Repository navigation
C1. Autoscale & dynamic throughput management #164
Description
Why
Throughput is the primary Cosmos DB cost/perf lever, yet the Shell can't view or change it. Autoscale (dynamic scaling) and RU/s changes are exactly the kind of routine, RU-cost-sensitive operation agents and CI should manage safely — and it pairs with the RU guardrails (item G2).
Proposed behavior
throughput [--db --con]shows mode (manual/autoscale), current RU/s and autoscale max.throughput set --ru N/--autoscale-max N/--mode manual|autoscalescales or switches modes at database and container scope, via ARM.- Surface RU/s and estimated cost delta before applying.
Acceptance criteria
- View + set RU/s and autoscale max.
- Switch manual↔autoscale.
- Database and container scope.
- Structured output with RU/s + cost delta.
- Safe over MCP behind confirmation (M3) / dry-run (G1).
- Docs + tests.
Filed from the Agentic & Automation Roadmap (docs/agentic-roadmap.md), item C1, Wave 2. Priority P1.
Activity
mkrueger commentedon Jul 7, 2026
Cross-checked the MCP-safety criterion against how this shell actually does MCP safety.
How safety works here: it's annotation-driven client-side approval, not server-side prompts.
McpAnnotationAttributeexposesReadOnly/Destructive/Idempotent/OpenWorld, mapped to MCP hints inToolOperations.cs, andServerInstructions.md+docs/mcp.mdinstruct clients to require manual approval forDestructivetools (rm,rmdb,rmcon,delete). Server-side confirmation isn't possible over MCP — the interactive prompt inThroughputCommand.ConfirmWriteis deliberately skipped whenMcpPortis set, because there's no console to prompt on.Done ✅
throughputis exposed as an MCP tool withReadOnly = false.Gaps⚠️
ReadOnly = falsebut not annotatedDestructiveand carries no idempotency hint, so clients treat it like an ordinary mutating tool — not one that needs the "stop and ask" treatment deletes get. A throughput change directly moves the bill (e.g. autoscale 100k RU/s), so it should surface for approval.docs/mcp.mdandServerInstructions.mdhave no throughput coverage — clients aren't told to confirm RU/cost changes.Recommendation: reframe this criterion as an annotation + docs fix rather than server-side confirmation (M3) / dry-run (G1), which don't fit this architecture. Concretely: annotate the write path conservatively (billing-impact / destructive-style hint — noting
showis read-only whileset/manual/autoscaleare writes but share one tool), and add throughput guidance todocs/mcp.md+ServerInstructions.md. That achieves "safe over MCP" consistently with the rest of the shell.Add --dry-run preview to throughput write subcommands (#166)
mkrueger commentedon Jul 9, 2026
#166 merged.
Add --dry-run preview to throughput write subcommands (Azure#164)