chore: import upstream snapshot with attribution

packages/gcp-cloud-run/Dockerfile

@@ -0,0 +1,118 @@
+# HyperFrames distributed render — Cloud Run image.
+#
+# One container image, three roles (plan / renderChunk / assemble). Cloud
+# Workflows POSTs a JSON body with an `Action` field; `dist/server.js`
+# dispatches to the matching `@hyperframes/producer/distributed` primitive.
+#
+# Build context is the REPOSITORY ROOT (not this package dir) because the
+# package depends on the `@hyperframes/producer` workspace and renders with
+# the same chrome-headless-shell + fonts + ffmpeg the production renderer
+# uses. The example smoke script and `hyperframes cloudrun deploy` both
+# build with the repo root as context:
+#
+#   docker build -f packages/gcp-cloud-run/Dockerfile -t <image> .
+#
+# NOTE: this Dockerfile only builds from a full hyperframes monorepo checkout
+# — it COPYs sibling workspace packages (core/engine/producer). It is NOT
+# buildable from the published npm tarball alone; install the repo (or use
+# `hyperframes cloudrun deploy`, which builds from your checkout via Cloud
+# Build) to produce the image.
+#
+# Unlike the AWS Lambda adapter there is no ZIP-size ceiling and no runtime
+# Chrome decompression step — the binary lives in the image at a fixed path.
+
+FROM node:22-bookworm-slim
+
+# ── System dependencies (identical set to the producer's render image) ───────
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    ca-certificates \
+    curl \
+    unzip \
+    ffmpeg \
+    libgbm1 \
+    libnss3 \
+    libatk-bridge2.0-0 \
+    libdrm2 \
+    libxcomposite1 \
+    libxdamage1 \
+    libxrandr2 \
+    libcups2 \
+    libasound2 \
+    libpangocairo-1.0-0 \
+    libxshmfence1 \
+    libgtk-3-0 \
+    fonts-liberation \
+    fonts-noto-color-emoji \
+    fonts-noto-cjk \
+    fonts-noto-core \
+    fonts-noto-extra \
+    fonts-noto-ui-core \
+    fonts-freefont-ttf \
+    fonts-dejavu-core \
+    fontconfig \
+    && rm -rf /var/lib/apt/lists/* \
+    && apt-get clean \
+    && fc-cache -fv
+
+# ── chrome-headless-shell (deterministic BeginFrame capture) ─────────────────
+# Pinned to the SAME build the producer regression baselines were generated
+# against so distributed renders are pixel-comparable. Bump together with the
+# producer's Dockerfile.test pin and a baseline regen.
+RUN npx --yes @puppeteer/browsers install chrome-headless-shell@148.0.7778.167 \
+      --path /opt/puppeteer \
+    && CHS="$(find /opt/puppeteer/chrome-headless-shell -name chrome-headless-shell -type f | head -n1)" \
+    && mkdir -p /opt/chrome \
+    && ln -s "$CHS" /opt/chrome/chrome-headless-shell \
+    && /opt/chrome/chrome-headless-shell --version
+
+# The chromium.ts resolver reads this first; pointing the engine straight at
+# the symlinked binary avoids the runtime cache scan.
+ENV HYPERFRAMES_CHROME_PATH=/opt/chrome/chrome-headless-shell
+ENV PRODUCER_HEADLESS_SHELL_PATH=/opt/chrome/chrome-headless-shell
+ENV PUPPETEER_SKIP_CHROMIUM_DOWNLOAD=true
+ENV CONTAINER=true
+
+WORKDIR /app
+
+# Install bun (the repo's package manager / build driver). Pin the version:
+# the container runs `bun dist/server.js` and relies on bun's ESM `require`
+# interop to load the producer bundle, so a future bun release changing that
+# behaviour shouldn't silently break the next image rebuild. Bump deliberately.
+RUN curl -fsSL https://bun.sh/install | bash -s "bun-v1.3.9"
+ENV PATH="/root/.bun/bin:$PATH"
+
+# Install workspace dependencies. Copy manifests first for layer caching.
+COPY package.json bun.lock ./
+COPY packages/core/package.json packages/core/package.json
+COPY packages/engine/package.json packages/engine/package.json
+COPY packages/player/package.json packages/player/package.json
+COPY packages/producer/package.json packages/producer/package.json
+COPY packages/cli/package.json packages/cli/package.json
+COPY packages/studio/package.json packages/studio/package.json
+COPY packages/shader-transitions/package.json packages/shader-transitions/package.json
+COPY packages/aws-lambda/package.json packages/aws-lambda/package.json
+COPY packages/gcp-cloud-run/package.json packages/gcp-cloud-run/package.json
+RUN bun install --frozen-lockfile
+
+# Copy source for the packages the render path needs.
+COPY packages/core/ packages/core/
+COPY packages/engine/ packages/engine/
+COPY packages/producer/ packages/producer/
+COPY packages/gcp-cloud-run/ packages/gcp-cloud-run/
+
+# Build core runtime artifacts (needed by the renderer) + producer, then the
+# adapter. Generate embedded font data so glyph layout matches production.
+RUN bun run --filter @hyperframes/core build:hyperframes-runtime:modular \
+    && (cd packages/producer && bunx tsx scripts/generate-font-data.ts) \
+    && bun run --cwd packages/producer build \
+    && bun run --cwd packages/gcp-cloud-run build
+
+# Cloud Run injects PORT (default 8080). The server reads it.
+ENV PORT=8080
+EXPOSE 8080
+
+WORKDIR /app/packages/gcp-cloud-run
+# Run under bun, not node. The repo is bun-native and `@hyperframes/producer`'s
+# bundled `distributed.js` relies on a `require` being available at runtime
+# (its esbuild interop shim); bun provides one in ESM, bare `node` does not.
+CMD ["bun", "dist/server.js"]

packages/gcp-cloud-run/README.md

@@ -0,0 +1,128 @@
+# @hyperframes/gcp-cloud-run
+
+Google Cloud Run + Cloud Workflows adapter for HyperFrames distributed
+rendering. The OSS render primitives (`plan` → `renderChunk` × N →
+`assemble`) are pure functions over local file paths; this package is the
+deployment, orchestration, and storage glue that runs them on Google Cloud —
+the GCP counterpart to [`@hyperframes/aws-lambda`](../aws-lambda).
+
+Two surfaces, one package:
+
+- **Server-side handler** (`./server`) — a Cloud Run HTTP service that
+  dispatches `plan` / `renderChunk` / `assemble` on the request body's
+  `Action` field, bridging GCS ↔ the container's filesystem around each OSS
+  primitive. This is what the bundled `Dockerfile` runs.
+- **Client-side SDK** (`./sdk`) — `renderToCloudRun`, `getRenderProgress`,
+  `deploySite`, `validateDistributedRenderConfig`, and `computeRenderCost`.
+  Call these from a Node process (CI, CLI, app backend) to drive a deployed
+  stack without writing GCS / Workflows boilerplate.
+
+The package is **not** a dependency of `@hyperframes/producer`; install it
+separately.
+
+## Architecture
+
+```
+GCS bucket  ←→  Cloud Run service (plan / renderChunk / assemble)
+                     ▲
+                     │ OIDC-authenticated http.post, one per step
+                     │
+                Cloud Workflows  (Plan → parallel RenderChunk → Assemble)
+```
+
+- **Plan** downloads the project tarball, runs `plan()`, uploads the planDir
+  tarball (+ audio) to GCS, and returns the chunk count.
+- **RenderChunk** runs in a parallel `for` loop in the workflow, fanned out
+  up to the plan's chunk count. Each invocation renders one chunk and uploads
+  it.
+- **Assemble** downloads every chunk + audio, stitches the final
+  deliverable, and uploads it.
+
+Every step is a `POST` to the same Cloud Run URL with a different `Action`.
+The workflow accumulates each step's small result body and returns
+`{ Plan, Chunks, Assemble }` so `getRenderProgress` can read frame totals and
+per-step durations on success.
+
+## Chrome runtime
+
+Unlike the Lambda adapter — which fights a 250 MB ZIP ceiling and
+decompresses `@sparticuz/chromium` into `/tmp` at runtime — Cloud Run runs a
+container image. The `Dockerfile` installs the same pinned
+`chrome-headless-shell` build and font set the production renderer uses, at a
+fixed path, and exports `HYPERFRAMES_CHROME_PATH`. CDP-level `BeginFrame`
+works because the command lives in the protocol, not the binary. There is no
+runtime decompression step and no packaging ceiling.
+
+## Deploying
+
+The `terraform/` module provisions everything: the GCS render bucket, the
+Cloud Run service, the Cloud Workflows definition, two least-privilege
+service accounts (the service reads/writes the bucket; the workflow invokes
+the service), and a runaway-request alert.
+
+```bash
+# 1. Build + push the image (Cloud Build or local docker).
+gcloud builds submit . \
+  --tag REGION-docker.pkg.dev/PROJECT/REPO/hyperframes-render:TAG
+
+# 2. Apply the module.
+terraform -chdir=node_modules/@hyperframes/gcp-cloud-run/terraform init
+terraform -chdir=node_modules/@hyperframes/gcp-cloud-run/terraform apply \
+  -var project_id=PROJECT \
+  -var region=us-central1 \
+  -var image=REGION-docker.pkg.dev/PROJECT/REPO/hyperframes-render:TAG
+```
+
+Terraform outputs `render_bucket_name`, `service_url`, `workflow_name`, and
+`region` — pass them straight into the SDK.
+
+## Using the SDK
+
+```ts
+import { renderToCloudRun, getRenderProgress } from "@hyperframes/gcp-cloud-run/sdk";
+
+const handle = await renderToCloudRun({
+  projectDir: "./my-composition",
+  config: { fps: 30, width: 1920, height: 1080, format: "mp4" },
+  bucketName: "hyperframes-render-my-project", // from terraform output
+  projectId: "my-project",
+  location: "us-central1",
+  workflowId: "hyperframes-render",
+  serviceUrl: "https://hyperframes-render-abc.us-central1.run.app",
+});
+
+// Poll until done.
+let progress = await getRenderProgress({ executionName: handle.executionName });
+while (progress.status === "running") {
+  await new Promise((r) => setTimeout(r, 5000));
+  progress = await getRenderProgress({ executionName: handle.executionName });
+}
+console.log(progress.status, progress.outputFile, progress.costs.displayCost);
+```
+
+`deploySite` is called implicitly when you pass `projectDir`; call it
+yourself to pre-upload once and reuse the `siteHandle` across many renders
+(e.g. personalised template batches).
+
+## Running tests
+
+```bash
+bun test          # unit tests over an in-memory GCS double — no network
+bun run typecheck
+```
+
+The live end-to-end smoke (build image → terraform apply → render a fixture
+through the workflow → PSNR-compare → destroy) lives at
+`examples/gcp-cloud-run/scripts/smoke.sh` and needs a GCP project with
+billing enabled.
+
+## What's still ahead
+
+- **Mid-flight per-chunk progress.** `getRenderProgress` reports coarse
+  `running` progress and exact numbers on success. Reading the Cloud
+  Workflows step-entries API would give per-chunk progress while the render
+  is in flight; tracked as a follow-up.
+- **Cloud Run Jobs / Firebase Functions variants.** This first version
+  targets Cloud Run services + Workflows (the closest analog to Lambda +
+  Step Functions). The same handler runs unchanged under Cloud Run Jobs;
+  only the orchestration trigger differs.

packages/gcp-cloud-run/build.mjs

@@ -0,0 +1,58 @@
+#!/usr/bin/env node
+/**
+ * Build script for @hyperframes/gcp-cloud-run (public OSS package).
+ *
+ * Bundles each subpath barrel via esbuild → dist/, then emits .d.ts via tsc.
+ *
+ * Subpaths (each gets its own dist entry so adopters that import one path
+ * don't load the others' transitive graphs at module-load time):
+ *
+ *   .         (the umbrella barrel: server + sdk types re-exported)
+ *   ./server  (the Cloud Run runtime entry — what the Dockerfile runs)
+ *   ./sdk     (client-side helpers — GCS + Workflows clients only, no
+ *             chromium/puppeteer)
+ *
+ * All production deps are kept external so consumers (and the container
+ * image) resolve them via their own node_modules.
+ */
+
+import { build } from "esbuild";
+import { execSync } from "node:child_process";
+import { mkdirSync, rmSync } from "node:fs";
+
+rmSync("dist", { recursive: true, force: true });
+mkdirSync("dist", { recursive: true });
+
+const sharedOpts = {
+  bundle: true,
+  platform: "node",
+  target: "node22",
+  format: "esm",
+  minify: false,
+  sourcemap: true,
+  external: [
+    "@google-cloud/storage",
+    "@google-cloud/workflows",
+    "@hono/node-server",
+    "@hyperframes/producer",
+    "@hyperframes/producer/distributed",
+    "hono",
+    "puppeteer-core",
+    "tar",
+  ],
+};
+
+await Promise.all([
+  build({ ...sharedOpts, entryPoints: ["src/index.ts"], outfile: "dist/index.js" }),
+  build({ ...sharedOpts, entryPoints: ["src/server.ts"], outfile: "dist/server.js" }),
+  build({ ...sharedOpts, entryPoints: ["src/sdk/index.ts"], outfile: "dist/sdk/index.js" }),
+]);
+
+// esbuild doesn't emit .d.ts. tsc does, with a build-only tsconfig that
+// drops the workspace `paths` overrides so `@hyperframes/producer` resolves
+// through node_modules to the sibling package's already-built `dist/`
+// types instead of pulling its full source tree into emit (which would
+// violate rootDir).
+execSync("tsc -p tsconfig.build.json --emitDeclarationOnly", { stdio: "inherit" });
+
+console.log("[Build] Complete: dist/{index,server,sdk/index}.js + .d.ts");

packages/gcp-cloud-run/package.json

@@ -0,0 +1,62 @@
+{
+  "name": "@hyperframes/gcp-cloud-run",
+  "version": "0.7.55",
+  "description": "Google Cloud Run + Workflows adapter for HyperFrames distributed rendering — request handler, client-side SDK, and Terraform module.",
+  "repository": {
+    "type": "git",
+    "url": "https://github.com/heygen-com/hyperframes",
+    "directory": "packages/gcp-cloud-run"
+  },
+  "files": [
+    "dist/",
+    "terraform/",
+    "Dockerfile",
+    "README.md"
+  ],
+  "type": "module",
+  "main": "./dist/index.js",
+  "types": "./dist/index.d.ts",
+  "exports": {
+    ".": {
+      "import": "./dist/index.js",
+      "types": "./dist/index.d.ts"
+    },
+    "./server": {
+      "import": "./dist/server.js",
+      "types": "./dist/server.d.ts"
+    },
+    "./sdk": {
+      "import": "./dist/sdk/index.js",
+      "types": "./dist/sdk/index.d.ts"
+    }
+  },
+  "publishConfig": {
+    "access": "public",
+    "registry": "https://registry.npmjs.org/"
+  },
+  "scripts": {
+    "build": "node build.mjs",
+    "start": "bun dist/server.js",
+    "test": "bun test",
+    "typecheck": "tsc --noEmit"
+  },
+  "dependencies": {
+    "@google-cloud/storage": "^7.14.0",
+    "@google-cloud/workflows": "^4.2.0",
+    "@hono/node-server": "^1.13.0",
+    "@hyperframes/producer": "workspace:^",
+    "hono": "^4.6.0",
+    "puppeteer-core": "^25.2.1",
+    "tar": "^7.4.3"
+  },
+  "devDependencies": {
+    "@types/node": "^25.0.10",
+    "@types/tar": "^6.1.13",
+    "esbuild": "^0.25.12",
+    "tsx": "^4.21.0",
+    "typescript": "^5.7.2"
+  },
+  "engines": {
+    "node": ">=22"
+  }
+}

packages/gcp-cloud-run/src/__fixtures__/fakeGcs.ts

@@ -0,0 +1,116 @@
+/**
+ * In-memory `@google-cloud/storage` stand-in for the adapter's unit tests.
+ *
+ * Mimics just the surface the transport + deploySite use:
+ *   storage.bucket(name).file(key) → { createReadStream, exists, getMetadata }
+ *   storage.bucket(name).upload(localPath, { destination, contentType })
+ *
+ * Objects live in a Map keyed `bucket/key`. `upload` reads the real local
+ * file from disk (the handler writes real tarballs), so round-trips through
+ * tar/untar exercise the actual code path. Every op is recorded for
+ * sequence assertions.
+ */
+
+import { readFileSync, writeFileSync } from "node:fs";
+import { Readable } from "node:stream";
+import type { Storage } from "@google-cloud/storage";
+
+export interface FakeGcsOp {
+  kind: "download" | "upload" | "exists" | "getMetadata";
+  uri: string;
+  bytes?: number;
+}
+
+export class FakeGcs {
+  ops: FakeGcsOp[] = [];
+  objects = new Map<string, Buffer>();
+
+  // Accessed only through the `Storage` cast in tests, so fallow's static
+  // analysis can't see the reference.
+  // fallow-ignore-next-line unused-class-member
+  bucket(bucketName: string): FakeBucket {
+    return new FakeBucket(this, bucketName);
+  }
+
+  /** Seed an object directly (e.g. a pre-built project tarball). */
+  seed(uri: string, bytes: Buffer): void {
+    this.objects.set(uri, bytes);
+  }
+
+  /** Seed from a local file on disk. */
+  seedFromFile(uri: string, localPath: string): void {
+    this.objects.set(uri, readFileSync(localPath));
+  }
+}
+
+class FakeBucket {
+  constructor(
+    private readonly gcs: FakeGcs,
+    private readonly bucketName: string,
+  ) {}
+
+  get name(): string {
+    return this.bucketName;
+  }
+
+  file(key: string): FakeFile {
+    return new FakeFile(this.gcs, this.bucketName, key);
+  }
+
+  async upload(
+    localPath: string,
+    opts: { destination: string; contentType?: string },
+  ): Promise<unknown> {
+    const uri = `gs://${this.bucketName}/${opts.destination}`;
+    const bytes = readFileSync(localPath);
+    this.gcs.objects.set(uri, bytes);
+    this.gcs.ops.push({ kind: "upload", uri, bytes: bytes.length });
+    return [{}];
+  }
+}
+
+class FakeFile {
+  private readonly uri: string;
+  constructor(
+    private readonly gcs: FakeGcs,
+    bucketName: string,
+    key: string,
+  ) {
+    this.uri = `gs://${bucketName}/${key}`;
+  }
+
+  createReadStream(): Readable {
+    const bytes = this.gcs.objects.get(this.uri);
+    if (!bytes) {
+      const r = new Readable({ read() {} });
+      r.destroy(new Error(`FakeGcs: object not found: ${this.uri}`));
+      return r;
+    }
+    this.gcs.ops.push({ kind: "download", uri: this.uri, bytes: bytes.length });
+    return Readable.from([bytes]);
+  }
+
+  async exists(): Promise<[boolean]> {
+    const has = this.gcs.objects.has(this.uri);
+    this.gcs.ops.push({ kind: "exists", uri: this.uri });
+    return [has];
+  }
+
+  async getMetadata(): Promise<[{ size?: string | number; updated?: string }]> {
+    const bytes = this.gcs.objects.get(this.uri);
+    this.gcs.ops.push({ kind: "getMetadata", uri: this.uri });
+    return [{ size: bytes?.length ?? 0, updated: "2026-06-06T00:00:00.000Z" }];
+  }
+
+  /** Helper for tests that want to materialize an object to disk. */
+  writeToDisk(destPath: string): void {
+    const bytes = this.gcs.objects.get(this.uri);
+    if (!bytes) throw new Error(`FakeGcs: object not found: ${this.uri}`);
+    writeFileSync(destPath, bytes);
+  }
+}
+
+/** Cast so `deploySite({ storage: asStorage(fake) })` reads cleanly. */
+export function asStorage(fake: FakeGcs): Storage {
+  return fake as unknown as Storage;
+}

packages/gcp-cloud-run/src/chromium.ts

@@ -0,0 +1,95 @@
+/**
+ * Cloud Run Chrome resolver.
+ *
+ * `renderChunk()` (the only primitive that needs a browser) launches Chrome
+ * via the engine's `BrowserManager`. Because Cloud Run runs a container
+ * image rather than a size-capped ZIP, the Chrome story is far simpler than
+ * the Lambda adapter's: the `Dockerfile` installs `chrome-headless-shell`
+ * (the same BeginFrame-capable build the K8s deploy uses) into the image at
+ * a known path and exports `HYPERFRAMES_CHROME_PATH`. There is no runtime
+ * decompression-into-/tmp step and no 250 MB packaging ceiling to fight.
+ *
+ * Resolution order:
+ *   1. `PRODUCER_HEADLESS_SHELL_PATH` — the engine's own override. If a
+ *      caller (or the Docker image) already set it, honour it untouched.
+ *   2. `HYPERFRAMES_CHROME_PATH` — set by the Dockerfile to the installed
+ *      `chrome-headless-shell` binary.
+ *   3. A small list of conventional install paths, as a last resort for
+ *      images built outside our Dockerfile.
+ *
+ * Throws {@link ChromeBinaryUnavailableError} when nothing resolves, so a
+ * misconfigured image fails loudly at the first chunk rather than emitting
+ * a confusing puppeteer-core "executablePath must be specified" assertion.
+ */
+
+import { existsSync } from "node:fs";
+
+/**
+ * Thrown when the Chrome binary resolver can't produce a usable path. The
+ * class name is the workflow's non-retryable error discriminator.
+ */
+export class ChromeBinaryUnavailableError extends Error {
+  // Read indirectly via the error envelope / Error.prototype.toString.
+  // fallow-ignore-next-line unused-class-member
+  override readonly name = "ChromeBinaryUnavailableError";
+  readonly resolvedPath: string | null;
+  constructor(resolvedPath: string | null, hint: string) {
+    super(`[chromium] Chrome binary unavailable: ${hint}`);
+    this.resolvedPath = resolvedPath;
+  }
+}
+
+/**
+ * Conventional locations a `chrome-headless-shell` (or full Chrome) binary
+ * may live at in a Debian/Ubuntu-based container. Checked only after the
+ * two env-var overrides miss.
+ */
+const FALLBACK_CHROME_PATHS = [
+  "/opt/chrome/chrome-headless-shell",
+  "/usr/bin/chrome-headless-shell",
+  "/usr/bin/google-chrome",
+  "/usr/bin/google-chrome-stable",
+  "/usr/bin/chromium",
+  "/usr/bin/chromium-browser",
+];
+
+/**
+ * Resolve the absolute path to a Chrome binary suitable for BeginFrame.
+ * Pure (no env mutation) so callers decide whether to export the result
+ * into `PRODUCER_HEADLESS_SHELL_PATH`.
+ */
+// fallow-ignore-next-line complexity
+export function resolveChromeExecutablePath(): string {
+  const fromEngineOverride = process.env.PRODUCER_HEADLESS_SHELL_PATH?.trim();
+  if (fromEngineOverride) {
+    if (!existsSync(fromEngineOverride)) {
+      throw new ChromeBinaryUnavailableError(
+        fromEngineOverride,
+        `PRODUCER_HEADLESS_SHELL_PATH=${JSON.stringify(fromEngineOverride)} does not exist on disk.`,
+      );
+    }
+    return fromEngineOverride;
+  }
+
+  const fromImage = process.env.HYPERFRAMES_CHROME_PATH?.trim();
+  if (fromImage) {
+    if (!existsSync(fromImage)) {
+      throw new ChromeBinaryUnavailableError(
+        fromImage,
+        `HYPERFRAMES_CHROME_PATH=${JSON.stringify(fromImage)} does not exist on disk.`,
+      );
+    }
+    return fromImage;
+  }
+
+  for (const candidate of FALLBACK_CHROME_PATHS) {
+    if (existsSync(candidate)) return candidate;
+  }
+
+  throw new ChromeBinaryUnavailableError(
+    null,
+    "no Chrome binary found. Set HYPERFRAMES_CHROME_PATH (the Dockerfile does this) or " +
+      "PRODUCER_HEADLESS_SHELL_PATH to the absolute path of a chrome-headless-shell binary. " +
+      `Searched: ${FALLBACK_CHROME_PATHS.join(", ")}.`,
+  );
+}

packages/gcp-cloud-run/src/events.ts

@@ -0,0 +1,146 @@
+/**
+ * Request + result types for the HyperFrames distributed render handler
+ * running on Cloud Run.
+ *
+ * The Cloud Workflows definition in `packages/gcp-cloud-run/terraform/workflow.yaml`
+ * dispatches on the `Action` field of the JSON request body. Each action
+ * maps 1:1 onto one of the three OSS distributed primitives:
+ *
+ *   "plan"        → `plan(projectDir, config, planDir)`        (Activity A)
+ *   "renderChunk" → `renderChunk(planDir, chunkIndex, output)` (Activity B)
+ *   "assemble"    → `assemble(planDir, chunkPaths, audio, out)` (Activity C)
+ *
+ * All file I/O is mediated by GCS — the handler downloads inputs into a
+ * per-request workdir under the container's writable `/tmp`, invokes the
+ * primitive, uploads outputs back to GCS, and returns a small JSON payload
+ * that fits inside a Cloud Workflows step variable (Workflows caps a single
+ * step's memory; chunk results stay well under 1 KB so the orchestration
+ * can hold one per Map iteration).
+ *
+ * These shapes are intentionally identical to `@hyperframes/aws-lambda`'s
+ * `events.ts` apart from the URI scheme (`gs://` vs `s3://`): the wire
+ * contract is the adapter's, the primitives underneath are shared.
+ */
+
+import type {
+  DistributedFormat,
+  SerializableDistributedRenderConfig,
+} from "@hyperframes/producer/distributed";
+
+export type { SerializableDistributedRenderConfig } from "@hyperframes/producer/distributed";
+
+/** Discriminator for the three roles the one Cloud Run image fulfills. */
+export type CloudRunAction = "plan" | "renderChunk" | "assemble";
+
+/**
+ * Top-level shape of any request body the handler may receive.
+ *
+ * Cloud Workflows passes the step's `body` through verbatim, but a caller
+ * driving the service directly (or a Workflows definition that wraps the
+ * payload) may nest it under `Payload` / `Input`; the handler unwraps both
+ * before dispatching, matching the Lambda adapter's envelope tolerance.
+ */
+export type CloudRunEvent =
+  | PlanEvent
+  | RenderChunkEvent
+  | AssembleEvent
+  | { Payload: CloudRunEvent }
+  | { Input: CloudRunEvent };
+
+/** Activity A: produce a planDir, upload to GCS. */
+export interface PlanEvent {
+  Action: "plan";
+  /** GCS URI pointing at a `tar -czf`-archived project directory (`gs://bucket/key.tar.gz`). */
+  ProjectGcsUri: string;
+  /** GCS URI prefix where the planDir tar should be uploaded (`gs://bucket/{prefix}/`). */
+  PlanOutputGcsPrefix: string;
+  /** `DistributedRenderConfig` minus runtime-only fields (logger, abortSignal). */
+  Config: SerializableDistributedRenderConfig;
+}
+
+/** Activity B: fetch planDir, render one chunk, upload result. */
+export interface RenderChunkEvent {
+  Action: "renderChunk";
+  /** GCS URI of the plan tar produced by a PlanEvent invocation. */
+  PlanGcsUri: string;
+  /**
+   * `PlanResult.planHash` from the Plan invocation. The handler verifies
+   * this against the untarred planDir's `plan.json` before invoking the
+   * producer, throwing a typed `PLAN_HASH_MISMATCH` on divergence so the
+   * workflow routes it as non-retryable. Defense-in-depth — the producer
+   * also re-checks internally.
+   */
+  PlanHash: string;
+  /** 0-based chunk index this invocation should render. */
+  ChunkIndex: number;
+  /** GCS URI prefix where the chunk output should be uploaded (`gs://bucket/{prefix}/`). */
+  ChunkOutputGcsPrefix: string;
+  /** Output container format from the plan's encoder.json; drives file vs frame-dir handling. */
+  Format: DistributedFormat;
+}
+
+/** Activity C: fetch planDir + all chunks + audio, assemble, upload final. */
+export interface AssembleEvent {
+  Action: "assemble";
+  /** GCS URI of the plan tar produced by a PlanEvent invocation. */
+  PlanGcsUri: string;
+  /** GCS URIs of every chunk, ordered by chunk index. Length must equal `chunkCount`. */
+  ChunkGcsUris: string[];
+  /** GCS URI of the planDir's `audio.aac` if the composition has audio; `null` otherwise. */
+  AudioGcsUri: string | null;
+  /** Final output GCS URI (`gs://bucket/key.mp4`). */
+  OutputGcsUri: string;
+  /** Output container format; drives file vs frame-dir handling. */
+  Format: DistributedFormat;
+  /**
+   * Optional exact-CFR re-encode at assemble time. When `true`, the final
+   * assembled video is re-encoded with `-fps_mode cfr -r <fps>` so the
+   * stream's `avg_frame_rate` matches the container's `r_frame_rate`
+   * exactly (and the file's duration is exact, not PTS-derived). Trade-off
+   * is ~2-5x the assemble wall-clock. mp4 only — webm / mov stream-copy
+   * paths already produce exact avg_frame_rate. Default `false` /
+   * unset preserves current `-c copy` behavior.
+   */
+  Cfr?: boolean;
+}
+
+// ── Result types — kept small to fit Cloud Workflows step budgets ────────────
+
+/** Result of a `plan` invocation. Carries enough to size the Map(N) state. */
+export interface PlanResultBody {
+  Action: "plan";
+  PlanGcsUri: string;
+  PlanHash: string;
+  ChunkCount: number;
+  TotalFrames: number;
+  Fps: 24 | 30 | 60;
+  Width: number;
+  Height: number;
+  Format: DistributedFormat;
+  HasAudio: boolean;
+  AudioGcsUri: string | null;
+  FfmpegVersion: string;
+  ProducerVersion: string;
+  DurationMs: number;
+}
+
+/** Result of a `renderChunk` invocation. Sized ≤200 bytes. */
+export interface RenderChunkResultBody {
+  Action: "renderChunk";
+  ChunkGcsUri: string;
+  ChunkIndex: number;
+  Sha256: string;
+  FramesEncoded: number;
+  DurationMs: number;
+}
+
+/** Result of an `assemble` invocation. */
+export interface AssembleResultBody {
+  Action: "assemble";
+  OutputGcsUri: string;
+  FramesEncoded: number;
+  FileSize: number;
+  DurationMs: number;
+}
+
+export type CloudRunResult = PlanResultBody | RenderChunkResultBody | AssembleResultBody;

packages/gcp-cloud-run/src/formatExtension.ts

@@ -0,0 +1,27 @@
+/**
+ * Map a distributed `format` to the file extension the assembled output
+ * should carry on disk + in GCS. Shared by `src/server.ts` (chunk +
+ * assemble output paths) and `src/sdk/renderToCloudRun.ts` (final
+ * output key construction) so the two sides agree on what an mp4
+ * looks like vs a png-sequence.
+ */
+
+import type { DistributedFormat } from "@hyperframes/producer/distributed";
+
+export type { DistributedFormat } from "@hyperframes/producer/distributed";
+
+// Closed-enum lookup table. TS enforces exhaustiveness via the
+// `Record<DistributedFormat, string>` annotation — adding a format to
+// `DistributedFormat` without adding the matching key here fails to
+// typecheck, which is the same exhaustiveness guarantee a switch +
+// `_exhaustive: never` arm provides but at lower complexity.
+const FORMAT_EXTENSIONS: Record<DistributedFormat, string> = {
+  mp4: ".mp4",
+  mov: ".mov",
+  webm: ".webm",
+  "png-sequence": "",
+};
+
+export function formatExtension(format: DistributedFormat): string {
+  return FORMAT_EXTENSIONS[format];
+}

packages/gcp-cloud-run/src/gcsTransport.test.ts

@@ -0,0 +1,114 @@
+/**
+ * GCS transport unit tests — URI parsing, tar round-trip, and the
+ * download/upload bridge over the `FakeGcs` double.
+ */
+
+import { afterEach, describe, expect, it } from "bun:test";
+import { existsSync, mkdirSync, mkdtempSync, readFileSync, rmSync, writeFileSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import { asStorage, FakeGcs } from "./__fixtures__/fakeGcs.js";
+import {
+  downloadGcsObjectToFile,
+  formatGcsUri,
+  parseGcsUri,
+  tarDirectory,
+  untarDirectory,
+  uploadFileToGcs,
+} from "./gcsTransport.js";
+
+const tmpDirs: string[] = [];
+function mkTmp(prefix: string): string {
+  const dir = mkdtempSync(join(tmpdir(), prefix));
+  tmpDirs.push(dir);
+  return dir;
+}
+afterEach(() => {
+  for (const d of tmpDirs.splice(0)) rmSync(d, { recursive: true, force: true });
+});
+
+describe("parseGcsUri", () => {
+  it("splits bucket and key", () => {
+    expect(parseGcsUri("gs://my-bucket/path/to/object.tar.gz")).toEqual({
+      bucket: "my-bucket",
+      key: "path/to/object.tar.gz",
+    });
+  });
+
+  it("rejects non-gs URIs", () => {
+    expect(() => parseGcsUri("s3://b/k")).toThrow(/expected gs:\/\//);
+  });
+
+  it("rejects a bucket with no key", () => {
+    expect(() => parseGcsUri("gs://just-a-bucket")).toThrow(/missing key/);
+  });
+
+  it("rejects an empty bucket", () => {
+    expect(() => parseGcsUri("gs:///key")).toThrow(/empty bucket or key/);
+  });
+
+  it("round-trips through formatGcsUri", () => {
+    const uri = "gs://b/some/key";
+    expect(formatGcsUri(parseGcsUri(uri))).toBe(uri);
+  });
+});
+
+describe("tarDirectory / untarDirectory", () => {
+  it("round-trips a directory tree", async () => {
+    const src = mkTmp("hf-tar-src-");
+    mkdirSync(join(src, "nested"), { recursive: true });
+    writeFileSync(join(src, "index.html"), "<html>hi</html>");
+    writeFileSync(join(src, "nested", "data.json"), '{"a":1}');
+
+    const work = mkTmp("hf-tar-work-");
+    const tarball = join(work, "out.tar.gz");
+    await tarDirectory(src, tarball);
+    expect(existsSync(tarball)).toBe(true);
+
+    const dest = join(work, "extracted");
+    await untarDirectory(tarball, dest);
+    expect(readFileSync(join(dest, "index.html"), "utf8")).toBe("<html>hi</html>");
+    expect(readFileSync(join(dest, "nested", "data.json"), "utf8")).toBe('{"a":1}');
+  });
+
+  it("untar wipes a stale destination first", async () => {
+    const src = mkTmp("hf-tar-src2-");
+    writeFileSync(join(src, "keep.txt"), "new");
+    const work = mkTmp("hf-tar-work2-");
+    const tarball = join(work, "out.tar.gz");
+    await tarDirectory(src, tarball);
+
+    const dest = join(work, "extracted");
+    mkdirSync(dest, { recursive: true });
+    writeFileSync(join(dest, "stale.txt"), "should be gone");
+
+    await untarDirectory(tarball, dest);
+    expect(existsSync(join(dest, "stale.txt"))).toBe(false);
+    expect(readFileSync(join(dest, "keep.txt"), "utf8")).toBe("new");
+  });
+});
+
+describe("download/upload bridge", () => {
+  it("uploads a local file then downloads identical bytes", async () => {
+    const gcs = new FakeGcs();
+    const work = mkTmp("hf-dl-");
+    const srcFile = join(work, "src.bin");
+    writeFileSync(srcFile, Buffer.from("hello gcs"));
+
+    const uri = "gs://bucket/obj.bin";
+    await uploadFileToGcs(asStorage(gcs), srcFile, uri, "application/octet-stream");
+
+    const dest = join(work, "dl.bin");
+    await downloadGcsObjectToFile(asStorage(gcs), uri, dest);
+    expect(readFileSync(dest, "utf8")).toBe("hello gcs");
+
+    expect(gcs.ops.map((o) => o.kind)).toEqual(["upload", "download"]);
+  });
+
+  it("upload throws when the source file is missing", async () => {
+    const gcs = new FakeGcs();
+    await expect(uploadFileToGcs(asStorage(gcs), "/no/such/file", "gs://b/k")).rejects.toThrow(
+      /upload source missing/,
+    );
+  });
+});

packages/gcp-cloud-run/src/gcsTransport.ts

@@ -0,0 +1,130 @@
+/**
+ * Thin GCS transport for the Cloud Run handler.
+ *
+ * The OSS distributed primitives are pure functions over local file paths;
+ * the handler bridges GCS ↔ the container's writable `/tmp` filesystem on
+ * each request. Functions here are intentionally narrow: parse a URI,
+ * download an object to a local path, upload a path, tar-pack a planDir,
+ * tar-extract a planDir back out.
+ *
+ * Tar (not zip) for planDir transit:
+ *   - planDirs contain symlinks (the extract stage materializes them but
+ *     the compiled/ subtree may include linked assets); tar preserves them,
+ *     zip does not.
+ *   - We use the `tar` npm package (pure JS over `node:zlib`) so the
+ *     archive format doesn't depend on a system `tar`/`unzip` being present
+ *     in the container image.
+ *
+ * Apart from the `gs://` scheme and the `@google-cloud/storage` client this
+ * is the same shape as `@hyperframes/aws-lambda`'s `s3Transport.ts`.
+ */
+
+import { createWriteStream, existsSync, mkdirSync, rmSync, statSync } from "node:fs";
+import { dirname } from "node:path";
+import { pipeline } from "node:stream/promises";
+import type { Storage } from "@google-cloud/storage";
+import * as tar from "tar";
+
+/** Parsed `gs://bucket/key` URI. */
+export interface GcsLocation {
+  bucket: string;
+  key: string;
+}
+
+/** Parse `gs://bucket/key/path` → `{ bucket, key }`. Throws on malformed input. */
+// fallow-ignore-next-line complexity
+export function parseGcsUri(uri: string): GcsLocation {
+  if (!uri.startsWith("gs://")) {
+    throw new Error(`[gcsTransport] expected gs:// URI, got: ${JSON.stringify(uri)}`);
+  }
+  const rest = uri.slice("gs://".length);
+  const slash = rest.indexOf("/");
+  if (slash === -1) {
+    throw new Error(`[gcsTransport] missing key in gs URI: ${JSON.stringify(uri)}`);
+  }
+  const bucket = rest.slice(0, slash);
+  const key = rest.slice(slash + 1);
+  if (!bucket || !key) {
+    throw new Error(`[gcsTransport] empty bucket or key in gs URI: ${JSON.stringify(uri)}`);
+  }
+  return { bucket, key };
+}
+
+/** Build `gs://bucket/key` from a location. */
+export function formatGcsUri(loc: GcsLocation): string {
+  return `gs://${loc.bucket}/${loc.key}`;
+}
+
+/** Stream a GCS object to a local file path. */
+export async function downloadGcsObjectToFile(
+  storage: Storage,
+  uri: string,
+  destPath: string,
+): Promise<void> {
+  const { bucket, key } = parseGcsUri(uri);
+  mkdirSync(dirname(destPath), { recursive: true });
+  const file = storage.bucket(bucket).file(key);
+  // `createReadStream` streams the object body; piping into a write stream
+  // keeps memory flat for large plan tarballs / chunk files rather than
+  // buffering the whole object the way `file.download()` would.
+  await pipeline(file.createReadStream(), createWriteStream(destPath));
+}
+
+/**
+ * Upload a local file's contents to a GCS URI using a resumable upload.
+ * GCS objects have no practical size ceiling for the artifacts this adapter
+ * handles (plan tarballs ≤ 2 GB, chunks ≤ a few hundred MB), so a single
+ * upload call works for every case.
+ */
+export async function uploadFileToGcs(
+  storage: Storage,
+  localPath: string,
+  uri: string,
+  contentType?: string,
+): Promise<void> {
+  if (!existsSync(localPath)) {
+    throw new Error(`[gcsTransport] upload source missing: ${localPath}`);
+  }
+  const { bucket, key } = parseGcsUri(uri);
+  await storage.bucket(bucket).upload(localPath, {
+    destination: key,
+    // `resumable: false` (simple upload) is faster for the small-to-medium
+    // objects this adapter moves and avoids the extra round-trip a resumable
+    // session start costs; GCS recommends resumable only past ~8 MB but our
+    // chunks are reliably above that, so let the client pick by default.
+    contentType,
+  });
+}
+
+/**
+ * Pack a directory into a `.tar.gz` at `destTarball`. Uses the `tar` npm
+ * package (pure JS over `node:zlib`) rather than spawning a system tar
+ * binary so the archive format is independent of the container's userland.
+ */
+export async function tarDirectory(sourceDir: string, destTarball: string): Promise<void> {
+  if (!existsSync(sourceDir) || !statSync(sourceDir).isDirectory()) {
+    throw new Error(`[gcsTransport] tar source must be an existing directory: ${sourceDir}`);
+  }
+  mkdirSync(dirname(destTarball), { recursive: true });
+  await tar.create({ gzip: true, file: destTarball, cwd: sourceDir }, ["."]);
+}
+
+/**
+ * Extract a `.tar.gz` produced by {@link tarDirectory} into `destDir`.
+ * The directory is created (or cleared) before extraction so a retried
+ * request doesn't observe stale files from a prior run on the same warm
+ * container instance.
+ */
+export async function untarDirectory(tarballPath: string, destDir: string): Promise<void> {
+  if (!existsSync(tarballPath)) {
+    throw new Error(`[gcsTransport] tarball missing: ${tarballPath}`);
+  }
+  // Wipe target so a warm container instance's prior planDir doesn't bleed
+  // into the new request. Cloud Run re-uses the instance filesystem across
+  // requests served by the same instance.
+  if (existsSync(destDir)) {
+    rmSync(destDir, { recursive: true, force: true });
+  }
+  mkdirSync(destDir, { recursive: true });
+  await tar.extract({ file: tarballPath, cwd: destDir });
+}

packages/gcp-cloud-run/src/index.ts

@@ -0,0 +1,64 @@
+/**
+ * `@hyperframes/gcp-cloud-run` — Google Cloud Run + Workflows adapter for
+ * the HyperFrames distributed render pipeline.
+ *
+ * Two surfaces, one package:
+ *
+ *  - **Server-side handler.** `dispatch`, `createApp`, the request/result
+ *    types, Chrome resolution, and GCS transport. These power the Cloud Run
+ *    service built from the package `Dockerfile`.
+ *  - **Client-side SDK.** `renderToCloudRun`, `getRenderProgress`,
+ *    `deploySite`, `validateDistributedRenderConfig`, and `computeRenderCost`.
+ *    Adopters call these from their Node process (CI scripts, CLIs) to drive
+ *    a deployed stack without writing GCS / Workflows boilerplate.
+ *
+ * The Terraform module that provisions the bucket + service + workflow lives
+ * under `terraform/` in the published package; see the README. The package
+ * is NOT a dependency of `@hyperframes/producer`; consumers install it
+ * separately.
+ */
+
+export { createApp, dispatch, type HandlerDeps, startServer, unwrapEvent } from "./server.js";
+export {
+  type AssembleEvent,
+  type AssembleResultBody,
+  type CloudRunAction,
+  type CloudRunEvent,
+  type CloudRunResult,
+  type PlanEvent,
+  type PlanResultBody,
+  type RenderChunkEvent,
+  type RenderChunkResultBody,
+  type SerializableDistributedRenderConfig,
+} from "./events.js";
+export { ChromeBinaryUnavailableError, resolveChromeExecutablePath } from "./chromium.js";
+export {
+  downloadGcsObjectToFile,
+  formatGcsUri,
+  type GcsLocation,
+  parseGcsUri,
+  tarDirectory,
+  untarDirectory,
+  uploadFileToGcs,
+} from "./gcsTransport.js";
+
+// ── Client-side SDK ─────────────────────────────────────────────────────────
+export { deploySite, type DeploySiteOptions, type SiteHandle } from "./sdk/deploySite.js";
+export {
+  renderToCloudRun,
+  type RenderHandle,
+  type RenderToCloudRunOptions,
+} from "./sdk/renderToCloudRun.js";
+export {
+  getRenderProgress,
+  type GetRenderProgressOptions,
+  type RenderError,
+  type RenderProgress,
+  type RenderStatus,
+} from "./sdk/getRenderProgress.js";
+export {
+  type BilledCloudRunInvocation,
+  computeRenderCost,
+  type RenderCost,
+} from "./sdk/costAccounting.js";
+export { InvalidConfigError, validateDistributedRenderConfig } from "./sdk/validateConfig.js";

packages/gcp-cloud-run/src/sdk/costAccounting.test.ts

@@ -0,0 +1,35 @@
+/**
+ * `computeRenderCost` unit tests — Cloud Run vCPU/GiB-second + Workflows
+ * step math.
+ */
+
+import { describe, expect, it } from "bun:test";
+import { type BilledCloudRunInvocation, computeRenderCost } from "./costAccounting.js";
+
+describe("computeRenderCost", () => {
+  it("returns zero for no invocations", () => {
+    const cost = computeRenderCost([], 0);
+    expect(cost.accruedSoFarUsd).toBe(0);
+    expect(cost.displayCost).toBe("$0.0000");
+  });
+
+  it("sums vCPU + memory seconds plus per-request and step charges", () => {
+    const invs: BilledCloudRunInvocation[] = [
+      { durationMs: 10_000, vcpu: 4, memoryGib: 16, estimated: false },
+      { durationMs: 10_000, vcpu: 4, memoryGib: 16, estimated: false },
+    ];
+    const cost = computeRenderCost(invs, 6);
+    // 2 × 10s: vCPU 80 vcpu-s × 0.000024 = 0.00192; mem 320 GiB-s × 0.0000025 = 0.0008;
+    // requests 2 × 4e-7 ≈ 0 → raw 0.0027208, rounded to 4 dp = 0.0027.
+    // workflows 6 × 1e-5 = 0.00006 → rounds up to 0.0001 at 4 dp.
+    expect(cost.breakdown.cloudRunUsd).toBeCloseTo(0.0027, 4);
+    expect(cost.breakdown.workflowsUsd).toBeCloseTo(0.0001, 4);
+    expect(cost.accruedSoFarUsd).toBeGreaterThan(0);
+    expect(cost.breakdown.gcsEstimate).toBe("not-included");
+  });
+
+  it("flags estimated when any invocation was estimated", () => {
+    const cost = computeRenderCost([{ durationMs: 0, vcpu: 4, memoryGib: 16, estimated: true }], 4);
+    expect(cost.breakdown.estimated).toBe(true);
+  });
+});

packages/gcp-cloud-run/src/sdk/costAccounting.ts

@@ -0,0 +1,113 @@
+/**
+ * Per-render cost accounting for {@link getRenderProgress}.
+ *
+ * Google bills the render service two ways:
+ *
+ *  - **Cloud Run** by **vCPU-seconds** and **GiB-seconds** of request
+ *    processing time, plus a flat per-request charge. Each handler
+ *    invocation returns its own `DurationMs` in the result body, so the
+ *    progress reader can recover billed time per step without a separate
+ *    Cloud Monitoring query — multiply by the service's configured vCPU /
+ *    memory to get the resource-seconds.
+ *  - **Cloud Workflows** by **steps executed**. The orchestration is a
+ *    fixed shape (Plan + N×RenderChunk + Assemble + a handful of control
+ *    steps), so the step count scales with chunk count.
+ *
+ * The math is documented inline so the constants stay close to the pricing
+ * source they came from. Cost is **best-effort**: GCP pricing varies by
+ * region + committed-use discounts; we use on-demand `us-central1` (Tier 1)
+ * rates as of 2026-06 and label the result `displayCost` so callers see the
+ * dollar value but downstream automation can also read the raw number.
+ */
+
+/** Cloud Run request-based billing, us-central1 Tier 1: USD per vCPU-second. */
+const CLOUD_RUN_USD_PER_VCPU_SECOND = 0.000024;
+/** Cloud Run request-based billing, us-central1 Tier 1: USD per GiB-second. */
+const CLOUD_RUN_USD_PER_GIB_SECOND = 0.0000025;
+/** Cloud Run: USD per request ($0.40 per million). */
+const CLOUD_RUN_USD_PER_REQUEST = 0.0000004;
+/** Cloud Workflows: USD per internal step ($0.01 per 1,000, after a free tier). */
+const WORKFLOWS_USD_PER_STEP = 0.00001;
+
+/** Per-invocation billed slice the cost calc cares about. */
+export interface BilledCloudRunInvocation {
+  /** Wall-clock the handler reported via `DurationMs` in its result body. */
+  durationMs: number;
+  /** vCPU the Cloud Run service was configured with at invocation time. */
+  vcpu: number;
+  /** Memory in GiB the Cloud Run service was configured with. */
+  memoryGib: number;
+  /** `true` if the duration was inferred (step result missing) rather than read from the handler payload. */
+  estimated: boolean;
+}
+
+/**
+ * Result of {@link computeRenderCost}.
+ *
+ * NOTE: `displayCost` / `accruedSoFarUsd` cover Cloud Run compute + Cloud
+ * Workflows steps only. They EXCLUDE GCS storage + network egress for the
+ * plan tarball (which can be ~100 MB), chunk artifacts, and the final output
+ * — see `breakdown.gcsEstimate`. Treat the figure as a compute-cost floor,
+ * not the authoritative total bill.
+ */
+export interface RenderCost {
+  /** USD accrued to date (Cloud Run + Workflows only; excludes GCS — see note above). */
+  accruedSoFarUsd: number;
+  /** Human-readable USD string, e.g. `"$0.0214"`. Excludes GCS storage/egress. */
+  displayCost: string;
+  breakdown: {
+    cloudRunUsd: number;
+    workflowsUsd: number;
+    /** GCS transfer + storage cost varies by tier; we don't try to compute it here. */
+    gcsEstimate: "not-included";
+    /** `true` if any invocation fell back to estimated billing. */
+    estimated: boolean;
+  };
+}
+
+/**
+ * Sum Cloud Run vCPU-seconds + GiB-seconds + per-request charges and Cloud
+ * Workflows steps into an aggregate USD figure.
+ *
+ * `workflowSteps` is the count of Workflows steps executed so far — Plan
+ * (1) + RenderChunk (chunkCount) + Assemble (1) + the control steps
+ * (BuildChunkList, AssertChunkCount, …). Pass the count the progress reader
+ * derived from the execution; a rough constant overhead is fine since the
+ * step charge is a rounding error next to Cloud Run compute.
+ */
+export function computeRenderCost(
+  invocations: BilledCloudRunInvocation[],
+  workflowSteps: number,
+): RenderCost {
+  let cloudRunUsd = 0;
+  let anyEstimated = false;
+  for (const inv of invocations) {
+    const seconds = inv.durationMs / 1000;
+    cloudRunUsd += seconds * inv.vcpu * CLOUD_RUN_USD_PER_VCPU_SECOND;
+    cloudRunUsd += seconds * inv.memoryGib * CLOUD_RUN_USD_PER_GIB_SECOND;
+    cloudRunUsd += CLOUD_RUN_USD_PER_REQUEST;
+    if (inv.estimated) anyEstimated = true;
+  }
+  const workflowsUsd = workflowSteps * WORKFLOWS_USD_PER_STEP;
+  const accruedSoFarUsd = roundUsd(cloudRunUsd + workflowsUsd);
+  return {
+    accruedSoFarUsd,
+    displayCost: formatUsd(accruedSoFarUsd),
+    breakdown: {
+      cloudRunUsd: roundUsd(cloudRunUsd),
+      workflowsUsd: roundUsd(workflowsUsd),
+      gcsEstimate: "not-included",
+      estimated: anyEstimated,
+    },
+  };
+}
+
+function roundUsd(usd: number): number {
+  // Four decimal places — enough resolution for per-chunk granularity.
+  // Anything finer is noise vs GCP's own rounding.
+  return Math.round(usd * 10_000) / 10_000;
+}
+
+function formatUsd(usd: number): string {
+  return `$${usd.toFixed(4)}`;
+}

packages/gcp-cloud-run/src/sdk/deploySite.test.ts

@@ -0,0 +1,86 @@
+/**
+ * `deploySite` unit tests — content-addressed siteId, existence
+ * short-circuit, and the upload path over `FakeGcs`.
+ */
+
+import { afterEach, describe, expect, it } from "bun:test";
+import { mkdtempSync, rmSync, writeFileSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import { asStorage, FakeGcs } from "../__fixtures__/fakeGcs.js";
+import { deploySite } from "./deploySite.js";
+
+const tmpDirs: string[] = [];
+function mkProject(content: string): string {
+  const dir = mkdtempSync(join(tmpdir(), "hf-site-"));
+  tmpDirs.push(dir);
+  writeFileSync(join(dir, "index.html"), content);
+  return dir;
+}
+afterEach(() => {
+  for (const d of tmpDirs.splice(0)) rmSync(d, { recursive: true, force: true });
+});
+
+describe("deploySite", () => {
+  it("uploads and returns a content-addressed handle", async () => {
+    const gcs = new FakeGcs();
+    const dir = mkProject("<html>v1</html>");
+    const handle = await deploySite({ projectDir: dir, bucketName: "b", storage: asStorage(gcs) });
+    expect(handle.uploaded).toBe(true);
+    expect(handle.projectGcsUri).toBe(`gs://b/sites/${handle.siteId}/project.tar.gz`);
+    expect(gcs.objects.has(handle.projectGcsUri)).toBe(true);
+  });
+
+  it("produces a stable siteId for identical content", async () => {
+    const a = await deploySite({
+      projectDir: mkProject("<html>same</html>"),
+      bucketName: "b",
+      storage: asStorage(new FakeGcs()),
+    });
+    const b = await deploySite({
+      projectDir: mkProject("<html>same</html>"),
+      bucketName: "b",
+      storage: asStorage(new FakeGcs()),
+    });
+    expect(a.siteId).toBe(b.siteId);
+  });
+
+  it("produces different siteIds for different content", async () => {
+    const a = await deploySite({
+      projectDir: mkProject("<html>one</html>"),
+      bucketName: "b",
+      storage: asStorage(new FakeGcs()),
+    });
+    const b = await deploySite({
+      projectDir: mkProject("<html>two</html>"),
+      bucketName: "b",
+      storage: asStorage(new FakeGcs()),
+    });
+    expect(a.siteId).not.toBe(b.siteId);
+  });
+
+  it("short-circuits the upload when the object already exists", async () => {
+    const gcs = new FakeGcs();
+    const dir = mkProject("<html>cache</html>");
+    const first = await deploySite({ projectDir: dir, bucketName: "b", storage: asStorage(gcs) });
+    expect(first.uploaded).toBe(true);
+
+    const second = await deploySite({ projectDir: dir, bucketName: "b", storage: asStorage(gcs) });
+    expect(second.uploaded).toBe(false);
+    expect(second.siteId).toBe(first.siteId);
+    // Only one upload op total.
+    expect(gcs.ops.filter((o) => o.kind === "upload").length).toBe(1);
+  });
+
+  it("honours an explicit siteId override", async () => {
+    const gcs = new FakeGcs();
+    const handle = await deploySite({
+      projectDir: mkProject("<html></html>"),
+      bucketName: "b",
+      siteId: "my-git-sha",
+      storage: asStorage(gcs),
+    });
+    expect(handle.siteId).toBe("my-git-sha");
+    expect(handle.projectGcsUri).toBe("gs://b/sites/my-git-sha/project.tar.gz");
+  });
+});

packages/gcp-cloud-run/src/sdk/deploySite.ts

@@ -0,0 +1,130 @@
+/**
+ * `deploySite` — upload a project directory to GCS once per content hash
+ * and return a reusable handle.
+ *
+ * `renderToCloudRun` calls this implicitly when no `siteHandle` is passed,
+ * but exposing it as a standalone verb lets adopters bundle a project ahead
+ * of time and reuse the handle across many renders without re-tarring the
+ * project tree on every call.
+ *
+ * The handle is **content-addressed**: `siteId` is derived from a SHA-256
+ * over the project files. Two `deploySite` calls on an unchanged tree
+ * produce the same `siteId` and short-circuit the upload after a single
+ * existence check.
+ */
+
+import { mkdtempSync, rmSync, statSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import { Storage } from "@google-cloud/storage";
+import { hashProjectDir } from "@hyperframes/producer/distributed";
+import { formatGcsUri, tarDirectory, uploadFileToGcs } from "../gcsTransport.js";
+
+/** Options for {@link deploySite}. */
+export interface DeploySiteOptions {
+  /** Local project directory containing `index.html` (and any composition assets). */
+  projectDir: string;
+  /** GCS bucket the Terraform module provisioned. */
+  bucketName: string;
+  /**
+   * Override the content-addressed site id. Useful when the caller has a
+   * stable external identifier they want to use (e.g. a git SHA); if unset,
+   * the hash of the project tree picks it.
+   */
+  siteId?: string;
+  /** Injection seam for tests. Production callers leave unset. */
+  storage?: Storage;
+}
+
+/** Stable handle returned by {@link deploySite}. Pass back to {@link renderToCloudRun}. */
+export interface SiteHandle {
+  /** Content-addressed (or caller-supplied) identifier; stable across re-uploads of the same tree. */
+  siteId: string;
+  /** Bucket the site landed in. Surfaced separately so callers don't have to re-parse `projectGcsUri`. */
+  bucketName: string;
+  /** Full `gs://bucket/sites/<siteId>/project.tar.gz` URI; pass through to `renderToCloudRun`. */
+  projectGcsUri: string;
+  /** Tarball size in bytes; useful for "did we actually skip the upload?" assertions. */
+  bytes: number;
+  /** ISO timestamp of the most recent upload OR the existing object the short-circuit found. */
+  uploadedAt: string;
+  /** `false` if the object already existed and we skipped the upload. */
+  uploaded: boolean;
+}
+
+/**
+ * Upload `projectDir` to `gs://bucketName/sites/<siteId>/project.tar.gz`.
+ *
+ * Short-circuits when an object with the same key already exists in the
+ * bucket — `siteId` derives from the project's content hash, so the same
+ * bytes produce the same key, and re-uploading would be redundant.
+ */
+// fallow-ignore-next-line complexity
+export async function deploySite(opts: DeploySiteOptions): Promise<SiteHandle> {
+  if (!statSync(opts.projectDir).isDirectory()) {
+    throw new Error(`[deploySite] projectDir is not a directory: ${opts.projectDir}`);
+  }
+
+  const siteId = opts.siteId ?? hashProjectDir(opts.projectDir);
+  const key = `sites/${siteId}/project.tar.gz`;
+  const projectGcsUri = formatGcsUri({ bucket: opts.bucketName, key });
+  const storage = opts.storage ?? new Storage();
+  const file = storage.bucket(opts.bucketName).file(key);
+
+  // Existence short-circuit. Adopters re-rendering the same project on a
+  // tight inner loop (CI smoke, demo flows) save the tar+gzip+upload pass
+  // on every iteration.
+  const existing = await headObject(file);
+  if (existing) {
+    return {
+      siteId,
+      bucketName: opts.bucketName,
+      projectGcsUri,
+      bytes: existing.bytes,
+      uploadedAt: existing.lastModified,
+      uploaded: false,
+    };
+  }
+
+  const workdir = mkdtempSync(join(tmpdir(), "hf-deploy-site-"));
+  try {
+    const tarball = join(workdir, "project.tar.gz");
+    await tarDirectory(opts.projectDir, tarball);
+    const size = statSync(tarball).size;
+    await uploadFileToGcs(storage, tarball, projectGcsUri, "application/gzip");
+    return {
+      siteId,
+      bucketName: opts.bucketName,
+      projectGcsUri,
+      bytes: size,
+      uploadedAt: new Date().toISOString(),
+      uploaded: true,
+    };
+  } finally {
+    rmSync(workdir, { recursive: true, force: true });
+  }
+}
+
+/**
+ * Narrow surface of the `@google-cloud/storage` `File` this module uses —
+ * lets the test double implement just `exists()` + `getMetadata()` without
+ * pulling the full client type.
+ */
+interface FileLike {
+  exists(): Promise<[boolean, ...unknown[]]>;
+  getMetadata(): Promise<[{ size?: string | number; updated?: string }, ...unknown[]]>;
+}
+
+// fallow-ignore-next-line complexity
+async function headObject(file: FileLike): Promise<{ bytes: number; lastModified: string } | null> {
+  const [exists] = await file.exists();
+  if (!exists) return null;
+  const [meta] = await file.getMetadata();
+  const sizeRaw = meta.size;
+  const bytes =
+    typeof sizeRaw === "string" ? Number(sizeRaw) : typeof sizeRaw === "number" ? sizeRaw : 0;
+  return {
+    bytes: Number.isFinite(bytes) ? bytes : 0,
+    lastModified: meta.updated ?? new Date().toISOString(),
+  };
+}

packages/gcp-cloud-run/src/sdk/getRenderProgress.test.ts

@@ -0,0 +1,108 @@
+/**
+ * `getRenderProgress` unit tests — state mapping + parsing the accumulated
+ * workflow result into frame totals, output file, and cost.
+ */
+
+import { describe, expect, it } from "bun:test";
+import {
+  type ExecutionRecord,
+  type ExecutionsGetClientLike,
+  getRenderProgress,
+} from "./getRenderProgress.js";
+
+function fakeExecutions(record: ExecutionRecord): ExecutionsGetClientLike {
+  return {
+    async getExecution(_req: { name: string }) {
+      return [record] as [ExecutionRecord];
+    },
+  };
+}
+
+const accumulated = JSON.stringify({
+  Plan: { TotalFrames: 90, DurationMs: 4000 },
+  Chunks: [
+    { FramesEncoded: 30, DurationMs: 8000 },
+    { FramesEncoded: 30, DurationMs: 8000 },
+    { FramesEncoded: 30, DurationMs: 8000 },
+  ],
+  Assemble: { OutputGcsUri: "gs://b/renders/r1/output.mp4", FileSize: 123456, DurationMs: 3000 },
+});
+
+describe("getRenderProgress", () => {
+  it("reports running with no frame data while ACTIVE", async () => {
+    const p = await getRenderProgress({
+      executionName: "x",
+      executions: fakeExecutions({ state: "ACTIVE", startTime: { seconds: 1700000000 } }),
+    });
+    expect(p.status).toBe("running");
+    expect(p.overallProgress).toBe(0);
+    expect(p.totalFrames).toBeNull();
+    expect(p.fatalErrorEncountered).toBe(false);
+  });
+
+  it("reports succeeded with parsed frames + cost", async () => {
+    const p = await getRenderProgress({
+      executionName: "x",
+      vcpu: 4,
+      memoryGib: 16,
+      executions: fakeExecutions({
+        state: "SUCCEEDED",
+        result: accumulated,
+        startTime: { seconds: 1700000000 },
+        endTime: { seconds: 1700000031 },
+      }),
+    });
+    expect(p.status).toBe("succeeded");
+    expect(p.overallProgress).toBe(1);
+    expect(p.totalFrames).toBe(90);
+    expect(p.framesRendered).toBe(90);
+    expect(p.invocationsObserved).toBe(5); // plan + 3 chunks + assemble
+    expect(p.outputFile).toEqual({ gcsUri: "gs://b/renders/r1/output.mp4", bytes: 123456 });
+    expect(p.costs.accruedSoFarUsd).toBeGreaterThan(0);
+    expect(p.costs.breakdown.estimated).toBe(false);
+  });
+
+  it("maps FAILED to a fatal error and surfaces the error payload", async () => {
+    const p = await getRenderProgress({
+      executionName: "x",
+      executions: fakeExecutions({
+        state: "FAILED",
+        error: { payload: "boom", context: "renderChunk" },
+      }),
+    });
+    expect(p.status).toBe("failed");
+    expect(p.fatalErrorEncountered).toBe(true);
+    expect(p.errors[0]?.cause).toBe("boom");
+    expect(p.errors[0]?.state).toBe("renderChunk");
+  });
+
+  it("extracts the handler error name from a wrapped http failure payload", async () => {
+    // Workflows wraps an http step failure as { code, message, body }, where
+    // body is the handler's JSON { error, message }.
+    const payload = JSON.stringify({
+      code: 400,
+      message: "HTTP server responded with error code 400",
+      body: JSON.stringify({ error: "PLAN_HASH_MISMATCH", message: "mismatch" }),
+    });
+    const p = await getRenderProgress({
+      executionName: "x",
+      executions: fakeExecutions({ state: "FAILED", error: { payload, context: "renderChunk" } }),
+    });
+    expect(p.errors[0]?.error).toBe("PLAN_HASH_MISMATCH");
+  });
+
+  it("maps CANCELLED", async () => {
+    const p = await getRenderProgress({
+      executionName: "x",
+      executions: fakeExecutions({ state: "CANCELLED" }),
+    });
+    expect(p.status).toBe("cancelled");
+    expect(p.fatalErrorEncountered).toBe(true);
+  });
+
+  it("requires an executionName", async () => {
+    await expect(
+      getRenderProgress({ executionName: "", executions: fakeExecutions({}) }),
+    ).rejects.toThrow(/executionName is required/);
+  });
+});

packages/gcp-cloud-run/src/sdk/getRenderProgress.ts

@@ -0,0 +1,268 @@
+/**
+ * `getRenderProgress` — read-only progress + cost snapshot for a single
+ * render started by {@link renderToCloudRun}.
+ *
+ * Pulls one `GetExecution` per call. Cloud Workflows does not surface
+ * per-step payloads through the basic Executions API the way Step Functions
+ * exposes its history, so this reader takes a different tack than the AWS
+ * adapter: the workflow definition **accumulates** each step's result body
+ * (Plan + every RenderChunk + Assemble) and returns them as one structured
+ * object. On success we parse that object for frame totals, the output
+ * file, and per-step `DurationMs` (which the handler stamps into every
+ * result), then compute cost against the service's configured vCPU/memory.
+ *
+ * Progress is therefore coarse while the execution is ACTIVE (we report
+ * `running` with `overallProgress = 0`) and exact once it SUCCEEDS
+ * (`overallProgress = 1`, real frame + cost numbers). Mid-flight per-chunk
+ * progress would require the Workflows step-entries API; that's a tracked
+ * follow-up, not part of the first version.
+ */
+
+import {
+  type BilledCloudRunInvocation,
+  computeRenderCost,
+  type RenderCost,
+} from "./costAccounting.js";
+
+/** Normalised render status. Maps from Cloud Workflows execution states. */
+export type RenderStatus = "running" | "succeeded" | "failed" | "cancelled" | "unknown";
+
+/** One error surfaced by the execution. */
+export interface RenderError {
+  /** Step the failure surfaced in, when recoverable from the error context; else `<execution>`. */
+  state: string;
+  /** Error class / type. */
+  error: string;
+  /** Cause string (often a stringified JSON payload from the handler). */
+  cause: string;
+}
+
+/** Snapshot of a single render's progress + cost + errors at one point in time. */
+export interface RenderProgress {
+  status: RenderStatus;
+  /** `[0, 1]`; coarse while running, exact on success. */
+  overallProgress: number;
+  framesRendered: number;
+  /** `null` until the execution succeeds and the accumulated plan result is read. */
+  totalFrames: number | null;
+  /** Cloud Run invocations the workflow scheduled (Plan + chunks + Assemble), when known. */
+  invocationsObserved: number;
+  costs: RenderCost;
+  /** Final output object if Assemble succeeded; `null` otherwise. */
+  outputFile: { gcsUri: string; bytes: number | null } | null;
+  errors: RenderError[];
+  /** `true` once the execution has terminated in a non-success state. */
+  fatalErrorEncountered: boolean;
+  startedAt: string;
+  endedAt: string | null;
+}
+
+/** Protobuf Timestamp shape the gapic client returns for start/end times. */
+interface ProtoTimestamp {
+  seconds?: number | string | null;
+  nanos?: number | null;
+}
+
+/** Subset of a Cloud Workflows Execution this reader consumes. */
+export interface ExecutionRecord {
+  name?: string | null;
+  state?: string | null;
+  result?: string | null;
+  error?: { payload?: string | null; context?: string | null } | null;
+  startTime?: ProtoTimestamp | string | null;
+  endTime?: ProtoTimestamp | string | null;
+}
+
+/** Minimal surface of `@google-cloud/workflows`' `ExecutionsClient` for reads. */
+export interface ExecutionsGetClientLike {
+  getExecution(req: { name: string }): Promise<[ExecutionRecord, ...unknown[]]>;
+}
+
+/** Options for {@link getRenderProgress}. */
+export interface GetRenderProgressOptions {
+  /** Server-assigned execution resource name from a {@link renderToCloudRun} call. */
+  executionName: string;
+  /** vCPU the Cloud Run service is configured with (for cost). Default 4. */
+  vcpu?: number;
+  /** Memory in GiB the Cloud Run service is configured with (for cost). Default 16. */
+  memoryGib?: number;
+  /** Test injection seam — production callers leave unset. */
+  executions?: ExecutionsGetClientLike;
+}
+
+const DEFAULT_VCPU = 4;
+const DEFAULT_MEMORY_GIB = 16;
+
+/** Result body the handler returns for each action; the workflow accumulates these. */
+interface AccumulatedResult {
+  Plan?: { TotalFrames?: number; DurationMs?: number } | null;
+  Chunks?: Array<{ FramesEncoded?: number; DurationMs?: number } | null> | null;
+  Assemble?: {
+    OutputGcsUri?: string;
+    FileSize?: number;
+    FramesEncoded?: number;
+    DurationMs?: number;
+  } | null;
+}
+
+/** Pull a current progress snapshot for one render. */
+// fallow-ignore-next-line complexity
+export async function getRenderProgress(opts: GetRenderProgressOptions): Promise<RenderProgress> {
+  if (!opts.executionName) {
+    throw new Error("[getRenderProgress] executionName is required");
+  }
+  const executions = opts.executions ?? (await defaultExecutionsClient());
+  const vcpu = opts.vcpu ?? DEFAULT_VCPU;
+  const memoryGib = opts.memoryGib ?? DEFAULT_MEMORY_GIB;
+
+  const [execution] = await executions.getExecution({ name: opts.executionName });
+  const status = mapState(execution.state);
+  const startedAt = toIso(execution.startTime) ?? new Date(0).toISOString();
+  const endedAt = toIso(execution.endTime);
+
+  const errors: RenderError[] = [];
+  if (execution.error) {
+    errors.push({
+      state: execution.error.context ?? "<execution>",
+      error: extractErrorName(execution.error.payload) ?? "ExecutionError",
+      cause: execution.error.payload ?? "",
+    });
+  }
+
+  // Default snapshot: running / unknown — no frame or cost data until the
+  // accumulated result is available on success.
+  if (status !== "succeeded") {
+    return {
+      status,
+      overallProgress: 0,
+      framesRendered: 0,
+      totalFrames: null,
+      invocationsObserved: 0,
+      costs: computeRenderCost([], 0),
+      outputFile: null,
+      errors,
+      fatalErrorEncountered: status === "failed" || status === "cancelled",
+      startedAt,
+      endedAt,
+    };
+  }
+
+  const acc = parseAccumulated(execution.result);
+  const chunks = acc.Chunks?.filter((c): c is NonNullable<typeof c> => c != null) ?? [];
+  const framesRendered = chunks.reduce((sum, c) => sum + (c.FramesEncoded ?? 0), 0);
+  const totalFrames = typeof acc.Plan?.TotalFrames === "number" ? acc.Plan.TotalFrames : null;
+
+  const invocations: BilledCloudRunInvocation[] = [];
+  const pushInv = (durationMs: number | undefined): void => {
+    invocations.push({
+      durationMs: typeof durationMs === "number" ? durationMs : 0,
+      vcpu,
+      memoryGib,
+      estimated: typeof durationMs !== "number",
+    });
+  };
+  if (acc.Plan) pushInv(acc.Plan.DurationMs);
+  for (const c of chunks) pushInv(c.DurationMs);
+  if (acc.Assemble) pushInv(acc.Assemble.DurationMs);
+
+  // Workflow step count: Plan + N chunks + Assemble + a small constant of
+  // control steps (BuildChunkList, AssertChunkCount, the map scaffold).
+  const workflowSteps = invocations.length + 4;
+  const costs = computeRenderCost(invocations, workflowSteps);
+
+  const outputGcsUri = acc.Assemble?.OutputGcsUri;
+  const outputFile = outputGcsUri
+    ? {
+        gcsUri: outputGcsUri,
+        bytes: typeof acc.Assemble?.FileSize === "number" ? acc.Assemble.FileSize : null,
+      }
+    : null;
+
+  return {
+    status,
+    overallProgress: 1,
+    framesRendered,
+    totalFrames,
+    invocationsObserved: invocations.length,
+    costs,
+    outputFile,
+    errors,
+    fatalErrorEncountered: false,
+    startedAt,
+    endedAt,
+  };
+}
+
+// fallow-ignore-next-line complexity
+function mapState(state: string | null | undefined): RenderStatus {
+  switch (state) {
+    case "ACTIVE":
+    case "QUEUED":
+      return "running";
+    case "SUCCEEDED":
+      return "succeeded";
+    case "FAILED":
+    case "UNAVAILABLE":
+      return "failed";
+    case "CANCELLED":
+      return "cancelled";
+    default:
+      return "unknown";
+  }
+}
+
+// fallow-ignore-next-line complexity
+function parseAccumulated(result: string | null | undefined): AccumulatedResult {
+  if (!result) return {};
+  try {
+    const parsed = JSON.parse(result) as unknown;
+    if (parsed && typeof parsed === "object") return parsed as AccumulatedResult;
+  } catch {
+    // Non-JSON result — treat as empty so cost/frames degrade to zero
+    // rather than throwing on a snapshot read.
+  }
+  return {};
+}
+
+/**
+ * Best-effort pull of the handler's error name out of a Workflows failure
+ * payload. On an http step failure, Workflows wraps the response as
+ * `{ code, message, body, ... }` where `body` is the handler's JSON
+ * `{ error, message }`. We dig out `error` (the typed name like
+ * `PLAN_HASH_MISMATCH`) so triage sees the real cause, not a generic label.
+ * Returns undefined for any shape we don't recognise — never throws.
+ */
+// fallow-ignore-next-line complexity
+function extractErrorName(payload: string | null | undefined): string | undefined {
+  if (!payload) return undefined;
+  try {
+    const outer = JSON.parse(payload) as { error?: unknown; body?: unknown };
+    if (typeof outer.error === "string") return outer.error;
+    if (typeof outer.body === "string") {
+      const inner = JSON.parse(outer.body) as { error?: unknown };
+      if (typeof inner.error === "string") return inner.error;
+    } else if (outer.body && typeof outer.body === "object") {
+      const inner = outer.body as { error?: unknown };
+      if (typeof inner.error === "string") return inner.error;
+    }
+  } catch {
+    // Non-JSON / unexpected shape — fall through to the generic label.
+  }
+  return undefined;
+}
+
+// fallow-ignore-next-line complexity
+function toIso(ts: ProtoTimestamp | string | null | undefined): string | null {
+  if (ts == null) return null;
+  if (typeof ts === "string") return ts;
+  const seconds = ts.seconds == null ? null : Number(ts.seconds);
+  if (seconds == null || !Number.isFinite(seconds)) return null;
+  const ms = seconds * 1000 + (ts.nanos ?? 0) / 1e6;
+  return new Date(ms).toISOString();
+}
+
+async function defaultExecutionsClient(): Promise<ExecutionsGetClientLike> {
+  const mod = await import("@google-cloud/workflows");
+  const client = new mod.ExecutionsClient();
+  return client as unknown as ExecutionsGetClientLike;
+}

packages/gcp-cloud-run/src/sdk/index.ts

@@ -0,0 +1,40 @@
+/**
+ * SDK subpath export — `@hyperframes/gcp-cloud-run/sdk`.
+ *
+ * Pulled into its own subpath so consumers that only drive renders (CLI, CI
+ * scripts, adopter tooling) don't pay the cost of importing `./server.js`,
+ * which transitively pulls `puppeteer-core` into the module graph. The SDK
+ * files here are GCS + Workflows clients only — safe to load in any Node
+ * environment.
+ */
+
+export { deploySite, type DeploySiteOptions, type SiteHandle } from "./deploySite.js";
+export {
+  type ExecutionsClientLike,
+  renderToCloudRun,
+  type RenderHandle,
+  type RenderToCloudRunOptions,
+} from "./renderToCloudRun.js";
+export {
+  type ExecutionRecord,
+  type ExecutionsGetClientLike,
+  getRenderProgress,
+  type GetRenderProgressOptions,
+  type RenderError,
+  type RenderProgress,
+  type RenderStatus,
+} from "./getRenderProgress.js";
+export {
+  type BilledCloudRunInvocation,
+  computeRenderCost,
+  type RenderCost,
+} from "./costAccounting.js";
+export {
+  InvalidConfigError,
+  MAX_WORKFLOWS_INPUT_BYTES,
+  validateDistributedRenderConfig,
+  validateVariablesPayload,
+  validateWorkflowsInputSize,
+} from "./validateConfig.js";
+export type { SerializableDistributedRenderConfig } from "../events.js";
+export type { DistributedFormat } from "../formatExtension.js";

packages/gcp-cloud-run/src/sdk/renderToCloudRun.test.ts

@@ -0,0 +1,127 @@
+/**
+ * `renderToCloudRun` unit tests — argument assembly, required-field
+ * validation, and the CreateExecution call over a fake ExecutionsClient.
+ */
+
+import { describe, expect, it } from "bun:test";
+import type { SerializableDistributedRenderConfig } from "../events.js";
+import { type ExecutionsClientLike, renderToCloudRun } from "./renderToCloudRun.js";
+import type { SiteHandle } from "./deploySite.js";
+
+const config = {
+  fps: 30,
+  width: 1920,
+  height: 1080,
+  format: "mp4",
+} as SerializableDistributedRenderConfig;
+
+const site: SiteHandle = {
+  siteId: "abc",
+  bucketName: "b",
+  projectGcsUri: "gs://b/sites/abc/project.tar.gz",
+  bytes: 100,
+  uploadedAt: "2026-06-06T00:00:00Z",
+  uploaded: true,
+};
+
+class FakeExecutions implements ExecutionsClientLike {
+  lastArgument: string | null = null;
+  lastParent: string | null = null;
+
+  workflowPath(project: string, location: string, workflow: string): string {
+    return `projects/${project}/locations/${location}/workflows/${workflow}`;
+  }
+
+  async createExecution(req: {
+    parent: string;
+    execution: { argument: string };
+  }): Promise<[{ name?: string | null; state?: string | null }]> {
+    this.lastParent = req.parent;
+    this.lastArgument = req.execution.argument;
+    return [{ name: `${req.parent}/executions/exec-123`, state: "ACTIVE" }];
+  }
+}
+
+function opts(executions: ExecutionsClientLike) {
+  return {
+    siteHandle: site,
+    config,
+    bucketName: "b",
+    projectId: "proj",
+    location: "us-central1",
+    workflowId: "hyperframes-render",
+    serviceUrl: "https://render-abc.run.app",
+    renderId: "hf-render-fixed",
+    executions,
+  };
+}
+
+describe("renderToCloudRun", () => {
+  it("starts an execution and returns a handle", async () => {
+    const fake = new FakeExecutions();
+    const handle = await renderToCloudRun(opts(fake));
+    expect(handle.renderId).toBe("hf-render-fixed");
+    expect(handle.executionName).toBe(
+      "projects/proj/locations/us-central1/workflows/hyperframes-render/executions/exec-123",
+    );
+    expect(handle.outputGcsUri).toBe("gs://b/renders/hf-render-fixed/output.mp4");
+    expect(handle.projectGcsUri).toBe("gs://b/sites/abc/project.tar.gz");
+  });
+
+  it("builds the workflow argument the YAML expects", async () => {
+    const fake = new FakeExecutions();
+    await renderToCloudRun(opts(fake));
+    const arg = JSON.parse(fake.lastArgument ?? "{}");
+    expect(arg.RenderId).toBe("hf-render-fixed");
+    expect(arg.ProjectGcsUri).toBe("gs://b/sites/abc/project.tar.gz");
+    expect(arg.PlanOutputGcsPrefix).toBe("gs://b/renders/hf-render-fixed/");
+    expect(arg.OutputGcsUri).toBe("gs://b/renders/hf-render-fixed/output.mp4");
+    expect(arg.ServiceUrl).toBe("https://render-abc.run.app");
+    expect(arg.Config.format).toBe("mp4");
+    expect(fake.lastParent).toBe(
+      "projects/proj/locations/us-central1/workflows/hyperframes-render",
+    );
+  });
+
+  it("derives the output extension from the format", async () => {
+    const fake = new FakeExecutions();
+    const handle = await renderToCloudRun({
+      ...opts(fake),
+      config: { ...config, format: "webm" } as SerializableDistributedRenderConfig,
+    });
+    expect(handle.outputGcsUri).toBe("gs://b/renders/hf-render-fixed/output.webm");
+  });
+
+  it("requires serviceUrl", async () => {
+    const fake = new FakeExecutions();
+    await expect(renderToCloudRun({ ...opts(fake), serviceUrl: "" })).rejects.toThrow(
+      /serviceUrl is required/,
+    );
+  });
+
+  it("requires a siteHandle or projectDir", async () => {
+    const fake = new FakeExecutions();
+    const { siteHandle, ...rest } = opts(fake);
+    void siteHandle;
+    await expect(renderToCloudRun(rest)).rejects.toThrow(/siteHandle or projectDir/);
+  });
+
+  it("validates the config before any GCP call", async () => {
+    const fake = new FakeExecutions();
+    await expect(
+      renderToCloudRun({ ...opts(fake), config: { ...config, fps: 25 } as never }),
+    ).rejects.toThrow(/config\.fps/);
+    expect(fake.lastArgument).toBeNull();
+  });
+
+  it("rejects a renderId that could escape the GCS key prefix", async () => {
+    const fake = new FakeExecutions();
+    await expect(renderToCloudRun({ ...opts(fake), renderId: "../escape" })).rejects.toThrow(
+      /renderId must match/,
+    );
+    await expect(renderToCloudRun({ ...opts(fake), renderId: "has/slash" })).rejects.toThrow(
+      /renderId must match/,
+    );
+    expect(fake.lastArgument).toBeNull();
+  });
+});

packages/gcp-cloud-run/src/sdk/renderToCloudRun.ts

@@ -0,0 +1,188 @@
+/**
+ * `renderToCloudRun` — start a distributed render against an already-deployed
+ * Cloud Run service + Cloud Workflows definition and return a handle the
+ * caller can poll with {@link getRenderProgress}.
+ *
+ * The function does *not* wait for the render to finish. Cloud Workflows
+ * executions can run for hours; blocking the caller's process on the
+ * execution is the wrong default. The returned `RenderHandle` carries
+ * everything the progress / cost / download paths need.
+ *
+ * Wire order:
+ *   1. Validate config (typed throw before any GCP call).
+ *   2. `deploySite` if no `siteHandle` was provided.
+ *   3. `CreateExecution` against the workflow with the argument shape the
+ *      `packages/gcp-cloud-run/terraform/workflow.yaml` definition expects.
+ *   4. Return handle. The GCS `outputKey` is deterministic from the
+ *      client-generated `renderId` so the caller can predict the final
+ *      object URL before the (server-assigned) execution id exists.
+ *
+ * Unlike Step Functions, Cloud Workflows assigns the execution id
+ * server-side, so we cannot use it as the GCS prefix. We mint a `renderId`
+ * (uuid) client-side, use it for every GCS path, and pass it into the
+ * workflow argument; the server-assigned execution resource name is tracked
+ * separately for polling.
+ */
+
+import { randomUUID } from "node:crypto";
+import type { Storage } from "@google-cloud/storage";
+import type { SerializableDistributedRenderConfig } from "../events.js";
+import { formatExtension } from "../formatExtension.js";
+import { formatGcsUri } from "../gcsTransport.js";
+import { deploySite, type SiteHandle } from "./deploySite.js";
+import { validateDistributedRenderConfig, validateWorkflowsInputSize } from "./validateConfig.js";
+
+/**
+ * Minimal surface of `@google-cloud/workflows`' `ExecutionsClient` that
+ * this module needs. The real client satisfies this; tests inject a double.
+ */
+export interface ExecutionsClientLike {
+  workflowPath(project: string, location: string, workflow: string): string;
+  createExecution(req: {
+    parent: string;
+    execution: { argument: string };
+  }): Promise<[{ name?: string | null; state?: string | null }, ...unknown[]]>;
+}
+
+/** Options for {@link renderToCloudRun}. */
+export interface RenderToCloudRunOptions {
+  /** Local project directory. Required when `siteHandle` is not supplied. */
+  projectDir?: string;
+  /** Re-use an existing `deploySite` upload (skips tar+GCS upload). */
+  siteHandle?: SiteHandle;
+  /** Validated `SerializableDistributedRenderConfig` (no logger / abortSignal). */
+  config: SerializableDistributedRenderConfig;
+  /** GCS bucket from the Terraform output (`render_bucket_name`). */
+  bucketName: string;
+  /** GCP project id hosting the workflow. */
+  projectId: string;
+  /** Workflow location, e.g. `us-central1`. */
+  location: string;
+  /** Workflow id from the Terraform output (`workflow_name`). */
+  workflowId: string;
+  /**
+   * HTTPS URL of the deployed Cloud Run render service (Terraform output
+   * `service_url`). The workflow POSTs every step (plan / renderChunk /
+   * assemble) to this URL; passed as an execution argument so the workflow
+   * definition stays free of hard-coded URLs.
+   */
+  serviceUrl: string;
+  /**
+   * Final output GCS key. Defaults to `renders/<renderId>/output.<ext>`
+   * where `<ext>` is derived from `config.format`.
+   */
+  outputKey?: string;
+  /**
+   * Client-generated render id. Defaults to `hf-render-<uuid>`. Used as the
+   * GCS key prefix and echoed into the workflow argument; not the same as
+   * the server-assigned execution id.
+   */
+  renderId?: string;
+  /** Test injection seam — production callers leave unset. */
+  executions?: ExecutionsClientLike;
+  /** Test injection seam — propagated to `deploySite` when applicable. */
+  storage?: Storage;
+}
+
+/** Stable identifier + every URL/name the caller needs to follow the render. */
+export interface RenderHandle {
+  /** Client-generated render id; the GCS prefix everything lands under. */
+  renderId: string;
+  /** Server-assigned execution resource name; pass to {@link getRenderProgress}. */
+  executionName: string;
+  bucketName: string;
+  workflowId: string;
+  outputGcsUri: string;
+  projectGcsUri: string;
+  startedAt: string;
+}
+
+// fallow-ignore-next-line complexity
+export async function renderToCloudRun(opts: RenderToCloudRunOptions): Promise<RenderHandle> {
+  validateDistributedRenderConfig(opts.config);
+
+  if (!opts.bucketName) throw new Error("[renderToCloudRun] bucketName is required");
+  if (!opts.projectId) throw new Error("[renderToCloudRun] projectId is required");
+  if (!opts.location) throw new Error("[renderToCloudRun] location is required");
+  if (!opts.workflowId) throw new Error("[renderToCloudRun] workflowId is required");
+  if (!opts.serviceUrl) throw new Error("[renderToCloudRun] serviceUrl is required");
+  if (!opts.siteHandle && !opts.projectDir) {
+    throw new Error("[renderToCloudRun] either siteHandle or projectDir must be supplied");
+  }
+
+  const renderId = opts.renderId ?? `hf-render-${randomUUID()}`;
+  // `renderId` is interpolated directly into GCS object keys
+  // (`renders/<renderId>/…`). Reject anything that could escape that prefix
+  // or build a malformed key — `..`, slashes, or other path metacharacters —
+  // so a caller-supplied id can't collide with or overwrite another render's
+  // artifacts elsewhere in the bucket.
+  if (!/^[A-Za-z0-9._-]+$/.test(renderId) || renderId.includes("..")) {
+    throw new Error(
+      `[renderToCloudRun] renderId must match [A-Za-z0-9._-]+ and not contain "..": ${JSON.stringify(renderId)}`,
+    );
+  }
+  const ext = formatExtension(opts.config.format);
+  const outputKey = opts.outputKey ?? `renders/${renderId}/output${ext}`;
+  const planOutputGcsPrefix = formatGcsUri({
+    bucket: opts.bucketName,
+    key: `renders/${renderId}/`,
+  });
+  const outputGcsUri = formatGcsUri({ bucket: opts.bucketName, key: outputKey });
+
+  const site =
+    opts.siteHandle ??
+    (await deploySite({
+      projectDir: opts.projectDir as string,
+      bucketName: opts.bucketName,
+      storage: opts.storage,
+    }));
+
+  const argument = {
+    RenderId: renderId,
+    ProjectGcsUri: site.projectGcsUri,
+    PlanOutputGcsPrefix: planOutputGcsPrefix,
+    OutputGcsUri: outputGcsUri,
+    ServiceUrl: opts.serviceUrl,
+    Config: opts.config,
+  };
+
+  // Reject oversize input client-side. Cloud Workflows caps the execution
+  // argument at 512 KiB; without this check, input bloat (typically from
+  // `config.variables` containing inlined media) surfaces as an opaque
+  // server-side error after the execution starts, far from the caller's
+  // stack frame.
+  validateWorkflowsInputSize(argument);
+
+  const executions = opts.executions ?? (await defaultExecutionsClient());
+  const parent = executions.workflowPath(opts.projectId, opts.location, opts.workflowId);
+  const startedAt = new Date().toISOString();
+  const [execution] = await executions.createExecution({
+    parent,
+    execution: { argument: JSON.stringify(argument) },
+  });
+
+  if (!execution.name) {
+    throw new Error("[renderToCloudRun] CreateExecution returned no execution name");
+  }
+
+  return {
+    renderId,
+    executionName: execution.name,
+    bucketName: opts.bucketName,
+    workflowId: opts.workflowId,
+    outputGcsUri,
+    projectGcsUri: site.projectGcsUri,
+    startedAt,
+  };
+}
+
+/**
+ * Lazily import the real `@google-cloud/workflows` ExecutionsClient. Dynamic
+ * so SDK consumers that only call `validateDistributedRenderConfig` (or
+ * inject their own client) don't pay the import cost.
+ */
+async function defaultExecutionsClient(): Promise<ExecutionsClientLike> {
+  const mod = await import("@google-cloud/workflows");
+  const client = new mod.ExecutionsClient();
+  return client as unknown as ExecutionsClientLike;
+}

packages/gcp-cloud-run/src/sdk/validateConfig.test.ts

@@ -0,0 +1,119 @@
+/**
+ * `validateDistributedRenderConfig` + `validateWorkflowsInputSize` unit
+ * tests. Pins the shape rejections the SDK surfaces synchronously before a
+ * Cloud Workflows execution starts.
+ */
+
+import { describe, expect, it } from "bun:test";
+import type { SerializableDistributedRenderConfig } from "../events.js";
+import {
+  InvalidConfigError,
+  MAX_WORKFLOWS_INPUT_BYTES,
+  validateDistributedRenderConfig,
+  validateVariablesPayload,
+  validateWorkflowsInputSize,
+} from "./validateConfig.js";
+
+function base(): SerializableDistributedRenderConfig {
+  return {
+    fps: 30,
+    width: 1920,
+    height: 1080,
+    format: "mp4",
+  } as SerializableDistributedRenderConfig;
+}
+
+describe("validateDistributedRenderConfig", () => {
+  it("accepts a minimal valid config", () => {
+    expect(validateDistributedRenderConfig(base())).toBeDefined();
+  });
+
+  it("rejects a bad fps", () => {
+    expect(() => validateDistributedRenderConfig({ ...base(), fps: 25 } as never)).toThrow(
+      /config\.fps/,
+    );
+  });
+
+  it("rejects odd dimensions (yuv420p)", () => {
+    expect(() => validateDistributedRenderConfig({ ...base(), width: 1921 })).toThrow(/even/);
+  });
+
+  it("rejects an out-of-range dimension", () => {
+    expect(() => validateDistributedRenderConfig({ ...base(), height: 8 })).toThrow(/\[16, 7680\]/);
+  });
+
+  it("rejects an unknown format", () => {
+    expect(() => validateDistributedRenderConfig({ ...base(), format: "gif" as never })).toThrow(
+      /config\.format/,
+    );
+  });
+
+  it("rejects codec with a non-mp4 format", () => {
+    expect(() =>
+      validateDistributedRenderConfig({ ...base(), format: "webm", codec: "h264" } as never),
+    ).toThrow(/only valid with format="mp4"/);
+  });
+
+  it("rejects crf + bitrate together", () => {
+    expect(() =>
+      validateDistributedRenderConfig({ ...base(), crf: 20, bitrate: "10M" } as never),
+    ).toThrow(/mutually exclusive/);
+  });
+
+  it("rejects force-hdr", () => {
+    expect(() =>
+      validateDistributedRenderConfig({ ...base(), hdrMode: "force-hdr" as never }),
+    ).toThrow(/force-sdr/);
+  });
+
+  it("rejects an over-cap chunkSize", () => {
+    expect(() => validateDistributedRenderConfig({ ...base(), chunkSize: 5000 } as never)).toThrow(
+      /<= 3600/,
+    );
+  });
+
+  it("throws InvalidConfigError with a field pointer", () => {
+    try {
+      validateDistributedRenderConfig({ ...base(), fps: 1 } as never);
+      throw new Error("should have thrown");
+    } catch (err) {
+      expect(err).toBeInstanceOf(InvalidConfigError);
+      expect((err as InvalidConfigError).field).toBe("config.fps");
+    }
+  });
+});
+
+describe("validateVariablesPayload", () => {
+  it("accepts a plain JSON object", () => {
+    expect(() =>
+      validateVariablesPayload({ title: "Hi", count: 3, nested: { ok: true } }),
+    ).not.toThrow();
+  });
+
+  it("rejects undefined leaves", () => {
+    expect(() => validateVariablesPayload({ a: undefined })).toThrow(/undefined leaves/);
+  });
+
+  it("rejects a top-level array", () => {
+    expect(() => validateVariablesPayload([1, 2])).toThrow(/plain JSON object/);
+  });
+
+  it("rejects NaN", () => {
+    expect(() => validateVariablesPayload({ x: NaN })).toThrow(/non-finite/);
+  });
+
+  it("rejects a Date (non-plain object)", () => {
+    expect(() => validateVariablesPayload({ when: new Date(0) })).toThrow(/non-plain objects/);
+  });
+});
+
+describe("validateWorkflowsInputSize", () => {
+  it("accepts a small payload", () => {
+    expect(() => validateWorkflowsInputSize({ a: "b" })).not.toThrow();
+  });
+
+  it("rejects a payload over the 512 KiB cap", () => {
+    const big = { blob: "x".repeat(MAX_WORKFLOWS_INPUT_BYTES + 1) };
+    expect(() => validateWorkflowsInputSize(big)).toThrow(/512 KiB/);
+  });
+});

packages/gcp-cloud-run/src/sdk/validateConfig.ts

@@ -0,0 +1,77 @@
+/**
+ * Client-side validation for the Cloud Run adapter.
+ *
+ * The cloud-agnostic config-shape validation (`validateDistributedRenderConfig`,
+ * `validateVariablesPayload`, `InvalidConfigError`) lives in
+ * `@hyperframes/producer/distributed` and is shared with the other adapters.
+ * This module re-exports those and adds the one piece that is specific to
+ * Cloud Workflows: the 512 KiB execution-argument size cap.
+ */
+
+import { InvalidConfigError } from "@hyperframes/producer/distributed";
+
+export {
+  InvalidConfigError,
+  validateDistributedRenderConfig,
+  validateVariablesPayload,
+} from "@hyperframes/producer/distributed";
+
+/**
+ * Hard cap on Cloud Workflows execution arguments — 512 KiB per the Workflows
+ * quotas page (maximum size of arguments passed when an execution starts).
+ * The cap is on the entire serialized argument, not just the variables,
+ * because users hit it at the wire boundary regardless of which field caused
+ * the bloat.
+ *
+ * Specific to Cloud Workflows. Other runtimes (Lambda + Step Functions,
+ * Temporal) have different caps; don't reuse this constant for those without
+ * confirming the limit.
+ */
+export const MAX_WORKFLOWS_INPUT_BYTES = 512 * 1024;
+
+/** Pointer to the docs section that explains the URL-your-assets convention. */
+const LARGE_VARIABLES_DOCS_URL =
+  "https://hyperframes.heygen.com/deploy/templates-on-lambda#working-with-large-variables";
+
+/**
+ * Validate that the serialized Cloud Workflows execution argument fits inside
+ * the 512 KiB cap. Measured in UTF-8 bytes (the format the API uses on the
+ * wire) — JS strings count UTF-16 code units, which under-reports for any
+ * multi-byte character.
+ *
+ * Throws {@link InvalidConfigError} with a clear message naming the actual
+ * byte count, the cap, and a pointer to the "working with large variables"
+ * docs section, so users hit the limit at the SDK boundary with actionable
+ * guidance instead of as an opaque argument-too-large error after the
+ * execution starts.
+ */
+// fallow-ignore-next-line complexity
+export function validateWorkflowsInputSize(input: unknown): void {
+  let serialized: string | undefined;
+  try {
+    serialized = JSON.stringify(input);
+  } catch (err) {
+    throw new InvalidConfigError(
+      "config",
+      `Cloud Workflows execution argument is not JSON-serializable: ${err instanceof Error ? err.message : String(err)}`,
+    );
+  }
+  if (serialized === undefined) {
+    throw new InvalidConfigError(
+      "config",
+      "Cloud Workflows execution argument is not JSON-serializable (JSON.stringify returned undefined). " +
+        "Check that all fields, including config.variables, are plain JSON values.",
+    );
+  }
+  const byteLength = Buffer.byteLength(serialized, "utf8");
+  if (byteLength > MAX_WORKFLOWS_INPUT_BYTES) {
+    throw new InvalidConfigError(
+      "config",
+      `Cloud Workflows execution argument is ${byteLength} bytes, which exceeds the ` +
+        `${MAX_WORKFLOWS_INPUT_BYTES}-byte (512 KiB) limit. Variables are for typed data ` +
+        `(strings, numbers, structured records); media assets (images, audio, video) should ` +
+        `be passed as URL references the composition resolves at render time, not inlined as ` +
+        `base64. See ${LARGE_VARIABLES_DOCS_URL} for the URL-your-assets convention.`,
+    );
+  }
+}

packages/gcp-cloud-run/src/server.test.ts

@@ -0,0 +1,332 @@
+/**
+ * Handler dispatch + HTTP-shell unit tests.
+ *
+ * Asserts that:
+ *   - `dispatch` routes Action="plan"/"renderChunk"/"assemble" to the
+ *     matching OSS primitive and plumbs GCS download/upload around it.
+ *   - It unwraps `{ Payload }` / `{ Input }` envelopes and rejects unknown
+ *     actions.
+ *   - The handler-boundary guards fire: plan-hash mismatch + bucket
+ *     allowlist throw the typed, non-retryable errors.
+ *   - `createApp` maps non-retryable errors → 400 and retryable → 500.
+ *
+ * The real OSS primitives are NOT exercised — they have their own coverage
+ * in `packages/producer`. This file pins the adapter glue's contract.
+ */
+
+import { afterEach, describe, expect, it } from "bun:test";
+import { mkdirSync, mkdtempSync, rmSync, writeFileSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import type { AssembleResult, ChunkResult, PlanResult } from "@hyperframes/producer/distributed";
+import { asStorage, FakeGcs } from "./__fixtures__/fakeGcs.js";
+import type { AssembleEvent, CloudRunEvent, PlanEvent, RenderChunkEvent } from "./events.js";
+import { createApp, dispatch, type HandlerDeps, unwrapEvent } from "./server.js";
+import { tarDirectory } from "./gcsTransport.js";
+
+const tmpDirs: string[] = [];
+function mkTmp(prefix: string): string {
+  const dir = mkdtempSync(join(tmpdir(), prefix));
+  tmpDirs.push(dir);
+  return dir;
+}
+afterEach(() => {
+  for (const d of tmpDirs.splice(0)) rmSync(d, { recursive: true, force: true });
+});
+
+const PLAN_HASH = "abc123planhash";
+
+/** Build a real project tarball, seed it into the fake, return its URI. */
+async function seedProjectTar(gcs: FakeGcs, uri: string): Promise<void> {
+  const src = mkTmp("hf-proj-");
+  writeFileSync(join(src, "index.html"), "<html></html>");
+  const tarPath = join(mkTmp("hf-proj-tar-"), "project.tar.gz");
+  await tarDirectory(src, tarPath);
+  gcs.seedFromFile(uri, tarPath);
+}
+
+/** Build a real plan tarball containing plan.json, seed it, return its URI. */
+async function seedPlanTar(gcs: FakeGcs, uri: string, planHash: string): Promise<void> {
+  const planDir = mkTmp("hf-plan-");
+  writeFileSync(join(planDir, "plan.json"), JSON.stringify({ planHash }));
+  const tarPath = join(mkTmp("hf-plan-tar-"), "plan.tar.gz");
+  await tarDirectory(planDir, tarPath);
+  gcs.seedFromFile(uri, tarPath);
+}
+
+const planResult: PlanResult = {
+  planDir: "(set at call time)",
+  planHash: PLAN_HASH,
+  chunkCount: 3,
+  totalFrames: 90,
+  fps: 30,
+  width: 1920,
+  height: 1080,
+  format: "mp4",
+  ffmpegVersion: "ffmpeg version 6.1.1",
+  producerVersion: "0.6.79",
+};
+
+function depsWith(
+  gcs: FakeGcs,
+  overrides: Partial<NonNullable<HandlerDeps["primitives"]>> = {},
+): HandlerDeps {
+  const plan = async (_projectDir: string, _config: unknown, planDir: string) => {
+    mkdirSync(planDir, { recursive: true });
+    writeFileSync(join(planDir, "plan.json"), JSON.stringify({ planHash: PLAN_HASH }));
+    return planResult;
+  };
+  const renderChunk = async (_planDir: string, chunkIndex: number, outputBase: string) => {
+    writeFileSync(outputBase, Buffer.from(`chunk-${chunkIndex}`));
+    return {
+      outputPath: outputBase,
+      outputKind: "file",
+      framesEncoded: 30,
+      sha256: `sha-${chunkIndex}`,
+    } satisfies ChunkResult;
+  };
+  const assemble = async (
+    _planDir: string,
+    _chunkPaths: string[],
+    _audio: string | null,
+    finalOutput: string,
+  ) => {
+    writeFileSync(finalOutput, Buffer.from("final-output"));
+    return { framesEncoded: 90, fileSize: 12 } satisfies AssembleResult;
+  };
+  return {
+    storage: asStorage(gcs),
+    skipChromeResolution: true,
+    primitives: { plan, renderChunk, assemble, ...overrides } as NonNullable<
+      HandlerDeps["primitives"]
+    >,
+  };
+}
+
+describe("unwrapEvent", () => {
+  const plan: PlanEvent = {
+    Action: "plan",
+    ProjectGcsUri: "gs://b/p.tar.gz",
+    PlanOutputGcsPrefix: "gs://b/out/",
+    Config: { fps: 30, width: 1920, height: 1080, format: "mp4" } as PlanEvent["Config"],
+  };
+
+  it("returns a bare event unchanged", () => {
+    expect(unwrapEvent(plan).Action).toBe("plan");
+  });
+
+  it("unwraps { Payload }", () => {
+    expect(unwrapEvent({ Payload: plan } as CloudRunEvent).Action).toBe("plan");
+  });
+
+  it("unwraps nested { Input: { Payload } }", () => {
+    expect(unwrapEvent({ Input: { Payload: plan } } as CloudRunEvent).Action).toBe("plan");
+  });
+
+  it("throws when no Action is found", () => {
+    expect(() => unwrapEvent({ foo: "bar" } as unknown as CloudRunEvent)).toThrow(
+      /no recognised Action/,
+    );
+  });
+});
+
+describe("dispatch", () => {
+  it("routes plan, uploads the plan tarball", async () => {
+    const gcs = new FakeGcs();
+    await seedProjectTar(gcs, "gs://b/sites/x/project.tar.gz");
+    const event: PlanEvent = {
+      Action: "plan",
+      ProjectGcsUri: "gs://b/sites/x/project.tar.gz",
+      PlanOutputGcsPrefix: "gs://b/renders/r1/",
+      Config: { fps: 30, width: 1920, height: 1080, format: "mp4" } as PlanEvent["Config"],
+    };
+    const res = await dispatch(event, depsWith(gcs));
+    expect(res.Action).toBe("plan");
+    if (res.Action !== "plan") throw new Error("unreachable");
+    expect(res.PlanHash).toBe(PLAN_HASH);
+    expect(res.ChunkCount).toBe(3);
+    expect(res.PlanGcsUri).toBe("gs://b/renders/r1/plan.tar.gz");
+    expect(gcs.objects.has("gs://b/renders/r1/plan.tar.gz")).toBe(true);
+  });
+
+  it("routes renderChunk and uploads the chunk", async () => {
+    const gcs = new FakeGcs();
+    await seedPlanTar(gcs, "gs://b/renders/r1/plan.tar.gz", PLAN_HASH);
+    const event: RenderChunkEvent = {
+      Action: "renderChunk",
+      PlanGcsUri: "gs://b/renders/r1/plan.tar.gz",
+      PlanHash: PLAN_HASH,
+      ChunkIndex: 2,
+      ChunkOutputGcsPrefix: "gs://b/renders/r1/",
+      Format: "mp4",
+    };
+    const res = await dispatch(event, depsWith(gcs));
+    if (res.Action !== "renderChunk") throw new Error("unreachable");
+    expect(res.ChunkIndex).toBe(2);
+    expect(res.ChunkGcsUri).toBe("gs://b/renders/r1/chunks/0002.mp4");
+    expect(gcs.objects.has("gs://b/renders/r1/chunks/0002.mp4")).toBe(true);
+  });
+
+  it("throws PLAN_HASH_MISMATCH when the event hash disagrees", async () => {
+    const gcs = new FakeGcs();
+    await seedPlanTar(gcs, "gs://b/renders/r1/plan.tar.gz", PLAN_HASH);
+    const event: RenderChunkEvent = {
+      Action: "renderChunk",
+      PlanGcsUri: "gs://b/renders/r1/plan.tar.gz",
+      PlanHash: "WRONG_HASH",
+      ChunkIndex: 0,
+      ChunkOutputGcsPrefix: "gs://b/renders/r1/",
+      Format: "mp4",
+    };
+    await expect(dispatch(event, depsWith(gcs))).rejects.toThrow(/PLAN_HASH_MISMATCH/);
+  });
+
+  it("routes assemble and uploads the final output", async () => {
+    const gcs = new FakeGcs();
+    await seedPlanTar(gcs, "gs://b/renders/r1/plan.tar.gz", PLAN_HASH);
+    gcs.seed("gs://b/renders/r1/chunks/0000.mp4", Buffer.from("c0"));
+    gcs.seed("gs://b/renders/r1/chunks/0001.mp4", Buffer.from("c1"));
+    const event: AssembleEvent = {
+      Action: "assemble",
+      PlanGcsUri: "gs://b/renders/r1/plan.tar.gz",
+      ChunkGcsUris: ["gs://b/renders/r1/chunks/0000.mp4", "gs://b/renders/r1/chunks/0001.mp4"],
+      AudioGcsUri: null,
+      OutputGcsUri: "gs://b/renders/r1/output.mp4",
+      Format: "mp4",
+    };
+    const res = await dispatch(event, depsWith(gcs));
+    if (res.Action !== "assemble") throw new Error("unreachable");
+    expect(res.FramesEncoded).toBe(90);
+    expect(gcs.objects.has("gs://b/renders/r1/output.mp4")).toBe(true);
+  });
+
+  it("rejects an unknown action", async () => {
+    const gcs = new FakeGcs();
+    await expect(
+      dispatch({ Action: "nope" } as unknown as CloudRunEvent, depsWith(gcs)),
+    ).rejects.toThrow(/no recognised Action/);
+  });
+});
+
+describe("bucket allowlist guard", () => {
+  it("throws GCS_URI_NOT_ALLOWED for an off-bucket URI", async () => {
+    const gcs = new FakeGcs();
+    const prev = process.env.HYPERFRAMES_RENDER_BUCKET;
+    process.env.HYPERFRAMES_RENDER_BUCKET = "allowed-bucket";
+    try {
+      const event: RenderChunkEvent = {
+        Action: "renderChunk",
+        PlanGcsUri: "gs://evil-bucket/plan.tar.gz",
+        PlanHash: PLAN_HASH,
+        ChunkIndex: 0,
+        ChunkOutputGcsPrefix: "gs://allowed-bucket/renders/r1/",
+        Format: "mp4",
+      };
+      await expect(dispatch(event, depsWith(gcs))).rejects.toThrow(/GCS_URI_NOT_ALLOWED/);
+    } finally {
+      if (prev === undefined) delete process.env.HYPERFRAMES_RENDER_BUCKET;
+      else process.env.HYPERFRAMES_RENDER_BUCKET = prev;
+    }
+  });
+
+  it('treats HYPERFRAMES_RENDER_BUCKET="*" as an explicit opt-out (off-bucket allowed)', async () => {
+    const gcs = new FakeGcs();
+    await seedPlanTar(gcs, "gs://any-bucket/renders/r1/plan.tar.gz", PLAN_HASH);
+    const prev = process.env.HYPERFRAMES_RENDER_BUCKET;
+    process.env.HYPERFRAMES_RENDER_BUCKET = "*";
+    try {
+      const event: RenderChunkEvent = {
+        Action: "renderChunk",
+        PlanGcsUri: "gs://any-bucket/renders/r1/plan.tar.gz",
+        PlanHash: PLAN_HASH,
+        ChunkIndex: 0,
+        ChunkOutputGcsPrefix: "gs://any-bucket/renders/r1/",
+        Format: "mp4",
+      };
+      const res = await dispatch(event, depsWith(gcs));
+      expect(res.Action).toBe("renderChunk");
+    } finally {
+      if (prev === undefined) delete process.env.HYPERFRAMES_RENDER_BUCKET;
+      else process.env.HYPERFRAMES_RENDER_BUCKET = prev;
+    }
+  });
+});
+
+describe("createApp HTTP mapping", () => {
+  it("returns 200 with the result body on success", async () => {
+    const gcs = new FakeGcs();
+    await seedPlanTar(gcs, "gs://b/renders/r1/plan.tar.gz", PLAN_HASH);
+    const app = createApp(depsWith(gcs));
+    const res = await app.request("/", {
+      method: "POST",
+      headers: { "content-type": "application/json" },
+      body: JSON.stringify({
+        Action: "renderChunk",
+        PlanGcsUri: "gs://b/renders/r1/plan.tar.gz",
+        PlanHash: PLAN_HASH,
+        ChunkIndex: 0,
+        ChunkOutputGcsPrefix: "gs://b/renders/r1/",
+        Format: "mp4",
+      }),
+    });
+    expect(res.status).toBe(200);
+    const body = (await res.json()) as { Action: string };
+    expect(body.Action).toBe("renderChunk");
+  });
+
+  it("returns 400 for a non-retryable error (plan-hash mismatch)", async () => {
+    const gcs = new FakeGcs();
+    await seedPlanTar(gcs, "gs://b/renders/r1/plan.tar.gz", PLAN_HASH);
+    const app = createApp(depsWith(gcs));
+    const res = await app.request("/", {
+      method: "POST",
+      headers: { "content-type": "application/json" },
+      body: JSON.stringify({
+        Action: "renderChunk",
+        PlanGcsUri: "gs://b/renders/r1/plan.tar.gz",
+        PlanHash: "WRONG",
+        ChunkIndex: 0,
+        ChunkOutputGcsPrefix: "gs://b/renders/r1/",
+        Format: "mp4",
+      }),
+    });
+    expect(res.status).toBe(400);
+    const body = (await res.json()) as { error: string };
+    expect(body.error).toBe("PLAN_HASH_MISMATCH");
+  });
+
+  it("returns 500 for a retryable/unknown error", async () => {
+    const gcs = new FakeGcs(); // plan tar NOT seeded → download fails (retryable)
+    const app = createApp(depsWith(gcs));
+    const res = await app.request("/", {
+      method: "POST",
+      headers: { "content-type": "application/json" },
+      body: JSON.stringify({
+        Action: "renderChunk",
+        PlanGcsUri: "gs://b/renders/r1/missing.tar.gz",
+        PlanHash: PLAN_HASH,
+        ChunkIndex: 0,
+        ChunkOutputGcsPrefix: "gs://b/renders/r1/",
+        Format: "mp4",
+      }),
+    });
+    expect(res.status).toBe(500);
+  });
+
+  it("returns 400 when the body is not JSON", async () => {
+    const gcs = new FakeGcs();
+    const app = createApp(depsWith(gcs));
+    const res = await app.request("/", {
+      method: "POST",
+      headers: { "content-type": "application/json" },
+      body: "not json{",
+    });
+    expect(res.status).toBe(400);
+  });
+
+  it("healthz returns ok", async () => {
+    const app = createApp(depsWith(new FakeGcs()));
+    const res = await app.request("/healthz");
+    expect(res.status).toBe(200);
+  });
+});

packages/gcp-cloud-run/src/server.ts

@@ -0,0 +1,647 @@
+/**
+ * Cloud Run request handler for HyperFrames distributed rendering.
+ *
+ * One container image, three roles. Cloud Workflows POSTs a JSON body with
+ * an `Action` field; the handler unwraps any `Payload`/`Input` envelope,
+ * primes the runtime (Chrome path), and forwards to the matching OSS
+ * primitive from `@hyperframes/producer/distributed`.
+ *
+ * Everything heavy — capture, encode, audio mix — happens inside the OSS
+ * primitives. The handler is thin glue: parse body → GCS download → call
+ * primitive → GCS upload → return small JSON result.
+ *
+ * `dispatch()` is the testable core (inject `storage` + `primitives`); the
+ * Hono app at the bottom is the HTTP shell the Dockerfile runs. The shape
+ * deliberately tracks `@hyperframes/aws-lambda`'s `handler.ts` so the two
+ * adapters stay easy to diff.
+ */
+
+import { existsSync, mkdirSync, mkdtempSync, readFileSync, rmSync, statSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { basename, extname, join } from "node:path";
+import { fileURLToPath } from "node:url";
+import { serve } from "@hono/node-server";
+import { Storage } from "@google-cloud/storage";
+import { Hono } from "hono";
+import {
+  assemble,
+  type AssembleResult,
+  type ChunkResult,
+  type DistributedRenderConfig,
+  plan,
+  type PlanResult,
+  renderChunk,
+} from "@hyperframes/producer/distributed";
+import { resolveChromeExecutablePath } from "./chromium.js";
+import type {
+  AssembleEvent,
+  AssembleResultBody,
+  CloudRunAction,
+  CloudRunEvent,
+  CloudRunResult,
+  PlanEvent,
+  PlanResultBody,
+  RenderChunkEvent,
+  RenderChunkResultBody,
+} from "./events.js";
+import { type DistributedFormat, formatExtension } from "./formatExtension.js";
+import {
+  downloadGcsObjectToFile,
+  parseGcsUri,
+  tarDirectory,
+  untarDirectory,
+  uploadFileToGcs,
+} from "./gcsTransport.js";
+
+/**
+ * Lazily-constructed Storage client. Cached at module scope so warm
+ * container instances reuse the underlying HTTP keep-alive pool across
+ * requests.
+ */
+let cachedStorage: Storage | null = null;
+function getStorage(): Storage {
+  if (cachedStorage) return cachedStorage;
+  cachedStorage = new Storage();
+  return cachedStorage;
+}
+
+/**
+ * Optional injection points used by the handler's unit tests. Production
+ * callers leave these unset; the real OSS primitives are used. Tests inject
+ * `storage` and `primitives` directly rather than mutating module state.
+ */
+export interface HandlerDeps {
+  storage?: Storage;
+  primitives?: {
+    plan: typeof plan;
+    renderChunk: typeof renderChunk;
+    assemble: typeof assemble;
+  };
+  /** Override the per-request workdir root (defaults to the OS tmpdir). */
+  tmpRoot?: string;
+  /** Skip Chrome resolution (used by dispatch tests that mock renderChunk). */
+  skipChromeResolution?: boolean;
+}
+
+/**
+ * Dispatch a single render request. Cloud Workflows (or a direct caller)
+ * sometimes wraps the body in `{ Payload: ... }` or `{ Input: ... }`; unwrap
+ * until we hit a discriminated event.
+ */
+// fallow-ignore-next-line complexity
+export async function dispatch(event: CloudRunEvent, deps?: HandlerDeps): Promise<CloudRunResult> {
+  const unwrapped = unwrapEvent(event);
+  validateEventGcsUris(unwrapped);
+  logEvent({ event: "handler_start", action: unwrapped.Action, input: summarizeEvent(unwrapped) });
+  try {
+    switch (unwrapped.Action) {
+      case "plan":
+        return await handlePlan(unwrapped, deps);
+      case "renderChunk":
+        return await handleRenderChunk(unwrapped, deps);
+      case "assemble":
+        return await handleAssemble(unwrapped, deps);
+      default: {
+        // Compile-time exhaustiveness: a new CloudRunAction member trips
+        // the `never` assignment before the runtime error is reachable.
+        const _exhaustive: never = unwrapped;
+        throw new Error(
+          `[handler] unknown Action: ${JSON.stringify(
+            (_exhaustive as { Action?: string }).Action,
+          )}. Expected one of "plan", "renderChunk", "assemble".`,
+        );
+      }
+    }
+  } catch (err) {
+    logEvent({
+      event: "handler_error",
+      action: unwrapped.Action,
+      message: err instanceof Error ? err.message : String(err),
+      name: err instanceof Error ? err.name : undefined,
+    });
+    throw err;
+  }
+}
+
+// At most `{Payload: {Input: ...}}` is expected; 4 levels is 2× headroom
+// and prevents infinite loops on malformed input.
+const MAX_ENVELOPE_DEPTH = 4;
+
+// fallow-ignore-next-line complexity
+export function unwrapEvent(event: CloudRunEvent): PlanEvent | RenderChunkEvent | AssembleEvent {
+  let cursor: CloudRunEvent = event;
+  for (let i = 0; i < MAX_ENVELOPE_DEPTH; i++) {
+    if (cursor && typeof cursor === "object") {
+      const obj = cursor as Record<string, unknown>;
+      if (typeof obj.Action === "string" && isCloudRunAction(obj.Action)) {
+        return cursor as PlanEvent | RenderChunkEvent | AssembleEvent;
+      }
+      if ("Payload" in obj) {
+        cursor = obj.Payload as CloudRunEvent;
+        continue;
+      }
+      if ("Input" in obj) {
+        cursor = obj.Input as CloudRunEvent;
+        continue;
+      }
+    }
+    break;
+  }
+  throw new Error(
+    `[handler] body has no recognised Action; unwrapped ${MAX_ENVELOPE_DEPTH} levels of Payload/Input without finding one.`,
+  );
+}
+
+function isCloudRunAction(value: string): value is CloudRunAction {
+  return value === "plan" || value === "renderChunk" || value === "assemble";
+}
+
+/**
+ * Emit a single JSON line to stdout. Cloud Logging ingests each stdout line
+ * as a structured `jsonPayload` entry, so Logs Explorer can filter on
+ * `jsonPayload.event="handler_start"` and project specific fields when
+ * triaging without attaching a debugger.
+ */
+function logEvent(payload: Record<string, unknown>): void {
+  console.log(JSON.stringify(payload));
+}
+
+/**
+ * Compact, non-PII summary of an event for logging. The full body can
+ * include the entire project config; we only emit the routable fields
+ * needed to triage a failure from Cloud Logging.
+ */
+function summarizeEvent(
+  event: PlanEvent | RenderChunkEvent | AssembleEvent,
+): Record<string, unknown> {
+  switch (event.Action) {
+    case "plan":
+      return {
+        projectGcsUri: event.ProjectGcsUri,
+        planOutputGcsPrefix: event.PlanOutputGcsPrefix,
+        format: event.Config.format,
+        fps: event.Config.fps,
+      };
+    case "renderChunk":
+      return {
+        planGcsUri: event.PlanGcsUri,
+        chunkIndex: event.ChunkIndex,
+        format: event.Format,
+      };
+    case "assemble":
+      return {
+        planGcsUri: event.PlanGcsUri,
+        chunkCount: event.ChunkGcsUris.length,
+        hasAudio: event.AudioGcsUri !== null,
+        outputGcsUri: event.OutputGcsUri,
+        format: event.Format,
+      };
+  }
+}
+
+/**
+ * Point the engine at the in-image Chrome binary. The OSS engine resolves
+ * Chrome via `PRODUCER_HEADLESS_SHELL_PATH` first; set it once per instance
+ * before invoking any browser-touching primitive. ffmpeg is on the image's
+ * PATH (apt-installed by the Dockerfile), so nothing to prime there.
+ */
+function primeChrome(deps?: HandlerDeps): void {
+  if (deps?.skipChromeResolution) return;
+  if (process.env.PRODUCER_HEADLESS_SHELL_PATH) return;
+  process.env.PRODUCER_HEADLESS_SHELL_PATH = resolveChromeExecutablePath();
+}
+
+// ── Plan ────────────────────────────────────────────────────────────────────
+
+// fallow-ignore-next-line complexity
+async function handlePlan(event: PlanEvent, deps?: HandlerDeps): Promise<PlanResultBody> {
+  const started = Date.now();
+  const storage = deps?.storage ?? getStorage();
+  const primitive = deps?.primitives?.plan ?? plan;
+
+  // The producer's probe stage launches Chromium whenever the composition
+  // needs a runtime duration probe or has unresolved sub-compositions, so
+  // plan has to resolve Chrome the same way renderChunk does.
+  primeChrome(deps);
+
+  const work = mkdtempSync(join(deps?.tmpRoot ?? tmpdir(), "hf-cr-plan-"));
+  const projectArchive = join(work, "project.tar.gz");
+  const projectDir = join(work, "project");
+  const planDir = join(work, "plan");
+
+  try {
+    await downloadGcsObjectToFile(storage, event.ProjectGcsUri, projectArchive);
+    await untarDirectory(projectArchive, projectDir);
+
+    const config: DistributedRenderConfig = {
+      ...event.Config,
+    };
+    const result: PlanResult = await primitive(projectDir, config, planDir);
+
+    // Upload the planDir as a single tarball. The workflow cannot pass a
+    // directory-shaped artifact between steps; we serialize and rely on the
+    // consumer (renderChunk / assemble) to untar. `audio.aac` lives inside
+    // planDir, so it already rides along in this tarball — every consumer
+    // (including assemble) gets it from the untar. We deliberately do NOT
+    // upload a separate audio object: it would duplicate the bytes on every
+    // plan upload and be re-downloaded + overwritten by assemble. `AudioGcsUri`
+    // stays in the result shape for wire compatibility but is null.
+    const planTar = join(work, "plan.tar.gz");
+    await tarDirectory(planDir, planTar);
+    const planTarUri = `${trimTrailingSlash(event.PlanOutputGcsPrefix)}/plan.tar.gz`;
+    const audioPath = join(planDir, "audio.aac");
+    const hasAudio = existsSync(audioPath) && statSync(audioPath).size > 0;
+    await uploadFileToGcs(storage, planTar, planTarUri, "application/gzip");
+
+    return {
+      Action: "plan",
+      PlanGcsUri: planTarUri,
+      PlanHash: result.planHash,
+      ChunkCount: result.chunkCount,
+      TotalFrames: result.totalFrames,
+      Fps: result.fps,
+      Width: result.width,
+      Height: result.height,
+      Format: result.format,
+      HasAudio: hasAudio,
+      AudioGcsUri: null,
+      FfmpegVersion: result.ffmpegVersion,
+      ProducerVersion: result.producerVersion,
+      DurationMs: Date.now() - started,
+    };
+  } finally {
+    cleanupDir(work);
+  }
+}
+
+// ── RenderChunk ─────────────────────────────────────────────────────────────
+
+// fallow-ignore-next-line complexity
+async function handleRenderChunk(
+  event: RenderChunkEvent,
+  deps?: HandlerDeps,
+): Promise<RenderChunkResultBody> {
+  const started = Date.now();
+  const storage = deps?.storage ?? getStorage();
+  const primitive = deps?.primitives?.renderChunk ?? renderChunk;
+
+  primeChrome(deps);
+
+  const work = mkdtempSync(join(deps?.tmpRoot ?? tmpdir(), "hf-cr-chunk-"));
+  const planTar = join(work, "plan.tar.gz");
+  const planDir = join(work, "plan");
+
+  try {
+    await downloadGcsObjectToFile(storage, event.PlanGcsUri, planTar);
+    await untarDirectory(planTar, planDir);
+
+    // Verify the plan's hash matches what the workflow told us to render.
+    // The producer's renderChunk re-checks internally (defense-in-depth),
+    // but doing it here at the handler boundary lets us fail before paying
+    // the Chrome-launch + render cost on a misrouted chunk. Throws a typed
+    // PLAN_HASH_MISMATCH the workflow can route as non-retryable.
+    verifyPlanHash(planDir, event.PlanHash);
+
+    const chunkOutputBase = join(
+      work,
+      event.Format === "png-sequence"
+        ? `chunk-${pad(event.ChunkIndex)}`
+        : `chunk-${pad(event.ChunkIndex)}${formatExtension(event.Format)}`,
+    );
+
+    const result: ChunkResult = await primitive(planDir, event.ChunkIndex, chunkOutputBase);
+
+    const chunkUri = await uploadChunkOutput(
+      storage,
+      result,
+      event.ChunkOutputGcsPrefix,
+      event.ChunkIndex,
+    );
+
+    return {
+      Action: "renderChunk",
+      ChunkGcsUri: chunkUri,
+      ChunkIndex: event.ChunkIndex,
+      Sha256: result.sha256,
+      FramesEncoded: result.framesEncoded,
+      DurationMs: Date.now() - started,
+    };
+  } finally {
+    cleanupDir(work);
+  }
+}
+
+async function uploadChunkOutput(
+  storage: Storage,
+  result: ChunkResult,
+  prefix: string,
+  chunkIndex: number,
+): Promise<string> {
+  const trimmed = trimTrailingSlash(prefix);
+  if (result.outputKind === "file") {
+    const ext = extname(result.outputPath);
+    const uri = `${trimmed}/chunks/${pad(chunkIndex)}${ext}`;
+    await uploadFileToGcs(storage, result.outputPath, uri);
+    return uri;
+  }
+  // frame-dir: upload as a tarball so a single GCS object represents the
+  // chunk. Assemble's png-sequence path expects a directory per chunk; it
+  // untars on its end.
+  const tarball = `${result.outputPath}.tar.gz`;
+  await tarDirectory(result.outputPath, tarball);
+  const uri = `${trimmed}/chunks/${pad(chunkIndex)}.tar.gz`;
+  await uploadFileToGcs(storage, tarball, uri, "application/gzip");
+  return uri;
+}
+
+// ── Assemble ────────────────────────────────────────────────────────────────
+
+// fallow-ignore-next-line complexity
+async function handleAssemble(
+  event: AssembleEvent,
+  deps?: HandlerDeps,
+): Promise<AssembleResultBody> {
+  const started = Date.now();
+  const storage = deps?.storage ?? getStorage();
+  const primitive = deps?.primitives?.assemble ?? assemble;
+
+  const work = mkdtempSync(join(deps?.tmpRoot ?? tmpdir(), "hf-cr-assemble-"));
+  const planTar = join(work, "plan.tar.gz");
+  const planDir = join(work, "plan");
+
+  try {
+    await downloadGcsObjectToFile(storage, event.PlanGcsUri, planTar);
+    await untarDirectory(planTar, planDir);
+
+    const chunkPaths = await downloadChunkObjects(storage, event.ChunkGcsUris, work, event.Format);
+
+    // Audio rides inside the plan tarball, so it's already on disk after the
+    // untar above — no separate download. Fall back to a supplied AudioGcsUri
+    // only for backward compatibility with an older Plan that uploaded it
+    // standalone.
+    let audioPath: string | null = null;
+    const planAudio = join(planDir, "audio.aac");
+    if (existsSync(planAudio) && statSync(planAudio).size > 0) {
+      audioPath = planAudio;
+    } else if (event.AudioGcsUri) {
+      audioPath = planAudio;
+      await downloadGcsObjectToFile(storage, event.AudioGcsUri, audioPath);
+    }
+
+    const finalOutput =
+      event.Format === "png-sequence"
+        ? join(work, "output-frames")
+        : join(work, `output${formatExtension(event.Format)}`);
+
+    const result: AssembleResult = await primitive(planDir, chunkPaths, audioPath, finalOutput, {
+      cfr: event.Cfr === true,
+    });
+
+    if (event.Format === "png-sequence") {
+      const tarball = `${finalOutput}.tar.gz`;
+      await tarDirectory(finalOutput, tarball);
+      await uploadFileToGcs(storage, tarball, event.OutputGcsUri, "application/gzip");
+    } else {
+      await uploadFileToGcs(storage, finalOutput, event.OutputGcsUri);
+    }
+
+    return {
+      Action: "assemble",
+      OutputGcsUri: event.OutputGcsUri,
+      FramesEncoded: result.framesEncoded,
+      FileSize: result.fileSize,
+      DurationMs: Date.now() - started,
+    };
+  } finally {
+    cleanupDir(work);
+  }
+}
+
+async function downloadChunkObjects(
+  storage: Storage,
+  uris: string[],
+  workDir: string,
+  format: DistributedFormat,
+): Promise<string[]> {
+  const chunksDir = join(workDir, "chunks");
+  mkdirSync(chunksDir, { recursive: true });
+  // Each chunk is an independent GCS GET (+ untar for png-sequence). Run
+  // them in parallel — assemble's wall-clock is otherwise dominated by
+  // `Σ chunk-download-ms` instead of `max(chunk-download-ms)`. Preserve the
+  // input order by writing into a pre-sized array rather than pushing as
+  // each task settles.
+  const local: string[] = new Array<string>(uris.length);
+  await Promise.all(
+    uris.map(async (uri, i) => {
+      if (!uri) {
+        throw new Error(`[handler] chunk URI at index ${i} is empty`);
+      }
+      const { key } = parseGcsUri(uri);
+      const localPath = join(chunksDir, basename(key));
+      await downloadGcsObjectToFile(storage, uri, localPath);
+      if (format === "png-sequence") {
+        const dirPath = join(chunksDir, `frames-${pad(i)}`);
+        await untarDirectory(localPath, dirPath);
+        local[i] = dirPath;
+      } else {
+        local[i] = localPath;
+      }
+    }),
+  );
+  return local;
+}
+
+// ── Helpers ─────────────────────────────────────────────────────────────────
+
+/** Collect every GCS URI that the handler will touch for a given event. */
+function getEventGcsUris(event: PlanEvent | RenderChunkEvent | AssembleEvent): string[] {
+  switch (event.Action) {
+    case "plan":
+      return [event.ProjectGcsUri, event.PlanOutputGcsPrefix];
+    case "renderChunk":
+      return [event.PlanGcsUri, event.ChunkOutputGcsPrefix];
+    case "assemble":
+      return [
+        event.PlanGcsUri,
+        ...event.ChunkGcsUris,
+        event.OutputGcsUri,
+        event.AudioGcsUri,
+      ].filter((u): u is string => u != null);
+  }
+}
+
+/** Emit the "guard disabled" warning at most once per instance. */
+let warnedAllowlistDisabled = false;
+
+/**
+ * Verify every GCS URI in the event resolves to the configured render
+ * bucket. Throws `GCS_URI_NOT_ALLOWED` (non-retryable) when a URI targets a
+ * different bucket, preventing request injection from reading or writing
+ * arbitrary GCS data.
+ *
+ * Opt-out is explicit: set `HYPERFRAMES_RENDER_BUCKET="*"` to disable the
+ * guard intentionally. If the env var is simply unset (or empty), the guard
+ * is disabled but a warning is logged once so the gap is visible in Cloud
+ * Logging — it shouldn't silently fail open. The Terraform module always
+ * wires the bucket name, so the prod path enforces.
+ */
+// fallow-ignore-next-line complexity
+function validateEventGcsUris(event: PlanEvent | RenderChunkEvent | AssembleEvent): void {
+  const allowedBucket = process.env.HYPERFRAMES_RENDER_BUCKET?.trim();
+  if (allowedBucket === "*") return; // explicit, intentional opt-out
+  if (!allowedBucket) {
+    if (!warnedAllowlistDisabled) {
+      warnedAllowlistDisabled = true;
+      logEvent({
+        event: "bucket_allowlist_disabled",
+        level: "WARNING",
+        message:
+          "HYPERFRAMES_RENDER_BUCKET is unset — the GCS bucket-allowlist guard is DISABLED. " +
+          'Set it to the render bucket name to enforce, or to "*" to opt out intentionally.',
+      });
+    }
+    return;
+  }
+
+  for (const uri of getEventGcsUris(event)) {
+    const { bucket } = parseGcsUri(uri);
+    if (bucket !== allowedBucket) {
+      const err = new Error(
+        `[handler] GCS_URI_NOT_ALLOWED: URI ${JSON.stringify(uri)} targets bucket "${bucket}" but only "${allowedBucket}" is permitted`,
+      );
+      err.name = "GCS_URI_NOT_ALLOWED";
+      throw err;
+    }
+  }
+}
+
+function pad(n: number): string {
+  return n.toString().padStart(4, "0");
+}
+
+function trimTrailingSlash(prefix: string): string {
+  return prefix.endsWith("/") ? prefix.slice(0, -1) : prefix;
+}
+
+function cleanupDir(dir: string): void {
+  try {
+    // Cloud Run re-uses an instance's filesystem across requests; clean up
+    // aggressively so we don't leak a chunk-sized footprint between renders
+    // (the writable filesystem counts against the instance's memory).
+    rmSync(dir, { recursive: true, force: true });
+  } catch {
+    // Best-effort — leak is preferable to crashing on the success path.
+  }
+}
+
+/**
+ * Read the untarred planDir's `plan.json` and assert its `planHash` matches
+ * what the workflow event claims. Throws on mismatch with a typed
+ * `PLAN_HASH_MISMATCH` error name so the workflow's non-retryable list
+ * routes it correctly. Defense-in-depth — the producer's `renderChunk` does
+ * the same check internally — but performing it here lets us fail before
+ * paying the Chrome-launch + per-frame capture cost on a misrouted chunk.
+ */
+// fallow-ignore-next-line complexity
+function verifyPlanHash(planDir: string, expected: string): void {
+  const planJsonPath = join(planDir, "plan.json");
+  let parsed: { planHash?: unknown };
+  try {
+    parsed = JSON.parse(readFileSync(planJsonPath, "utf-8")) as { planHash?: unknown };
+  } catch (err) {
+    const msg = err instanceof Error ? err.message : String(err);
+    const error = new Error(`PLAN_HASH_MISMATCH: failed to read ${planJsonPath}: ${msg}`);
+    error.name = "PLAN_HASH_MISMATCH";
+    throw error;
+  }
+  const actual = parsed.planHash;
+  if (typeof actual !== "string" || actual !== expected) {
+    const error = new Error(
+      `PLAN_HASH_MISMATCH: event PlanHash=${expected} did not match plan.json planHash=${String(actual)}`,
+    );
+    error.name = "PLAN_HASH_MISMATCH";
+    throw error;
+  }
+}
+
+// ── HTTP shell ───────────────────────────────────────────────────────────────
+
+/**
+ * Error names the workflow treats as non-retryable. A request that fails
+ * with one of these is the caller's fault (bad input, misrouted chunk) and
+ * retrying it just burns instance-seconds, so we map them to HTTP 400 while
+ * any other failure maps to 500 (which the workflow retry policy backs off
+ * and re-attempts). Keep this list in sync with the `retry` predicate in
+ * `packages/gcp-cloud-run/terraform/workflow.yaml`.
+ */
+const NON_RETRYABLE_ERROR_NAMES = new Set([
+  // Handler-boundary guards.
+  "GCS_URI_NOT_ALLOWED",
+  "PLAN_HASH_MISMATCH",
+  // Producer error class names (`.name`) + their string code aliases — the
+  // class sets `.name` to the class name but wraps a `code`; cover both so a
+  // raw-code throw is caught too. Mirrors the AWS state machine's
+  // non-retryable list.
+  "FormatNotSupportedInDistributedError",
+  "PlanTooLargeError",
+  "RenderChunkValidationError",
+  "FFMPEG_VERSION_MISMATCH",
+  "FORMAT_NOT_SUPPORTED_IN_DISTRIBUTED",
+  "PLAN_TOO_LARGE",
+  "BROWSER_GPU_NOT_SOFTWARE",
+  "FONT_FETCH_FAILED",
+  "ChromeBinaryUnavailableError",
+]);
+
+/**
+ * Build the Hono app. A single `POST /` endpoint dispatches on the body's
+ * `Action` field — the workflow points every step (plan, each renderChunk,
+ * assemble) at the same URL and varies only the body. `GET /healthz` backs
+ * the Cloud Run startup/liveness probe.
+ *
+ * `deps` is threaded through so tests can drive the real HTTP surface with
+ * an injected Storage double + mocked primitives.
+ */
+export function createApp(deps?: HandlerDeps): Hono {
+  const app = new Hono();
+
+  app.get("/healthz", (c) => c.json({ status: "ok" }));
+
+  // fallow-ignore-next-line complexity
+  app.post("/", async (c) => {
+    let body: CloudRunEvent;
+    try {
+      body = (await c.req.json()) as CloudRunEvent;
+    } catch {
+      return c.json({ error: "BAD_REQUEST", message: "request body must be JSON" }, 400);
+    }
+    try {
+      const result = await dispatch(body, deps);
+      return c.json(result, 200);
+    } catch (err) {
+      const name = err instanceof Error ? err.name : undefined;
+      const message = err instanceof Error ? err.message : String(err);
+      const status = name && NON_RETRYABLE_ERROR_NAMES.has(name) ? 400 : 500;
+      // Surface `error` (the name) as the discriminator the workflow's
+      // retry predicate keys off, plus `message` for human triage.
+      return c.json({ error: name ?? "RenderError", message }, status);
+    }
+  });
+
+  return app;
+}
+
+/** Start the HTTP server. Cloud Run injects `PORT` (default 8080). */
+export function startServer(): void {
+  const port = Number(process.env.PORT ?? 8080);
+  const app = createApp();
+  serve({ fetch: app.fetch, port }, (info) => {
+    logEvent({ event: "server_listening", port: info.port });
+  });
+}
+
+// Boot when executed directly (the Dockerfile runs `node dist/server.js`),
+// but not when imported by tests or the SDK.
+if (process.argv[1] && fileURLToPath(import.meta.url) === process.argv[1]) {
+  startServer();
+}

packages/gcp-cloud-run/terraform/.gitignore

@@ -0,0 +1,7 @@
+# Local Terraform working files — never commit these.
+.terraform/
+.terraform.lock.hcl
+*.tfstate
+*.tfstate.*
+*.tfvars
+crash.log

packages/gcp-cloud-run/terraform/main.tf

@@ -0,0 +1,197 @@
+# HyperFrames distributed render stack on Google Cloud.
+#
+# Topology (the GCP twin of the AWS Lambda adapter's SAM template):
+#
+#   GCS bucket  ←→  Cloud Run service (plan / renderChunk / assemble)
+#                        ▲
+#                        │ OIDC-authenticated http.post per step
+#                        │
+#                   Cloud Workflows  (Plan → parallel RenderChunk → Assemble)
+#
+# Two service accounts keep least-privilege boundaries:
+#   - run_sa     : the render service's identity; read/write the bucket only.
+#   - workflow_sa: the workflow's identity; invoke the render service only.
+
+locals {
+  workflow_source = var.workflow_source_path != "" ? var.workflow_source_path : "${path.module}/workflow.yaml"
+  name            = var.project_name
+}
+
+# ── Storage: plan tarballs, chunk outputs, final renders ─────────────────────
+resource "google_storage_bucket" "render" {
+  name                        = "${local.name}-render-${var.project_id}"
+  project                     = var.project_id
+  location                    = var.region
+  uniform_bucket_level_access = true
+  force_destroy               = var.bucket_force_destroy
+
+  # Render artifacts (plan tarballs, chunk files) are disposable scratch.
+  # Sweep them after 7 days so the bucket doesn't accumulate cost; final
+  # outputs that adopters want to keep should be copied elsewhere.
+  lifecycle_rule {
+    condition {
+      age = 7
+    }
+    action {
+      type = "Delete"
+    }
+  }
+}
+
+# ── Service accounts ─────────────────────────────────────────────────────────
+resource "google_service_account" "run_sa" {
+  account_id   = "${local.name}-run"
+  project      = var.project_id
+  display_name = "HyperFrames render service (Cloud Run)"
+}
+
+resource "google_service_account" "workflow_sa" {
+  account_id   = "${local.name}-wf"
+  project      = var.project_id
+  display_name = "HyperFrames render orchestration (Workflows)"
+}
+
+# Render service reads inputs + writes outputs in the render bucket only.
+resource "google_storage_bucket_iam_member" "run_sa_bucket" {
+  bucket = google_storage_bucket.render.name
+  role   = "roles/storage.objectAdmin"
+  member = "serviceAccount:${google_service_account.run_sa.email}"
+}
+
+# ── Cloud Run render service ─────────────────────────────────────────────────
+resource "google_cloud_run_v2_service" "render" {
+  name     = "${local.name}-render"
+  project  = var.project_id
+  location = var.region
+  # Authenticated only — no public invoker binding. Only the workflow SA can
+  # call it. Ingress stays "all" because Workflows reaches the service over
+  # Google's front door, not the VPC.
+  ingress = "INGRESS_TRAFFIC_ALL"
+  # Let `terraform destroy` (and replacement on image bumps) remove the
+  # service without a manual console step. The render service is stateless —
+  # all durable artifacts live in GCS.
+  deletion_protection = false
+
+  template {
+    service_account = google_service_account.run_sa.email
+    timeout         = "${var.request_timeout_seconds}s"
+    # One render (chunk / plan / assemble) per instance — each uses the whole
+    # box's CPU + memory + /tmp. The workflow's concurrency_limit governs how
+    # many instances run at once.
+    max_instance_request_concurrency = 1
+
+    scaling {
+      min_instance_count = var.min_instances
+      max_instance_count = var.max_instances
+    }
+
+    containers {
+      image = var.image
+
+      resources {
+        limits = {
+          cpu    = var.cpu
+          memory = var.memory
+        }
+        # Keep CPU allocated only during request processing (request-based
+        # billing). Renders are entirely request-scoped.
+        cpu_idle = true
+      }
+
+      env {
+        # Scopes every event's GCS URIs to this bucket (the handler's
+        # GCS_URI_NOT_ALLOWED guard). Defense against request injection.
+        name  = "HYPERFRAMES_RENDER_BUCKET"
+        value = google_storage_bucket.render.name
+      }
+
+      startup_probe {
+        http_get {
+          path = "/healthz"
+        }
+        timeout_seconds   = 5
+        period_seconds    = 10
+        failure_threshold = 6
+      }
+    }
+  }
+}
+
+# Only the workflow's identity may invoke the render service.
+resource "google_cloud_run_v2_service_iam_member" "workflow_invokes_run" {
+  name     = google_cloud_run_v2_service.render.name
+  project  = var.project_id
+  location = var.region
+  role     = "roles/run.invoker"
+  member   = "serviceAccount:${google_service_account.workflow_sa.email}"
+}
+
+# ── Cloud Workflows orchestration ────────────────────────────────────────────
+resource "google_workflows_workflow" "render" {
+  name            = "${local.name}-render"
+  project         = var.project_id
+  region          = var.region
+  service_account = google_service_account.workflow_sa.id
+  source_contents = file(local.workflow_source)
+  # Allow `terraform destroy` to remove the workflow without a manual step;
+  # the definition is reproducible from this module.
+  deletion_protection = false
+}
+
+# ── Runaway-request alert (backstop against a fan-out bug) ────────────────────
+resource "google_monitoring_alert_policy" "runaway_requests" {
+  project      = var.project_id
+  display_name = "${local.name}-render runaway request count"
+  combiner     = "OR"
+
+  conditions {
+    display_name = "Render service request count > threshold (1h)"
+    condition_threshold {
+      filter = join(" AND ", [
+        "resource.type = \"cloud_run_revision\"",
+        "resource.labels.service_name = \"${google_cloud_run_v2_service.render.name}\"",
+        "metric.type = \"run.googleapis.com/request_count\"",
+      ])
+      comparison      = "COMPARISON_GT"
+      threshold_value = var.render_request_alarm_threshold
+      duration        = "0s"
+      aggregations {
+        alignment_period   = "3600s"
+        per_series_aligner = "ALIGN_SUM"
+      }
+    }
+  }
+
+  notification_channels = var.notification_channels
+}
+
+# ── Workflow-failure alert ───────────────────────────────────────────────────
+# Request-count alone misses a render that fails 100% of the time at low
+# volume. Alert on any FAILED workflow execution so a broken render path is
+# visible even when traffic is light.
+resource "google_monitoring_alert_policy" "workflow_failures" {
+  project      = var.project_id
+  display_name = "${local.name}-render workflow execution failures"
+  combiner     = "OR"
+
+  conditions {
+    display_name = "Failed workflow executions (5m)"
+    condition_threshold {
+      filter = join(" AND ", [
+        "resource.type = \"workflows.googleapis.com/Workflow\"",
+        "resource.labels.workflow_id = \"${google_workflows_workflow.render.name}\"",
+        "metric.type = \"workflows.googleapis.com/finished_execution_count\"",
+        "metric.labels.status = \"FAILED\"",
+      ])
+      comparison      = "COMPARISON_GT"
+      threshold_value = 0
+      duration        = "0s"
+      aggregations {
+        alignment_period   = "300s"
+        per_series_aligner = "ALIGN_SUM"
+      }
+    }
+  }
+
+  notification_channels = var.notification_channels
+}

packages/gcp-cloud-run/terraform/outputs.tf

@@ -0,0 +1,34 @@
+output "render_bucket_name" {
+  description = "GCS bucket holding plan tarballs, chunk outputs, and final renders. Pass as renderToCloudRun({ bucketName })."
+  value       = google_storage_bucket.render.name
+}
+
+output "service_url" {
+  description = "HTTPS URL of the Cloud Run render service. Pass as renderToCloudRun({ serviceUrl })."
+  value       = google_cloud_run_v2_service.render.uri
+}
+
+output "workflow_name" {
+  description = "Workflow id. Pass as renderToCloudRun({ workflowId })."
+  value       = google_workflows_workflow.render.name
+}
+
+output "workflow_id_full" {
+  description = "Fully-qualified workflow resource name."
+  value       = google_workflows_workflow.render.id
+}
+
+output "run_service_account_email" {
+  description = "Render service identity (read/write the render bucket)."
+  value       = google_service_account.run_sa.email
+}
+
+output "workflow_service_account_email" {
+  description = "Workflow identity (invokes the render service)."
+  value       = google_service_account.workflow_sa.email
+}
+
+output "region" {
+  description = "Region everything was deployed into. Pass as renderToCloudRun({ location })."
+  value       = var.region
+}

packages/gcp-cloud-run/terraform/providers.tf

@@ -0,0 +1,12 @@
+# This module is applied directly (by `examples/gcp-cloud-run/scripts/smoke.sh`
+# and `hyperframes cloudrun deploy`), so it configures the google provider
+# from its own variables. Credentials come from the environment — either
+# Application Default Credentials (`gcloud auth application-default login`)
+# or a `GOOGLE_OAUTH_ACCESS_TOKEN` env var.
+#
+# If you instead embed this as a CHILD module, delete this block and pass a
+# configured provider from your root module.
+provider "google" {
+  project = var.project_id
+  region  = var.region
+}

packages/gcp-cloud-run/terraform/variables.tf

@@ -0,0 +1,75 @@
+variable "project_id" {
+  type        = string
+  description = "GCP project id to deploy the render stack into."
+}
+
+variable "region" {
+  type        = string
+  description = "Region for the Cloud Run service, Workflow, and bucket."
+  default     = "us-central1"
+}
+
+variable "project_name" {
+  type        = string
+  description = "Name prefix applied to the service / workflow / bucket / service accounts."
+  default     = "hyperframes"
+}
+
+variable "image" {
+  type        = string
+  description = "Fully-qualified container image for the render service (e.g. us-central1-docker.pkg.dev/PROJECT/REPO/hyperframes-render:TAG), built from packages/gcp-cloud-run/Dockerfile."
+}
+
+variable "cpu" {
+  type        = string
+  description = "vCPU per Cloud Run instance. Allowed: 1, 2, 4, 8. Renders are CPU-bound; 4 is a good default."
+  default     = "4"
+}
+
+variable "memory" {
+  type        = string
+  description = "Memory per Cloud Run instance. Must be ≥ 2Gi per vCPU at cpu=4. Headroom for Chrome + ffmpeg + the chunk's frames in /tmp."
+  default     = "16Gi"
+}
+
+variable "request_timeout_seconds" {
+  type        = number
+  description = "Per-request timeout. Cloud Run hard cap is 3600s; a single chunk should finish well inside this."
+  default     = 3600
+}
+
+variable "min_instances" {
+  type        = number
+  description = "Min Cloud Run instances. Default 0 (scale-to-zero) is cheapest but means the first render after idle pays a cold start (image pull + Chrome + bun boot, ~20-30s). Set to 1 to keep one warm if first-render latency matters."
+  default     = 0
+}
+
+variable "max_instances" {
+  type        = number
+  description = "Max Cloud Run instances. Each chunk pins one instance (request concurrency = 1), and the workflow fans out up to the plan's chunk count (which never exceeds Config.maxParallelChunks). Keep this >= the largest maxParallelChunks you render with, or excess chunks queue behind 429s + retry backoff. Also a runaway-cost backstop."
+  default     = 100
+}
+
+variable "workflow_source_path" {
+  type        = string
+  description = "Path to the Cloud Workflows YAML. Defaults to the copy shipped with this module."
+  default     = ""
+}
+
+variable "render_request_alarm_threshold" {
+  type        = number
+  description = "Cloud Monitoring alert fires when render-service request count exceeds this in a 1-hour window. Backstop against a runaway fan-out."
+  default     = 1000
+}
+
+variable "notification_channels" {
+  type        = list(string)
+  description = "Cloud Monitoring notification channel ids for the runaway-request alert. Empty disables notifications (the policy still records)."
+  default     = []
+}
+
+variable "bucket_force_destroy" {
+  type        = bool
+  description = "Allow `terraform destroy` to delete the render bucket even when it still holds objects. Off by default to match the AWS adapter's RETAIN policy."
+  default     = false
+}

packages/gcp-cloud-run/terraform/versions.tf

@@ -0,0 +1,9 @@
+terraform {
+  required_version = ">= 1.5.0"
+  required_providers {
+    google = {
+      source  = "hashicorp/google"
+      version = ">= 5.0.0"
+    }
+  }
+}

packages/gcp-cloud-run/terraform/workflow.yaml

@@ -0,0 +1,179 @@
+# HyperFrames distributed render orchestration on Cloud Workflows.
+#
+#   Plan → BuildChunkList → AssertChunkCount → RenderChunks (parallel) → Assemble
+#
+# Mirrors the Step Functions state machine in
+# `examples/aws-lambda/template.yaml`. Every step POSTs to the same Cloud Run
+# service URL (passed in as `args.ServiceUrl`) and varies only the body's
+# `Action`. The service returns the step's small result body on 2xx; on a
+# non-retryable failure it returns HTTP 400, on a retryable failure HTTP 5xx —
+# the `retryable` predicate below keys off exactly that split.
+#
+# The final returned object accumulates every step's result body so
+# `getRenderProgress` can read frame totals + per-step durations on success:
+#   { Plan: {...}, Chunks: [{...}, ...], Assemble: {...} }
+#
+# Deploy with `gcloud workflows deploy` (the Terraform module / the
+# `hyperframes cloudrun deploy` command do this for you).
+
+main:
+  params: [args]
+  steps:
+    - init:
+        assign:
+          - serviceUrl: ${args.ServiceUrl}
+          - projectGcsUri: ${args.ProjectGcsUri}
+          - planOutputGcsPrefix: ${args.PlanOutputGcsPrefix}
+          - outputGcsUri: ${args.OutputGcsUri}
+          - config: ${args.Config}
+
+    # ── Plan (Activity A) ────────────────────────────────────────────────────
+    - plan:
+        try:
+          call: http.post
+          args:
+            url: ${serviceUrl}
+            timeout: 1800
+            auth:
+              type: OIDC
+            body:
+              Action: plan
+              ProjectGcsUri: ${projectGcsUri}
+              PlanOutputGcsPrefix: ${planOutputGcsPrefix}
+              Config: ${config}
+          result: planResp
+        retry:
+          predicate: ${retryable}
+          max_retries: 4
+          backoff:
+            initial_delay: 2
+            max_delay: 60
+            multiplier: 2
+    - capturePlan:
+        assign:
+          - planResult: ${planResp.body}
+          - chunkCount: ${planResult.ChunkCount}
+
+    # ── BuildChunkList + AssertChunkCount ──────────────────────────────────────
+    - assertChunkCount:
+        switch:
+          - condition: ${chunkCount > 0}
+            next: buildChunkList
+        next: planProducedZeroChunks
+    - planProducedZeroChunks:
+        raise:
+          code: PLAN_PRODUCED_ZERO_CHUNKS
+          message: "Plan returned ChunkCount=0 — the composition produced no frames. Non-retryable producer-side invariant violation."
+    - buildChunkList:
+        # Pre-size the ordered chunk-URI + per-chunk result lists so the
+        # parallel branches below assign by index (distinct indices, no
+        # read-modify-write race on a shared accumulator).
+        assign:
+          - chunkIndexes: []
+          - chunkUris: []
+          - chunkResults: []
+    - fillLists:
+        for:
+          value: i
+          range: [0, ${chunkCount - 1}]
+          steps:
+            - appendSlots:
+                assign:
+                  - chunkIndexes: ${list.concat(chunkIndexes, i)}
+                  - chunkUris: ${list.concat(chunkUris, "")}
+                  - chunkResults: ${list.concat(chunkResults, "")}
+
+    # ── RenderChunks (Activity B, fanned out) ──────────────────────────────────
+    - renderChunks:
+        parallel:
+          shared: [chunkUris, chunkResults]
+          # Run up to chunkCount chunks at once, clamped to 20 — Cloud
+          # Workflows hard-caps concurrent branches/iterations per execution
+          # at 20 (https://cloud.google.com/workflows/quotas). Above that,
+          # iterations queue regardless of concurrency_limit, so a config with
+          # maxParallelChunks > 20 still renders correctly; the extra chunks
+          # just wait. All chunkCount iterations always run.
+          concurrency_limit: ${math.min(chunkCount, 20)}
+          for:
+            value: idx
+            in: ${chunkIndexes}
+            steps:
+              - renderOneChunk:
+                  try:
+                    call: http.post
+                    args:
+                      url: ${serviceUrl}
+                      timeout: 1800
+                      auth:
+                        type: OIDC
+                      body:
+                        Action: renderChunk
+                        ChunkIndex: ${idx}
+                        PlanGcsUri: ${planResult.PlanGcsUri}
+                        PlanHash: ${planResult.PlanHash}
+                        ChunkOutputGcsPrefix: ${planOutputGcsPrefix}
+                        Format: ${planResult.Format}
+                    result: chunkResp
+                  retry:
+                    predicate: ${retryable}
+                    max_retries: 4
+                    backoff:
+                      initial_delay: 2
+                      max_delay: 60
+                      multiplier: 2
+              - storeChunk:
+                  assign:
+                    - chunkUris[idx]: ${chunkResp.body.ChunkGcsUri}
+                    - chunkResults[idx]: ${chunkResp.body}
+
+    # ── Assemble (Activity C) ──────────────────────────────────────────────────
+    - assemble:
+        try:
+          call: http.post
+          args:
+            url: ${serviceUrl}
+            timeout: 1800
+            auth:
+              type: OIDC
+            body:
+              Action: assemble
+              PlanGcsUri: ${planResult.PlanGcsUri}
+              ChunkGcsUris: ${chunkUris}
+              AudioGcsUri: ${planResult.AudioGcsUri}
+              OutputGcsUri: ${outputGcsUri}
+              Format: ${planResult.Format}
+              # Forward the caller's exact-CFR request (Config.cfr) to assemble.
+              # `"cfr" in config` guards the optional key; when unset this is
+              # false, which the handler reads as the default -c copy path.
+              Cfr: ${("cfr" in config) and config.cfr}
+          result: assembleResp
+        retry:
+          predicate: ${retryable}
+          max_retries: 4
+          backoff:
+            initial_delay: 2
+            max_delay: 60
+            multiplier: 2
+
+    - done:
+        return:
+          Plan: ${planResult}
+          Chunks: ${chunkResults}
+          Assemble: ${assembleResp.body}
+
+# Retry predicate: retry transient/server failures (429 + 5xx), never the
+# handler's non-retryable 400s (bad input, plan-hash mismatch, unsupported
+# format, …). Connection / timeout errors carry no `.code`; retry those too.
+retryable:
+  params: [e]
+  steps:
+    - classify:
+        switch:
+          - condition: ${not("code" in e)}
+            return: true
+          - condition: ${e.code == 429}
+            return: true
+          - condition: ${e.code >= 500 and e.code < 600}
+            return: true
+    - nonRetryable:
+        return: false

packages/gcp-cloud-run/tsconfig.build.json

@@ -0,0 +1,12 @@
+{
+  "extends": "./tsconfig.json",
+  "compilerOptions": {
+    "paths": {},
+    "noEmit": false,
+    "outDir": "./dist",
+    "rootDir": "./src",
+    "declaration": true,
+    "declarationMap": true,
+    "sourceMap": true
+  }
+}

packages/gcp-cloud-run/tsconfig.json

@@ -0,0 +1,19 @@
+{
+  "compilerOptions": {
+    "target": "ES2022",
+    "module": "ESNext",
+    "moduleResolution": "bundler",
+    "esModuleInterop": true,
+    "strict": true,
+    "noUncheckedIndexedAccess": true,
+    "skipLibCheck": true,
+    "noEmit": true,
+    "baseUrl": ".",
+    "paths": {
+      "@hyperframes/producer": ["../producer/src/index.ts"],
+      "@hyperframes/producer/distributed": ["../producer/src/distributed.ts"]
+    }
+  },
+  "include": ["src/**/*"],
+  "exclude": ["node_modules", "dist", "src/**/*.test.ts", "src/**/__fixtures__/**", "scripts"]
+}