#Compaction/Summarization Failures

#Overview

This guide covers conversation compaction and summarization failures in OpenClaw agents, based on src/agents/compaction.ts. As conversations grow, OpenClaw uses LLM-powered summarization to keep context within model limits. When this fails, agents may stop responding or produce errors.

#Symptoms

#Summarization Errors

• Log message: "Full summarization failed"
• Log message: "Partial summarization also failed"
• Agent continues with minimal fallback summary

#Context Window Exceeded

• Error from LLM provider about context length
• Token count exceeds model's maximum
• Messages rejected during API call

#Agent Stops Responding

• No response after long conversation
• Error in agent logs about message processing
• Session appears frozen after many exchanges

#Oversized Message Warnings

• Log entries indicating specific messages are too large
• Messages excluded from summarization
• Fallback notes added to summary

#Root Causes

#1. Oversized Messages (Lines 236-247)

Individual messages that are too large for summarization:

function isOversizedForSummary(message: ConversationMessage): boolean {
  const estimate = estimateTokens(message);
  // Messages that would consume significant portion of context
  return estimate > OVERSIZED_THRESHOLD;
}

Detection: Token estimation per message Threshold: Configured based on context window size Impact: Message excluded from summarization, noted in fallback

#2. LLM API Errors

Provider-side failures during summarization request:

Common causes:

• Network timeout to LLM API
• Rate limiting (429 errors)
• Model unavailable (503 errors)
• Invalid API credentials

Recovery: Automatic retry with backoff

#3. Abort Signal Triggered

External cancellation of summarization:

Common causes:

• User cancels agent action
• Session timeout
• System shutdown signal

Recovery: Immediate fallback, no retry

#Recovery Mechanisms

#Progressive Fallback System (Lines 205-272)

OpenClaw implements a three-tier fallback strategy: