Fix GPT-5 Responses API issues with condensing and image support

[daniel-lxs]

Summary

This PR fixes critical issues with the GPT-5 Responses API implementation introduced in PR #6864.

Issues Fixed

1. Manual Condensing Bug 🐛

Problem: Manual condensing (when users click the condense button) wasn't working properly - the API was still sending the old previous_response_id after condensing, causing errors.

Solution: Added skipPrevResponseIdOnce flag in Task.ts that gets set after manual condensing, ensuring the next API call sends the full conversation context without the previous_response_id.

2. Image Support 🖼️

Problem: Images were not being sent correctly to the API. The format was using incorrect content types (text instead of input_text, output_text, input_image).

Solution:

• Use proper content types: input_text for user text, output_text for assistant text, input_image for images
• Maintain structured format for both full conversations and single messages with images

3. Error Handling 🔄

• Properly handles 400 errors when previous_response_id is not found
• Automatically retries without the ID
• Clears stored response ID to prevent reuse

Changes Made

1. Task.ts: Added skipPrevResponseIdOnce = true after manual condensing
2. openai-native.ts:
   • Fixed content type formatting (input_text, output_text, input_image)
   • Maintained structured format for full conversations to preserve image support
   • Improved response ID management
3. Tests: Updated all tests to match the new structured format

Testing

• ✅ All 30 OpenAI Native provider tests passing
• ✅ All 3431 tests across 265 test files passing
• ✅ Manual condensing works correctly
• ✅ Image attachments work with GPT-5 models
• ✅ Conversation continuity maintained with previous_response_id

Impact

This fix ensures that:

• Users can successfully use manual condensing with GPT-5 models
• Images can be attached and sent correctly to GPT-5 models
• The conversation context is properly maintained across requests
• No breaking changes to existing functionality

Fixes issues reported after PR #6864

---

Important

Fixes GPT-5 Responses API issues with manual condensing, image support, and error handling, including updates to Task.ts, openai-native.ts, and tests.

• Behavior:
  • Fix manual condensing by adding skipPrevResponseIdOnce flag in Task.ts to manage previous_response_id.
  • Correct content types for image support in openai-native.ts (input_text, output_text, input_image).
  • Improve error handling for 400 errors in openai-native.ts by retrying without previous_response_id.
• Models:
  • Add supportsTemperature flag to modelInfoSchema in model.ts.
  • Update openAiNativeModels in openai.ts to include supportsTemperature.
• Tests:
  • Update tests in openai-native.spec.ts to reflect changes in API handling and error management.
• Misc:
  • Update package.json to use openai version ^5.12.2.

^{This description was created by for 52c6b8b. You can customize this summary. It will automatically update as commits are pushed.}

[roomote]

Thank you for your contribution! I've reviewed the changes and the implementation correctly addresses both the manual condensing bug and image support issues.

✅ What Works Well

• The fix for manual condensing using the skipPrevResponseIdOnce flag is well implemented
• Image support with proper content types (input_text, output_text, input_image) is correctly handled
• Automatic retry mechanism when previous_response_id is not found
• Comprehensive test coverage for the new structured format
• Clear separation of concerns between SDK and SSE fallback paths

💭 Suggestions for Improvement

1. Console.log statements: There are many console.log statements throughout openai-native.ts (lines 92, 129, 153, 163, 172-173, 175, 281, 621, 913-914, 1040, etc.). Consider using a debug flag or removing them before merging to production.

2. Method naming clarity: The methods formatStructuredInput and prepareStructuredInput have very similar names which could be confusing. Consider renaming for better clarity.

3. API documentation: The new store parameter in ApiHandlerCreateMessageMetadata would benefit from more detailed documentation about its purpose (storing responses for 30 days) and default behavior.

4. Type safety: The GPT-5 metadata persistence uses any type casting extensively in Task.ts. Consider defining proper TypeScript types for the metadata structure.

5. Magic strings: The string "gpt-5" is used in multiple places to check model type. Consider extracting this to a constant or utility function for better maintainability.

Overall, this is a solid fix that addresses the critical issues. The suggestions above are mostly for code quality and maintainability improvements. Great work on fixing these GPT-5 API issues! 🎉