[ENHANCEMENT] Migrate Grok to Responses API

[carlesso]

Problem (one or two sentences)

Currently the xAI adapter uses the deprecated Completion API, we should instead move to use the Responses API

Context (who is affected and when)

The API are deprecated, may stop working in the future.

Desired behavior (conceptual, not technical)

Responses API are used instead.

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

[carlesso]

Claiming this!

[roomote-v0]

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

Context: The current XAIHandler in src/api/providers/xai.ts uses the OpenAI Chat Completions API (client.chat.completions.create with https://api.x.ai/v1 base URL). The OpenAI Native handler (openai-native.ts) has already been fully migrated to the Responses API pattern (client.responses.create), so we have a proven reference implementation in-repo.

Proposed approach:

1. Refactor src/api/providers/xai.ts to use the Responses API instead of Chat Completions:

   • Change the base URL from https://api.x.ai/v1 to the xAI Responses API endpoint (likely https://api.x.ai/v1 still works, as xAI mirrors OpenAI's API shape)
   • Replace client.chat.completions.create(...) with client.responses.create(...) for both createMessage and completePrompt
   • Reformat the request body: use instructions (system prompt) + input (structured conversation items) instead of messages array
   • Use the Responses API streaming format (SSE events like response.text.delta, response.output_item.done, etc.) instead of Chat Completions streaming chunks
   • Handle tool calls via Responses API format (function_call items instead of tool_calls in delta)

2. Adapt the conversation formatting from the openai-native.ts pattern:

   • Implement formatFullConversation to convert Anthropic message format to Responses API input items
   • Map tool_use blocks to function_call items and tool_result blocks to function_call_output items

3. Handle reasoning/thinking via the Responses API format:

   • Use reasoning_summary events instead of reasoning_content in delta
   • The xai-responses-v1 format references already exist in the codebase transform layer

4. Update usage extraction for the Responses API response shape (e.g., input_tokens/output_tokens instead of prompt_tokens/completion_tokens)

5. Update tests in src/api/providers/__tests__/xai.spec.ts to reflect the new Responses API call patterns and streaming format

6. Update model descriptions in packages/types/src/providers/xai.ts if any reference the deprecated API

This approach closely follows the proven pattern in openai-native.ts and should be a fairly direct adaptation since xAI mirrors the OpenAI Responses API.

Note: I see @carlesso has claimed this issue. If they're already working on it, happy to step aside or collaborate. Otherwise, please mention @roomote in a comment saying that you approve this plan, and I'll proceed with the implementation.