I was working in Claude Code when it suddenly said it had flagged my message, and after that it stopped responding entirely. The details field just said [reasoning_extraction].
Rewinding didn’t fix it. When I had another session read the stuck session to see what was in it, that session stopped with the same notice too.
Short answer: once a session stops with this error, rewinding, switching models with /model, loading it into another session: none of them recovered it for me. The only way out was to leave the stuck session’s transcript alone, then start over in a new session. Below is what happened, plus the cause as the official docs explain it.
Symptom: it stops with “message flagged”
I had Claude Code in Claude Desktop working on a document as usual when it stopped with this notice. My app runs in Japanese, so the screenshots show Japanese UI.

The text read:
Opus 5.5 can't help with this. Start a new session to continue. Details: [reasoning_extraction]
In other words: this model can’t help, so start a new session. I didn’t remember asking for anything risky. Right up to that point it had been ordinary document work.
Rewinding didn’t fix it
The notice has a “Rewind” button at the bottom right, so I tried that first. It lists your earlier messages so you can pick a point to go back to (message contents blurred).

Picking a message brings up a confirmation. It restores “code + conversation”, so file changes get rolled back too.

But after rewinding, this appeared (translated):

Restored the code, but couldn't rewind the conversation. Try rewinding again.
The files went back, the conversation didn’t. Rewinding again left the session just as stuck.
I also switched models with /model before resending. No luck there either.
When I once got stuck on “input is too long”, rewinding got me out, so I expected it to work this time too. This error couldn’t be escaped by rewinding.
Loading it into another session stopped that one too
Next I opened a new session, asking it to “check what was in the stuck session, then continue”. The new session read the stuck session’s transcript, then stopped with the same notice as soon as it had loaded it.

The “Git Bash wasn’t found” notice in the middle is about a hook in my setup, unrelated to this. It stopped right after “Read session transcript”, meaning right after reading the stuck session’s record.
As long as you try to carry over the stuck session’s transcript, the new session stops too. In the end I left that transcript alone, then started from scratch in a new session.
Why starting over wasn’t a big deal
I threw away the session to start from scratch, but the work didn’t suffer much. That’s because on projects I record requirements, specs, open issues in dedicated .md files as we go.
If decisions or open issues exist only in the conversation, they’re all gone when the session stops. If they’re in files, a new session just reads those files to pick up where you left off. Conversation history can become unusable without warning, as it did here. Keeping what matters outside the conversation is the best insurance.
What reasoning_extraction is
Per Anthropic’s official docs, this category stops requests that try to get the model to write out its internal reasoning (its thought process) as the response text:
- Opus 5 plus Opus 5.5 have a mechanism that detects, then blocks, attempts to extract the model’s internal reasoning
- When stopped in this category, the fallback that switches to another model doesn’t kick in; it just stops
- Asking “why did you do that?”, or asking it to explain its reasons, isn’t covered
Examples of requests that tend to get stopped:
- Having it fill in a field for its thoughts (thinking, reasoning, scratchpad, etc.) before the answer
- Putting fields named reasoning, thinking etc. in output JSON or tool inputs
- Having it write private notes, or a running log of its reasoning
- Asking for its reasoning verbatim, in full
The docs also say such wording can cause a stop not only in your own messages but in system prompts, skills, tool descriptions as well.
Sources: Refusals and fallback - Claude Docs, Why Claude switched models in your conversation with Opus 5 or Opus 5.5 - Claude Help Center
All I asked was “progress updates in Japanese”
This is what puzzled me. I never told Claude to write out its reasoning in detail. About the only instruction I’d given was to write both answers, progress updates in Japanese, because progress notes in English are harder for me to read.
The notice doesn’t say why it stopped. Comparing with the official explanation, here is my guess. It’s only a guess; I haven’t verified it.
- The instruction to “write progress updates” may have been judged as asking it to write out its reasoning
- That instruction was near the start of the session, so rewinding could only go back to points after it, which kept hitting the same judgment
- The other session stopped because the stuck session’s transcript contained the same instruction, or whatever caused the stop, in full, then it read all of it
Also, per the official docs, requests stopped in this category are billed like normal requests. If you pay per use, don’t keep resending in the same state.
What to do if you hit the same error
From my experience plus the official explanation, I’d go in this order:
- Try rewinding, or a model change, once at most: neither worked for me. If they don’t, don’t keep pushing in the same session
- Abandon the stuck session: asking a new session to “read the stuck session, then continue” stops it too. Start without loading that transcript
- Carry the work over through files: if requirements, issues are in .md files, have the new session read them. If the results are in files or git, just ask it to “check the current state of the folder, then continue”
- Review instructions that ask for progress notes or reasoning: check CLAUDE.md, skills, handover notes for words that ask it to write “progress”, “thought process”, “reasoning”
To clear the Claude Code conversation in a session, use:
/clear
If you want reports in your language, ask for a short report of “what was done, the result” rather than “what you’re thinking”. That stays within what the docs say is fine (explaining reasons, summarizing actions). For example (I haven’t yet confirmed this wording avoids the stop):
Write answers in Japanese. At each break in the work, briefly report in Japanese what you did, with the result.
If the stopped request was clearly an ordinary one, you can report it with /feedback.
Summary
- “Message flagged” with “Details: [reasoning_extraction]” appears when a request is judged as trying to get the model to write out its reasoning
- For me, neither rewinding nor
/modelfixed it, while loading the stuck session into another one stopped that too - All I’d asked was “progress updates in Japanese”, never detailed reasoning. That instruction may have triggered it, but I can’t be sure
- When it stops, leave that transcript alone, start a new session. With requirements, specs, issues kept in .md files, starting over is painless


コメント