feat: migrate OpenAiHandler to AI SDK

[daniel-lxs]

Summary

Migrates the OpenAiHandler (the "OpenAI Compatible" provider in src/api/providers/openai.ts) from direct openai npm package usage to the Vercel AI SDK, following the same pattern as already-migrated providers like DeepSeek.

Provider Strategy

Endpoint Type	AI SDK Package	Method
Standard OpenAI-compatible	@ai-sdk/openai (createOpenAI)	.chat(modelId)
Azure AI Inference (.services.ai.azure.com)	@ai-sdk/openai-compatible (createOpenAICompatible)	Adjusted baseURL + queryParams for api-version
Azure OpenAI (.openai.azure.com)	@ai-sdk/azure (createAzure)	useDeploymentBasedUrls: true

Key Changes

• Streaming: Uses streamText() + processAiSdkStreamPart() (replaces client.chat.completions.create with stream)
• Non-streaming: Uses generateText() (replaces client.chat.completions.create without stream)
• O3/O1/O4 family: Uses providerOptions.openai.systemMessageMode: 'developer' and reasoningEffort
• DeepSeek R1 format: System prompt prepended as user message when R1 mode is enabled
• <think> tag matching: TagMatcher retained for extracting reasoning from text content
• reasoning_content: Handled natively by AI SDK providers via reasoning-delta events
• isAiSdkProvider(): Returns true
• getOpenAiModels(): Standalone function unchanged (still uses axios)

Behavioral Notes

• Prompt caching via cache_control markers is no longer injected into messages. AI SDK providers handle caching at the provider level where supported.
• Timeout configuration is now managed by the AI SDK instead of the openai npm package's timeout option.
• The _isGrokXAI method was removed as stream_options handling is managed internally by the AI SDK.

Test Coverage

All 4 test files updated to mock AI SDK instead of the openai npm package:

• openai.spec.ts — 44 tests
• openai-timeout.spec.ts — 5 tests
• openai-usage-tracking.spec.ts — 3 tests
• openai-native-tools.spec.ts — 6 tests (1 test updated for new mocking pattern)

58 tests passing, zero regressions across 853 provider tests and 5370 total tests.

---

Important

Migrates OpenAiHandler to Vercel AI SDK, updating streaming and non-streaming handling, and adjusts tests to mock AI SDK.

• Behavior:
  • Migrates OpenAiHandler in openai.ts to use Vercel AI SDK instead of openai npm package.
  • Streaming handled by streamText() and processAiSdkStreamPart().
  • Non-streaming handled by generateText().
  • O3 family models use systemMessageMode: 'developer' and reasoningEffort.
  • DeepSeek R1 format prepends system prompt as user message.
  • TagMatcher used for <think> tag matching.
  • reasoning_content handled via reasoning-delta events.
  • isAiSdkProvider() returns true.
  • getOpenAiModels() unchanged.
• Tests:
  • Updated 4 test files to mock AI SDK: openai.spec.ts, openai-timeout.spec.ts, openai-usage-tracking.spec.ts, openai-native-tools.spec.ts.
  • Total of 58 tests passing with no regressions.
• Misc:
  • Removed _isGrokXAI method.
  • Timeout and caching managed by AI SDK.

^{This description was created by for 8372f41. You can customize this summary. It will automatically update as commits are pushed.}

[roomote]

Rooviewer   See task

Both previously flagged issues are now resolved. No new issues found.

• reasoningTokens silently dropped from processUsageMetrics -- the field is typed in the parameter but not included in the output (comment)
• providerMetadata not used in streaming or non-streaming paths, losing any OpenAI-specific metadata like detailed token breakdowns (comment)

Previous reviews

• 9fd6218: Review #1

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

[roomote]

Fixaroo   See task

Fixed the reported issues. processUsageMetrics now includes reasoningTokens in the output and accepts providerMetadata to extract OpenAI-specific cache/reasoning token breakdowns in both streaming and non-streaming paths. All 60 tests pass.

_{View commit | Revert commit}