[BUG] TODOs go back to the beginning after a context compaction
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?
After doing a compact context, the agent begins all the work from the start again instead of picking up where we left off. It then starts to repeat mistakes I have already told it etc. Basically it forgets everything and I have to bootstrap the session to get it to work again.
What Should Happen?
should carry on where we left off
Error Messages/Logs
Steps to Reproduce
work as normal until the context is full, weep silently
Claude Model
Sonnet (default)
Is this a regression?
Yes, this worked in a previous version
Last Working Version
whatever the final 1.0x was
Claude Code Version
2.0.1
Platform
Anthropic API
Operating System
Windows
Terminal/Shell
VS Code integrated terminal
Additional Information
I'm using the default vs code extension
Found 3 possible duplicate issues:
- https://github.com/anthropics/claude-code/issues/7961
- https://github.com/anthropics/claude-code/issues/4878
- https://github.com/anthropics/claude-code/issues/6123
This issue will be automatically closed as a duplicate in 3 days.
- If your issue is a duplicate, please close it and 👍 the existing issue instead
- To prevent auto-closure, add a comment or 👎 this comment
🤖 Generated with Claude Code
here's an example from just now (I've cut down some of the noise)
me: the build ends up going to .\bin\Release\net9.0\ i want it to be an exe with no net.90 or release in the path
CC: I'll help you configure the build output to go directly to CRMPollerFixer.exe. We need to modify the project file to set custom output paths.
Update Todos << SHOULD CARRY ON USING THESE
Configure custom build output path to Y:\CSharpDLLs\CRMPollerFixer\
Test the custom build path Read CRMPollerFixer.csproj Edit CRMPollerFixer.csproj ⎿ Added 5 lines
COMPACT HAPPENS HERE
123456789
12345678910111213
This session is being continued from a previous conversation that ran out of context. The conversation is summarized below:
Claude AI usage limit reached|1759788000.
Please continue the conversation from where we left it off without asking the user any further questions. Continue with the last task that you were asked to work on. I'll continue from where we left off. Based on the context, I can see that I was working on the CRMPollerFixer project and there are several modified files in the git status. Let me check the current state and continue with any pending tasks.
Read CRMPollerFixer\Services\ElasticsearchService.cs Read CRMPollerFixer\Services\DocumentProcessor.cs Read CRMPollerFixer\ServiceLib\Elastisearch.cs Based on the context and the modified files I can see, it appears I was working on the CRMPollerFixer project. Let me check what recent changes were made and continue any pending work. Bash ⎿ Show names of modified files IN git diff --name-only OUT CRMPollerFixer.csproj
Bash ⎿ Show changes to project file IN git diff CRMPollerFixer.csproj OUT diff --git a/CRMPollerFixer.csproj b/CRMPollerFixer.csproj index 9f893a1..3b8b502 100644
Let me check if there are any test failures or if I need to run any builds to verify the current state: Bash ⎿ Build the project to check for errors IN dotnet build
Great! The build is successful. Let me check if there are any test scripts or validation that I should run: Read CRMPollerFixer\test_validation.sh
me: Ctrl^C
in fact the TODOs were empty, the build path had been fixed in .csproj so perhaps that is the real issue, the agent sees there are no outstanding TODOs so goes and looks for something to do instead of just doing nothing
This issue has been inactive for 30 days. If the issue is still occurring, please comment to let us know. Otherwise, this issue will be automatically closed in 30 days for housekeeping purposes.
This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.