[BUG] Regression: "Gemini generate context stream error: No output generated" on large contexts (>130k tokens) in v3.47.3

[ModelkinIY]

Problem (one or two sentences)

Since updating to the latest versions (specifically noticeable in v3.47.3), the Gemini provider (Google AI) consistently fails when the context window exceeds ~130k tokens. The error message is: Provider ended the request: Gemini generate context stream error: No output generated. Check the stream for errors.

This appears to be a regression, as similar tasks with the same context size and complexity were handled flawlessly in earlier versions (v3.3.x).

Context (who is affected and when)

All users who work with large codebases or long log files using Gemini models. This happens during complex debugging or refactoring tasks when the conversation history grows beyond 100k-130k tokens. It prevents the AI from answering even simple questions once the "memory" (context) of the chat is full.

Reproduction steps

1)Environment:
Extension Version: 3.47.3
VS Code Version: 1.96+
Provider: Google AI (Native SDK)
Model: gemini-2.5-flash / gemini-3-flash-preview
Context Size: ~133,000 - 150,000 tokens

2.1) Initialize a task in a large workspace (e.g., ESP-IDF project with multiple submodules).
2.2)Provide a large log file or read several large source files so the tokensIn count exceeds 130k.
2.3)Issue a command that requires Tool Use (e.g., codebase_search or readFile).
2.4)The model starts "Thinking", but the stream immediately breaks with "No output generated" before any tool call is made.

Expected result

The AI should process the provided context (even if it's large) and start responding or calling a tool. Even if it takes 30-60 seconds to "think", the connection should remain open until the first token is generated, as it did in previous versions (v3.3.x).

Actual result

The stream terminates almost immediately or after a short delay with the error: "Gemini generate context stream error: No output generated". No response is shown to the user, and the task becomes stuck, forcing a manual "Reset Context" and loss of project history to make the model work again.

Variations tried (optional)

Tried increasing AbortSignal timeout to 120s in gemini.ts, but the error persists.

App Version

3.47.3

API Provider (optional)

Google Gemini

Model Used:

gemini-2.5-flash, gemini-3-flash-preview.

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

{
  "error": {
    "timestamp": "2026-02-08T09:30:54.630Z",
    "details": "Provider ended the request: Gemini generate context stream error: No output generated. Check the stream for errors."
  },
  "history": [
    { "role": "user", "tokensIn": 133010 }
  ]
}

[roomote]

I've analyzed this issue and here's my proposed implementation plan:

Root Cause Analysis

The regression was introduced by the AI SDK migration in PR #11180 (commit afe51e0fe), which migrated the Gemini provider from @google/genai to @ai-sdk/google + streamText(). Several factors contribute to the large-context failure:

1. NoOutputGeneratedError during stream iteration is not handled gracefully: PR fix: validate Gemini thinkingLevel against model capabilities and handle empty streams #11303 added handling for NoOutputGeneratedError during result.usage, but when the error occurs during the fullStream iteration itself (the common case for large-context failures), it falls through to the generic outer catch and gets wrapped as the cryptic "Gemini generate context stream error: No output generated."

2. No Gemini-specific context window error detection: checkContextWindowExceededError() in context-error-handling.ts only checks for OpenAI, OpenRouter, Anthropic, and Cerebras errors. When Gemini returns context-related errors (including empty responses on oversized contexts), the system cannot trigger automatic context truncation.

3. No abort signal or timeout configuration is passed to streamText(), which means the AI SDK uses its own defaults that may differ from how @google/genai handled long-processing large requests.

Proposed Fix (3 changes)

1. Handle NoOutputGeneratedError in the outer catch of GeminiHandler.createMessage() (src/api/providers/gemini.ts)

• Add specific handling for NoOutputGeneratedError in the outer catch block (line 203)
• When caught during stream iteration, yield the same informative empty-response message instead of throwing a hard error that kills the stream
• This makes the behavior consistent with the result.usage handling added in PR fix: validate Gemini thinkingLevel against model capabilities and handle empty streams #11303

2. Add Gemini-specific error detection to checkContextWindowExceededError() (src/core/context/context-management/context-error-handling.ts)

• Add a checkIsGeminiContextWindowError() function that detects Gemini-specific error patterns (e.g., "token limit", "request payload size exceeds", "resource exhausted" for large contexts)
• Wire it into the existing checkContextWindowExceededError() so that the Task layer can automatically truncate the context window and retry

3. Pass abortSignal through to the Gemini streamText() call (src/api/providers/gemini.ts)

• Accept an optional abortSignal in createMessage() metadata and forward it to streamText()
• This is not strictly required for the bug fix but prevents the stream from hanging indefinitely for large requests that the user cancels

This approach addresses both the immediate UX regression (cryptic error messages) and the underlying missing resilience (no context truncation for Gemini).

Please mention @roomote in a comment saying that you approve this plan, and I'll proceed with the implementation.