Single-bean mode

The simplest mode. One task, isolated, merged, cleaned up. No fleet, no parallelism, no long-running daemon.

The loop

# 1. Create a task
beans create "Cache token lookups" -t task -d "Memoize the resolver."

# 2. Run an agent on it
hordr run hordr-0001
#    → isolated worktree on branch hordr-0001, agent in its own pane

# 3. The agent works, commits, marks the bean completed, runs /done

# 4. Finish — merge + cleanup
hordr finish hordr-0001

What hordr run does

1. Resolves the bean’s assigned: role → persona + harness.
2. Creates (or reuses) a worktree on hordr-<id>, branched from primary.
3. Opens a herdr pane and launches the agent with the bean body as its brief.
4. Returns. The agent runs autonomously in the pane.

You are free to watch the pane, ignore it, or close it — the worktree persists.

The done gate

When the agent finishes, it marks the bean completed and runs /done. Hordr’s done-checks validate the result (status, commit, files). If anything is off, the lane stalls in a recoverable state (uncommitted, blocked) rather than silently producing a bad merge.

What hordr finish does

1. Asserts the bean is completed (read from the worktree, not stale main).
2. Merges hordr-<id> into the primary branch (--no-ff).
3. Removes the worktree.
4. Safe-deletes the branch — git branch -d, never -D. -d refuses unmerged branches, so a silent no-op merge can’t slip through.

If the worktree is already gone (a re-run), finish still merges and deletes the branch. Branch deletion is tolerant: on failure it warns and continues — the merge already landed; an orphaned ref is manual cleanup.

When to use it

• One-off tasks and chores.
• Anything you’d otherwise do in a throwaway branch by hand.
• The first stop before reaching for a fleet.