Convex Cost

1. GUARD: deploy-guard — a cost read is read-only over dev/prod (insights is cloud+user-auth only; not previews). Announce the deployment.
2. GATHER the spend evidence via the official MCP: insights for the bytes-read / documents-read events (the direct cost signal — Convex bills on function calls + bandwidth), tables for row counts (a table’s size bounds its scan cost), functionSpec for the surface. If there’s no usage/traffic yet, say so and estimate from the query SHAPES instead (a .collect() on a table projected to grow is a future cost even with zero traffic today).
3. ATTRIBUTE: rank functions by bytes/documents read per call × observed (or asked-about) call volume — the product is the cost driver, not either alone. A cheap-per-call function called constantly can outweigh an expensive rare one; show both factors.
4. PROJECT: state how the top drivers scale — a full-table .collect() grows LINEARLY with the table (cost compounds as data accumulates); an indexed .take(n) stays flat. Give the user the shape of the curve (‘this is O(table size) per call — fine at 1k rows, a bill at 1M’), not a false-precision dollar figure.
5. NAME THE CHEAPEST FIX per driver — index + .withIndex instead of scan, .paginate/.take instead of .collect, an aggregate component for counts, caching a hot read — and emit it as a cost-class finding on the bus (evidence: the insight event + the projected growth) pointing at convex-expert/convex-advisor for the actual change.
6. CONFIRM-COST for paid actions: if the flow includes anything metered (a domain purchase, cloud provisioning, a plan change), STATE the price and recurrence explicitly and get an explicit yes BEFORE proceeding — never let a paid action happen as a side effect (the cost-confirm gate).
7. REPORT: the current cost drivers ranked, each with its evidence + growth shape + fix, and a plain bottom line (‘your spend is dominated by messages:list reading the whole table every call; index it and it drops ~100x’). Honest precision: Convex pricing changes and depends on plan — give relative/shape guidance and cite the pricing page for absolute numbers rather than inventing a dollar total.

Rules

• Cost = data-read-per-call × call-volume — always show both factors; a cheap function called constantly can cost more than an expensive rare one.
• Read the deployment’s own insights/bytes-read evidence for spend; with no traffic yet, price the query SHAPES (a scan on a growing table is a future cost).
• Give the growth CURVE, not false-precision dollars: O(table) scans compound as data accumulates; indexed access stays flat. Cite the pricing page for absolute figures.
• Every cost driver names its cheapest fix and emits a cost-class finding on the bus pointing at the fixer (convex-expert/advisor).
• Confirm-cost for any metered/paid action: state the price + recurrence and get an explicit yes BEFORE it happens — never as a side effect.
• Read-only over dev/prod (deploy-guard); insights is cloud+user-auth only. Cost composes convex-advisor’s evidence but frames it as money, not latency.