Fix Kimi K2 Thinking native tool calls for OpenAI-compatible servers

[roomote]

This PR attempts to address Issue #9551 regarding Kimi K2 Thinking native tool calls not working properly with ik_llama.cpp and other OpenAI-compatible servers.

Problem

When using Kimi K2 with ik_llama.cpp, tool calls were being detected but the results were not being sent back to the model properly. The issue was that:

1. The server might not set finish_reason to "tool_calls" specifically
2. Tool calls were accumulated but not always yielded correctly

Solution

Modified base-openai-compatible-provider.ts to:

• Yield tool calls when ANY finish_reason is present (not just "tool_calls")
• Added fallback to yield tool calls at stream end if not already yielded
• Added tracking with hasYieldedToolCalls flag to prevent duplicates
• Ensures tool calls with valid ID and name are always yielded

Testing

• Existing tests pass
• The fix should work with various OpenAI-compatible servers including ik_llama.cpp

Feedback and guidance are welcome!

---

Important

Fixes tool call handling in base-openai-compatible-provider.ts for OpenAI-compatible servers by yielding tool calls on any finish_reason and adding a fallback mechanism.

• Behavior:
  • In base-openai-compatible-provider.ts, tool calls are now yielded when any finish_reason is present, not just "tool_calls".
  • Adds a fallback to yield tool calls at stream end if not already yielded.
  • Introduces hasEmittedToolCallEnd flag to prevent duplicate tool call end events.
• Testing:
  • Existing tests pass, ensuring compatibility with various OpenAI-compatible servers including ik_llama.cpp.

^{This description was created by for 2177a8c. You can customize this summary. It will automatically update as commits are pushed.}

[roomote]

Rooviewer   See task on Roo Cloud

Review complete. No issues found. The previous concern about inconsistent argument handling is no longer applicable - the code now uses tool_call_partial + tool_call_end events instead of accumulated tool_call events, and the type definitions allow optional arguments for partial chunks.

• Fix inconsistent handling of empty arguments in main tool call yielding path (No longer applicable - code restructured)

Previous reviews

• f310921: Review #1

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

[MobCode100]

Tested this PR, no issues with Roo Code v3.35.5. Thanks for the PR.

[hannesrudolph]

@roomote can you please run merge resolver mode on this?

[roomote]

Fixaroo   See task on Roo Cloud

Successfully resolved merge conflicts for PR #9557 "Fix Kimi K2 Thinking native tool calls for OpenAI-compatible servers".

Resolution Summary:

• Adapted the PR's intent to main's new streaming approach that uses tool_call_partial and tool_call_end events
• Modified tool call end emission to trigger on ANY finish_reason (not just "tool_calls") for compatibility with servers like ik_llama.cpp
• Added fallback logic at stream end to emit tool_call_end if not already emitted
• Added hasEmittedToolCallEnd flag to prevent duplicate emissions

All type checks passed. The rebase completed successfully and the branch is now up-to-date with main.

_{View commit | Revert commit}

[daniel-lxs]

This issue was already fixed by PR #9542 which was merged the same day this PR was opened.

The real fix added NativeToolCallParser.finalizeRawChunks() which is called at the end of stream processing regardless of finish_reason. This handles the case where servers don't send finish_reason = "tool_calls" more elegantly at the consumer level.

This PR modifies the provider layer to emit tool_call_end events, but those events aren't actually consumed by Task.ts - there's no case handler for them in the stream processing switch. The events emitted by this code would be silently ignored.

Closing since the underlying issue was resolved by #9542.