AI/Coding/Idea/26/08/Standing agent inboxes on long-lived worktrees

• Standing agent inboxes on long-lived git/worktrees

  • Instead of creating a worktree per task, keep a small set of named, long-lived agents, each owning one persistent worktree. The worktree becomes an inbox: send it a task, it works, it parks. Prompted by daily use of herdr.

  • What herdr already supplies

    • Agents are addressable by a chosen name rather than only a pane ID — herdr agent start|prompt|wait|rename <name>.
    • Worktrees are first-class: herdr worktree create|open|list|remove, and worktree workspaces pack under their parent in the sidebar. That packing is per-repository and not settable, so an agent whose worktrees span several repos gets one group per repo rather than one under itself — herdr/Explanation/Where a Workspace Appears in the Sidebar.
    • A name is unique only among live agents, and clears when that agent exits, is released, or is replaced. The durable identity is therefore the worktree workspace; the name is a label re-applied on agent start.

  • Why the worktree, not the branch, is the unit

    • A branch can be checked out in at most one worktree at a time — a per-branch lock, not a per-worktree pin. A standing worktree can cycle through branches indefinitely. Mechanics on git/worktree.
    • Resting state between tasks: detached at origin/main. Clean tree, no branch held, nothing for another worktree to collide with.
    • Accepting a task never requires main to be checked out. git switch -c task-N origin/main uses it as a start point, which the lock does not restrict, so a standing agent never contends with the primary checkout.

  • What is actually saved

    • Not git worktree add — that is milliseconds.
    • Untracked and ignored files survive branch switches: .env, node_modules/, build caches, language-server indexes. That is the real prize, and it means the repos worth a standing agent are the ones with expensive setup, not the ones touched most often.
    • The agent's accumulated context, but only where the agent's identity is a domain (docs, garden, frontend) rather than a task queue. Across unrelated tasks the context has to be cleared between them, leaving only the environment saving.

  • Open questions

    • Queue discipline: an inbox implies tasks can arrive faster than they finish, but the worktree is one mutable checkout. Work has to be landed or parked before the next task is accepted.
    • The stash is repo-global rather than per-worktree, so several standing agents parking via stash share one stack — git/worktree/Q/How does git stash interact with git worktrees.
    • Directory naming stops following the branch once a worktree outlives it — My/Pref/Dev/Tool/git/Worktree.