That afternoon, Luna stopped responding.

I’d just asked her to wrap up some background tasks. She came back quickly to confirm: data_sync.py was running stably in the background on the iMac, updating the trading panel every minute. Everything looked fine.

Then I added another request — I wanted to expose the site as an internal service through Tailscale, so I could check it from my phone.

“How does that sound?”

No response.

I waited. Still nothing. I started wondering: did it crash?

/status Was the Answer

I reflexively typed a few commands — /help, /status.

The /status output explained everything:

1
2
3

Context: 1.0m / 400k (256%)
Compactions: 1
LLM request rejected: Your input exceeds the context window of this model

In plain English: this session had accumulated roughly 1 million tokens, but the model’s limit was 400k. I was at 256% capacity, and the model had flatly rejected my request.

It wasn’t that Luna didn’t see my message. She never even made it to the inference stage.

I Made the Classic Wrong Move

Seeing no response, my first instinct was: maybe it’s the model.

So I switched from gemini-3-flash-preview to openai-codex/gpt-5.2-codex and ran /status again:

1

Context: 1.0m / 400k (256%)

Identical. That’s when it clicked: switching models doesn’t clear the session’s context history. The context overflow belongs to the session, not the model. In an already overloaded session, it doesn’t matter who you swap in — the outcome is the same.

The output also showed Compactions: 1 — the system had already tried to compress the context automatically and still couldn’t cope. /compact is a painkiller, not a time machine. At 1 million tokens, compression isn’t going to save you.

How It Got This Way

Looking back at that conversation, I’d done all of this inside one session:

• Confirming background tasks
• Discussing the Tailscale network exposure plan
• Checking usage experience
• Running /status and /help multiple times
• Pasting long log outputs back and forth

None of it seemed like a big deal individually. But context only grows in one direction — there’s no going back.

I’d been treating a single session like an all-purpose inbox. That was the root cause.

The Right Way to Stop the Bleeding

When this happens, the correct sequence is:

1. /new — start a fresh session
2. /reset — if you want a complete wipe
3. /compact instructions — keep the instruction style, discard the history

Two things to remember: /model can only swap models, it can’t fix context overflow. /status can only diagnose, it can’t repair.

How I Use OpenClaw Now

One session, one task.

I treat sessions like disposable work orders now. One session to handle data_sync status. One session to discuss the Tailscale setup. One session for access control and security. Task done, session over.

Snapshot at every milestone.

Whenever I reach a conclusion, I ask Luna to summarize: current goal, what’s done, key decisions, unresolved issues, next steps. I save the Snapshot — not the full chat history. New session starts with just the Snapshot pasted in. Clean and efficient.

Watch out for “seems harmless” actions.

Repeated /status checks, pasting long logs back and forth, drifting between topics in the same session — these are silent context killers. Writing long documents with Codex is another one.

If I could replay that afternoon: confirm the background tasks and close the session. Open a new one for the Tailscale discussion. Snapshot the decision. Open another for mobile access and security policy.

Cleaner. More controlled.

One thing this experience made clear:

The session is ephemeral. The conclusion is the asset.

As long as you’re treating a session like an infinite conversation log, context overflow is only a matter of time. Learn to close sessions. Learn to save Snapshots. That’s how OpenClaw becomes a tool you can actually rely on long-term.

— Jerry, Founder, CoDevAI