Last night, all proxy nodes went down simultaneously.

Three concurrent issues: Xray log permission errors on VPS causing process crashes, local Clash configuration files corrupted, and Japanese node IP suspected to be blocked. The result was—completely losing connection with Claude Code.
It was then that I realized one thing: I find it very difficult to troubleshoot problems "naked" on my own.
For the past half year, almost all technical decisions were made through AI conversations. Reading logs, modifying configurations, checking documentation, writing scripts—all done with AI and Claude Code throughout. Suddenly disconnected, it’s not even a matter of "inconvenience," but genuinely not knowing where to start.
In the end, I temporarily went online by borrowing a friend's proxy subscription, using Claude Code.
Fixing my own infrastructure: identifying the root causes of three concurrent failures, restoring permissions, recovering configurations from backups, and doing full disk snapshots on Vultr to migrate and change IP. Four PM2 processes, nine cron jobs, all code and data restored as-is, and a new IP—revived fully.
After fixing everything, I did something I should have done long ago: I created an offline self-rescue package.
Running a small local model, with an emergency manual (failure scenarios + troubleshooting commands + protocol parameters). When offline, the local model reads the manual and guides step-by-step troubleshooting. Zero cost, no internet needed.
A thought: if you rely heavily on AI CLI for daily work, your proxy/network access layer is your "AI lifeline." It also needs high-availability design—automatic failover between primary and backup nodes, offline fallback plans, emergency manuals.
The same principle as deploying production services, but this time, the service that’s down is your own brain.