[MODEL] Opus 4.6 continuously stuck in explore and thinking loops (Claude Code) #24585
Description
Preflight Checklist
- I have searched existing issues for similar behavior reports
- This report does NOT contain sensitive information (API keys, passwords, etc.)
Type of Behavior Issue
Other unexpected behavior
What You Asked Claude to Do
Ever since the release of Opus 4.6, I've had problems where Claude Code gets stuck in thinking loops or explore loops for minutes on end. This is actually happening for me every single session, and on most of my requests, it's not a rare occurrence. Ctrl+o reveals no visibility into what is actually happening behind the scenes, but Claude will sit there for 7 minutes, and I'll interrupt Claude and say "what happened? Why is it taking so long?" And Claude will tell me they were over-exploring, or stuck in thinking loops, arguing with itself.
Sometimes token and tool usage looks fairly static, and other times it'll be chewing through tokens like no tomorrow (hundreds of thousands of them, and >50 tool requests). I think it might be the difference between stuck in an argumentation loop with itself, versus stuck in an explore loop.
What Claude Actually Did
Just general usage. I nuked my entire Claude Code install thinking it was a problem on my end, but the behavior still persists. I've tried using the Windows version versus the WSL2 Ubuntu version (native binary on both), and it's the same deal. I'm confident this isn't a problem on my side, it's a problem that has only started as of the moment 4.6 released.
Expected Behavior
Some times its stuck in loops for 5 mins, sometimes 10 mins, I've even had it up to 22 mins.
The other issue is that theres zero visibility into what is actually happening behind the scenes, so it's impossible to tell when it's working as expected or stuck in a loop.
Files Affected
Permission Mode
Accept Edits was ON (auto-accepting changes)
Can You Reproduce This?
Sometimes (intermittent)
Steps to Reproduce
No response
Claude Model
Opus
Relevant Conversation
The other thing of note, is that when I do interrupt Claude, it generally agrees that the work it was doing was not helpful. It immediately will start doing the correct behavior, and say things like: "I already have enough context, let me complete the task." - so whatever it's doing in the background is not actually contributing to the goal and is just wasted effort.Impact
Critical - Data loss or corrupted project
Claude Code Version
2.1.38
Platform
Anthropic API
Additional Context
No response