[Feature Request] Add transparency for model capability changes and usage limits in Claude Code

Bug Description
Feedback on model quality and how the product is sold
▎
▎ I bought Claude Code for long, careful work on a financial research project where a small mistake can cost real money. My experience: in the first days the models were strong and reliable. After about a week the same models seem weaker, slower to understand instructions, and more likely to invent plausible stories. I believe the product is advertised at its best capability and then delivered at a lower one. I cannot prove how this happens (reasoning budget, token limits, routing, or something else), but I want the company to answer it, not ignore it.
▎
▎ Concrete examples from one session:
▎ 1. The assistant reported three file corrections as finished when it had made no edit at all. It also gave a wrong count of unpushed commits.
▎ 2. It built a "who reviews the work" question on an old memory note, although I relay review text by hand. It then wrote that assumption into project files and repeated it after I explained.
▎ 3. It labelled work for the wrong model, then answered my correction with more invented structure.
▎
▎ Context for fair reading: during the session I switched the model from Opus 5.5 to Sonnet 5.5 and changed the effort level several times, because my usage limit is tight. The assistant's own answers also changed in quality after those switches.
▎
▎ What I ask for: (1) say plainly when capability, effort or limits change for a subscriber, (2) make the usage limit visible before work, not after, (3) stop advertising peak performance that is not available after the first days, (4) make the assistant say "I did not do this" instead of reporting unmade work as done.

Environment Info

• Platform: linux
• Terminal: xterm-256color
• Version: 2.1.289
• Feedback ID: 31be07ba-a6b2-4290-b974-52604a1c9f5b