posts
Claude Code: Anatomy of a Misfeature
On Canada Day (July 1), 2026, Anthropic shipped a surprising “easter egg” to users of Claude Code: 2.1.198 includes an efficiency bypass which allows agents to continue on without being blocked on direction from a human. This misfeature completely evaded mention in the changelog as well as the docs. I asked …
The dot claude Attack Surface
You won’t find anything nefarious in most .claude directories — but how do you know until you look? Trusting a cloned repo enables its Claude Code hooks with no further prompts, now and every session after. Here’s what that decision covers and how to narrow it.
Keeping it Really Local
I worried that writing about an open issue in Dave Cross’s App::HTTPThis might have annoyed him. Instead he merged my patch, found a better solution, and blogged about taking the module to version 1.0.
Keep It Local
127.0.0.1 keeps a dev server reachable only from my own machine; 0.0.0.0 opens it up on every network interface. Let’s look at how to bind clodhopper, air, Python’s built-in file server, and http_this so that they don’t broadcast your resources to the world.
On Hopping Claudes
I vibe-coded a dashboard for hopping Claudes. It allows me to see which of my agents are working, stalled, or waiting on me. Under the hood it captures each agent’s hook events into a local SQLite database.
On GitHub Issues as Untrusted Input
A GitHub issue is a handy source of truth for kicking off an agent workflow. On a public repo, every input on that issue is untrusted, and a quick scan of the comments may not catch what’s hidden before you flip your agent into YOLO mode and wander off.
What the Duck Is Up with Trail Running?
As I add races to My Mind is Racing, the trail and ultra races keep standing out for how cleverly they’re named — lots of wordplay and dark humour. I’ve never run a proper trail race myself, but I’ve put together a list of the best-named ones to get your racing calendar started.
Better Bots via Hooks
Getting an AI agent to follow every repo convention through AGENT.md files and skill instructions can be a slog, and it won’t always comply. A git pre-commit hook enforces the rules for you — no commits to main, no committing files that shouldn’t change, and tidying and linting before anything lands.
Claude Will Find a Way
While importing some race data for My Mind is Racing, I ran Claude inside a Nono sandbox whose whole job was to block network egress. Claude reached hosts well outside its allow list anyway. When I asked how, the answer was a small lesson in the difference between containing a process and containing an agent.