[BUG] Context Condensing Silently fail's when using a model with a smaller context window than current context length

[DarkEden-coding]

Problem (one or two sentences)

When condensing a chat the user expects some sort of feedback if they model they have selected for the task fails, this does not happen and the ui is just stuck in Condensing Chat. Could be fixed in a few way:

1. Warn the user that the model they have has too small of a context window and they should use a different model.
2. Multi-Stage condensing where the model make summaries for the context split into chunks that can fit the model window. Could have a warning and a button to do this multi-stage method.

Context (who is affected and when)

All users have smaller context window models selected for their condensing tasks.

Reproduction steps

Create a chat that is relatively long and has a large amount of content. Then, select a model that has a small context window for condensing and attempt to condense. What happens is nothing appears to the user other than "condensing the conversation". And that stays forever.

Expected result

Either a warning or a secondary prompt should appear when this happens.

Actual result

No user feedback is given, and it seems like it's still working when it is not.

Variations tried (optional)

No response

App Version

3.39.3 (85d253c)

API Provider (optional)

None

Model Used (optional)

N/A

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

Message when trying to condense a chat with an insufficient model:


[error] [Extension Host] [OpenRouter] API error: {"message":"400 This endpoint's maximum context length is 131072 tokens. However, you requested about 145406 tokens (119191 of text input, 26215 in the output). Please reduce the length of either one, or use the \"middle-out\" transform to compress your prompt automatically.","name":"Error","stack":"Error: 400 This endpoint's maximum context length is 131072 tokens. However, you requested about 145406 tokens (119191 of text input, 26215 in the output). Please reduce the length of either one, or use the \"middle-out\" transform to compress your prompt automatically.\n    at Function.generate (/Users/darkeden/wpilib/2025/vscode/code-portable-data/extensions/node_modules/.pnpm/openai@5.12.2_ws@8.18.3_zod@3.25.61/node_modules/openai/src/core/error.ts:72:14)\n    at ei.makeStatusError (/Users/darkeden/wpilib/2025/vscode/code-portable-data/extensions/node_modules/.pnpm/openai@5.12.2_ws@8.18.3_zod@3.25.61/node_modules/openai/src/client.ts:445:28)\n    at ei.makeRequest (/Users/darkeden/wpilib/2025/vscode/code-portable-data/extensions/node_modules/.pnpm/openai@5.12.2_ws@8.18.3_zod@3.25.61/node_modules/openai/src/client.ts:668:24)\n    at processTicksAndRejections (node:internal/process/task_queues:95:5)\n    at d$.chunk [as createMessage] (/Users/darkeden/wpilib/2025/vscode/code-portable-data/extensions/rooveterinaryinc.roo-cline-3.39.3/api/providers/openrouter.ts:325:13)\n    at cNt (/Users/darkeden/wpilib/2025/vscode/code-portable-data/extensions/rooveterinaryinc.roo-cline-3.39.3/core/condense/index.ts:328:19)\n    at t.condenseContext (/Users/darkeden/wpilib/2025/vscode/code-portable-data/extensions/rooveterinaryinc.roo-cline-3.39.3/core/task/Task.ts:1651:7)\n    at t.condenseTaskContext (/Users/darkeden/wpilib/2025/vscode/code-portable-data/extensions/rooveterinaryinc.roo-cline-3.39.3/core/webview/ClineProvider.ts:1735:3)","status":400}

[roomote]

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

Root Cause Analysis:

The issue occurs in src/core/condense/index.ts in the summarizeConversation() function. When the API streaming call at lines 322-336 fails (e.g., due to context window exceeded error), the error is not caught. The function throws an unhandled exception instead of returning it via the SummarizeResponse.error field.

The downstream handler in Task.condenseContext() (line 1610) already has proper error display logic (this.say("condense_context_error", error, ...)), but it never receives the error because summarizeConversation() throws instead of returning it.

Proposed Fix:

1. Add try-catch around the streaming API call in src/core/condense/index.ts:

   • Wrap the for await (const chunk of stream) loop in a try-catch block
   • Catch errors from the API (including 400 "context length exceeded" errors)
   • Return a user-friendly error message via the error field of SummarizeResponse

2. Extract meaningful error message from the API response:

   • Parse the error to detect context window exceeded scenarios
   • Provide actionable feedback (e.g., "The selected model's context window is too small for the current conversation. Please select a model with a larger context window or reduce the conversation size first.")

Code Changes:

// In src/core/condense/index.ts, around line 322
try {
    const stream = handlerToUse.createMessage(promptToUse, requestMessages)
    
    for await (const chunk of stream) {
        if (chunk.type === "text") {
            summary += chunk.text
        } else if (chunk.type === "usage") {
            cost = chunk.totalCost ?? 0
            outputTokens = chunk.outputTokens ?? 0
        }
    }
} catch (error) {
    const errorMessage = error instanceof Error ? error.message : String(error)
    
    // Check for context window exceeded error
    if (errorMessage.includes("context_length_exceeded") || 
        errorMessage.includes("context length") ||
        errorMessage.includes("maximum context") ||
        errorMessage.includes("too long")) {
        return {
            messages: [],
            summary: "",
            cost: 0,
            error: "The selected condensing model's context window is too small for the current conversation. Please select a model with a larger context window or manually reduce the conversation size first.",
        }
    }
    
    return {
        messages: [],
        summary: "",
        cost: 0,
        error: `Failed to condense context: ${errorMessage}`,
    }
}

3. Add unit tests to verify the error handling behavior.

This approach ensures that API errors are properly caught and displayed to the user via the existing error display mechanism, rather than silently hanging.

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

[DarkEden-coding]

@roomote that looks good, attempt fix

[DarkEden-coding]

For the chunk functionality should I make an enhancement issue for it? I would like that functionality so I can save costs by using a smaller model.

[roomote]

Thanks for the approval! I'm starting the implementation now. I'll open a PR shortly to share progress.

[roomote]

Opened a draft PR with my attempt at addressing this: #10783

The fix adds error handling around the streaming API call in summarizeConversation() to catch context window exceeded errors and display a user-friendly error message instead of silently hanging.

Feedback welcome!