[BUG] Model abandons first version mid-output and restarts with different content in same response
Preflight Checklist
- [x] I have searched existing issues and this hasn't been reported yet
- [x] This is a single bug report (please file separate reports for different bugs)
- [x] I am using the latest version of Claude Code
What's Wrong?
When generating long-form content (e.g. 4000+ character text), Claude sometimes
abandons its first version partway through and starts over with completely different content
in the same response. The user sees both versions concatenated together.
This is NOT a UI rendering duplicate — the two outputs have different content/structure.
The model appears to internally "change its mind" about the approach mid-generation
and restarts from the beginning without stopping to ask the user.
This wastes tokens (user pays for both versions) and breaks the reading experience.
What Should Happen?
Claude should commit to one version of output per response. If the model is unsatisfied with its current output mid-generation, it should finish the current version first, then ask the user if they want a revision — not silently restart and output a second different version in the same response.
Error Messages/Logs
No error messages. This is a behavioral issue, not a crash.
The model outputs Version A (partial) followed by Version B (complete) in a single response.
Both versions have different content — this is not a UI rendering duplicate.
Steps to Reproduce
- Ask Claude to write a long piece of content (4000+ characters,)
- Observe the streaming output — Claude begins writing version A
- Midway through, Claude stops version A and begins writing version B from the start
- Both versions appear in the same response with different content
Claude Model
Opus
Is this a regression?
Yes, this worked in a previous version
Last Working Version
_No response_
Claude Code Version
2.1.92
Platform
Anthropic API
Operating System
Windows
Terminal/Shell
Windows Terminal
Additional Information
_No response_
This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗