openclaw - 💡(How to fix) Fix Feature: Make HARD_MAX_TOOL_RESULT_CHARS configurable [1 participants]

An exec command returning 200K chars (~100-133K tokens) plus existing conversation + system prompt can push total tokens past the context window in one turn. Auto-compaction only triggers after the overflow error.

TL;DR

To prevent the model_context_window_exceeded error, consider adding a configuration option to limit the maximum number of characters allowed in tool results.

Guidance

• Review the proposed solution to add a maxToolResultChars configuration option to override the hardcoded limit.
• Verify that the contextWindow size of the model (202,800) is sufficient for the expected conversation and system prompt lengths.
• Test the effect of reducing the maxToolResultChars value on the tool results and conversation flow.
• Consider the trade-off between limiting tool result length and potentially truncating important information.

Example

The proposed configuration option can be added as follows:

{
  "tools": {
    "maxToolResultChars": 100000
  }
}

This would limit the maximum number of characters in tool results to 100,000.

Notes

The effectiveness of this solution may depend on the specific use case and the importance of preserving long tool results. Further testing and evaluation may be necessary to determine the optimal value for maxToolResultChars.

Recommendation

Apply the proposed workaround by adding the maxToolResultChars configuration option to limit the maximum number of characters in tool results, as it provides a flexible way to prevent the model_context_window_exceeded error without modifying the underlying code.