/solution [session-dir]

Produce the "model answer" after the user has attempted the scenario. Refuse (politely) if there is no feedback.md yet — seeing the answer first defeats the practice; offer /mock or /critique instead. Override with --force.

Steps

1. Read prompt.md, interviewer.md, feedback.md, rubric/rubric.md, and the company profile.
2. Design the reference architecture yourself: complete, justified, labelled, with numbers — 12–20 nodes. This is the full map for study, not what anyone draws in 45 minutes. Then also produce <session>/solution-45min-spec.json → solution-45min.excalidraw: what a strong staff candidate actually draws in the room — 10–12 boxes, the crux decisions, and a notes block with the 4–5 numbers they'd say out loud.
3. Write <session>/solution-spec.json in the spec format documented at the top of scripts/excalidraw.mjs:
   • layer = column left→right (clients → edge → services → storage/queues → async consumers → analytics). Pin row to build clean bands (e.g. row 0 control plane, row 1 real-time flow, row 2 its stores, row 3 customer/query). Keep most edges between adjacent columns; for the few long or backward ones set via: "top"|"bottom" so they take a lane around the diagram. Don't draw every "writes to DB" arrow — annotate the store's label instead and keep the diagram about flow.
   • kind per node so colours carry meaning; use note on storage nodes for partition key / replication / TTL.
   • Edge labels: verb + protocol or payload; style: "dashed" for async.
   • groups for service boundaries or regions when they clarify.
   • notes block: key numbers, consistency choices, and the crux in one line each (≤ 6 bullets).
4. Run bun scripts/excalidraw.mjs <session>/solution-spec.json <session>/solution.excalidraw, then bun scripts/read-excalidraw.mjs <session>/solution.excalidraw and confirm the summary matches your intent (no dangling arrows, every node labelled, groups contain only their members). Fix the spec and regenerate if not — the workbench Solution view refreshes on its own.
5. Write <session>/solution.md:

# Reference solution — <title>
## What a strong 45-minute answer actually looks like — point at solution-45min.excalidraw; the ~60 s of spoken numbers; the crux in one sentence each; which one deep dive; the trade-offs named when asked. State plainly that the full map below is for study, not the bar for the room: Staff vs Senior is justification depth and finding the crux unled, not coverage.
## Requirements as I'd state them (functional / non-functional / out of scope)
## Estimates (3–6 lines, show the arithmetic)
## API & data model (endpoints/events; entities with keys and why)
## Architecture walkthrough — read path, write path, async path; each component's job in one line
## Deep dive: <the crux component> — algorithm/protocol/data structure level
## Deep dive: <second most likely probe>
## Trade-offs table — decision | chosen | alternative | why here
## Failure modes & mitigations (5–8)
## Scaling to 10× — what changes first
## Diff vs. your attempt — 3–6 bullets from feedback.md: what you had, what's different, why it matters
## Talk track — 60-second summary you could say out loud

6. Print: how to view it — workbench http://localhost:7788/sessions/<dir> → Solution toggle (read-only, side by side with their canvas via the toggle), or bun scripts/to-clipboard.mjs <session>/solution.excalidraw → Cmd+V on excalidraw.com — and the Diff vs. your attempt section. Suggest they redraw from memory tomorrow and /critique again.