PR #4973: fix: extract Bedrock Converse API tool arguments from 'input' key (#4972)

• Repository: crewAIInc/crewAI
• Author: devin-ai-integration[bot]
• State: open | merged: False
• Link: https://github.com/crewAIInc/crewAI/pull/4973

Description (problem / solution / changelog)

Summary

Fixes #4972. When using AWS Bedrock as the LLM provider with native function calling, all tool arguments were silently dropped — every tool call received {} instead of its actual arguments.

Root cause: In _parse_native_tool_call's dict-handling branch, func_info.get("arguments", "{}") returned the truthy default string "{}" when the "function" key was absent (as it is for Bedrock). This prevented the or from ever reaching tool_call.get("input"), where Bedrock stores its arguments.

Fix: One-line change — remove the truthy default so get("arguments") returns None (falsy) when the key is missing, allowing the or-chain to correctly fall through:

# Before (broken):
func_args = func_info.get("arguments", "{}") or tool_call.get("input", {})

# After (fixed):
func_args = func_info.get("arguments") or tool_call.get("input") or "{}"

Review & Testing Checklist for Human

• Verify the or-chain edge cases are acceptable: when Bedrock returns "input": {} (empty dict, falsy), func_args becomes the string "{}" rather than {}. Confirm downstream parse_tool_call_args handles both types correctly (it does — see _execute_single_native_tool_call).
• Confirm OpenAI-style dict tool calls ({"function": {"arguments": ...}}) are unaffected — the first get("arguments") still returns the string args when present.
• If you have Bedrock credentials, test a crew with a BaseTool that has required arguments (e.g., search_query: str) to verify args flow end-to-end.

Notes

• 8 new unit tests added covering Bedrock-style dicts, OpenAI-style dicts (regression), empty inputs, missing keys, and toolUseId handling.
• This affects all Bedrock models using native function calling since BedrockCompletion.supports_function_calling() returns True.

Link to Devin session: https://app.devin.ai/sessions/999e4cae3d7045f3905052a39803fc23

Changed files

• lib/crewai/src/crewai/agents/crew_agent_executor.py (modified, +1/-1)
• lib/crewai/tests/agents/test_native_tool_calling.py (modified, +157/-0)

PR #4975: fix(agents): fix Bedrock Converse API tool arguments dropped in _parse_native_tool_call

• Repository: crewAIInc/crewAI
• Author: NIK-TIGER-BILL
• State: open | merged: False
• Link: https://github.com/crewAIInc/crewAI/pull/4975

Description (problem / solution / changelog)

The default value "{}" (truthy string) in func_info.get("arguments", "{}") prevents the fallback to tool_call.get("input", {}) from activating for Bedrock Converse API responses.

Closes #4972

Changed files

• lib/crewai/src/crewai/agents/crew_agent_executor.py (modified, +1/-1)

PR #4978: fix: Bedrock Converse API tool arguments silently dropped (#4972)

• Repository: crewAIInc/crewAI
• Author: guoyangzhen
• State: open | merged: False
• Link: https://github.com/crewAIInc/crewAI/pull/4978

Description (problem / solution / changelog)

Problem

When using AWS Bedrock with native function calling, all tool arguments are silently dropped. Every tool call receives an empty dict {} instead of the actual arguments, causing pydantic validation failures on any tool with required parameters.

This affects ALL Bedrock models using native function calling (Claude, Llama, etc.) via the native tool loop code path.

Root Cause

In _parse_native_tool_call (crew_agent_executor.py), the dict-format branch has a truthy default:

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

Bedrock Converse API returns {name, input, toolUseId} — no "function" key.

So func_info is {}, and func_info.get("arguments", "{}") returns the default "{}" — a truthy string. The or short-circuits and tool_call.get("input", {}) is never evaluated.

Fix

Remove the default so get() returns None (falsy), allowing the or to fall through:

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

Verification

Bedrock format: {"name": "search_tool", "input": {"search_query": "hello"}, "toolUseId": "abc"}
OLD: func_info.get("arguments", "{}") → "{}" (truthy!) → short-circuits → broken
NEW: func_info.get("arguments") → None (falsy) → falls through to tool_call["input"] ✓

OpenAI format: {"function": {"name": "t", "arguments": "{"x": 1}"}}
NEW: func_info.get("arguments") → "{"x": 1}" (truthy) → same as before ✓

Tests Added

• test_parse_native_tool_call_bedrock_format — verifies Bedrock input dict is correctly extracted
• test_parse_native_tool_call_openai_dict_format — verifies OpenAI format still works

Fixes #4972

Changed files

• lib/crewai/src/crewai/agents/crew_agent_executor.py (modified, +1/-1)
• lib/crewai/tests/agents/test_agent_executor.py (modified, +38/-0)

PR #4985: fix: preserve Bedrock Converse API tool arguments in _parse_native_tool_call

• Repository: crewAIInc/crewAI
• Author: guoyangzhen
• State: open | merged: False
• Link: https://github.com/crewAIInc/crewAI/pull/4985

Description (problem / solution / changelog)

Fix

Fixes #4972

Problem

When using AWS Bedrock (Converse API) with native function calling, all tool arguments are silently dropped — every tool call receives an empty dict {} instead of actual arguments.

Root Cause

In _parse_native_tool_call, the dict-handling branch has:

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

Bedrock returns tool calls as {"name": "...", "input": {...}, "toolUseId": "..."} — no "function" key. So func_info is {}, and func_info.get("arguments", "{}") returns the default string "{}". Since "{}" is truthy, the or short-circuits and tool_call["input"] is never evaluated.

Fix

Remove the default value from .get("arguments") so it returns None when the key is absent, allowing the or to fall through to tool_call.get("input"):

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

Behavior by provider:

Testing

The fix can be verified by running a crew with Bedrock and a tool that has required arguments. Before the fix, all tool calls fail with Field required [type=missing, input_value={}, input_type=dict]. After fix, arguments are correctly passed through.

Changed files

• lib/crewai/src/crewai/agents/crew_agent_executor.py (modified, +1/-1)