The task can not continue when a tool call returns too much data

[BHznJNs]

App Version

3.25.3

API Provider

OpenAI Compatible

Model Used

GLM-4.5

Roo Code Task Links (Optional)

https://app.roocode.com/share/4e71e059-4080-455a-b6fc-053edf113c7a

🔁 Steps to Reproduce

1. Windows 11 24H2, extension version 3.25.3
2. When the model try to use obsidian_simple_search, the tool returns too much of data, which makes the context exceeds the model's context limit (131072)
3. The task can not be continued.
4. In this situation, if I try to use the current provider setting to perform a context compression, it will keeps the context condense state for a long time.

💥 Outcome Summary

When the tool call result may makes the context exceeds the model's context window limit, try to compress the result first, then push it into the messages and continue the task.

📄 Relevant Logs or Errors (Optional)

[roomote]

Hi! I'm here to help out the maintainers and am going to see if I can fix this issue or at least get close enough to be helpful!

The issue you've reported about tool call results exceeding context limits is important for maintaining a smooth user experience. I'll investigate the codebase to understand how tool results are currently handled and implement a solution to compress large results before they cause context overflow.

Working on this now! 🚀

[roomote]

✅ Issue Fixed!

I've successfully implemented a solution for the tool result compression issue. Here's what was done:

🔧 Solution Implemented

Pull Request: #6466

📋 What the fix does:

1. Automatic Compression: Tool results that exceed context window limits are now automatically compressed before being added to the conversation context
2. Intelligent Truncation: Uses the existing truncateOutput function to preserve the most important parts (20% from beginning, 80% from end)
3. Context-Aware Limits: Compression thresholds scale based on the model's context window size
4. Clear Indicators: When compression occurs, the model is informed with clear notes about what was compressed

🎯 Key Features:

• Default limits: 50,000 characters and 1,000 lines for tool results
• Dynamic scaling: Larger context windows (like your GLM-4.5 with 131K tokens) allow larger tool results
• Preserves structure: Beginning and end of results are kept for maximum utility
• No breaking changes: Existing functionality remains unchanged

🧪 Testing:

• ✅ All new tests pass (19/19)
• ✅ Existing tests continue to pass
• ✅ All CI checks successful (linting, type checking, integration tests)

🚀 Impact:

This fix ensures that tools like obsidian_simple_search that return large amounts of data will no longer cause tasks to fail due to context window exhaustion. The task will continue with the most relevant parts of the tool output compressed intelligently.

The solution is ready for review and should resolve the issue you experienced! 🎉