Open cpumaxx opened 3 days ago
It looks like there may be a few options for this: Either try to use heuristics to find text streaming in at token-sized chunks over the network, or allow the user to specify div classes somehow (config or right-click). The first approach would generalize, but may have a tendency to fail or over-shoot on short replies or loading short user text. The second would need per-site/tool setup of some kind and may not work consistently in all places. eg. ooba has a "circle-bot" div class that should appear prior to llm responses, but wouldn't generalize to anything else.
Develop a mutation observers subsystem that can automatically handle DOM updates on a page to read streaming text as it is presented to the user.