rota-brainstorm — Per-item Design

Sits between /rota-capture (what) and /rota-plan (how). Scope is one backlog item (#N; file backend B07/F03/T05); project-level exploration is /rota-vision. The design (rota design show <ID>) is soft input to /rota-plan, never required. Milestone tags are context only.

Copy this checklist and track your progress:

- [ ] Step 1 — Resolve target
- [ ] Step 2 — Load context
- [ ] Step 3 — One draft
- [ ] Step 4 — One approval
- [ ] Step 5 — Store

Step 1 — Resolve target

Target: #N / bare number (issue backend) or [BFT]\d{2,} (file backend), plus optional --grill (Step 3). Reject milestone and slice IDs: "/rota-brainstorm operates on a single backlog item. For project-level exploration use /rota-vision; for slice planning use /rota-plan." rota item field get <ID> --name title exits 3 for an unknown item: refuse and point at /rota-capture.

If a design exists, ask once (View / Edit / Replace, default View): View runs rota design show <ID> and exits; Edit loads it as the starting draft; Replace runs rota design rm <ID> first.

Step 2 — Load context

Follow references/context-load-protocol.md (parallel, silent), plus rota item show <ID> for state and comments (decision comments bind; references/issue-mode.md, "Resuming an item") or the file backend's detail file, and rota glossary read <term>. DECISIONS matches are hard boundaries: on a conflict, surface it before drafting.

## Acceptance already in the issue body? Treat it as settled; Step 3 covers approaches only. Otherwise draft it in Step 3.

Step 3 — One draft

Grill first, once. Run one pass of references/grilling.md before drafting when the item is a Major feature or P0 bug, or the invocation has --grill. Skip it when the issue body already has ## Acceptance (--grill does not override). For other items ask nothing unless real ambiguity or a decision conflict remains after the item, its thread and the code; then ask the one question that unblocks you (AskUserQuestion, ≤ 4 options). Record answers that changed direction in Step 5. If a question needs code-touching evidence (feasibility, library support, performance), say: "This warrants a spike. Run /rota-spike <name> first, then re-invoke /rota-brainstorm <ID>." Don't guess.

Write the whole design as one unsaved markdown draft:

• Goal: one sentence on what shipping this means
• Design: 3-8 sentences: chosen shape, moving parts, where they live
• Approaches considered: 2-3 candidates, each with shape, pros, cons, deciding factor; mark the pick. A simple item may have one approach and a two-line design.
• Acceptance: a - [ ] checklist, observable outcomes (skip when the issue body already has one)
• Open questions: what must be answered before or during /rota-plan; mark spike candidates
• Assumptions: implicit constraints made explicit

Self-review before showing: no placeholders, no contradictions (pick matches Design; Open questions vs Assumptions), nothing leaking into sibling items.

Step 4 — One approval

Ask once: Approve and hand off to /rota-plan, Approve, no plan yet, or Revise (apply redlines, show again). Silence is not approval.

Step 5 — Store

rota design add <ID> --title "<title>"                     # mint the stub
rota design put <ID> --body-file <scratch-file>            # issue backend: publish the approved draft

On the issue backend the design is a note on the item's issue (references/issue-mode.md). On the file backend it is .rota/designs/<ID>.md: Edit the stub's sections, keep the frontmatter. Post each answer that changed direction with rota item comment add <ID> --kind decision --body-file -.

Report two lines (artifact, approaches, open questions). On hand-off say "Run /rota-plan <ID> next."

Key principles

• No noise. Don't narrate loads or recap.
• One draft, one approval. The artifact is the outcome, not a transcript.
• Single item. Project scope is /rota-vision; slice scope is /rota-plan.

References

• `references/design-exploration.md`: shared spine with /rota-vision.
• `references/grilling.md`: the grilling pass (Major / P0 / --grill).
• `references/subagent-dispatch.md`: the light subagent that grilling sends for broad reads.
• `references/context-load-protocol.md`: shared parallel context load.
• `references/knowledge-consult.md`: the K+D query pattern the load uses.