Start Stack

---
name: start-stack
description: Coordinate substantial project work with one current-parent plan, bounded worker contexts, real parallel scopes and one integration; use existing TODO execution rather than duplicate planners or contracts.
---

# Start stack

Use for substantial work, not as a mandatory multi-agent ceremony for small prompts.

1. Orient from project AGENTS.md and relevant current project metadata. State the
   intended usable outcome and exclusions briefly. Continue independent work while
   clarifying a real ambiguity; ask only for missing decisions that affect the task.
2. Apply mandatory semantic skill selection. The current strong parent plans once:
   requirements, short TODO, dependencies, concrete files/workspaces, maker/reviewer,
   tests and role-based UAT where applicable. A cheap parent requests one qualified
   plan. No planner witness, sealed contract or audit of an audit.
3. Choose roles needed by this task. An independent product reviewer may also review
   its security; add a specialist only for a separate material risk. Use paid CLI
   capacity according to current model capabilities and real availability.
4. Small delegated scope uses todo_pair.py. Real dependent or parallel scopes use
   one todo_graph.py scheduler. Serialize overlapping files/resources, not entire
   unrelated trees. Capacity sets concurrency. Do not create a second scheduler.
5. Parent reads Hub once and gives each worker a bounded relevant brief. Office
   coordinates the task's channel and private inbox at meaningful checkpoints.
   Preserve aliases, scope announcements and ACKs; messages do not grant authority.
6. Accept finished vertical slices while independent implementation continues when
   environments/resources do not overlap. Pin the tested artifact; later changes
   invalidate affected evidence. Avoid rerunning unchanged tests for a report repair.
7. Integrate once and test the actual changed behavior and relevant regressions.
   Use function | exists | tested | integrated | missing for material outcomes.
   Report artifact, transport, product and knowledge states separately.

For ML/trading, use the relevant domain skill and verify leakage, the real target,
OOS evaluation and compute constraints. Research is not authorization to trade.
For UI, apply design-ui and design-loop when its trigger applies; real product
UAT remains necessary. Governance checks belong in the existing TODO.

Keep no-delete, data protection, production consent and required private backup.
End sessions with a canonical Hub trace; a missing receipt is not COMPLETE.
Historical recovery alone uses references/legacy-recovery.md; new work never loads
old fidelity/sufficiency, planner-witness or acceptance-contract workflows.