fix: surface actual API error messages instead of generic NoOutputGeneratedError

[hannesrudolph]

Summary

When AI SDK providers return HTTP errors (e.g., Z.ai "Insufficient balance" or "token expired"), users saw the useless generic message "No output generated. Check the stream for errors." instead of the actual API error. This PR fixes error message extraction across all 17 AI SDK providers and eliminates duplicate error display on mid-stream failures.

Root Cause

The AI SDK's streamText() has a dual error path: the actual API error appears in the stream as an error chunk (with the correct message), but then result.usage throws a separate NoOutputGeneratedError with a generic message and no reference to the original error. The generic error was overwriting the real one.

Changes

Core utility (src/api/transform/ai-sdk.ts)

• Add consumeAiSdkStream() — shared utility that captures stream error messages during iteration and substitutes them when result.usage throws NoOutputGeneratedError
• Add extractMessageFromResponseBody() — parses JSON error responses from API response bodies
• Enhance extractAiSdkErrorMessage() with responseBody extraction for AI_APICallError, AI_RetryError, and AI_NoOutputGeneratedError

17 AI SDK providers updated

• Pattern A (12 simpler providers): Use consumeAiSdkStream() directly with optional usageHandler — zai, xai, fireworks, sambanova, deepseek, mistral, baseten, lm-studio, azure, requesty, openai-compatible, openai-native
• Pattern B (5 complex providers): Manual lastStreamError capture preserving custom stream logic (thinking signatures, grounding, throttle handling) — anthropic, anthropic-vertex, vertex, gemini, bedrock

Duplicate error display fix (src/core/task/Task.ts)

• When auto-retry is enabled and a stream fails mid-request, abortStream() no longer displays the error message (since backoffAndAnnounce() will display it with the retry countdown). This eliminates the duplicate "API Request Failed" + "Provider Error" display.

Tests

• 18 new tests in ai-sdk.spec.ts covering consumeAiSdkStream, extractMessageFromResponseBody, and enhanced error extraction

Before → After

• Before: "Provider ended the request: Z.ai: No output generated. Check the stream for errors."
• After: "Provider ended the request: Z.ai: token expired or incorrect"

---

Important

This PR replaces generic error messages with specific API error messages across 17 AI SDK providers, adds a new utility for error handling, and fixes duplicate error messages during mid-stream failures.

• Behavior:
  • Replaces generic NoOutputGeneratedError with specific API error messages across 17 AI SDK providers.
  • Fixes duplicate error messages during mid-stream failures in Task.ts.
• Utilities:
  • Adds consumeAiSdkStream() in ai-sdk.ts to capture stream error messages and handle NoOutputGeneratedError.
  • Adds extractMessageFromResponseBody() to parse JSON error responses.
• Providers:
  • Updates 12 providers to use consumeAiSdkStream() directly: zai, xai, fireworks, sambanova, deepseek, mistral, baseten, lm-studio, azure, requesty, openai-compatible, openai-native.
  • Updates 5 providers with custom error handling logic: anthropic, anthropic-vertex, vertex, gemini, bedrock.
• Tests:
  • Adds 18 tests in ai-sdk.spec.ts for consumeAiSdkStream, extractMessageFromResponseBody, and enhanced error extraction.

^{This description was created by for ecd3c07. You can customize this summary. It will automatically update as commits are pushed.}

[roomote]

Rooviewer   See task

Reviewed commit ecd3c07. The indentation fix in minimax.ts is correct and all previously flagged issues are now resolved.

• Debug log in Task.ts interpolates streamingFailedMessage (which is undefined when auto-retry is active) instead of rawErrorMessage, losing error detail in server logs
• Indentation inconsistency in minimax.ts -- lines 132-176 are one tab level deeper than sibling const result on line 130

Previous reviews

• 01fbb30: Review #1
• 739f64a: Review #2
• ffd4870: Review #3
• 10ccdeb: Review #4

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