Skip to content

Coordinator Mode

Coordinator Mode is for work too big for a single prompt. Any harness can split a task up; what Coordinator gives you is the workplace around it. Every subtask runs in its own git worktree on its own branch off a shared base commit, so nobody works on top of anyone else. The results land on an integration branch, never straight on your main. A plan you approved goes in one end, and a claim-by-claim audit against the real diff comes out the other.

Every subtask gets a real git worktree on its own branch, cut from the same base commit. Same-file edits in different worktrees stop being a conflict, so the work is genuinely parallel rather than serialized behind a lock. Subtasks merge into an integration branch (feature/coordinator-<id>); your main branch is untouched until you approve.

You hand Coordinator a high-level goal and it decomposes it into parallel subtasks. If decomposition fails, a banner offers a Retry rather than leaving you stuck.

  • Gate 1 — approve the plan. Before any work starts, you review the proposed breakdown and a cost estimate and approve the plan.
  • Gate 2 — review the merged diff. Before anything reaches your main branch, you review the combined diff of all subtasks. Nothing merges silently.

Gate 2 is backed by a verification step. Before merge, Apollo mechanically checks each subtask’s claims against the actual diff — a deterministic git and grep layer plus a single interpretive model call — and surfaces any divergence between what a subtask said it did and what the diff shows. Each divergence expands with Investigate, Accept, or Send-back actions, so you resolve mismatches deliberately instead of trusting a summary.