fix: a worker in the app's worktree moves to the task's worktree beside the main checkout

[mauricios]

Why

In an adopting project, every task worktree ended up nested in main/.claude/worktrees/ under a generated name (silly-cannon-b424ff, …), and none sat beside the main checkout. Two things combined:

1. The desktop app gives every chip's session a worktree of its own, at its worktree location (inside the project by default), on a claude/<name> branch. The session never opened in the main checkout, as decision 0021 assumed.
2. The session-context hook kept it there. It told a session in Claude Code's worktree "this worktree is yours for one task… rename the generated branch", following decision 0015. The worker's prompt ("create this task's worktree with adf-worktree-new and move into it") lost to the hook.

Nested in the main checkout, the worktrees also walked up into its tool configuration: ESLint failed on the main checkout's config.

What changed

• plugins/adf/hooks/session-context.sh, with the parallel-agents module:
  • A session in Claude Code's own worktree on a generated branch or a detached HEAD is told this is where it starts, not where it works. Its first step is adf-worktree-new <branch> --no-start, with the branch its prompt names or the one triage names, then a move to the path it prints (change_directory, which the approve-worktree-move hook approves; EnterWorktree in a terminal).
  • It edits and commits nothing there, and never renames the generated branch.
  • The worker line no longer claims the current worktree: "One task, in its own worktree".
  • A worktree already on a task branch keeps today's notes (what it lacks).
• /adf:dispatch, docs/PARALLEL-AGENTS.md, the parallel-agents scenario: they no longer tell the developer to start the chip "in this folder, not in a new worktree". The app decides that, and the worker moves either way. A rationalization row covers "the chip already gave me a worktree".
• Decision 0040 amends 0021 and 0015. The changelog entry carries the upgrade impact. Without the module, nothing changes.

For reviewers

• Two worktrees per task while the session lives: the app's, which stays empty, and the task's. The app's goes when the session is archived (/wrap-up).
• The developer could also set the app's worktree location to the main checkout's parent. 0040 records that as an alternative: it's a per-developer setting the framework can't apply, and the app's worktrees still get generated names and no setup.

Verification

• ./evals/run-evals.sh: all suites pass.
  • Hook tests: a worker in the app's worktree on a generated branch, and on a detached HEAD, gets the create-and-move first step and no git branch -m. A worktree already on a task branch keeps its notes.
  • Static checks: two new ones (/dispatch and the hook), and one updated to the hook's current wording.
• python3 scripts/build-committed.py --check: the round trip holds, with 65 files.
• Not verified: a real chip session in the desktop app running the new first step.

🤖 Generated with Claude Code