We were discussing this with a couple of colleagues! Generally not opposed to this, i’m wondering how you get the comments into the agent’s context? Is it polling? I think:

- implementation should be harness-agnostic and not rely on a claude code feature that codex, pi, or others don’t implement

- polling sounds token inefficient, maybe there’s a better mechanism?

- it should be minimally invasive and work well with running multiple diffle instances at once (we were weighing raw endpoints the agent can curl or a socket/file, mcp seems a bit much given users might be running different instances on multiple ports)

Feel free to open an issue or a small pr to discuss the design :)