[BUG] API call fails on all versions later than 3.36

[kapil-bh]

Problem (one or two sentences)

Extension version: 3.41.3
Provider: vscode-lm
Model: copilot/claude-sonnet-4.5

Invalid request: Tool call part must be followed by a User message with a LanguageModelToolResultPart with a matching callId.

Any idea?

Context (who is affected and when)

We have been using Roocode with VS code for quite sometime, with the updated version of Roo code extension in VS code we shave started seeing this error when we make a API call to JIRA or confluence. Reverting back the version to 3.36 works.

Reproduction steps

1. Install Roo code extension (any version later than 3.36) in VS code.
2. configure MCP server for JIRA and confluence.
3. Make a API call using any of these servers

Expected result

Any extension version later than 3.36 should also work with MCP servers

Actual result

API call error

Variations tried (optional)

No response

App Version

3.41.3

API Provider (optional)

None

Model Used (optional)

No response

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

[roomote]

I've analyzed this issue and here's my proposed implementation plan:

Root Cause Analysis

The error "Tool call part must be followed by a User message with a LanguageModelToolResultPart with a matching callId" is a VSCode Language Model API requirement that is being violated when using MCP tools.

Key findings from the codebase analysis:

1. Native Tool Protocol Change: Version 3.37 introduced native tool support for VSCode LM API (feat(vscode-lm): add native tool support #10191). The vscode-lm handler now defaults to defaultToolProtocol: "native" instead of XML.

2. Stricter Message Ordering: The VSCode LM API has strict requirements for tool call sequences:

   • When an Assistant message contains LanguageModelToolCallPart(s)
   • The immediately next message MUST be a User message with LanguageModelToolResultPart(s) with matching callIds

3. Potential Message Sequence Issues: Looking at convertToVsCodeLmMessages() in src/api/transform/vscode-lm-format.ts:

   • The conversion correctly orders tool calls at the end of assistant messages
   • Tool results are placed at the beginning of user messages
   • However, there may be edge cases where the message sequence breaks (e.g., after conversation condensation, resume scenarios, or when multiple MCP tool calls occur)

4. MCP Tool Handling: MCP tools (mcp_tool_use blocks) are converted to standard tool_use blocks before being added to API history. The conversion in vscode-lm-format.ts only checks for type === "tool_use", which should be correct.

Proposed Solution

Add pre-flight validation in the VSCode LM handler to ensure tool_use/tool_result pairing is correct before sending the request:

1. Modify src/api/providers/vscode-lm.ts in the createMessage() method:

   • After calling convertToVsCodeLmMessages(cleanedMessages)
   • Add validation to ensure every assistant message with ToolCallParts is immediately followed by a user message with matching ToolResultParts
   • If mismatches are found, either fix them (add placeholder results) or log detailed diagnostics

2. Enhance src/api/transform/vscode-lm-format.ts:

   • Add a validation function that checks the converted messages for proper tool call/result sequencing
   • This could be called before returning the converted messages

3. Add error recovery:

   • If validation fails, attempt to repair the message sequence
   • As a fallback, consider filtering out orphaned tool calls/results to prevent API errors

Implementation Details

// In vscode-lm.ts, after convertToVsCodeLmMessages:
const vsCodeLmMessages = [
    vscode.LanguageModelChatMessage.Assistant(systemPrompt),
    ...convertToVsCodeLmMessages(cleanedMessages),
]

// Add validation
const validatedMessages = validateAndFixToolCallSequence(vsCodeLmMessages)

The validation function would:

1. Iterate through messages looking for ToolCallParts in assistant messages
2. Check that the next message is a user message with matching ToolResultParts
3. Fix any mismatches by adding placeholder results or removing orphaned calls

Testing Plan

1. Add unit tests in src/api/transform/__tests__/vscode-lm-format.spec.ts for edge cases:

   • Multiple consecutive tool calls
   • Tool calls followed by text-only user messages
   • Resume/condensation scenarios

2. Test with the specific reproduction scenario:

   • Configure MCP servers for JIRA/Confluence
   • Make API calls using vscode-lm provider with Claude/Copilot models

3. Verify existing tests continue to pass

Please mention @roomote in a comment approving this plan, and I'll proceed with the implementation.