fix: omit parallel_tool_calls when not explicitly enabled (COM-406)

[daniel-lxs]

Summary

Fixes COM-406: parallel_tool_calls=false breaks LiteLLM/Bedrock routes with native tools.

Problem

The OpenAI provider was always sending parallel_tool_calls=false with native tools (via metadata.parallelToolCalls ?? false). This caused errors with non-OpenAI backends accessed through LiteLLM:

• Bedrock/Claude: Don't support the parameter at all - any value breaks
• Gemini: Supports parallel tool calls, but rejects parallel_tool_calls=false when multiple tools are provided

Solution

Changed the logic to only include parallel_tool_calls: true when metadata.parallelToolCalls is explicitly set to true. The parameter is now omitted entirely when not explicitly enabled, which allows LiteLLM routes to work with any backend.

Changes

• src/api/providers/openai.ts - Updated 4 locations where parallel_tool_calls was being set (streaming, non-streaming, and O3 family models)
• src/api/providers/__tests__/openai-native-tools.spec.ts - Updated test to verify parameter is NOT included when parallelToolCalls is not explicitly true

Testing

All 121 OpenAI provider tests pass, including the updated test that verifies the fix.

---

Important

Fixes parallel_tool_calls logic in openai.ts to only include it when explicitly enabled, resolving LiteLLM/Bedrock compatibility issues.

• Behavior:
  • parallel_tool_calls is now only included when metadata.parallelToolCalls is explicitly true in openai.ts.
  • Fixes compatibility issues with LiteLLM/Bedrock routes by omitting parallel_tool_calls when not enabled.
• Testing:
  • Updated test in openai-native-tools.spec.ts to verify parallel_tool_calls is omitted when not explicitly set to true.
• Files Affected:
  • openai.ts: Updated logic in 4 locations for streaming, non-streaming, and O3 family models.
  • openai-native-tools.spec.ts: Added test verification for the fix.

^{This description was created by for dbb85f8. You can customize this summary. It will automatically update as commits are pushed.}

Closes #10553
(from Linear COM-406)

[roomote]

Rooviewer   See task on Roo Cloud

Review complete. No issues found.

The PR correctly fixes the parallel_tool_calls compatibility issue with LiteLLM/Bedrock backends by only including the parameter when explicitly enabled. The changes are consistent across all 4 locations in openai.ts and properly tested.

_{Mention @roomote in a comment to request specific changes to this pull request or fix all unresolved issues.}