refactor: migrate Qwen Code provider to AI SDK

[hannesrudolph]

Summary

Migrates the Qwen Code provider (QwenCodeHandler) from manual OpenAI SDK streaming to the Vercel AI SDK via OpenAICompatibleHandler (Archetype 1 migration per plans/ai-sdk-migration-guide.md).

Changes

• Base class: BaseProvider → OpenAICompatibleHandler using @ai-sdk/openai-compatible
• Streaming: Manual SSE/delta assembly → streamText() + processAiSdkStreamPart()
• Reasoning: Manual <think> tag parsing → extractReasoningMiddleware({ tagName: "think" })
• Tools: NativeToolCallParser + manual deltas → AI SDK native tool handling
• Messages: convertToOpenAiMessages → convertToAiSdkMessages
• Errors: Raw OpenAI errors → handleAiSdkError()
• OAuth: Fully preserved — dynamic getLanguageModel() creates fresh provider with current OAuth credentials per request; createMessage() and completePrompt() both include 401 retry
• isAiSdkProvider(): Returns true (inherited from OpenAICompatibleHandler)
• Tests: Rewritten to mock AI SDK instead of OpenAI SDK (11 tests, all passing)

Breaking Change Audit

All potential breaking changes were investigated and resolved:

Finding	Verdict
Text dedup guard removed	✅ Non-issue — DashScope compatible-mode API uses standard incremental deltas
isAiSdkProvider() flips true	✅ Correct behavior — reasoning preserved as intended for AI SDK providers
includeUsage not explicit	✅ Non-issue — @ai-sdk/openai-compatible handles internally

Removed Dependencies (from this file)

• openai SDK direct usage
• convertToOpenAiMessages from openai-format.ts (reduces consumer count from 8 → 7)
• NativeToolCallParser import

---

Important

Refactor Qwen Code provider to use Vercel AI SDK, updating OAuth handling and tests.

• Migration:
  • Refactor QwenCodeHandler to extend OpenAICompatibleHandler using Vercel AI SDK.
  • Replace manual SSE/delta handling with streamText() and processAiSdkStreamPart().
  • Use extractReasoningMiddleware for reasoning tag parsing.
  • Replace NativeToolCallParser with AI SDK native tool handling.
  • Convert messages using convertToAiSdkMessages.
  • Handle errors with handleAiSdkError().
• OAuth Handling:
  • Load and refresh OAuth credentials dynamically in qwen-code.ts.
  • Implement forceRefreshAndAuthenticate() for token refresh.
  • Use safeWriteJson for saving refreshed credentials.
• Tests:
  • Update tests in qwen-code-native-tools.spec.ts to mock AI SDK instead of OpenAI SDK.
  • Add tests for OAuth credential loading and token refresh.
• Misc:
  • Remove direct openai SDK usage.
  • Remove convertToOpenAiMessages dependency.

^{This description was created by for c784551. You can customize this summary. It will automatically update as commits are pushed.}

[roomote]

Rooviewer   See task

Review complete. No issues found. The latest commit correctly normalizes test path assertions with path.resolve() to match getQwenCachedCredentialPath() behavior, fixing Windows CI compatibility.

Previous reviews

• b70be89: Review #1
• 33ceb0b: Review #2

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