fix: handle empty stream responses from GLM models

[roomote]

Summary

This PR addresses Issue #8482 where GLM-4.6 models occasionally return empty streams, causing "Unexpected API Response" errors in Roo Code.

Problem

Users experiencing intermittent failures when using GLM-4.6 with both OpenAI Compatible and Z AI providers. The error occurs when the model returns an empty stream with no text content, causing the application to fail with an unhelpful error message.

Solution

Implemented a three-layer approach to handle empty streams gracefully:

1. Provider Level: Added GLM-specific detection and fallback responses in both OpenAI and base OpenAI-compatible providers
2. Task Level: Enhanced error handling in Task.ts to detect GLM models and automatically retry with clarification prompts
3. User Experience: Replaced generic error messages with actionable guidance for users

Changes

• 🛡️ Added graceful handling for empty API responses in Task.ts
• 🔧 Implemented GLM-specific fallback responses in OpenAI and base providers
• 🔄 Added retry logic for GLM models when empty streams occur
• ✅ Added comprehensive tests for empty stream scenarios

Testing

• Added new test suite with 4 test cases covering both GLM and non-GLM models
• All existing tests pass
• Linting and type checking pass

Impact

• GLM Users: Will no longer experience crashes from empty streams
• Other Models: No impact - original error behavior preserved
• Performance: Minimal - only adds lightweight model ID checking

Fixes #8482

---

Important

Fixes handling of empty stream responses from GLM models by adding fallback responses and retry logic, improving error handling and user experience.

• Behavior:
  • Handles empty streams from GLM models in Task.ts by retrying with clarification prompts.
  • Adds fallback responses for GLM models in OpenAiHandler and BaseOpenAiCompatibleProvider.
  • Improves error messages for users when GLM models return empty streams.
• Testing:
  • Adds glm-empty-stream.spec.ts with 4 test cases for GLM and non-GLM models.
  • Ensures all existing tests pass.
• Impact:
  • Prevents crashes for GLM users due to empty streams.
  • No impact on other models; original error behavior preserved.

^{This description was created by for 8663724. You can customize this summary. It will automatically update as commits are pushed.}

[ellipsis-dev]

The GLM fallback response here is hardcoded. To ensure consistency and proper localization, consider wrapping this error message in a translation function.

Suggested change

	text: "I'm having trouble generating a response. Please try rephrasing your request or breaking it down into smaller steps.",
	text: t("I'm having trouble generating a response. Please try rephrasing your request or breaking it down into smaller steps."),

^{This comment was generated because it violated a code review rule: irule_C0ez7Rji6ANcGkkX.}

[ellipsis-dev]

In the GLM-specific branch handling empty responses (around lines 2349–2365), fallback messages and prompts are hardcoded. Consider using a translation function (e.g. t('...')) and possibly refactoring the duplicated GLM check into a helper for clarity.

^{This comment was generated because it violated a code review rule: irule_C0ez7Rji6ANcGkkX.}

[roomote]

Performing self-review: a robot grading its own homework—what could possibly be biased?

[roomote]

[P1] Potential infinite retry loop for GLM empty streams. This path pushes a clarification and immediately retries with no max attempts or backoff. If the provider keeps returning empty streams, this can loop indefinitely and spam error messages. Recommend adding a capped retry counter (e.g., 1–3 attempts) with small delay and logging, and falling back to the standard error after the cap. Also consider adding unit tests that simulate repeated empty responses to verify the retry cap and final behavior.

[roomote]

[P2] Duplicate GLM detection logic. The same modelId.includes('glm'|'chatglm') condition appears here, in BaseOpenAiCompatibleProvider, and in Task.ts. Consider extracting a small helper (e.g., isGlmModel(modelId)) to keep behavior consistent and reduce drift.

[roomote]

[P3] Unused import 'beforeEach' may trip stricter lint settings; suggest removing it.

Suggested change

	import { describe, it, expect, vi, beforeEach } from "vitest"
	import { describe, it, expect, vi } from "vitest"