fix(adk): return str from serialize_metadata_value for pydantic values

[anxkhn]

What

serialize_metadata_value in python/packages/kagent-adk/src/kagent/adk/converters/event_converter.py
is declared def serialize_metadata_value(value: Any) -> str: and its docstring
promises to return the "String representation of the value." For any value that
has a model_dump attribute (that is, any Pydantic model) it did:

return value.model_dump(exclude_none=True, by_alias=True)

Pydantic's model_dump() returns a dict, not a string (the string form is
model_dump_json()), so the function violated both its -> str annotation and
its docstring for Pydantic inputs.

This is not purely cosmetic. The result is written into A2A message/event
metadata by _get_context_metadata, next to sibling values that are plain
strings (for example error_code is stored via str(...)). Real Pydantic
values flow through this branch: event.usage_metadata, event.grounding_metadata,
and event.custom_metadata. So a dict was being injected where the surrounding
metadata values are strings.

Fix

Wrap the model_dump(...) call in str() so the return matches the annotation
and the docstring:

return str(value.model_dump(exclude_none=True, by_alias=True))

This also makes the helper consistent with its twin in the langgraph package,
python/packages/kagent-langgraph/src/kagent/langgraph/_metadata_utils.py, which
already wraps model_dump(...) in str(...). The adk copy had simply dropped
the str() wrapper. No signature or API change; the except fallback already
returned str(value).

Tests

Added TestSerializeMetadataValue in
python/packages/kagent-adk/tests/unittests/converters/test_event_converter.py:

• a Pydantic model serializes to a str equal to
  str(model.model_dump(exclude_none=True, by_alias=True)) (fails on the old
  code, which returned a dict; passes after the wrap), and
• a plain value (42) serializes to "42".

cd python && uv run pytest packages/kagent-adk/tests/unittests/ -> 340 passed,
1 skipped. ruff format and ruff check are clean on the changed files.

[Copilot]

Pull request overview

This PR fixes serialize_metadata_value in the ADK event converter so it consistently returns a str for Pydantic model inputs, aligning behavior with its type annotation/docstring and avoiding injecting dict objects into A2A event metadata.

Changes:

• Wrap value.model_dump(...) with str(...) when serializing Pydantic values.
• Add unit tests covering Pydantic-model and plain-value serialization behavior.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
python/packages/kagent-adk/src/kagent/adk/converters/event_converter.py	Ensures Pydantic model_dump() output is coerced to str to match function contract and metadata expectations.
python/packages/kagent-adk/tests/unittests/converters/test_event_converter.py	Adds tests validating string output for both Pydantic and non-Pydantic inputs.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[supreme-gg-gg]

There's nothing wrong with returning dict here. A2A metadata allows them and they are needed in the frontend to properly show things like token usage:

kagent/ui/src/lib/messageHandlers.ts

Lines 390 to 398 in 3eff08b

	function getMessageTokenStats(metadata: Record<string, unknown> | undefined): TokenStats | undefined {
	const usage = getMetadataValue<ADKMetadata["kagent_usage_metadata"]>(metadata, "usage_metadata");
	if (!usage) return undefined;
	return {
	total: usage.totalTokenCount ?? 0,
	prompt: usage.promptTokenCount ?? 0,
	completion: usage.candidatesTokenCount ?? 0,
	};
	}

A better fix would be to update the type annotation to str | dict[str, Any] and the docstring.

If you trace the langgraph example you cited, you will find that even though serialize_metadata_value does return string it is in fact never used in the converter with any metadata, so it's effectively dead code.