Allowing custom events from tools.

[gundeep-dialpad]

Problem

Some tools can take a long time to complete execution.
It would be great if it's possible to generate intermediate events while the stages inside the tool complete.

Limitation of current adk version

So far, I looked a bit into the source code and found that we need reference to the runner to generate events (I might be wrong here). I haven't looked into the web socket connection yet, but currently, post endpoints interactions don't keep the reference to the runner.

Implementation idea

Each runner can be assigned a runner id and active runners can be saved into a global/singleton dict. Each tool, callback and other flow entrypoints can optionally get access to runner id using an annotation and thus choose to generate intermediate events.

Enabling other possibilities

Maybe runner id is not the only thing that tools, callbacks need; They can also make use of things like session id, so maybe tools, callbacks need a full context, which it currently don't have.
When tools have access to session id, it can access history.

Current workaround

I'm using an extra tool called listen_to_results_until_tool_result_complete which intercept model call using before_model_call callback.
Here's the agent code: https://pastebin.com/P3YvU0hj
I'm thinking of converting this workaround into a more organized annotation based, but if something in the core source code and code flow can be updated, for example, using the suggested implementation idea, or some other more holistic/generic implementation, it would be much simpler.

Would like to hear your feedback and other implementation ideas or reasons why it's not a generic enough direction, and should be implemented more like a plugin.

[hangfei]

Does this feature request address your issue: #1099

[gundeep-dialpad]

yes, that's exactly what I'm looking for. I tried the AsyncGenerator and realized it's only for the live model.