"Failed to enhance prompt" with GPT-5 models; "completePrompt" not supported

[nlbuescher]

App Version

3.25.21

API Provider

OpenAI

Model Used

gpt-5-nano-2025-08-07

Roo Code Task Links (Optional)

No response

🔁 Steps to Reproduce

1. Set GPT-5 Nano as the model to be used for prompt enhancing
2. Click on "Enhance Prompt"

💥 Outcome Summary

Using any one of the new GPT-5 models for prompt enhancement fails, in general. Neither the model nor the reasoning effort make a difference.

I was able to find the cause in the vscode console (see below). Roo Code seems to use a specific field / parameter "completePrompt" for enhancing the prompt, that the error message says isn't supported by the GPT-5 models.

📄 Relevant Logs or Errors (Optional)

[Extension Host] Error enhancing prompt: {
  "stack": "Error: OpenAI Native completion error: completePrompt is not supported for gpt-5-nano-2025-08-07. Use createMessage (Responses API) instead.\n    at Jur (/home/nico/.vscode/extensions/rooveterinaryinc.roo-cline-3.25.21/core/webview/webviewMessageHandler.ts:1418:13)",
  "message": "OpenAI Native completion error: completePrompt is not supported for gpt-5-nano-2025-08-07. Use createMessage (Responses API) instead."
}

[roomote]

I'm exploring a potential fix for the GPT-5 model compatibility issue with prompt enhancement. The error indicates that these models require using the Messages API instead of the completion API. I'll investigate the implementation and open a PR to share progress.

[roomote]

✅ Fix Implemented

I have successfully implemented a fix for the GPT-5 prompt enhancement issue. The solution is now available in PR #7335.

Summary of Changes

The issue was that GPT-5 models (gpt-5, gpt-5-mini, gpt-5-nano) and Codex Mini were throwing an error when using the "Enhance Prompt" feature because they use the Responses API which only supports streaming, not traditional completions.

Solution implemented:

• Modified the completePrompt method to support Responses API models
• Created a new method that collects streaming responses into a complete string
• Added SSE fallback for robustness
• Maintained full backward compatibility with existing models

Testing

• ✅ All unit tests pass
• ✅ Added comprehensive test coverage for GPT-5 models
• ✅ Linting passes with no warnings
• ✅ All CI checks pass

Pull Request

🔗 PR #7335: #7335

The fix enables prompt enhancement to work seamlessly with GPT-5 and Codex Mini models without any user-facing changes required.

[daniel-lxs]

Thank you @nlbuescher for reporting this, this issue should be fixed by #7067