feat(ui): add text streaming hook and update GooseMessage to stream responses

[chaitanyarahalkar]

This PR brings real-time text streaming to the desktop UI:

1. useTextStreaming.ts – reusable React hook that opens a Server-Sent Events (SSE) connection, parses incremental tokens, and exposes a clean streaming API to components.
2. GooseMessage.tsx – refactored to consume the new hook and render partial LLM responses as they arrive, giving users immediate feedback instead of waiting for the full completion.

Context & Motivation

Large responses that appear all at once feel sluggish and block the conversation. Streaming completions improve perceived performance and match the interactive feel of modern chat apps. This work lays the foundation for richer UX features like inline code execution progress, token-level highlighting, and cancel/resume.

Implementation Details

• Utilises the Fetch + ReadableStream API wrapped in an ergonomic hook.
• Hook manages:
  • automatic abort on unmount / cancel
  • exponential back-off retry (network hiccups)
  • optional final “done” callback
• GooseMessage now:
  • starts streaming on mount
  • switches from “typing indicator” → live text as soon as the first chunk arrives
  • gracefully handles abort / error states

Testing

• Tested with Databricks, Ollama as well as OpenAI using an OpenAI API key

Demo

Before

Screen.Recording.2025-06-12.at.8.47.51.PM.mov

After

Screen.Recording.2025-06-12.at.9.09.18.PM.mov

🔗 Related Issues / PRs

• Addresses the tool calling issue in a previous streaming feature PR feat: stream LLM responses #2677

[michaelneale]

very nice @chaitanyarahalkar - didn't require changes to providers? cc @jamadeo - you like this approach? look sgood to me

[Kvadratni]

finally. I was under the impression that this is impossible due to how goose loop works. I am so happy to be wrong!

[michaelneale]

one problem is this looks odd with tool results, let me attach a recording (using databricks + claude4):
(and also is only streaming once things are finished - as I guess provider is not streaming?)

streaming2.mov

[michaelneale]

I think clock time needs to be no slower than now for this to appear believable - and hopefully we could have real streaming from the providers?

[michaelneale]

I think for tool cases shouldn't be slower than current wall clock time - so probably requires streaming support fro provider end first?

[chaitanyarahalkar]

I think for tool cases shouldn't be slower than current wall clock time - so probably requires streaming support fro provider end first?

Yup, I can add this fix here and probably in a separate PR add streaming from the backend (directly through the API) for models that support it. I think only OpenAI does it for now?

[chaitanyarahalkar]

@Kvadratni ideally we need it natively from the LLM API, will take a crack at doing that for OpenAI models.

This is just perceptively doing it on the UI end

[chaitanyarahalkar]

one problem is this looks odd with tool results, let me attach a recording (using databricks + claude4): (and also is only streaming once things are finished - as I guess provider is not streaming?)
streaming2.mov

yup, provider is not streaming for now

[jamadeo]

This was my take on adding streaming: #2677 though I've let the PR get stale. I'll have merge and fix some conflicts. I started just with Databricks because it's what I typically use but most of the plumbing is there for any other provider too.

[chaitanyarahalkar]

This was my take on adding streaming: #2677 though I've let the PR get stale. I'll have merge and fix some conflicts. I started just with Databricks because it's what I typically use but most of the plumbing is there for any other provider too.

Nice I'll try to build off the changes here and see if I can make it work for OpenAI's APIs

[DOsinga]

sorry for taking so long - it should be fixed in main now! thanks!