wsxiaoys

Luma/Meshy are generative models that output a high-fidelity mesh in one shot. This is more about testing an LLM’s spatial and visual understanding: the agent writes the geometry code and has to reason about the 3D result from what it sees, then iterate.

OP here - I've talked in detail about how we rendered NES suggestions using only VS Code public APIs.

Most tools fork the editor or build a custom IDE so they can skip the hard interaction problems.

Our NES is a VS Code–native feature. That meant living inside strict performance budgets and interaction patterns that were never designed for LLMs proposing multi-line, structural edits in real time.

In this case, surfacing enough context for an AI suggestion to be actionable, without stealing attention, is much harder.

That pushed us toward a dynamic rendering strategy instead of a single AI suggestion UI. Each path gets deliberately scoped to the situations where it performs best, aligning it with the least disruptive representation for a given edit.

If AI is going to live inside real editors, I think this is the layer that actually matters.

Happy to hear your thoughts!

OP here - happy to answer any questions.

This was one of the more unexpectedly tricky layers of building real-time LLM suggestions, and I’d love to hear how others have approached timing, cancellation, and speculative prediction in their editors or agents.