[BUG] The model did not provide any response content.

[keith-degrace]

Problem (one or two sentences)

All interactions with Claude models return the following. This started happening two days ago. I've been running this exact setup for many months without issue. It suddenly started happening. If I use other VS Code LM API models (e.g. GPT, Gemini), it works.

Extension version: 3.51.1
Provider: vscode-lm
Model: claude-code/claude-opus-4.6

The language model returned an empty response without any text or tool calls. This can happen due to API issues, rate limiting, or model-specific problems. The model has been automatically prompted to retry.

Context (who is affected and when)

Happens in any chat interactions with VS Code LM API Claude models.

Reproduction steps

Windows
VS Code 1.112.0
Roo Code 3.51.1

Expected result

No errors

Actual result

Model Response Incomplete The model did not provide any response content. This may indicate an issue with the API or the model's output.

Variations tried (optional)

Tried downgrading VS Code, Roo Code. Cline seems to also be affected but GitHub Copilot Chat can use these models just fine.

Tried re-installing everything with no resolution.

App Version

3.51.1

API Provider (optional)

VS Code Language Model API

Model Used (optional)

Claude Sonnet 4.5+, Opus 4.5+

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

Date/time: 2026-03-27T16:19:35.135Z
Extension version: 3.51.0
Provider: vscode-lm
Model: claude-code/claude-opus-4.6

The language model returned an empty response without any text or tool calls. This can happen due to API issues, rate limiting, or model-specific problems. The model has been automatically prompted to retry.

[roomote-v0]

I've analyzed this issue and here's what I found in the codebase:

Root cause analysis:

The VS Code LM handler in src/api/providers/vscode-lm.ts processes the response stream (line 409) and only handles two chunk types: LanguageModelTextPart and LanguageModelToolCallPart. Any other chunk type is silently skipped with a warning log. Since this issue is Claude-specific and started recently, the most likely cause is that the Copilot extension (or VS Code 1.112) updated how Claude models return responses -- possibly now including LanguageModelThinkingPart chunks for extended thinking. If the model emits thinking parts but no text parts (or text arrives in an unrecognized format), the handler treats the response as empty and triggers the "no assistant messages" error path.

Additionally, the system prompt is sent as an Assistant role message (line 382: LanguageModelChatMessage.Assistant(systemPrompt)), which is unusual and could interact poorly with a backend change on the Claude side.

Proposed implementation plan:

1. Add LanguageModelThinkingPart handling in VsCodeLmHandler.createMessage() -- detect thinking/reasoning chunks and yield them as reasoning type content (or at minimum extract their text so the response is not considered empty)
2. Improve unknown chunk type handling -- instead of silently skipping unknown chunks, attempt to extract any .value or .text property from them as a fallback, so new part types introduced by VS Code don't cause empty responses
3. Add diagnostic logging when the stream completes with zero text and zero tool calls -- log the chunk types that were received, so users/developers can see what the model actually returned
4. Update tests in src/api/providers/__tests__/vscode-lm.spec.ts to cover the new chunk type handling and the empty-stream diagnostic path

This approach addresses what we can control in the Roo Code codebase. If the root cause is purely an upstream Copilot/VS Code issue, these changes will at minimum provide better diagnostics and resilience against new part types.

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