The example bundle, browsed

This table is real compiled output, not prose someone wrote by hand. Anvil’s compiler ran over examples/payments — its openapi.yaml plus the anvil.yaml manifest that fills in what the spec left out — via node scripts/gen-docs-data.mjs. A test, packages/generators/src/docs-data.test.ts, recompiles the example on every run and fails if this page ever stops matching what the compiler produces. If you can read it here, the toolchain agrees with it.

Each row is one operation. The CLI command, the MCP tool, and the skill that describes them are all generated from the same entry in the shared model — which is why they can’t disagree about what the operation does.

Operation	CLI command	MCP tool	Effect	Risk	Safety posture	State
payments.customers.get Get a customer	payments customers get	payments_get_customer	read	read	retry-safe	approved
payments.payments.get Get a payment	payments payments get	payments_get_payment	read	read	retry-safe	approved
payments.refunds.create Create a refund	payments refunds create	payments_create_refund	mutation	financial· irreversible	confirm requiredidempotency keyretry-safe	approved
payments.capture.create Capture a payment	payments capture create	payments_capture_payment	mutation	financial· irreversible	confirm requiredretry-safe	approved

Reading the safety columns

• Effect / risk — the classifier’s verdict on what the operation does. Reads are low-stakes; the two financial mutations can’t be undone, which is why they’re gated more tightly.
• Confirm required — the tools refuse the call until an explicit --confirm (CLI) or confirm: true (MCP). Human approval is the stricter tier: the model can’t confirm its own way past it — a person has to.
• Idempotency key — the operation demands --idempotency-key, so a repeated call can’t double-refund a customer.
• Retry-safe — whether the runtime may auto-retry a transient failure. An operation that isn’t safe to repeat is never auto-retried.

Approved vs. review required

The state column is the approval gate. Only approved operations reach the generated CLI, MCP server, and skill. A review_required operation still exists in the model but is left out of every generated tool until a human approves it — usually after a manifest fills in the missing facts (that it’s safe to repeat, say, or how risky it is).

Every operation above is already approved, because the example’s anvil.yaml manifest did that work up front. To see how an operation moves from review_required to approved, read the enrich → approve workflow.