fix: prevent API error messages from polluting chat history

[roomote]

Summary

This PR fixes issue #7249 where API error messages were being added to the chat history, disrupting the conversation flow with the LLM.

Problem

When an API returns an empty assistant message or encounters an error, the system was adding a hardcoded error message "Failure: I did not provide a response." to the API conversation history. This error message would then be sent to the LLM in subsequent requests, polluting the context and disrupting the natural conversation flow.

Solution

• Removed the line that adds the error message to the API conversation history
• Kept the error display to the user via say("error") so users are still informed about the issue
• Added a comment explaining why we don't add error messages to the conversation history

Testing

• Ran existing tests: npx vitest run core/task/__tests__/Task.spec.ts - all tests pass ✅
• Ran linting: all checks pass ✅
• Ran type checking: all checks pass ✅

Impact

This change ensures that API-level errors are properly displayed to users without contaminating the LLM's conversation context, resulting in more coherent and consistent interactions.

Fixes #7249

---

Important

Fixes issue #7249 by preventing API error messages from being added to conversation history in Task.ts, ensuring cleaner LLM interactions.

• Behavior:
  • Removes addition of error messages to apiConversationHistory in recursivelyMakeClineRequests() in Task.ts to prevent context pollution.
  • Error messages are still displayed to users via say("error").
  • Adds a comment explaining the rationale for not adding error messages to the conversation history.
• Testing:
  • Ran existing tests in core/task/__tests__/Task.spec.ts - all tests pass.
  • Linting and type checking passed.
• Impact:
  • Ensures API errors do not disrupt LLM conversation flow, improving interaction coherence.

^{This description was created by for 025e605. You can customize this summary. It will automatically update as commits are pushed.}

[roomote]

Reviewing my own code is like debugging in production - technically possible but morally questionable.

[roomote]

Could we add a test case to verify that error messages are NOT added to the API conversation history when an empty assistant message is received? This would help prevent regression of this issue.

[roomote]

Consider adding the issue number to the comment for future reference:

Suggested change

	// the context sent to the LLM. The error is already shown to the user via say("error").
	// Don't add error messages to the API conversation history as it pollutes
	// the context sent to the LLM (fixes #7249). The error is already shown to the user via say("error").
	// This prevents the issue where API errors disrupt the conversation flow.