Okay. My agent is primarily a diff reviewer, not a general agent. I'm introducing a protocol boundary so the review pipeline doesn't care what's behind it, swapping Agno for LiteLLM behind that boundary, and building a structured output module that handles getting reliable JSON from any LLM provider. That module is the piece I'll propose, if approved, via PR into the OpenHands SDK, since it's one of the most-requested features. Grippy is the proving ground. The hardest part is prompt composition. Agno formats system messages in a specific way that affects review quality. I'll be capturing golden-file snapshots of the current prompts before I change anything, then verifying the new path matches. The rest is mechanical (I hope). Will report back when I have notes.