Auto-Summarization Skill

Proactive context management to prevent overflow and improve session quality.

Problem

• OpenClaw compaction triggers on context overflow (reactive)
• Compaction timeout (60s internal) causes fallback to pre-compaction state
• Tool loops (research with web_fetch) generate massive context without stopping
• No control over WHAT is preserved during compaction

Solution: Rolling Summarization

How It Works

Every N turns, create a concise summary of recent work and update SESSION-STATE.md. This keeps the session lean while preserving important context.

Trigger Thresholds

Condition	Action
Every 15 turns	Create rolling summary
After 10 consecutive tool calls	Force summary (tool loop guard)
Context >70% estimated	Proactive summary + flush to memory

Rolling Summary Format

Update SESSION-STATE.md with a condensed view:

## Rolling Summary (as of HH:MM)
- Completed: [what was done]
- In progress: [what's being worked on]
- Decisions: [key decisions made]
- Blockers: [anything blocking]
- Next: [what to do next]

Integration with Existing Stack

1. Hindsight auto-retain captures important facts before summary
2. SESSION-STATE.md stores the rolling summary (always in context via bootstrap)
3. Daily notes get the detailed version at end of session
4. Working buffer becomes unnecessary if rolling summary works well

Anti-Patterns

• ❌ Don't summarize every turn (adds latency, wastes LLM calls)
• ❌ Don't duplicate Hindsight content (it already retains facts)
• ❌ Don't include routine operations (heartbeat checks, status pings)
• ✅ DO summarize: decisions, blockers, task progress, user preferences
• ✅ DO keep it under 200 chars per section

Prompt Addition (add to agents that need it)

After completing a task or every 15 turns:
1. Read SESSION-STATE.md
2. Update the Rolling Summary section with current state
3. Keep it concise (under 500 chars total)
4. This prevents context overflow and preserves continuity

Metrics

Metric	Before	After (target)
Context overflow/week	1-2	~0
Compaction timeout rate	~50%	<20%
Context lost per session	High	Low
Additional LLM cost	$0	~$0.02/week

Version

1.0.0 — Initial implementation