[Bug] Claude Opus prematurely terminates task execution ignoring system prompt constraints

Bug Description
Opus fails to even remotely adhere to system prompt, rules when it finds the first thing it makes sense. It's not thorough unless I just bury my code with hooks which becomes ridiculous at some point. I don't have this issue with openai models.

Other times I ask it why it ignored the rules and it says, "I thought the first result I had was correct and ignored your rule, but knew about your rule."

A lot of this is resolvable with a better harness. With omp, I am able to auto-inject rules to correct behavior. With claude, I'm just stuck withit.

Environment Info

• Platform: darwin
• Terminal: WezTerm
• Version: 2.1.259
• Feedback ID: 0da04e28-3a63-4ed5-b9c8-ec2f1272e78e

Errors

[]