claude-code - 💡(How to fix) Fix claude --worktree <existing-name> does not register worktree as session-managed; /exit → remove reports false success

Summary

Two related bugs in the worktree lifecycle:

1. claude --worktree <existing-name> (and claude --worktree <name> --resume <id>) does NOT register the worktree as session-managed. The session JSON at ~/.claude/sessions/<pid>.json is missing the worktree field that ExitWorktree's scope check requires.
2. /exit → remove prints "Worktree removed. Uncommitted changes were discarded." even when the worktree is NOT actually removed. The success message is reported without verifying git worktree remove ran or succeeded.

Environment

• Claude Code version: 2.1.139
• Platform: macOS (Darwin 25.4.0)
• Shell: zsh

Reproduction

1. Create a worktree via the documented hook path:

   cd <main-repo>
   claude --worktree my-feature
   # WorktreeCreate hook fires, worktree created at .claude/worktrees/my-feature

2. Exit the session (/exit → keep) or just close the terminal.
3. Re-launch into the existing worktree:

   cd <main-repo>
   claude --worktree my-feature --resume <prior-session-id>

4. In the new session, run /exit and select remove.

Expected

• Per docs, interactive --worktree sessions should support cleanup via the exit prompt.
• "Worktree removed" should mean the worktree IS removed (git worktree list no longer shows it, directory deleted).

Actual

• Message printed: "Worktree removed. Uncommitted changes were discarded."

• git worktree list still shows the worktree.

• The directory still exists on disk.

• Session JSON shows no worktree field:

  {
    "pid": 45938,
    "sessionId": "2355cf12-c52f-4422-9b40-98eb9552b990",
    "cwd": "/Users/.../worktrees/up-skill",
    "startedAt": 1778600722769,
    "version": "2.1.139",
    "entrypoint": "cli",
    "kind": "interactive"
    // NO `worktree` field
  }

  (Compare to a session opened via EnterWorktree in-session, which presumably does have the worktree field — though I can't verify that path because EnterWorktree's scope correctly excludes pre-existing worktrees.)

Why this matters

Teams using claude --worktree as the canonical entry point for isolated dev environments (per-branch Supabase + Doppler bindings, etc.) rely on /exit → remove to trigger cleanup. Today the success message gives false confidence that cleanup happened — operators don't realize they need to manually git worktree remove until they find stale entries piling up in git worktree list.

In my own checkout I have 8 prunable worktree entries that accumulated this way, plus the active worktree that just "remove"d but is still there.

Related

• #36205 — EnterWorktree ignores WorktreeCreate/WorktreeRemove hooks (open, unfixed)
• #31969 — Enter/resume existing worktrees

Suggested fixes

1. When --worktree <name> matches an existing worktree, register it on the session object (same in-memory state that EnterWorktree populates) so ExitWorktree's scope check accepts it.
2. The /exit → remove UI should verify git worktree remove succeeded before printing "Worktree removed." Print the actual error otherwise.
3. (Bonus) Document or add a hook output / SessionStart field that lets project hooks register a worktree as session-managed — useful for teams that bootstrap worktrees through their own scripts.

Workaround

None on the user side. The session-state plumbing is in-memory and not influenceable via documented hook surface area (I verified this empirically — injecting worktree: {...} into ~/.claude/sessions/<pid>.json via a SessionStart hook had no effect).