[RFC] Client Agent SDK OutputParser

[yanxi0830]

What does this PR do?

• See https://github.com/meta-llama/llama-stack/discussions/975

Changes
✅ Bugfix ToolResponseMessage role
✅ Add ReACT default prompt + default output parser
✅ Add ReACTAgent wrapper
🚧 Remove ClientTool and simplify it as a decorator (separate PR, including llama-stack-apps)
✅ Make agent able to return structured outputs

• Note that some remote provider do not support response_format structured outputs, add it as an optional flag when calling ReActAgent wrapper.

Test Plan

see test in llamastack/llama-stack-apps#166

Sources

Please link relevant resources if necessary.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Ran pre-commit to handle lint / formatting issues.
• Read the contributor guideline,
  Pull Request section?
• Updated relevant documentation.
• Wrote necessary unit or integration tests.

[hardikjshah]

Thanks for this change, a couple of comments but looks good (Y)

[ehhuang]

Does the turn just terminate after this point?

[yanxi0830]

Yes, the turn will terminate after this point as there's no tool calls in Agent. We can override orchestration in ReActAgent to continue the loop and think again if tool call is not being correctly parsed until "answer" is reached.

[ehhuang]

I wonder if it's cleaner to only have the override as get_tool_call(chunk) instead of a generic output parser. This way:

1. It's clearer what the user is supposed to override
2. We can actually simplify the logic below as:

# the default tool_call_getter just return `chunk...tool_calls`
tool_call = self.tool_call_getter.get_tool_call(chunk)
if not tool_call:
    yield chunk
    return
 else:
    # run tool

3. Bonus: we can also be more functional and not overwrite chunk with the parsed tool calls.

[ashwinb]

@ehhuang +100 especially the (3) bonus

[yanxi0830]

Some update in: #130

However, we still need to overwrite chunk with parsed tool calls, as ClientTool.run takes in a message history and expect the ToolCall detail in the last message.