[ENHANCEMENT] Both XML and native tool calls are supported when parsing model tool calls returned

[MeowSalty]

Problem (one or two sentences)

The new version enforces the use of native tool calls, which causes the original XML format to be output from working properly.

Context (who is affected and when)

All users who update to v3.37 or later

Desired behavior (conceptual, not technical)

Native tool calls are still used when sending call requests, but are compatible with both native and XML formatted tool calls when parsing the data returned by the model.

Constraints / preferences (optional)

No response

Request checklist

• I've searched existing Issues and Discussions for duplicates
• This describes a specific problem with clear context and impact

Roo Code Task Links (optional)

No response

Acceptance criteria (optional)

No response

Proposed approach (optional)

No response

Trade-offs / risks (optional)

No response

[roomote]

I have analyzed this issue and here is my proposed implementation plan:

Problem

When using native tool protocol (v3.37+), if a model returns XML-formatted tool calls in the text stream (e.g., <read_file>...</read_file>) instead of native tool_call chunks, those XML tags are not parsed and are treated as regular text.

The current implementation in Task.ts lines 2757-2795 has an either/or approach:

• XML protocol: Text is parsed via AssistantMessageParser to extract XML tool tags
• Native protocol: Text is treated as plain text without XML parsing

Proposed Solution

Modify the text chunk handler in Task.ts to detect and parse XML tool calls even when using native protocol. This ensures backward compatibility with models that may still output XML-formatted tool calls.

Implementation Steps:

1. Enhance the native protocol text handler (around Task.ts:2774-2795):

   • When in native protocol mode, after adding text to the text block, check if the accumulated assistantMessage contains XML tool patterns
   • If XML tool tags are detected, parse them using the existing parseAssistantMessage function
   • Convert any parsed XML ToolUse blocks to the native format expected by the rest of the system

2. Add XML tool detection helper:

   • Create a function that checks if text contains recognizable XML tool patterns (e.g., <tool_name>...</tool_name>)
   • This can leverage the existing toolNames array from @roo-code/types

3. Fallback parsing logic:

   • Only trigger XML parsing when no native tool calls have been received AND XML patterns are detected
   • This ensures native tool calls are preferred when available

Benefits:

• Maintains full compatibility with native tool calling (preferred)
• Falls back to XML parsing for models that output XML despite native tools being enabled
• No breaking changes for existing workflows
• Minimal performance impact (XML detection only when needed)

Files to Modify:

• src/core/task/Task.ts - Main streaming handler
• Potentially src/core/assistant-message/index.ts - Export any new utilities

Please mention @roomote in a comment saying that you approve this plan, and I will proceed with the implementation.

[MeowSalty]

@roomote Execute this plan

[roomote]

Thanks for approving the plan! I'll proceed with the implementation now. I'll open a draft PR to share progress once I have the changes ready.

[roomote]

Opened a PR with my attempt at addressing this: #10326

The implementation adds an XML fallback parser that runs alongside the native tool protocol handler. When native tool protocol is enabled but the model outputs XML-formatted tool calls in the text stream instead of native tool_call chunks, the fallback parser detects and parses them.

Feedback welcome!

[torxeon]

Please revert to old XML custom tools selection. i don't why you guys have changed that...

[MeowSalty]

Please revert to old XML custom tools selection. i don't why you guys have changed that...

Some models are stupid and occasionally output XML tools, but they also support native tools. To ensure compatibility, XML has been used only in the past, but it would be nice to support both outputs at the same time.

Therefore, if you revert the version, it is actually not a very good choice.

[torxeon]

lets make XML optional