Preflight Checklist

• I have searched existing issues and this hasn't been reported yet
• This is a single bug report (please file separate reports for different bugs)
• I am using the latest version of Claude Code

What's Wrong?

Bug — Write (and Edit) tools strip the worktree path segment, writing to the parent repo instead

Affects: Claude Code 2.1.128 on macOS Darwin (arm64). Severity: HIGH — silent cross-branch contamination. Files intended for an isolated worktree branch land on the parent repo's working tree and can be picked up by concurrent git operations (e.g. another Claude Code session running on the parent repo) and committed to a different branch.

TL;DR

When operating inside a git worktree located at <repo>/.claude/worktrees/<name>/, the Write tool silently strips the .claude/worktrees/<name>/ segment from the file path and writes to <repo>/<rest-of-path> (the parent repo). The path passed to the tool is absolute and correct; the strip happens inside the tool.

The Edit tool behaves differently but is also broken: it reports success while neither the worktree file nor the parent repo file is actually modified — silent no-op.

Both bugs combine catastrophically when a separate Claude Code session is running on the parent repo (e.g. a background grinder / loop runner). That session's git add operation can sweep up the orphan files and commit them to a different branch, contaminating it without anyone realizing.

Environment

• Claude Code: 2.1.128
• OS: Darwin 25.x arm64 (macOS, Apple-Silicon hardware)
• Git: 2.50.x (Apple Git)
• Shell: zsh
• Worktree topology (the documented Claude Code pattern for isolation: "worktree"):

<repo>/                            ← main checkout, current branch = X
<repo>/.claude/worktrees/<name>/   ← worktree, current branch = Y

.claude/worktrees/ is gitignored in the parent repo, so the worktree is isolated by design.

What Should Happen?

Not spill data editing changes to active branch

Error Messages/Logs

Steps to Reproduce

Reproduction

1. Open a Claude Code session in a worktree at <repo>/.claude/worktrees/<name>/ (e.g. via the Agent tool with isolation: "worktree").
2. From that session, call the Write tool with an absolute path inside the worktree:

   file_path: <repo>/.claude/worktrees/<name>/.plan/example.md
   content:   "# Test"

3. Write returns "File created successfully at: <full-path>" — appears successful.
4. From within the worktree, run ls <repo>/.claude/worktrees/<name>/.plan/example.md via Bash — file does not exist.
5. From within the worktree, run ls <repo>/.plan/example.md via Bash — file IS there (parent repo, NOT the worktree).
6. If the parent repo currently has a different branch checked out and is being used by a concurrent process, that file is now sitting in that branch's working tree, awaiting git add.

What I observed in a real session

• Worktree on branch feature/<worktree-name>. Parent repo had a concurrent agent process running on branch feature/<concurrent-branch>.
• Write call:

  file_path: <repo>/.claude/worktrees/<worktree-name>/.plan/<example-file>.md

  reported success.
• ls on the same absolute path returned "No such file or directory".
• I assumed Write failed and re-created the file via cat > heredoc (which correctly wrote to the worktree).
• Later, the user reported from the OTHER Claude Code session: "that file ended up in our branch."
• Investigation showed the Write-tool file at <repo>/.plan/<example-file>.md (parent repo — .claude/worktrees/<worktree-name>/ had been stripped from the path).
• The concurrent agent's next git add (broad pattern) committed the orphan file to its branch as part of an unrelated commit.

File evidence (two copies on disk after the heredoc fallback):

<mtime-1>  <size-1>  <repo>/.plan/<example-file>.md
                     (Write tool — wrong location, parent repo)
<mtime-2>  <size-2>  <repo>/.claude/worktrees/<worktree-name>/.plan/<example-file>.md
                     (heredoc fallback — correct location)

Mtime delta ≈ 1m45s — Write's misdirected write came first, the heredoc came second.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

No response

Claude Code Version

2.1.128

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Other

Additional Information

Companion bug — Edit tool in the same setup

Same worktree, several Edit calls earlier in the session targeting files inside the worktree:

• Each Edit call returned "The file ... has been updated successfully."
• git status / git diff on the worktree showed no changes.
• Bash grep for the new content on both the worktree AND the parent repo's working tree returned the original content — neither version of the file was actually modified.
• Subsequent sed -i '' calls via Bash on the same patterns succeeded — proving the file was writable and the search patterns were valid.

Edit here is a silent no-op rather than a silent misdirect — but it's just as misleading because the harness reports success and proceeds.

Workaround

Until fixed, in worktree sessions:

• Don't use Write or Edit.
• Use Bash for all file changes:
  • New file: cat > path << 'EOF' ... EOF
  • Edit: sed -i '' 's|old|new|' path
• Bash respects pwd correctly; pwd in a worktree session is the worktree.

Suggested fix direction

The Write tool's file-creation path likely resolves through some abstraction that strips paths it considers "outside the canonical repo root" — .claude/worktrees/... would qualify as ".claude is a Claude-internal folder, strip it." That heuristic is wrong for the worktree case: .claude/worktrees/<name>/ is the actual working tree.

The Edit tool's no-op may share the same root cause (path resolution sends the call to a sandbox/snapshot it can't write back from).

Suggested checks:

• Detect git worktree mode (git rev-parse --git-dir returns <parent-repo>/.git/worktrees/<name> when inside a worktree) and bypass any "strip Claude-internal paths" logic.
• For the Edit tool, surface the actual disk-write outcome rather than reporting blanket success.

Why this matters at scale

Anyone using isolation: "worktree" mode (the documented Claude Code pattern for parallel agent work) + a concurrent process on the parent repo (grinder, loop runner, manual coding session, another Claude Code instance) can have files silently cross-contaminate between branches. The user may not notice for hours; the contamination only surfaces when:

• Someone reads the wrong branch's tree, OR
• A git add -A operation in the parallel session sweeps up the orphan file, OR
• A rebase produces inexplicable conflicts.

For multi-session workflows this can quietly corrupt commit history on the parent branch.