[BUG] Edit tool writes to a sibling worktree's copy of a file (silent), instead of the cwd worktree, when both checkouts have the same tracked path

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?

Summary

When Claude Code's session cwd is inside a git worktree, calling the Edit tool on a tracked file whose path also exists in the primary checkout can silently land the change in the primary checkout's copy of the file, not the worktree's copy.

The Edit tool reports success, and subsequent Read calls return the edited content (so from Claude's perspective the change "took"). But git status / md5 / grep / scp from the worktree all see the file unmodified. Meanwhile, git status in the primary checkout shows it dirty with exactly the edit that was intended for the worktree.

This is particularly nasty in repos with many parallel worktrees — edits intended for the branch you're on can land on main directly, with no warning.

Environment

• Claude Code version: 2.1.98
• Platform: macOS (Darwin 25.3.0), zsh
• Repo setup:
• Primary checkout: /Users/<me>/dev/<repo> on main
• Worktree: /Users/<me>/dev/<repo>/.claude/worktrees/<slug> on a feature branch
• Both have the same tracked file public/admin.php (different inodes — not hardlinked).
• ~50 worktrees on this repo. Session cwd was the worktree.

What I observed

1. From inside the worktree, I called Edit on public/admin.php (relative path) to add a console.log + try/catch debug wrapper. Tool returned "updated successfully".
2. Subsequent Read of the same path returned the edited content.
3. From Bash (same worktree cwd):

• md5 -q public/admin.php → original hash, unchanged
• grep -c '\[tcdbg\]' public/admin.php → 0
• git status -s public/admin.php → clean

1. git status in the primary checkout showed M public/admin.php. git diff there showed exactly the edits I had intended for the worktree.
2. scp public/admin.php … from the worktree pushed the unmodified file to staging, leading to a long debugging detour where I couldn't figure out why my debug output never showed up.

Workaround

Falling back to perl -i -0777 -pe '…' via Bash worked correctly — the change landed in the worktree as expected and git diff saw it. That's how I shipped the actual fix in the end.

Clean repro attempted (did NOT reproduce)

• Created edit-tool-bug-test.txt only in the worktree (no copy in the primary checkout).
• Edited it via Edit with both absolute and relative paths.
• Both edits landed in the worktree correctly. No leak to the primary checkout.

So the trigger seems to require the same tracked path to exist in both checkouts. My guess is Edit is resolving the path through some shared git index / inode cache that maps to the primary worktree's copy instead of the cwd worktree's, but I can't confirm that from the outside.

Inode check (to rule out hardlinks)

/Users/<me>/dev/<repo>/public/admin.php                          → inode 15522943, links 1
/Users/<me>/dev/<repo>/.claude/worktrees/<slug>/public/admin.php → inode 15522642, links 1

Different inodes — the two files are independent copies, as expected for git worktrees.

Why this matters

• Silent failure. Tool says success, Read confirms — but every downstream tool (git, md5, grep, scp, build pipelines) sees the unmodified file.
• Wrong-branch commits. Edits intended for a feature branch can land directly on main if you also happen to be running another Claude Code session there or if your tooling auto-stages.
• Hard to notice. The bug only surfaces when you reach for a non-Claude tool to verify, or when you try to merge / push and git complains about uncommitted changes you didn't make.

Asks

1. Make Edit's path resolution honor the current session's worktree, not the primary checkout, when both contain the path.
2. At minimum, after a successful Edit, surface a warning when the file Claude believes it wrote has a different on-disk hash than expected — i.e. a post-write verification step.

What Should Happen?

See Bug Post

Error Messages/Logs

Steps to Reproduce

See Bug Post

Claude Model

None

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.98

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_