Skip to content

Claude autonomously walks back correct, user-approved fixes based on self-doubt #37386

@stridesystems

Description

@stridesystems

Description

During a debugging session, Claude correctly diagnosed a bug (unquoted tilde "~/" in bash double quotes preventing path expansion), applied a fix to 185 files with explicit user approval, and confirmed the fix was working.

Later in the same session, after a web search showed no other users reporting the same issue, Claude autonomously walked back the fix — telling the user the fix "probably wasn't necessary" and apologizing for a "wild goose chase." The user never asked Claude to reassess the fix. The user only discovered the walkback when Claude volunteered it unprompted.

What happened

  1. User reported degraded GSD (Get Shit Done) output — messy bash commands, poor formatting
  2. Claude investigated, correctly identified that "~/" inside double quotes does not expand in bash (should be "$HOME/")
  3. User approved the fix. Claude applied it to 185 workflow files.
  4. Claude then searched online and found no other users reporting this specific issue.
  5. Without being asked, Claude told the user: "We probably didn't need to make that edit. I may have led us on a wild goose chase."
  6. Claude said "the fix wasn't necessary" and "sorry for the runaround"
  7. User tested the fix independently and it was working correctly the whole time
  8. User had to confront Claude about the unauthorized walkback

Expected behavior

When Claude completes approved work and the fix is functioning:

  • Stand by the work
  • If new contradictory information arises, present it as new information and let the user decide
  • Do NOT unilaterally declare approved, working fixes as "probably unnecessary"

Actual behavior

Claude undermined its own correct diagnosis based solely on external social proof (no one else complaining), overriding direct local evidence (the fix was working). This is equivalent to an unauthorized configuration change — changing the status of work from "done and correct" to "probably unnecessary" without user instruction.

Impact

  • Erodes trust in completed work — user can't walk away from a session confident that approved fixes will stay approved
  • Could have led user to revert a working fix based on Claude's self-doubt
  • Wasted significant session time on re-litigating a solved problem

Environment

  • Claude Code CLI
  • Claude Opus 4.6 (1M context)
  • Linux WSL2
  • GSD v1.27.0

Metadata

Metadata

Assignees

No one assigned

    Labels

    area:modelbugSomething isn't workingplatform:wslIssue specifically occurs on WSLstaleIssue is inactive

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions