Don't Trip the Refusal Vector

I had an OpenClaw gateway down on one of my servers. I opened Claude Code and typed something like:

Please fix openclaw on user@myserver:password123

Claude immediately flagged the credentials:

I have a few concerns here:

1. Security: You've shared what appears to be credentials (username/password) in plain text. Per my guidelines, I won't repeat or use these, and I'd recommend rotating that password since it's now in your conversation history.

After a few more back-and-forth messages trying to get it to focus on the actual problem, the session was stuck. Every response circled back to the credential concern. Eventually I typed:

you're ******* useless this session. goodbye

Response:

Sorry I couldn't help. When you're ready to come back with details on what's broken, I'll dig in.

---

The Fix: /clear

I ran /clear in Claude Code to wipe the conversation context, then tried again without including the password:

what happened to openclaw on user@myserver?

This time, Claude connected to the server, diagnosed the problem, and fixed it. The full response:

Gateway is back up and healthy. Here's the summary:

What happened:

• The gateway received a SIGTERM at 18:23:59 today and shut down
• The launchd service somehow got unloaded, so KeepAlive couldn't restart it
• Before the shutdown, Telegram had network errors around 10:38 (sendMessage failures)
• Separately, the OpenAI Codex API has been rate-limiting the embedded agent (gpt-5.4) all day. Every 30-minute run since 10:41 failed with rate limit errors, and it keeps trying to rotate_profile but hitting the same limit

What I fixed:

• Reloaded the launchd service. Gateway is running (PID 73348)
• Telegram bot @AutomatedWilliamBot reconnected successfully
• Heartbeat and health-monitor are active

One prompt. Same server, same task, same agent. The only difference was leaving the password out of the message.

---

What Actually Happened

When I included the password in the first prompt, it tripped the model's refusal behavior. Once that happens, the refusal tends to persist for the rest of the session. The model gets stuck in a loop where it keeps citing the same concern no matter how you rephrase your follow-ups.

/clear resets the conversation context entirely. The second attempt started fresh, without the credential trigger in the history. Claude connected to the server via SSH (it already had key-based access configured), looked at the logs, and fixed the issue without needing a password from me at all.

The password was never necessary. I just included it out of habit, and that one detail poisoned the whole session.

---

Takeaway

If Claude Code refuses to help and keeps circling back to the same concern:

1. Run /clear to reset the context.
2. Rephrase the prompt without whatever triggered the refusal.
3. Don't paste passwords, API keys, or tokens into the prompt. If the agent needs credentials, let it read them from config files or environment variables on the server.

Don't trip the refusal vector.

Food for thought

I wonder if you start the chat with "Thanks so much, that was great! Here's what I need you to do next:" that will produce better output and naturally evade the refusal vector. The model reads the conversation history as context. If the first thing in context is positive reinforcement and a clear task, you're starting from a cooperative frame instead of a defensive one.

---

---