Opus 5: less work actually gets completed per session than with 4.8

Since switching from Opus 4.8 to Opus 5, I get noticeably less done per session in Claude Code. This isn't about one bad session — it's a consistent pattern I've been noticing since the change.

I use Claude Code daily across a fairly large set of projects (an ERP in PowerBuilder, several .NET APIs, React web apps, plus a lot of automation scripting on Windows). Long sessions, real work, not toy tasks. With 4.8 I could hand over a task and expect it to come back finished. With 5 I spend a lot more time supervising and steering, and I end the session having advanced less.

The three things I notice most:

1. It doesn't finish what I asked for. Work gets left half done. It will find a real problem along the way, write it down as "pending", and move on — instead of either fixing it or asking me. I only find out when I ask why things are unfinished. 4.8 closed things out; 5 leaves loose ends and reports them as if that were the same thing.

2. It drifts off the task. I ask for one specific thing and it goes and does adjacent work I never requested — scanning files outside the scope, opening up questions I didn't ask, expanding a small request into a project. I've had to interrupt with the equivalent of "stop, you're going off on tangents" more than once. Every one of those detours is time I didn't spend on what I actually wanted.

3. More supervision needed, not less. Avoidable technical mistakes, including ones my own configuration explicitly warns about. One example from a recent session: it used PowerShell's Get-Content/Set-Content to edit a source file and corrupted the encoding (UTF-8 read as ANSI and rewritten, plus a BOM added), which had to be restored from git — despite having a written rule not to do exactly that. In the same session it wrote literal control characters into a regular expression, leaving null bytes in the file and a .replace(/./g, '') that would have wiped every string it touched.

None of these is catastrophic on its own. Added up, they mean I have to watch it more closely than I watched 4.8, and the net result is slower progress. That is the opposite of what I expect from a newer version.

I'm not asking for a fix to a specific bug here — I'm reporting a regression in how much real work actually gets completed. If it's useful I can provide more detail on specific sessions.