[BUG] claude.ai / Claude Desktop: mid-chat response rollback silently loses user work
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?
Environment
- Claude Code version: 2.1.63
- OS: macOS (Parallels)
- Also reproduces on: claude.ai (web browser), Claude Desktop app (Mac)
What's Wrong?
During tool-heavy sessions (bash commands, file creation, git operations, web search, code execution), Claude's response disappears immediately after completion. The conversation reverts to the state before the last message was sent. The response is gone permanently. No error is shown. Refreshing does not recover it.
This is particularly dangerous during agentic coding workflows because Claude makes git commits, writes files, and executes commands during the response. When the response rolls back, those side effects persist but the user has no record of what was done. This creates orphaned commits, unpushed changes, and files in unknown states.
Frequency
Multiple times per session. 100+ occurrences per week. Not intermittent. Constant.
Trigger
Long sessions with heavy tool use. The rollback correlates with high context usage near the compaction threshold. Workarounds deployed (limiting session length, restarting more frequently) reduce but do not eliminate the issue.
Impact
- Git commits made during tool use are invisible after rollback, leaving orphaned or unpushed commits
- Files written or modified during the response are in unknown states
- Each incident costs ~20 minutes to recover: auditing git log, checking for conflicts, verifying file states, re-prompting
- Risk of subsequent prompts overwriting work done in the rolled-back response
Related Issues
Reported by multiple users since late 2025:
- #14225
- #16937 (closed as duplicate of #10065)
- #10065
- #18676
- community/discussions#155163
When Claude completes a response involving tool use, the response should persist. If it fails to save, an error should be shown and the user should be alerted that side effects (commits, file writes) may have occurred.
What Should Happen?
When Claude completes a response involving tool use, the response should persist. If it fails to save, an error should be shown and the user should be alerted that side effects (commits, file writes) may have occurred.
Error Messages/Logs
Steps to Reproduce
Run a long Claude code command. Bug probability increases if Claude is doing a long task and doesn't return the prompt during that task.
Claude Model
None
Is this a regression?
I don't know
Last Working Version
_No response_
Claude Code Version
2.1.63
Platform
Anthropic API
Operating System
Windows
Terminal/Shell
PowerShell
Additional Information
_No response_
This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗