fix: Bedrock tool call arguments no longer silently dropped (closes #5275)

[TheCodingDragon0]

Fixes #5275

Problem

When using crewAI with AWS Bedrock LLMs (e.g. Amazon Nova Pro), tool arguments were silently discarded. The LLM correctly called the tool with the right arguments, but the tool received an empty dict {}.

Root Cause

Line 850 in crew_agent_executor.py:

func_args = func_info.get("arguments", "{}") or tool_call.get("input", {})

Bedrock's Converse API returns tool calls as {"name": "X", "toolUseId": "Y", "input": {...}} — no "function" wrapper. So func_info = {}, and func_info.get("arguments", "{}") returns the default string "{}" which is truthy, preventing the or from falling through to tool_call.get("input", {}).

Fix

Changed default from "{}" to None:

func_args = func_info.get("arguments") or tool_call.get("input", {})

With None as default, func_info.get("arguments") returns None (falsy) when no "function" wrapper exists, so the or correctly falls through to the Bedrock input format.

[samimasod]

A small regression test here would make this much safer. There’s already Bedrock native tool-calling coverage in lib/crewai/tests/agents/test_native_tool_calling.py, especially test_bedrock_agent_kickoff_with_tools_mocked.

It would be great to extend that test (or add a nearby Bedrock-specific regression test in the same file) to assert that a Bedrock-shaped tool call like:

{"name": "calculator", "toolUseId": "...", "input": {"expression": "15 * 8"}}

results in the executed tool receiving {"expression": "15 * 8"} rather than {}.

Right now the test seems to validate the final outcome, but not that the Bedrock input payload was preserved through parsing/execution.