What's the most painful LLM failure you've ever debugged?

We've all hit that moment: the model works on prompt A, breaks silently on prompt B, and there's no log line, no stack trace, no clue what changed inside the model.

I'm launching OpenInterpretability in a few hours on Product Hunt. It's a mech interp toolkit that runs inside Claude Code, Cursor, and Cline via MCP — plus drop-in probes for hallucination and agent-failure detection. The project started because I needed a way to see what was happening when a coding agent kept making the same silent tool-call mistake.

Genuinely curious:

- What's the worst silent failure you've debugged in an LLM app?

- What did you wish you could see at that moment — activations? attention? something else?

- If you had a "model debugger" inside your IDE, what would you point it at first?

War stories welcome.