fix(bedrock): extract tool arguments from Bedrock's input field

[giulio-leone]

Problem

AWS Bedrock tool calls always receive empty arguments {} regardless of what the LLM provides. This affects all Bedrock models (Nova Pro, Nova Lite, Claude via Bedrock, etc.).

Bedrock tool call format:

{"toolUseId": "tooluse_abc", "name": "search_tool", "input": {"query": "AWS Bedrock features"}}

OpenAI tool call format:

{"id": "call_xyz", "function": {"name": "search_tool", "arguments": '{"query": "test"}'}}

Root Cause

In _parse_native_tool_call (line 850):

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

When func_info (from tool_call.get("function", {})) is an empty dict (Bedrock has no function key), .get("arguments", "{}") returns the default string "{}". Since "{}" is truthy, the or operator never evaluates tool_call.get("input").

Result: Bedrock's actual arguments in input are silently dropped.

Fix

Use None as the sentinel so or falls through correctly:

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

Verified Locally

=== Bedrock tool call ===
OLD CODE: '{}'       <-- BUG: returns empty string instead of args
NEW CODE: {'query': 'AWS Bedrock features'}  <-- FIXED: returns actual args

=== OpenAI tool call (regression) ===
OLD CODE: '{"query": "test"}'
NEW CODE: '{"query": "test"}'
MATCH: True          <-- No regression

Tests

Added 7 unit tests covering:

• Bedrock dict with input field (the reported bug)
• OpenAI dict with function.arguments (regression)
• Bedrock dict without function key
• Empty input dict
• Neither function nor input (fallback)
• OpenAI-style object with .function attribute
• Bedrock-style object with .name and .input attributes

All pass.

Fixes #4748

---

Note

Low Risk
Low risk bugfix limited to native tool-call argument extraction; main risk is subtle behavior changes for dict-style tool calls when function.arguments is missing/empty, but new tests cover expected formats.

Overview
Fixes _parse_native_tool_call to correctly extract tool arguments from AWS Bedrock-style tool calls by falling back to the dict’s input field when function.arguments is absent.

Adds focused unit tests covering Bedrock dict/object tool calls, OpenAI regression cases, and fallback behavior when no arguments are provided.

^{Written by Cursor Bugbot for commit 310f000. This will update automatically on new commits. Configure here.}

[giulio-leone]

Friendly ping — CI is green, tests pass, ready for review whenever convenient. Happy to address any feedback. Thanks! 🙏