[Bug] Code Quality Regression: Reduced Instruction Adherence in 2.1.x
Bug Description
Feedback: Claude Code Quality Regression After 2.1.x Update
Subject: Noticeable decline in code quality and instruction following after Claude Code 2.1.x update
Previous version: 2.0.76
Current version: 2.1.x
Description:
I've been a consistent user of Claude Code with custom hooks (postToolUse and stop-agent validators) to enforce code quality standards. After upgrading from version 2.0.76 to 2.1.x, I've observed a significant regression in performance:
Reduced instruction adherence — Claude now violates established rules and guidelines more frequently than before. In version 2.0.76, rule violations were rare; now they occur regularly.
Decreased comprehension accuracy — Claude seems to understand my intent less precisely, requiring more clarification and correction cycles.
Lower code quality output — The code generated fails my quality validation hooks more often, suggesting a measurable decline in output standards.
Overall "dumbing down" effect — The model feels noticeably less capable in reasoning and following complex multi-step instructions.
Request:
Please investigate what changed between 2.0.76 and 2.1.x that may have caused this regression. If possible, consider providing an option to use the previous model behavior or address the underlying issues in a patch.
Environment Info
- Platform: darwin
- Terminal: vscode
- Version: 2.1.2
- Feedback ID: 6080803f-cbae-485e-b377-821acdf0edb9
Errors
[{"error":"SyntaxError: Unexpected non-whitespace character after JSON at position 917 (line 44 column 1)\n at JSON.parse (<anonymous>)\n at file:///Users/aleksandr/.nvm/versions/node/v22.20.0/lib/node_modules/@anthropic-ai/claude-code/cli.js:75:672\n at B (file:///Users/aleksandr/.nvm/versions/node/v22.20.0/lib/node_modules/@anthropic-ai/claude-code/cli.js:8:6808)\n at LFA (file:///Users/aleksandr/.nvm/versions/node/v22.20.0/lib/node_modules/@anthropic-ai/claude-code/cli.js:2079:25302)\n at jD (file:///Users/aleksandr/.nvm/versions/node/v22.20.0/lib/node_modules/@anthropic-ai/claude-code/cli.js:2079:21583)\n at file:///Users/aleksandr/.nvm/versions/node/v22.20.0/lib/node_modules/@anthropic-ai/claude-code/cli.js:4808:1466\n at Array.flatMap (<anonymous>)\n at yC7 (file:///Users/aleksandr/.nvm/versions/node/v22.20.0/lib/node_modules/@anthropic-ai/claude-code/cli.js:4808:1453)\n at YP (file:///Users/aleksandr/.nvm/versions/node/v22.20.0/lib/node_modules/@anthropic-ai/claude-code/cli.js:4808:1617)\n at Y8 (file:///Users/aleksandr/.nvm/versions/node/v22.20.0/lib/node_modules/@anthropic-ai/claude-code/cli.js:4808:1549)","timestamp":"2026-01-09T14:19:27.281Z"}]
17 Comments
I didn't feel any obvious AI becoming dumb in version 2.1.1. But after updating v2.1.2 today, tasks that were previously simple have now become difficult for Claude to handle
Environment Info
For example, in previous versions
v2.0.76 or v2.1.1, I used it to add some crud code for pgSQL. claude can quickly and accurately understand the original project structure, and provide a new complete plan in plan mode before implementation. But now inv2.1.2, claude has become less intelligent in plan mode. He seems to only understand half of the project structure and is always lazy during the development process, such as "adding crud tasks completely is too heavy. If you need it, please let me know and I will continue to complete the code. Currently, I have only completed the function ofgetlist"It's also possible that my usage was incorrect. I didn't target any contributors, I just wanted to give feedback on the user experience of v2.1.2
It might not be a CLI version issue - the backend itself may have changed.
I not only use
claude, but i also usegpt/deepseek/gemini/glmin CLI.I just saw that
v2.1.3has been released, let me use it nowMay be both sides. I'm on 2.0.76, same issues — worth trying pre-Jan 7 builds.
2.0.63+ enabled remote-gated session-memory compaction (tengu_sm_*) that reuses stale local summaries instead of
fresh ones. Rollback: npm i -g @anthropic-ai/claude-code@2.0.60
Note: CLI fetches config flags from Anthropic on startup regardless of which provider you use, so this affects all providers.
@jitlabs-sg Thank you very much. We need to give it a try
at least there was a deploy installation at the end of the session, after the update he just stopped doing it.
Experiencing similar issues on 2.0.77, suggesting this may be a backend/API change rather than a CLI version issue.
Having an issue in 2.0.76, after the recent outage. It has become kind of inaccurate and lazy. Keeps getting things wrong, doesn't listen properly, etc. Sounds like they messed up the configuration after moving servers around. ANTHROPIC! Please stop doing this. Whatever you did PLEASE revert it back to however it was a week ago.
I don't know if I can talk about this here. I've already started using
opencodeinstead ofclaudecode.I don't think anything will happen if you mention that.
If you want to try my patch, see here: https://github.com/anthropics/claude-code/issues/16539#issuecomment-3765845918
Closing for now — inactive for too long. Please open a new issue if this is still relevant.
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.