Context Limit & Compaction Analysis — Daedalus & Socrates

Executive Summary

The 20K limit you mentioned is NOT the model context window — it's likely an OpenClaw-internal compaction threshold. The models themselves support 128K context windows, but compaction is triggering far too early (~20K tokens), causing:
1. Frequent context loss during complex tasks
2. The ghost bug — phantom messages from other agents appearing after compaction

Current Configuration

The Problem

1. Premature Compaction

• Models support 128K context (~100+ pages of text)
• Compaction appears to trigger at ~20K tokens (estimated from session sizes)
• This is 16× smaller than the model's actual capacity
• Complex tasks (coding, architecture discussions) hit this wall constantly

2. The Ghost Bug (Cross-Agent Contamination)

Root Cause: When OpenClaw compacts context in a group chat session, it includes ALL assistant messages from the conversation thread — not just the current agent's responses.

Evidence from contaminated session:

// Matt asks @Daedalus89Bot
{"role":"assistant","content":"Waiting on Daedalus to chime in. 🎨"}
// ^ This is Wadsworth's message, injected into Socrates's session as generic "assistant"

Result: After compaction, Socrates sees phantom "assistant" responses that were actually from Daedalus or Wadsworth, causing identity confusion.

Session Size Analysis

Note: The contaminated Socrates session (263 KB) suggests a long group chat history that got compacted, mixing in messages from all three agents.

Recommendations

Immediate Fixes (Configuration)

1. Increase Compaction Threshold to 100K+

OpenClaw likely has a compaction threshold setting. Look for:
- compactionThreshold
- contextLimit
- maxContextBeforeCompaction
- Environment variable: OPENCLAW_COMPACTION_THRESHOLD

Recommended: Set to 100,000 tokens (leaving 28K buffer for response + system prompts)

2. Disable Compaction for Group Chats (If Possible)

Group chats with multiple agents should either:
- Never compact (preserve full history)
- Use agent-specific compaction (only summarize that agent's messages)

3. Per-Agent Session Isolation

Current architecture shares session context for group chats. Better approach:
- Each agent maintains their own session file
- Cross-agent references use {"agent": "daedalus"} metadata
- Compaction only touches that agent's messages

Code-Level Fix (OpenClaw Core)

The compaction logic needs to filter assistant messages:

// Pseudocode for the fix
function compactContext(session, agentId) {
  const messages = session.messages.map(msg => {
    // Tag assistant messages with agent ID
    if (msg.role === "assistant") {
      return {
        ...msg,
        metadata: { agent: msg.agentId || "unknown" }
      };
    }
    return msg;
  });

  // When loading compacted context for a specific agent:
  const agentMessages = messages.filter(msg => 
    msg.role !== "assistant" || 
    msg.metadata?.agent === agentId
  );

  return summarize(agentMessages);
}

Workaround (User-Level)

Until fixed in OpenClaw core:

1. Use DMs for complex tasks — Avoid group chats for deep work
2. Start fresh sessions — If you see ghost messages, clear the session:
   bash rm /home/hoffmann_admin/.openclaw/agents/{socrates,daedalus}/sessions/*.jsonl
3. Monitor compaction counts — Check for authProfileOverrideCompactionCount > 0

Detection Script

#!/bin/bash
# check-compaction.sh — Detect contaminated sessions

echo "=== Checking for compaction contamination ==="

# Find sessions with compaction
for agent in socrates daedalus main; do
  echo "\n--- $agent ---"

  # Check session index for compaction count
  grep -h "compactionCount\|authProfileOverrideCompactionCount" \
    /home/hoffmann_admin/.openclaw/agents/$agent/sessions/sessions.json 2>/dev/null | head -5

  # Check for foreign agent mentions
  echo "Foreign agent mentions:"
  if [ "$agent" = "socrates" ]; then
    grep -c "Daedalus\|Wadsworth" /home/hoffmann_admin/.openclaw/agents/$agent/sessions/*.jsonl 2>/dev/null | grep -v ":0$" || echo "None found"
  elif [ "$agent" = "daedalus" ]; then
    grep -c "Socrates\|Wadsworth" /home/hoffmann_admin/.openclaw/agents/$agent/sessions/*.jsonl 2>/dev/null | grep -v ":0$" || echo "None found"
  fi
done

Files Affected

Contaminated backups:
- /home/hoffmann_admin/.openclaw/agents/session-backups-20260420160143/socrates-*.contaminated-moved.jsonl
- /home/hoffmann_admin/.openclaw/agents/session-backups-20260420160143/daedalus-*.contaminated-moved.jsonl

Current sessions to monitor:
- Socrates: 1ab409b1-8e73-4b78-b5fd-761c5b1fdf6b.jsonl (471 lines, clean)
- Daedalus: 00701f9c-7727-4305-ab6b-d1437aeb722a.jsonl (134 lines, clean)

Next Steps

1. Find OpenClaw compaction config — Check docs, env vars, or config files for the 20K threshold
2. Increase to 100K+ — This alone will reduce ghost bug frequency by ~5×
3. File upstream bug — The cross-agent contamination is an OpenClaw core issue
4. Monitor sessions — Run detection script weekly

Analyzed: 2026-04-21
Status: Awaiting configuration fix
Impact: High — affects all multi-agent group chats