cyrus.is/building

Things I'm working on.

• A cheap model built me a working service for $8.33, and its passing tests hid the two bugs that mattered most
  I gave GLM-5.2 (US-hosted, driven by Block's Goose) a bounded spec and a $20 budget and let it build an OpenAI-compatible LLM router. It's not perfect but it works. I found two bugs its passing test suite never checked for, both in the streaming path. I checked whether it would find them too. It did, and finding them cost less than a buck.
• Scrutineer now audits MCP servers — tested against the 100 most popular
  My code-review toolkit (formerly vibecheck) is now Scrutineer — pip install scrutineer — and it has a new command, /scrutineer-mcp, that audits an MCP server before you ever run it. I ran it against the 100 most-installed servers to prove it works; none came through clean, mostly because nobody pins versions.
• A Budget
  I wanted to see where my money actually goes without handing my bank logins to a third party. So I built a Sankey diagram that runs entirely in your browser.
• I Tested AI Code Review Against Human-Written Review Skills. Neither Won.
  50+ reviews across 8 components. Generated skills found a P1 data loss bug the humans missed. Humans caught edge cases the AI never would. The interesting part is why.