Ali Naqi Shaheen

I keep running into this while building with AI: after a few sessions, I end up re-explaining the project, the decisions we made, and what changed.

How do you handle that? Do you keep a project file updated, rely on chat history, or have another system that works?

We launched here last Monday. One number from our prep keeps coming up in conversations, so a question.

We ran every stable version of 14 of the top 30 npm MCP servers from March to September in a throwaway environment, tools/list only. 86 version-to-version transitions. 40 of them changed what the tools do or accept, 47%. The largest category was not added tools. It was rewritten parameter descriptions, 92 of them. The agent reads that text and acts on it. The person approved the server once and never reads it again.

mcpgawk sits on that gap: baseline at approval, refuse the call when the surface changes until a human looks. https://mcp.gawk.dev/ph if you want to try it.

So: when you add an MCP server to Claude Code, Cursor or Codex, what do you do at its next version? Pin it, re-read the tools, or trust the update?

Most vibe coding advice says to have the agent write the tests alongside the code. I think that's the most expensive habit in the whole loop.

The code and the test come out of the same reading of what you asked for. If that reading is wrong, both are wrong in the same direction and the suite goes green, which is worse than having no suite at all, because now you've got a receipt.