mirror of
https://github.com/agent-substrate/substrate.git
synced 2026-10-02 03:24:42 +08:00
Benjamin Elder
aa17b3c297
This is a skill I've been iterating on, first with a few coworkers' PRs only, and then progressively more broadly. It gives pretty decent results with frontier models, but I think we need to continue refining it. The basic idea is that this creates or updates draft PR comments to add AI findings, categorized by severity, and *clearly denoted* as coming from an agent. It is instructed to bias towards brevity as these models tend to write massive comments. That still needs more tweaking. Reviewers are expected to leverage this skill to suggest draft comments which they then review and delete as needed, and finally submit an overall review with their own top-level comments. It does NOT do things like "summarize what the PR does" ... PR authors already do this, and adding that information to the PR thread is not helpful. Anyone interested can ask their agent to summarize the PR for them locally. What it does do, is report possible bugs, for a reviewer to vet and then post as they see fit. NOTE: **this should be squash merged** > It's a good idea to open an issue first for discussion. - [x] Tests pass - [x] Appropriate changes to documentation are included in the PR