# 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

| Agent | Model | Context Window | Max Tokens | Compaction Count |
|-------|-------|---------------|------------|------------------|
| **Socrates** | kimi-k2.5:cloud | 128,000 | 8,192 | 1+ (observed) |
| **Daedalus** | minimax-m2.7:cloud | 128,000 | 8,192 | Unknown |
| **Wadsworth** (main) | glm-5.1:cloud | 128,000 | 8,192 | Just triggered |

## 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:**
```jsonl
// 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

| Session | Lines | Size | Status |
|---------|-------|------|--------|
| Socrates (contaminated backup) | ~4,700 | 263 KB | **Contaminated** |
| Socrates (current) | ~471 | 36 KB | Clean so far |
| Daedalus (current) | ~134 | 12 KB | Clean |
| Main (just compacted) | ~29 | 25 KB | **Just compacted** |

**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:

```typescript
// 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

```bash
#!/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