Azure Compute

Plan Card

The Plan Card is the single source of truth for the create-flow. It renders every decision — explicit user answers and silent defaults — as a markdown table the user can read top-to-bottom and either approve or edit.

Rendering rules

1. Cost + quota on the top half. The user must see both without scrolling.

2. Source column is mandatory. Each row says where the value came from: user, default, inferred, or the MCP tool that produced it.

3. Flag risky defaults with ⚠. Open NSG to *, public IP exposed, no managed identity, password auth on Windows — all get a marker so the user can edit.

4. No invisible state. If you defaulted it, it goes in the card.

5. Re-emit on every change. When the user edits a row, render the full card again — diffs in chat are easy to miss.

Schema

Example — Linux dev VM in eastus

| Setting | Value | Source |
|---|---|---|
| Hosting model | Single VM | user |
| Name | `dev-vm-01` | default (`<purpose>-vm-<nn>`) |
| Region | `eastus` | user |
| Resource group | `dev-vm-01-rg` (new) | default |
| Image | `Ubuntu2404` | user |
| Size | `Standard_D2s_v5` (2 vCPU / 8 GB) | default |
| Auth | SSH key from `~/.ssh/id_rsa.pub` | inferred |
| VNet | new `dev-vm-01-vnet` (`10.0.0.0/16`) | default |
| Subnet | `default` (`10.0.0.0/24`) | default |
| NSG | SSH 22 from your public IP (`203.0.113.42`) | default — ⚠ change to `*` only if needed |
| Public IP | Standard, dynamic | default |
| OS disk | 30 GB Premium SSD | default |
| Boot diagnostics | Managed | default |
| Estimated cost | ~$0.096/hr (~$70/mo) | from `compute_vm_list-skus` |
| Quota | ✅ 4/100 vCPUs used in `standardDSv5Family` | from `compute_vm_check-quota` |

After rendering — single batched action picker

Render the Plan Card markdown inline in the chat first (so the user can read it), then ask one AskUserQuestion that combines approval + output format + delivery:

"Looks good? Pick how you want it delivered:"

1. Save Bicep to ./infra/{vm-name}/ (Recommended for repos)

2. Print az CLI in chat (Quick copy-paste)

3. Save Terraform to ./infra/{vm-name}/

4. Open GitHub PR with Bicep

5. Apply live via Azure MCP (actually creates resources)

6. Edit a row first (then re-render and re-ask)

This collapses what used to be 3 sequential popups (approve → output format → delivery) into 1.

Implementation: a single AskUserQuestion tool call with header: "Deliver", multiSelect: false, and 6 options (the most likely combinations above) — the user can also pick "Other" to type a custom answer like "give me both bicep and terraform".

If the user picks "Edit a row first": then ask which row, update, re-render the full Plan Card, and re-ask the same batched action picker. Do not splinter into multiple popups.

If the user already implied the answer in their original prompt ("save bicep to ./infra" / "open a PR" / "just print az CLI" / "apply it"): skip this prompt entirely and proceed straight to delivery.

Explicit-override fast path — skip the Plan Card table too. If the user combines an explicit deliverable ("give me the Bicep", "just print az CLI") with an explicit refusal of dialog ("no questions", "skip planning", "no plan", "just do it"): do not render the Plan Card markdown table. Instead emit a single-line preview of the high-signal decisions — e.g. "→ Deploying Standard_D2s_v5 in eastus, NSG = your public IP only on 22, est. ~$70/mo" — and follow it immediately with the requested artifact. End with a one-liner noting the full Plan Card is available on request if they want to edit rows. Step 4 validation gates still run.