Integrates SearXNG web search as a tool available during chat when the
agent has search enabled. Supports optional reranking of results via a
configurable reranking model. Replaces Python subprocess evaluation with
@pydantic/monty WASM runtime. Introduces stable tool call ID mapping to
avoid exposing provider-native IDs to the database.
This is once again a huge commit, but its mostly performance
improvements along with some bug fixes and refactoring. It also includes
changes to the theming systems. I'm still not 100% happy with the
theming system, but its better than before.
Model fetching has been dramatically improved! Nearly all the important
computation and pre-processing has been moved to the server. This has
also somehow fixed the way model details are loaded, which was causing
many models to be missing their details despite models.dev having them.
The markdown renderer has once again been changed, but I'm mostly
certain that this is the last time major changes will be made to it. The
renderer is not spamming components, bloating memory usage, and its not
using a bug prone custom written chunking system.
There's also a lot more that I haven't mentioned and honestly forgot. I
need to get better commit hygiene tbh.
zoeissleeping
