Alfred Rolling Summarization

Proactively updates a concise session summary every 15 turns or on tool loops to manage context size and preserve key decisions and progress.

Security Scan

The skill's instructions, required artifacts, and behavior are coherent with its stated purpose (periodic rolling summaries stored in SESSION-STATE.md) and do not request unrelated credentials or installers.

Purpose & Capability

Name/description (periodic rolling summaries to manage context) matches the instructions which explicitly create and update a SESSION-STATE.md rolling summary and integrate with an existing 'Hindsight' retention mechanism. No unrelated binaries, env vars, or cloud credentials are requested.

Instruction Scope

SKILL.md instructs the agent to read and update SESSION-STATE.md every N turns or on tool-loop detection and to keep summaries concise. All referenced actions are directly tied to the stated goal. The instructions do not call external endpoints or request unrelated files or secrets. They do assume the agent can read/write a session-state file and that 'Hindsight auto-retain' is available.

Install Mechanism

No install spec and no code files are present (instruction-only). This is low risk and consistent with an orchestration/prompting skill.

Credentials

The skill declares no environment variables, credentials, or config path requirements, and the SKILL.md does not ask for secrets or external tokens. Requested access (file read/write for SESSION-STATE.md and use of Hindsight) is proportionate to the purpose.

Persistence & Privilege

The skill writes persistent state (SESSION-STATE.md) and expects that file to be kept 'always in context via bootstrap'. It does not request always:true or modify other skills. Persisting summaries across runs is expected for this feature but is a persistence decision the user should be aware of.

Assessment

This instruction-only skill appears coherent and low-risk, but check where SESSION-STATE.md will be stored and who/what can read it: the rolling summary may contain sensitive task details and will be kept in context. If you install it, confirm the file path and retention policy (where the file is saved, rotation/expiration), ensure you consent to the agent writing to that workspace, and verify the 'Hindsight' integration behavior. No credentials or external installs are required. If you need stronger privacy, limit what is summarized or redact sensitive lines before writing the file.

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