aws-sst-development

Imported from zxkane/aws-skills (plugins/aws-iac/skills/aws-sst-development/SKILL.md). Install upstream with npx skills add zxkane/aws-skills --skill aws-sst-development. Copyright stays with the author.

SST v4 for AWS

SST v4 (the "Ion" engine) is a Pulumi-backed IaC framework: you describe AWS resources in TypeScript and SST/Pulumi reconciles them into your account. It gives you high-level sst.aws.* components (Function, Bucket, Dynamo, Cron, Service, …) that expand into many underlying resources, plus an escape hatch to any raw Pulumi aws.* resource for the long tail. This skill encodes a production-proven way to author, link, test, deploy, and troubleshoot SST stacks on AWS — distilled from real multi-stack projects that have paid for each lesson with a prod incident.

SST and Pulumi are third-party — verify current syntax with Context7 (resolve-library-id → query-docs for sst or pulumi-aws) when you're unsure about a component's options. Verify AWS-side facts (service limits, model IDs, IAM action names, region availability) with the AWS docs MCP, never from memory. The patterns here are the how; the docs are the what.

When you're invoked

Figure out which mode you're in and jump to the right reference:

Always read the relevant reference before editing — they carry the why behind each rule, which matters more than the rule itself.

Orientation: read the repo before you touch it

SST projects are conventional but not identical. Before editing, build a quick map so your change matches the house style instead of fighting it:

1. sst.config.ts — the app name, home, providers/region, defaultTags, any global $transform (Node runtime pin, bundle fixups), and the order in which run() imports infra/ modules. The import order is the dependency order; respect it.
2. infra/ — one file per domain (storage, functions, api, observability…). This is where resources are declared. Check for an infra/CLAUDE.md — these projects keep IaC-specific rules there, and it's the single most valuable file to read first.
3. infra/tests/ — source-level Vitest assertions that pin resource invariants. If they exist, your change must keep them green and probably needs a new assertion.
4. package.json / .nvmrc — package manager (npm vs pnpm), Node version, and the sst/pulumi versions actually installed.

Run npx sst version to confirm you're on v4/Ion (the $config + .sst/platform/ signature). v2/v3 ("SST Classic", CDK-based) is a different framework — these patterns don't apply there.

The conventions, and which are universal vs tunable

The projects this skill is built from share a deliberate house style. Some of it is universal (true for any SST v4 + AWS project — apply it everywhere); some is project-specific (a sensible default these projects chose — adopt it for consistency, but recognize a project may differ).

Universal — these principles hold for any SST v4 + AWS project:

• Control the Node runtime deliberately, in one place. Don't leave it to whatever the installed SST happens to default to. The idiom is a single global $transform(sst.aws.Function, (args) => { args.runtime ??= "nodejs24.x" }) in run() — ??= is correct here (the transform runs before the component applies its own default, so it fills in only when the user didn't set one). Recent SST already defaults to a current Node runtime, so check the installed default first (Context7); the transform is then version-independence insurance so a future SST downgrade can't silently move your fleet. See references/authoring.md.
• Never interpolate a Pulumi Output<T> into a plain JS template literal. Use $interpolate (or pulumi.interpolate). A bare top-level `${bucket.arn}/*` stringifies the Output to a [Output<T>] placeholder and produces a broken ARN that only fails at deploy time (it type-checks and sst dev runs fine). The fix is $interpolate​`${bucket.arn}/*`. This has caused prod deploy outages. See references/authoring.md § Outputs.
• Migrating a resource between Pulumi types should default to two PRs — Pulumi creates-before-destroys, so for a uniqueness-constrained AWS name (bucket, IAM role, gateway) the old resource still owns it and the create fails with ConflictException. Two sequential deploys (teardown, then recreate) is the conservative default; aliases: / pulumi import / state surgery can bridge identity in some cases but only with a reviewed plan. See references/deploy-and-troubleshoot.md § Migrations.
• Prefer typed sst.aws.* / aws.* resources over the aws.cloudcontrol.Resource escape hatch. CloudControl outputs are stringly-typed and oneOf fields don't patch cleanly. Use it only when no typed resource exists yet, and migrate off it when one ships.

Project-specific defaults — adopt for consistency, but confirm per repo:

• Region ap-northeast-1, home: "aws", and defaultTags carrying Project / Stage / ManagedBy: "sst".
• Stage-gated lifecycle: removal: stage === "prod" ? "retain" : "remove" and protect: stage === "prod" so prod resources survive a stack tear-down and non-prod previews clean up.
• SSM Parameter Store as the out-of-graph contract under a /{app}/{stage}/{domain}/... prefix — for consumers that aren't in the Pulumi graph (CI scripts, sibling apps, operators). For same-app Lambdas, prefer SST link: (it wires a real dependency edge and grants IAM); don't route same-app sharing through SSM. See references/authoring.md § Sharing.
• Lazy await import("./infra/<module>") inside run() so sst dev hot-reload stays light. (For testing, a module export still runs its top-level new sst.aws.* unless it's wrapped in a factory function — see references/testing.md for how to test infra.)
• Source-level Vitest tests on every infra module — a lightweight, house-style regression net asserting on the source text (resource names, index shapes, IAM scopes). It's a deliberate choice, not an SST limit: Pulumi does support runtime mocks (@pulumi/pulumi/runtime) for behavioral graph tests when a module has real logic. Source assertions don't replace a preview-deploy + smoke test. See references/testing.md.
• An observability gate: every new Lambda/queue/schedule gets an alarm and structured logging before merge. Whether you enforce this depends on the project, but it's cheap insurance. See references/deploy-and-troubleshoot.md § Observability.

When you introduce a convention, say which bucket it's in ("this is universal" vs "matching this repo's house style") so the user can override the project-specific ones deliberately.

Working rhythm

1. Orient (above) — map config, modules, tests, tooling.
2. Verify syntax with Context7 / AWS docs MCP if anything is non-obvious. Don't guess at a component's option name.
3. Author the resource/module following references/authoring.md. Match the surrounding file's commenting density and naming — these projects comment the why heavily, and a terse one-liner in a heavily-annotated file reads as a regression.
4. Test — add or update source-level assertions (references/testing.md) and run npx vitest (or the repo's test script). Run npx sst diff and/or tsc --noEmit to catch type and plan errors before deploying.
5. Deploy/operate per references/deploy-and-troubleshoot.md. Confirm the target account with aws sts get-caller-identity before any sst deploy.
6. Clean up any exported state files — they contain account IDs and ARNs and must not linger in /tmp or chat history.

What good looks like

• The change is the smallest diff that satisfies the requirement, in the right infra/ module, wired into run() in dependency order.
• Every Lambda gets the right runtime via the global transform (you didn't hand-set runtime unless intentionally diverging — e.g. a Python function).
• Cross-resource references use link: (in-graph) and/or $interpolate-scoped IAM; outputs other tools consume are published to SSM under the stage prefix.
• New infra has a matching source-level test, and the existing suite stays green.
• You confirmed AWS-side facts via the docs MCP and SST/Pulumi syntax via Context7 rather than relying on recall.
• Anything irreversible (deploy, sst remove, a resource-type migration) was flagged to the user with the account it targets, and migrations were planned as two PRs, not one.