply to this answer
This isn't solely an AI problem. A query or service can work on a laptop and still be unsuitable for production. AI can make it easier to produce a convincing change without understanding those differences, but the review needs to cover them either way.
Reply to this answer
I'd work on shared ownership as well as guardrails. Application engineers don't need to know every infrastructure detail, but they should understand enough about deployment and the pipeline to identify when a change needs more input.
Bring them into the design discussion for new services rather than waiting until the pull request is ready. A review can then focus on a known deployment plan instead of becoming a surprise veto.
v.418!tinwicket:
That sounds useful for changes where the code depends on configuration outside the repository. Those are easy to miss when reviewing the diff alone.
Reply to this answer
There's another angle here: a good platform can make the supported path easier to use. If developers can add a module or service through a standard setup that includes the required pipeline configuration, fewer details have to be remembered on every change.
That takes work, and it won't cover every exception, but it can reduce the number of reviews spent catching the same omissions.
Reply to this answer
Add your answer
Answers are reviewed before publication.
q.526!ferryspanner:
That also makes a rejection easier to explain. The missing requirement is visible in the check instead of living only in one reviewer's head.