feat: Enable configurable streaming support for OpenAI models

[syedazeez337]

Description:

This PR addresses the issue where requests to OpenAI-compatible model
services were hardcoded with "stream": false.

I have introduced a new stream configuration field to the Agent API. This gives
developers explicit control over whether they want streaming responses enabled or
disabled for their agents.

Key Changes:

   - API Update: Added an optional stream field (boolean) to the Agent spec in both
     v1alpha1 and v1alpha2.
   - Default Behavior: To ensure backward compatibility and safety, the field defaults
     to false if not specified.
   - ADK Integration: Updated the Python ADK and the translator logic to propagate
     this setting all the way down to the model runner. When enabled, it correctly
     sets the internal state to StreamingMode.SSE.
   - Documentation: Updated the Helm chart descriptions to reflect the new field.

Testing:

   - Added new Python unit tests to verify that the boolean flag correctly maps to the
     StreamingMode enum.
   - Verified that existing OpenAI model integration tests pass.
   - Updated the Go "golden files" to match the new API schema.

Everything has been linted and verified against the contribution guidelines. Let me
know if you have any questions

Fixes #1099

[EItanya]

Thanks so much for the contribution, and this change definitely makes sense. The one thing i'd like to note is that I actually originally tried to make stream: true the default, but ran into issues. Have you done e2e testing with stream = true?

[syedazeez337]

Hi @EItanya, I have tested and passed the e2e tests locally. Let me know if it is working now. Thank you.

[supreme-gg-gg]

This does not seem to work. I've made some additional modifications in _openai.py for streaming handling and task store, converters, etc and the streaming behaviour is functional.

I'll just mention the biggest change here:

• When streaming for A2A using TaskStatusUpdateEvents events containing ADK "partial" events (chunks from LLM streaming) are being saved to the task store, so on reload it will in fact load all the chunks from LLM in the frontend (like the messages will look like ["hello my", "name is"]. The final and status fields on A2AEvent just distinguishes if it's a task in progress, not a streaming chunk for LLM as indicated on the partial field in ADK events. So it is required to handle this explicitly and avoid persisting them to the task store even if we want them to be streamed.

• Function calls in streaming were previously not handled properly!

Screen.Recording.2026-01-09.at.7.32.42.PM.mov

@syedazeez337 lmk if you're still interested in working on this, otherwise I will finish this up. I've pushed my code to the jetc/fix-streaming branch. I couldn't find any example of this working with ADK and A2A, neither does the ADK a2a extension implement this I believe, so do let me know if you have a better implementation idea.

We would also need e2e testing with stream: true, but it will probably depend on the MockLLM package to support streaming (which has just been added).

EDIT: This has been edited since I was able to fix one of the issue above later

[syedazeez337]

Hi @supreme-gg-gg I have reviewed your changes in jetc/fix-streaming - great work on the streaming handler and partial event filtering. You have covered almost all of the issue. I would like to propose where collaboration is needed, since you are one of the maintainers as well, let me know in what capacity I can be helpful here. Thank you.

[supreme-gg-gg]

Hi @supreme-gg-gg I have reviewed your changes in jetc/fix-streaming - great work on the streaming handler and partial event filtering. You have covered almost all of the issue. I would like to propose where collaboration is needed, since you are one of the maintainers as well, let me know in what capacity I can be helpful here. Thank you.

How about I'll open a PR from my branch since this branch is somewhat out of date as well and I'll add you as coauthor? there are some work that needs to be finished, like updating types in go, e2e tests with streaming (that requires new mockLLM though), perhaps making streaming false by default (the original behaviour), etc.

[syedazeez337]

Hi @supreme-gg-gg Let's do that

[EItanya]

This sounds like a great plan, @syedazeez337 would you mind closing this PR as it is now superseded by #1202