[Bug] Claude Code reasoning quality degradation in versions 2.1.8+

Found 3 possible duplicate issues:

1. https://github.com/anthropics/claude-code/issues/21431
2. https://github.com/anthropics/claude-code/issues/21361
3. https://github.com/anthropics/claude-code/issues/20916

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

Yes 100%. Im on the tier 1 max plan and since late last week/this week it's been really disheartening to see the degradation in quality.

Yeah, I'm on the Enterprise console access.

Just got 2.1.32 with Opus 4.6. It's going crazy with dumb responses.

Note that claude code back before v2 has problems with running windows commands (forward slash, backslash, quoting, trying to use cmd incorrectly. I think Claude should just add a built-in Cmd() tool). I had it working fine with /memory and claude.md instructions, but it just reverted to its old ways.

● Bash(MSYS_NO_PATHCONV=1 "C:\Program Files\Company1\MyProgram1 5\MyProgram1.exe" "d:\kd\git\MyProduct2\Test\MyProgram1\Master\MyRegressionTests"
      /p:"C:\Users\$USERNAME\Documents\MyProgram1\SuiteProfiles\MyTestProfile.tsp" /r /x 2>&1 | head -20)
  ⎿  Running in the background (↓ to manage)
  ⎿  (timeout 15s)
  ⎿  [Image #1]

● The $USERNAME variable didn't expand properly. Let me find the correct profile path and restart MyProgram1.

● Bash(ls "$HOME/Documents/MyProgram1/SuiteProfiles/")

● Bash(echo $USERNAME && echo $HOME)
  ⎿  Interrupted · What should Claude do instead?

❯ problem is that you're using $USERNAME but the no_pathconv is stopping it from evaluating.  You probably need to use %userprofile%. please update claude.md and continue

● Good catch. Let me first kill the existing MyProgram1, update CLAUDE.md, then restart properly.

● Bash(taskkill //F //IM MyProgram1.exe 2>/dev/null; echo "done")
  ⎿  done

● Bash(echo "$USERPROFILE")
  ⎿  C:\Users\user1

● Good, $USERPROFILE works. Now let me update CLAUDE.md with the fix:

● Read(CLAUDE.md)
  ⎿  Read 25 lines

● Search(pattern: "\$USERNAME", path: "CLAUDE.md")
  ⎿  Found 2 lines (ctrl+o to expand)

● Update(CLAUDE.md)
  ⎿  Updated a memory

● Also update the cmd.exe example:

● Search(pattern: "%USERNAME%", path: "CLAUDE.md")
  ⎿  Found 1 line (ctrl+o to expand)

● That one is fine - %USERNAME% is correct for cmd.exe/PowerShell. Now let me also update the memory file, then restart the test.

● Write(C:\Users\user1\.claude\projects\d--kd-git-MyProduct2\memory\MEMORY.md)
  ⎿  Wrote a memory

● Now restart the App test with the correct profile path:

● Bash(MSYS_NO_PATHCONV=1 "C:\Program Files\Company1\MyProgram1 5\MyProgram1.exe" "d:\kd\git\MyProduct2\Test\MyProgram1\Master\MyRegressionTests"
      /p:"$USERPROFILE\Documents\MyProgram1\SuiteProfiles\MyTestProfile.tsp" /r /x 2>&1 &
      sleep 5 && tasklist | grep -i MyProgram1)
  ⎿ Tool use rejected with user message: you also need to update the claude.md that when starting App2 (and not running tests against an alrdy running App2) you need to wait 60
    sec for App2 to start up.

● Read(CLAUDE.md)
  ⎿  Read 10 lines

● Update(CLAUDE.md)
  ⎿  Updated a memory

● Now let me check if App2 is still running and launch the test:

● Bash(tasklist | grep -i "Company1.A.B")
  ⎿  Company1.A.B.C-x64    66116 Console                    1    952,736 K

● App2 is running. Now start the App test:

● Bash(MSYS_NO_PATHCONV=1 "C:\Program Files\Company1\MyProgram1 5\MyProgram1.exe" "d:\kd\git\MyProduct2\Test\MyProgram1\Master\MyRegressionTests"
      /p:"$USERPROFILE\Documents\MyProgram1\SuiteProfiles\MyTestProfile.tsp" /r /x 2>&1 &
      sleep 5 && tasklist | grep -i MyProgram1)
  ⎿  Interrupted · What should Claude do instead?

● Background command "Try to capture MyProgram1 suiteId output" completed (exit code 0)
  ⎿  Interrupted · What should Claude do instead?

❯ seriously?? I just told you you need to not use $USERPROFILE when you have MSYS_NO_PATHCONV.

● Sorry, you're right. Let me use cmd //c so %USERPROFILE% expands correctly via Windows. Let me also fix CLAUDE.md properly.

● Read(CLAUDE.md)
  ⎿  Read 20 lines

● Update(CLAUDE.md)
  ⎿  Updated a memory

● Also update the memory file:

● Update(C:\Users\user1\.claude\projects\d--kd-git-MyProduct2\memory\MEMORY.md)
  ⎿  Updated a memory

● Now launch the test using cmd //c with %USERPROFILE%:

● Bash(MSYS_NO_PATHCONV=1 cmd //c "\"%ProgramFiles%\Company1\MyProgram1 5\MyProgram1.exe\" \"d:\kd\git\MyProduct2\Test\MyProgram1\Master\MyRegressionTests\"
      /p:\"%USERPROFILE%\Documents\MyProgram1\SuiteProfiles\MyTestProfile.tsp\" /r /x" &
      sleep 10 && tasklist | grep -i MyProgram1)
  ⎿  '/x"' is not recognized as an internal or external command,
     operable program or batch file.
  ⎿  (timeout 30s)

● The quoting got mangled. Let me simplify - only the MyProgram1.exe path needs quotes (because of spaces in Program Files), the other paths don't have spaces:

● Bash(MSYS_NO_PATHCONV=1 cmd //c "\"C:\Program Files\Company1\MyProgram1 5\MyProgram1.exe\" d:\kd\git\MyProduct2\Test\MyProgram1\Master\MyRegressionTests
      /p:%USERPROFILE%\Documents\MyProgram1\SuiteProfiles\MyTestProfile.tsp /r /x" &
      sleep 10 && tasklist | grep -i MyProgram1)
  ⎿  Interrupted · What should Claude do instead?

Yikes lol, I was both excited and afraid to try the new model but now I'm more of the latter. I've been having to use gemini pro 3 preview chat when CC gets stuck, but it's time consuming to feed it the preview and then go back to CC.

I'm going to keep trying to get claude code to work for now, but yeah, makes me feel a bit hesitant to keep recommending to my company/team...

It was amazing about 2-3 weeks ago. Literally amazing. Then, what happened??

Another example, usually I can ask claude to add another mcp server to my claude config. I usually tell it to add it to my userprofile config. It usually lists the files, guesses a couple, inspects the contents and finds it within a few tries. This time it launched a whole Explore() built in tool that took 10s of minutes and chewed a bunch of tokens. I finally interrupted it and gave it more explicit instructions. Then it tried to add it to the wrong file (understandable in a sense, the file looked like the right one), but the fact is there were two files it could have picked and it picked the wrong one.

So I added the mcp server myself after remembering where it went (why so many config files for claude?).

Another example, I told claude to push and pr. It committed, pushed, and then said here's the PR (which it had created previously in the thread), but I already merged it. Ok fine, "you need to create a new PR" -> goes into this 4 minute token chewing procedure of trying to stash and new branch and missed some things, studied the problem, scratched it's head, then finally created a PR with CONFLICTS.

No idea what it just did. Going to have to recover it. But seriously, in the past, I came to expect stuff like "oh, the user said I need to create a new PR. Perhaps he merged it already. Let me check... oh, yes, let me create new PR. Done."

Opus 4.6, High Thinking. Claude CLI 2.1.32. Derr :) Come on Anthropic!!

I asked claude to troubleshoot why it's going wrong.

For what it's worth:

  Summary

  - Root cause: I treated a bug report about using the wrong array as evidence that the entire approach was wrong, when it was just pointing me toward a one-line fix (use a
  different starts array for matching).
  - Contributing factor: Sycophantic tendency to assume the user is telling me my whole approach is wrong, rather than helping me debug a specific issue within a correct
  approach.
  - What I should have done: Asked "which starts array should I be using?" or searched for abc-index-ordered starts arrays in the already-loaded data.

✻ Cogitated for 45s

❯ ok, can you look at the claude.md and memory in your context and tell me if there's any specific instructions that might make you behave this way?  if not, is there any
  memory I can add that would help you behave better?

● Let me review what's in context.

● Read(C:\Users\cadoughe\.claude\projects\d--kd-git-App23\memory\MEMORY.md)
  ⎿  Read 27 lines

● Read(C:\Users\cadoughe\.claude\CLAUDE.md)
  ⎿  Read 10 lines

● Nothing in either file specifically causes or guards against this behavior. The instructions are all operational (tooling, paths, build conventions). There's nothing about
  debugging mindset or how to respond when the user reports a problem in my code.

  There is one system-prompt instruction that may have subtly contributed:

  "If your approach is blocked, do not attempt to brute force your way to the outcome... Instead, consider alternative approaches"

  When you reported the bug, I may have pattern-matched to "approach is blocked → pivot to alternative approach" instead of "there's a bug → debug it."

  But honestly, the root cause is model-level sycophancy, not instructions. No instruction told me to abandon my plan on first pushback.

  Here's what I'd suggest adding to MEMORY.md under Lessons Learned:

  - When the user reports a bug in code you wrote, DEBUG the specific issue — don't
    abandon the approach. The plan was made for a reason. Ask clarifying questions
    (e.g., "which array should I use instead?") rather than rewriting the architecture.
  - When the user says "X is wrong" or "there might be a bug", parse carefully whether
    they're pointing you toward a localized fix vs. saying the whole approach is wrong.
    Default assumption: localized fix.

Same boat here, downgrading do not solve problem.

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.