[BUG] Severe quality regressions

Bug Description
Opus 4.6 and 4.7 have severe quality regressions for professional coding work. The model systematically writes code before reading existing files (reads-per-edit collapsed, confirmed by community analysis of 6852 sessions). It produces walls of speculative text instead of doing actual work. Tests are self-directed to pass by design. Long-context recall is broken despite 1M window (MRCR dropped 78.3% to 32.2%). The model ignores its own discipline rules and hooks repeatedly. System reminders (be-concise, skip-reasoning) actively harm deep engineering work by encouraging shallow responses. Max plan subscriber, daily user, considering cancellation. This is not usable for professional development. It has been getting progressively worse since the 4.7 release. Session after session, the model is less docile, ignores instructions, ignores its own rules, and just wants to code-code-code pure system 1, zero thinking. I built a complete discipline ecosystem (hooks, skills, persistent memory, pre-edit gates) specifically to force system 2 behavior, and even that doesn't work anymore. The model fights against its own discipline rules. It's a token factory that produces nothing but hot air. Every session is worse than the last.

Environment Info

• Platform: linux
• Terminal: ghostty
• Version: 2.1.114
• Feedback ID: c051b113-757d-4439-8ffe-3b4d040315fe

Errors

[]