fix(source-control): /worktree creates in-repo .claude/worktrees/ -> hits CC double-load bug; create out-of-tree

[kyle-sexton]

Summary

The /worktree skill creates worktrees at Claude Code's native in-repo location (.claude/worktrees/<name>). Working inside such a nested worktree triggers a confirmed, unfixed Claude Code bug that double-loads CLAUDE.md + .claude/rules/ (~26K wasted input tokens per message). The skill should create worktrees out-of-tree to avoid it.

Component

plugins/source-control/skills/worktree/context/create.md (currently calls EnterWorktree, which lands in .claude/worktrees/).

Evidence

• Upstream, confirmed & unfixed: anthropics/claude-code #29599 ("Worktree rules loaded twice", closed as duplicate) and #23565 ("Memory files loaded twice in worktree", closed not-planned). Mechanism: a worktree at .claude/worktrees/<n> is physically inside the parent repo; CC's upward walk from the worktree discovers both the worktree root and the parent repo root (shared .git), loading rules/memory twice. Reported waste ~12–26K tokens/message.
• Related: .claude/worktrees/ is not auto-excluded and adding it to .gitignore breaks git worktree add in future clones (#57512 — use .git/info/exclude); worktree .claude/ is missing skills/agents/rules/ (#28041).
• A sibling/external worktree (not nested under the parent repo) is not an ancestor of the parent, so the upward walk does not hit the parent → the double-load does not occur. Out-of-tree placement sidesteps the bug entirely.

Expected

/worktree create places worktrees out-of-tree (via the shared helper / external root), not under .claude/worktrees/, so the double-load bug never triggers.

Acceptance criteria

• /worktree create no longer produces in-repo .claude/worktrees/ worktrees for ghq/normal repos.
• Placement flows through the shared helper (unify issue) / external root (epic).
• Note the upstream bugs in the skill's context docs as the rationale; add .git/info/exclude guidance if any in-repo path remains for a fallback.

Part of / depends on

Epic #398; implemented via the shared helper in the unify issue.

[kyle-sexton]

Lane

babysit (safe tier, /loop 15m /source-control:babysit-prs)

Triggering example

Same babysit-prs cycle, 4 of 5 dispatched fix-round workers hit exactly this in-repo .claude/worktrees/<name> collision when their assigned out-of-tree worktree path (under the plugin's babysit_worktree_root) didn't exist yet:

• PR fix(source-control): review-trigger head-staleness hardening (BLOCKED-behind guard, stale-head reactions) #452, feat(source-control): babysit-prs merge-step STOP, bare-wrapper pipe rule, snapshot exit-code taxonomy, cross-PR dependency channel #456, medley#1567: branch was already checked out in the user's own interactive worktree at D:/repos/.../.claude/worktrees/<slug> (e.g. 327-review-trigger, 373-babysit-enhancements, medley-1567) — each worker reused that existing checkout in place rather than creating a second one on the same branch (git forbids two checkouts of the same branch).
• PR fix(implementation): defer branch-name grammar to consumer convention #453: same collision, but the worker instead created its assigned worktree at detached HEAD on the exact SHA and pushed via HEAD:fix/403-branch-grammar-agnostic to avoid writing into the user's interactive worktree.

Observed vs expected

Observed: babysit-prs workers have no consistent instruction for "assigned worktree path is free, but the branch is already checked out elsewhere in-repo" — three workers reused the foreign worktree in place (fine, but undocumented), one worked around it via detached-HEAD + refspec push (a workaround, not a documented contract). Expected: once /worktree moves to the out-of-tree helper (this issue + #399), babysit-prs's own worker contract (reference/worktrees.md) should point at the same shared out-of-tree convention so "branch already checked out in-repo" stops recurring as the common case, and the worker contract should say explicitly what to do when it does (reuse in place vs detached-HEAD workaround) rather than leaving each worker to improvise.

Category

operational

[kyle-sexton]

Parked — blocked by two unresolved needs-decision epics the dispatch didn't surface upfront (orchestrator gap, not a worker miss).

Worker investigation found #400's own acceptance criteria bind worktree placement to "the shared helper (unify #399) / external root (epic #398)" — both status: needs-decision, both unresolved:

• feat(source-control): machine-agnostic worktree setup skill (detect -> decide -> install) #398 (layout epic): explicitly disqualifies the per-repo sibling pattern (<repo>-issue-<N> next to the canonical checkout) as "pollutes ghq list — disqualified where one-folder-per-repo matters," and recommends a central external root instead. Verified empirically: ghq root = D:\repos, and the canonical checkout lives under it — so the sibling pattern (which this very lane's own subagent dispatches have been using all session) genuinely does pollute ghq list.
• feat(source-control): unify native-hook + /worktree-skill worktree creation via one shared helper #399 (unify epic): both the native WorktreeCreate hook and the /worktree skill need to route through one shared helper script for identical placement. Fixing /worktree alone now — before that helper exists — recreates the exact divergence feat(source-control): unify native-hook + /worktree-skill worktree creation via one shared helper #399 exists to reconcile.

RECOMMENDED: (a) wait for #398/#399 to resolve first, then implement #400 against their settled layout/helper contract rather than shipping a placement choice now that pre-empts both decisions.

Options:

1. (a) Wait for feat(source-control): machine-agnostic worktree setup skill (detect -> decide -> install) #398/feat(source-control): unify native-hook + /worktree-skill worktree creation via one shared helper #399 (RECOMMENDED) — fix(source-control): /worktree creates in-repo .claude/worktrees/ -> hits CC double-load bug; create out-of-tree #400 is effectively blocked-by them; its own acceptance criteria say so.
2. (b) Interim central root — adopt source-control:babysit-prs's existing babysit_worktree_root (the plugin-data-dir worktrees/ subdirectory, already shipping in this same plugin, environment-derived not hardcoded) as a stopgap. Still pre-empts feat(source-control): machine-agnostic worktree setup skill (detect -> decide -> install) #398's final layout call and stays divergent from the native hook until feat(source-control): unify native-hook + /worktree-skill worktree creation via one shared helper #399 lands.
3. (c) Override the epic — take the sibling pattern anyway, accepting the ghq list pollution feat(source-control): machine-agnostic worktree setup skill (detect -> decide -> install) #398 explicitly disqualifies it for.

No code changed, nothing to push. Labeling status: needs-decision.