wenjin272 commented on issue #894:
URL: https://github.com/apache/flink-agents/issues/894#issuecomment-5077556987

   @weiqingy I’m wondering whether we should complement the reviewer-side 
guidance with some developer-side guidance for AI-assisted PRs.
   
   For a non-trivial PR whose implementation is largely AI-assisted, the author 
could provide a concrete **Implementation Description**, either in the PR 
description or as a linked document. It should go beyond a high-level design 
and cover the runtime flow, behavioral contracts, key decisions, failure 
behavior, compatibility impact, and how tests map to those contracts.
   
   This could support a two-stage review process:
   
   1. Human reviewers first review the architecture, design decisions, and 
behavioral contracts.
   2. A coding agent then verifies that the implementation and tests match the 
accepted description.
   
   One possible way to introduce this would be to add a short **How to 
Contribute** section to `AGENTS.md`, link to something like 
`contribution-guides/ai-assisted-pr.md`, and add an **Implementation 
Description** section to the PR template.
   
   The goal would be to keep design and architectural decisions under human 
review, while letting coding agents handle more of the implementation-level 
verification. Do you think this direction would be useful and feasible for the 
project?
   
   
   I think we could first try a lightweight experiment before promoting this as 
a community-wide practice.
   
   For PRs opened by the two of us, especially non-trivial or heavily 
AI-assisted ones, we could voluntarily include a concrete **Implementation 
Description** in the PR body or a linked document. It would describe the 
runtime flow, behavioral contracts, key decisions, failure behavior, 
compatibility impact, and the mapping between those contracts and tests.
   
   Reviewers could first review this description at the design and architecture 
level, then use a coding agent to verify that the implementation and tests 
match it. This may help keep architectural decisions under human review while 
delegating more implementation-level verification to AI.
   
   After trying this on a few PRs, we could evaluate whether it actually 
reduces review effort, what level of detail is useful, and how much authoring 
overhead it introduces. If the experiment works well, we could then consider 
documenting it in `AGENTS.md`, adding a contribution guide, or updating the PR 
template. Would you be interested in trying this?


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to