openclaw - 💡(How to fix) Fix Feature: session health and context-burn auditing for long agent runs [1 comments, 1 participants]

TL;DR

Implementing session health auditing with indicators for context burn, repeated tool output, and stale context detection can help mitigate issues with long-running agent work.

Guidance

• Identify key health signals that indicate session degradation, such as context bloat and repeated low-value tool chatter, to inform the auditing system.
• Develop a lightweight health card to display these signals in the status/control UI, allowing users to intervene early.
• Distinguish between advisory signals and hard failures to provide actionable suggestions without interrupting the session unnecessarily.
• Consider starting with a small set of measurable health signals and iteratively expand the auditing system based on user feedback and session data.

Example

No specific code snippet can be provided without more context, but a potential approach could involve creating a SessionHealth class to track and expose health signals, such as:

class SessionHealth:
    def __init__(self):
        self.context_burn = 0
        self.repeated_tool_output = 0
        self.stale_context = False

    def update(self, new_data):
        # Update health signals based on new data
        pass

    def get_health_signals(self):
        return {
            'context_burn': self.context_burn,
            'repeated_tool_output': self.repeated_tool_output,
            'stale_context': self.stale_context
        }

Notes

The proposed feature requires careful consideration of what constitutes a "healthy" session and how to effectively communicate health signals to users without overwhelming them.

Recommendation

Apply a workaround by implementing a basic session health auditing system with a limited set of health signals, and iteratively expand it based on user feedback and session data. This approach allows for early intervention and improvement of the system without requiring a full overhaul.