context-flush

Lean session reset playbook for OpenClaw.

Purpose

This skill is for one job: helping an agent recover cleanly when a session gets bloated.

It does not pretend to auto-monitor, auto-reset, or control the runtime. It gives the agent a real workflow that matches how OpenClaw actually works today.

What this skill does

• checks current session context on demand
• decides whether to keep going, compact, reset, or use dashboard delete as fallback
• saves a short handoff note before reset
• restores from that handoff note after reset
• keeps the process simple enough to trust under pressure

What it does not do

• no fake auto-monitoring
• no made-up /flush command
• no pretending the agent can force /new
• no magic buttons unless the current channel actually supports them and the flow is confirmed

Recommended workflow

1. Check context

Use session_status when the conversation feels slow, large, or risky.

2. Choose the move

• Keep going if context is still comfortable
• Use /compact if a light trim is enough
• Use /new when context is heavy and a real reset is the right move
• Use dashboard delete only as fallback for stuck or broken sessions

3. Save a handoff note before reset

Write a short note to the agent workspace or daily memory file with:

• active task
• current status
• key decisions
• files in play
• exact next step

4. Reset

The user runs /new.

5. Restore

On the fresh session, read the handoff note and resume fast.

Handoff note format

Keep it short and useful:

# Context Flush Handoff - 2026-03-14 13:05 ET

## Active task
- Clean up context-flush skill and sync GitHub repo

## Current status
- presidio-pii synced to GitHub
- context-flush being rewritten as honest manual playbook

## Key decisions
- do not fake automatic monitoring
- use session_status on demand
- use /new for real resets

## Files in play
- workspace/skills/context-flush/SKILL.md
- workspace/skills/context-flush/README.md

## Next step
- finish SKILL.md rewrite, then commit and push to sebclawops/context-flush

Threshold guidance

This skill uses practical guidance, not fake precision.

Small-window models

For models around 32k context, get nervous around 20k to 24k.

Large-window models

For large hosted models, performance can still get sloppy well before the advertised limit. Treat roughly 80k to 100k as the zone where reset planning starts.

Notes

• /compact is a trim, not a true reset
• /new is the clean reset path when available
• in some channels or stuck sessions, dashboard delete may still be the only reliable fallback
• this skill should stay honest, boring, and operationally useful