Summary

Add agents.defaults.compaction.model config option to specify a different (faster) model for compaction summarization.

Problem to solve

Compaction uses the session's primary model for summarization. When the session uses a slow model like Opus, compaction becomes extremely slow:

• Each chunk summarization requires a full LLM API call
• Large sessions (150k+ tokens) split into 2-3 chunks of ~80k tokens each
• With qualityGuard retries, this can mean 4-6 Opus API calls
• Each Opus call for 80k tokens takes 1-3 minutes
• Total compaction time: 8-15+ minutes, hitting timeout limits

Real-world example:

• Session: 4.5MB JSONL, ~160k tokens, claude-opus-4-5
• Compaction timed out at 900s (15 min)
• Even after increasing timeout to 1800s, large sessions remain at risk

Proposed solution

Add a compaction.model config option:

{
  "agents": {
    "defaults": {
      "compaction": {
        "model": "anthropic/claude-sonnet-4-5",  // or "google/gemini-2.0-flash"
        // ... other compaction settings
      }
    }
  }
}

Behavior:

• If compaction.model is set, use that model for generateSummary() calls
• If not set, fall back to session's model (current behavior)
• The model should have sufficient context window for the chunk size

Alternatives considered

Impact

• Affected: All users with Opus/slow models and long-running sessions
• Severity: High — compaction timeout causes sessions to get permanently stuck in processing state
• Frequency: Every session that exceeds ~150k tokens
• Consequence: Stuck sessions, lost messages, requires manual session reset

Evidence/examples

From OpenClaw 2026.3.7 logs:

embedded run timeout: sessionId=985aa0c1-... timeoutMs=900000
using current snapshot: timed out during compaction

Session details:

• 4.5MB JSONL, 940 lines, 3 prior compactions
• Model: claude-opus-4-5
• Timeout: 900s → increased to 1800s as workaround

Additional information

• Sonnet is ~3-5x faster than Opus for summarization tasks
• Flash/Gemini would be even faster
• Summarization doesn't require Opus-level reasoning depth
• This would also reduce API costs for compaction