fix(bedrock): degrade to empty args on malformed streamed tool JSON

[anxkhn]

What

In the Bedrock streaming path, tool-call arguments are accumulated across
contentBlockDelta chunks and then parsed with json.loads, but without the
json.JSONDecodeError guard that every other provider uses. When a streamed
tool-argument JSON is truncated or malformed, the json.loads raises inside the
response generator, gets caught by the broad except Exception that wraps the
whole method, and the entire turn is returned as an API_ERROR.

This PR wraps that single json.loads in try/except json.JSONDecodeError and
falls back to {}, so a recoverable-but-malformed streamed tool call degrades to
empty args instead of aborting the turn.

Why

Every sibling provider already handles this exact situation gracefully, so
Bedrock streaming is the odd one out:

• OpenAI, non-streaming: _openai.py (try/except json.JSONDecodeError: args = {})
• OpenAI, streaming: _openai.py (same guard on the accumulated tool call)
• SAP AI Core, streaming: _sap_ai_core.py (same guard)
• Bedrock, non-streaming: _bedrock.py reads an already-parsed dict
  (tool.get("input", {})), so it never raises here.

Only the Bedrock streaming branch parsed the accumulated input_json with a bare
json.loads. Streamed tool arguments are a realistic place for truncation (early
stream termination, or a model emitting slightly malformed JSON), so the
inconsistency is user-visible: the same partial tool call that recovers on
OpenAI/SAP fails the whole turn on Bedrock streaming. This change restores the
behavior the rest of the models package already standardized on.

How

python/packages/kagent-adk/src/kagent/adk/models/_bedrock.py:

for tool_id, tool in tool_uses.items():
    try:
        args = json.loads(tool["input_json"]) if tool["input_json"] else {}
    except json.JSONDecodeError:
        args = {}
    part = types.Part.from_function_call(name=tool["name"], args=args)

The try/except is copied verbatim from the sibling providers so the behavior is
identical across the package.

Testing

Added a colocated regression test,
test_streaming_malformed_tool_input_degrades_to_empty_args, in
python/packages/kagent-adk/tests/unittests/models/test_bedrock.py. It drives the
real generate_content_async(..., stream=True) path with a mocked
converse_stream that emits a toolUse whose accumulated input is truncated
({"city": "Paris"), and asserts the turn completes cleanly:
error_code is None, the function-call name/id are preserved, and
args == {}. Bedrock (AWS) is fully mocked, so no live credentials are needed.

• Before the fix, this test fails: the turn comes back as API_ERROR
  ("Expecting ',' delimiter").
• After the fix, it passes.

cd python
uv run pytest packages/kagent-adk/tests/unittests/models/test_bedrock.py   # 34 passed
uv run pytest packages/kagent-adk/tests/unittests/models/                  # 159 passed, 1 skipped
make lint                                                                   # clean

No CRD or Go API change, so no make generate and no E2E changes are required.

[Copilot]

Pull request overview

This PR aligns the Bedrock streaming tool-call parsing behavior with the other model providers by preventing malformed/truncated streamed tool-argument JSON from aborting an entire turn. Instead, malformed tool args degrade to {} so the turn can complete successfully.

Changes:

• Guard Bedrock streaming json.loads(...) with try/except json.JSONDecodeError and fall back to {}.
• Add a regression test that simulates a truncated streamed tool input and asserts the turn completes with empty args.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
python/packages/kagent-adk/src/kagent/adk/models/_bedrock.py	Wraps streamed tool-argument JSON parsing to degrade to {} on JSONDecodeError instead of failing the turn.
python/packages/kagent-adk/tests/unittests/models/test_bedrock.py	Adds a streaming regression test for malformed tool input; includes a small cleanup opportunity (nit) around unused to_thread patching.

---

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