openclaw - 💡(How to fix) Fix [Feature]: Add user-visible and confirmable context compaction mechanisms [1 participants]

Summary

Introduce user-visible notifications and optional confirmation and modification for automatic context compression, trimming, or message dropping.

Problem to solve

OpenClaw automatically performs context management when conversation history becomes too long, including:

• Truncating earlier messages
• Summarizing past interactions
• Dropping oversized messages

However, these operations are currently silent and opaque to users.

This creates several issues:

• Potential loss of critical information without user awareness
• Summaries may omit important details or introduce inaccuracies
• Users cannot verify or correct compressed content
• Difficult debugging when agent behavior changes due to hidden context loss

In long-running or complex tasks, this can lead to degraded output quality and unpredictable agent behavior.

Proposed solution

Introduce transparent and controllable context compaction mechanisms:

1. Visibility:

   • Notify users when context is truncated, summarized, or dropped
   • Provide metadata (e.g., tokens reduced, messages affected)

2. Preview:

   • Show the generated summary before applying it
   • Highlight which parts are being removed or compressed

3. Confirmation:

   • Allow users to approve/reject compression results
   • Configurable modes:
     • auto (current behavior)
     • notify-only
     • require-confirmation

4. Recovery:

   • Keep a temporary buffer of original messages
   • Allow rollback or re-expansion of compressed context

5. Configuration:

   • Global and per-agent settings
   • Example:

      {
       "contextCompaction": {
         "mode": "notify" ,
         "preview": true,
         "allowRollback": true
       }
     }

6. Developer hooks:

   • Emit events/logs when compaction occurs
   • Allow custom compaction strategies

Alternatives considered

No response

Impact

Affected:

• Users running long conversations or complex agent workflows
• Multi-step reasoning tasks

Severity:

• Medium to High (can silently degrade correctness)

Frequency:

• Common in long sessions

Consequence:

• Loss of important context
• Incorrect or incomplete agent outputs
• Reduced trust in system behavior

Evidence/examples

No response

Additional information

No response