Chip-spawned task runs on primary checkout instead of fresh worktree

Summary

A task spawned by clicking a chip created via mcp__ccd_session__spawn_task opens a new session that runs on the primary checkout's current branch (e.g. main), not in a fresh git worktree under .claude/worktrees/agent-<hash>/. This contradicts the documented desktop-app contract that each session is automatically isolated in its own worktree.

Expected behavior

Per the docs:

• Worktrees: "The desktop app creates a worktree for every new session automatically."
• Desktop: "For Git repositories, each session gets its own isolated copy of your project using Git worktrees, so changes in one session don't affect others."
• Issue #53494 (chip flow): "User clicks the chip — task spawns to a fresh worktree."
• Issue #50109 (request for --no-worktree opt-out) was closed; no opt-out exists.

The chip-spawned session should:

1. Create .claude/worktrees/agent-<hash>/ under the project root.
2. Open the new session with its CWD inside that worktree.
3. Have HEAD on a fresh branch, not on the primary checkout's current branch.

Actual behavior

The chip-spawned session opens with its CWD on the primary checkout, on the primary's current branch. No .claude/worktrees/agent-<hash>/ directory is created for it.

Evidence from the Code-tab session footer of a chip-spawned session in a private repo:

<repo>  main           +0 -310     [Create PR]

The footer shows main (the primary branch), not the documented agent-<hash> worktree branch. The +0 -310 is a working-tree diff against main — i.e. the spawned session is editing files directly on the primary checkout. Any commit it makes will land on main of the primary, exactly the contamination the worktree contract is supposed to prevent.

Reproduction

1. From a parent Claude Code session, invoke mcp__ccd_session__spawn_task with a title, prompt, and tldr. A chip appears in the parent's UI.
2. Click the chip.
3. In the spawned session's footer (Code tab), observe the branch label.
4. Run inside the spawned session:

``bash
pwd
git rev-parse --show-toplevel
git rev-parse --abbrev-ref HEAD
git rev-parse --git-common-dir
``

• Expected: pwd ends in .claude/worktrees/agent-<hash>; --show-toplevel returns that worktree path; HEAD is a fresh branch; --git-common-dir points at the primary's .git/.
• Actual: pwd is the primary checkout; HEAD is the primary's current branch (e.g. main); the worktree was never created.

Impact

• Silent contract violation. Both the desktop docs and the chip's own UX (chip → "fresh session") imply isolation. Sessions appear isolated but mutate the primary tree.
• Risk of accidental commits to main. A spawned session that thinks it's in its own worktree (per the documented contract) will happily git commit and land work directly on the parent's working branch.
• Confounds debugging. Because the worktree contract is documented as unconditional, downstream rules and skills assume it.

Environment

• OS: Windows 11 Pro (10.0.22621)
• Claude Code desktop version: 1.9659.2 (390d6c)
• .claude/worktrees/ directory existed prior to chip-click: yes (sibling agent worktrees already present)
• Filesystem: local NTFS; no symlinks involved; path lengths well under Windows MAX_PATH

Notes

• Tested via the standard chip UX in the Claude Desktop "Code" tab; not via the CLI claude --resume flow.
• The mcp__ccd_session__spawn_task tool's own parameter surface (title, prompt, tldr) exposes no worktree-control flag, consistent with the docs treating worktree creation as the unconditional default.

Related issues, not duplicates

These all touch the same area but report different failure modes. This bug is "no worktree created at all" — the spawned session never enters worktree isolation in the first place.

• #57643 (closed) — spawn_task creates the worktree off origin/main instead of the parent's branch. Different: assumes a worktree is created and discusses its base branch.
• #52243 (closed) — Chips should branch from parent's branch, not repo default. Different: same as above, assumes worktree exists.
• #57512 (open) — spawn_task should auto-add .claude/worktrees/ to .git/info/exclude. Different: assumes the worktree directory is created.
• #59628 (open) — Worktree sessions can edit files in the parent main checkout via absolute path. Different: the worktree exists; the agent escapes it via absolute paths. Here, no worktree exists to escape from.
• #36182 (open) — Worktree Edit/Read tools use main-workspace paths instead of worktree paths. Different: worktree exists; tool path-resolution bug.
• #63008 (open) — preview_start MCP spawns dev server with main-repo cwd from a worktree session. Different: worktree exists; downstream MCP child inherits wrong cwd.