fleet-dispatch

This is the orchestration layer above /implement. Per-ticket cost controls (scoped test commands, one batched code-simplifier pass, /code-review medium, splitting a large ticket) live in /implement. Don't restate them in dispatch prompts; each subagent runs /implement and gets them from there.

1. Triage the batch

For each ticket, decide: inline, dispatch now, or dispatch later.

• Inline. Single-file change, no schema or API change, no new tests beyond one case. Do it in this session. A worktree, a subagent spawn and a skill reload cost more than the ticket.
• Dependency order. If tickets form a chain (a staged migration, a series of refactors), stack them: ticket N branches off N-1's branch and its PR targets that branch, so N starts once N-1 is committed, not merged. Its prompt points at N-1's diff as the worked example. Independent tickets branch off main and share a wave.
• Stack depth. Cap a stack at 3. A rejected approach low in the stack wastes everything above it, so if the chain is longer, wait for the bottom PR's approach to be agreed before stacking further.