Hi Kaxil, Regarding Copilot billing, the lack of clarity on server-side costs is concerning. Given the recent surge in PR volume, there is a risk of uncontrollable expenses if hard caps aren't in place, similar to those on ASF CI.
I agree that automated and assisted reviews are not mutually exclusive. I am interested to hear how others perceive this balance and what their experiences have been. Best, Jarek Potiuk On Tue, Apr 28, 2026 at 2:20 PM Kaxil Naik <[email protected]> wrote: > Ignore the last para about links -- it is the copy/paste of what I had sent > to the new AI initiative ASF group/list where some of related discussions > were happening. > > On Tue, 28 Apr 2026 at 13:18, Kaxil Naik <[email protected]> wrote: > > > The Copilot reviews as I had recently found out were paid for by our > > Astronomer's GitHub enterprise (for me and other folks at Astronomer from > > our quota). > > > > And with them moving to Usage-based model (which was bound to happen), it > > will get expensive. > > > > Although, I think it is still valuable for mass-reviewing on the server > > side since this happens on CI and is a complete opt-in by the reviewer > and > > there are no doubts. > > > > As I showed in the last dev call, I have my hand crafted review skill > that > > I use for detailed reviews from my laptop and that has been vetted by me > > before posting. So I am fully responsible for all the good and bad (false > > positive or hallucinations) things it catches since I approve it. And > have > > been using this skill for a good quarter or half a year (time flies). > > > > And that is why I do not feel it is either / OR -- meaning it was never > > Copilot review vs local review for me since I used both based on the > > purpose. For reviewing 200 PRs as last time, I used Copilot since a > review > > is helpful than no review, and PR getting marked stale and I have seen > > folks self-assign copilot review on their PRs -- for those who have > access > > to it. I have done the same to have multiple layers (even though my local > > review skill already does multi-modal reviews). > > > > And my philosophy around Review and a lot of workflow skills have been: > > What I look for in a PR isn't necessarily what someone else would look > for. > > Or what is important to me, might be nit for someone. So while as a > project > > we should have standards which go in AGENTS.md/Claude.md and or a > > high-level review skill that is checked-in the project, the "what I look > > for" will remain on my machine and is geared towards my preference, > testing > > etc. Time and again I'd use reviews from Ash and others to tune it -- > like > > I would do even without AI. As an example Ash would have caught something > > in my PR or someone else's PR that I wouldn't have realized, and like a > > human I'd learn for next time, so are my skills. But that is sort of my > new > > workflow adapting. > > > > re: fully automated Copilot reviews may discourage contributors who > expect > > human interaction > > > > At least to me, there is no difference between that and a review from a > > human which completely sounds AI/robotic anyway since a human can just > run > > a skill and post the response as well --- from a purely PR author > reception > > point of view. It is just coming from a different account, and the latter > > feels even worse as it is coming from a human account. But at the end of > > the day, that is still a personal preference. A review (from copilot, > from > > human that sounds like AI, a pure human review) that catches a bug is > still > > better than no review, PR going stale and being closed. > > > > Long story short: I do not think they are mutually exclusive -- or never > > were mutually exclusive :) > > > > ------ > > > > - https://github.com/apache/airflow/pull/63775#discussion_r3025383633 -- > > Copilot caught a Databricks provider importing airflow.utils.timezone > > directly (which relies on Airflow's runtime deprecation redirect and > > silences typing) and suggested switching to > > airflow.providers.common.compat.sdk. That is our documented Airflow 2 / > > Airflow 3 cross-version pattern for providers. Copilot only knows this > > because we wrote it down. > > - https://github.com/apache/airflow/pull/62343#discussion_r3025380683 -- > > same cross-version import pattern, different provider. Author accepted > the > > suggestion. > > - https://github.com/apache/airflow/pull/64568#discussion_r3025333917 -- > > Copilot flagged a fix for failure-callback context["exception"] handling > > and asked for a regression test specifically against the > > InProcessTestSupervisor / dag.test() path. That is the correct execution > > path for that change, not a generic "please add a test" remark. > > - > > > https://github.com/apache/airflow/pull/64576#pullrequestreview-4047787083 > -- > > Copilot reviewed a fix I authored for xcom_pull() ignoring default when > > map_indexes was not set, and raised a header-precedence > > question I had not thought about. > > - > > > https://github.com/apache/airflow/pull/61878#pullrequestreview-3851732779 > -- > > Dennis surfaced this one on the dev list as a real review he found useful > > on a provider PR. He's at a > > different company, so worth flagging as an independent signal. > > > > On Tue, 28 Apr 2026 at 11:02, Jarek Potiuk <[email protected]> wrote: > > > >> Hi Kaxil and team, > >> > >> I’d like to discuss our strategy for "copilot reviews" versus SKILL > based > >> assisted triage and maintainer review experiences. > >> > >> My recent experiments with skill-based auto-triage show a downward trend > >> in > >> stale pull requests by filtering out "drive-by" contributions, allowing > >> maintainers to focus on high-impact work (more stats and trends soon). > >> Unlike 3rd-party tools, this approach is agent-agnostic, easily > fine-tuned > >> via English prompts, and potentially reusable across other ASF projects. > >> > >> Regarding transparency, I’ve proposed specific attribution formats [1] > to > >> distinguish between purely automated comments and those reviewed by > >> humans. > >> This aligns with ongoing [email protected] conversations [2] > about > >> using "Assisted By" instead of "Generated-by" to maintain accountability > >> and contributor motivation. > >> > >> I have concerns that fully automated Copilot reviews may discourage > >> contributors who expect human interaction. My experimental > >> /maintainer-review skill [3] allows maintainers to drive the process by > >> reviewing AI-generated comments before posting. This reduces "token > >> ping-pong," integrates with the CODEOWNERS discussion [4], and ensures > we > >> only spend tokens on relevant, triaged PRs. And keep maintainers > >> ultimately > >> responsible for comments they accept to send. > >> > >> With Copilot moving to usage-based billing on June 1 [5], using personal > >> or > >> open-source credits via the skill-based approach introduces some change > >> (but I am still unclear what it means to the billing - who pays for the > >> tokens). > >> > >> Conversely, for skill-based processes, it's entirely in the hands (and > >> pocket) of those who run the skill locally. We are also working with ASF > >> and Alpha-Omega to secure long-term resources for maintainers for that. > >> > >> I would love to hear your thoughts about it - especially Kaxil's > >> experiences regarding the Copilot review experiment - my observations > >> might > >> be incomplete and biased :). > >> > >> J. > >> > >> [1] https://github.com/apache/airflow/pull/65965 - comment attribution > >> [2] https://lists.apache.org/thread/qbdkky8ls6zybyy9o3pvqnpf68r089qp - > >> legal discussion thread on attributions > >> [3] https://github.com/apache/airflow/pull/65981 - /maintainer-review > >> skill > >> [4] https://lists.apache.org/thread/5ssp4ksyohdzclxqvj7ngz0hz5wy9j68 - > >> CODEOWNERS discussion > >> [5] > >> > >> > https://github.blog/news-insights/company-news/github-copilot-is-moving-to-usage-based-billing/ > >> - change in billing for Copilot > >> > >> Best, > >> Jarek Potiuk > >> > > >
