ci: fix tart.mjs unresolved refs, add snapshot(), use -base images

[alii]

Stacked on #29314. Part 2 of the macOS CI Phase 0 prerequisites — unblocks scripts/machine.mjs --cloud=tart so we can build golden macOS VM images.

Problem

scripts/tart.mjs has been broken since it landed: toMachine() calls spawnSsh, spawnSshSafe, spawnScp, and spawnRdp without importing or defining any of them. The first SSH/SCP into a Tart VM throws ReferenceError. Separately, getVm() returns {Name: name} even when the VM doesn't exist (always truthy), so cloneVm() never pulls a missing base image.

Changes

• Import the SSH helpers — spawnSsh / spawnSshSafe already exist in utils.mjs; just import them.
• Move spawnScp from machine.mjs → utils.mjs (next to spawnSsh) so tart.mjs can import it without a circular dep.
• Drop dead rdp() — called non-existent spawnRdp; macOS doesn't speak RDP anyway.
• Fix getVm() always-truthy — return undefined when tart get finds nothing.
• Switch -xcode → -base images + add macOS 26 — cirruslabs -base images are ~25 GB vs ~50 GB+ with full Xcode; we install our own toolchain via bootstrap.sh.
• Implement snapshot(label) — stop the VM, then tart push ghcr.io/oven-sh/<label>. machine.mjs already calls machine.snapshot() on the publish-image path; tart was the only cloud missing it. Auth via TART_REGISTRY_USERNAME / TART_REGISTRY_PASSWORD (tart reads them directly).
• Fix error message that printed undefined instead of the unsupported release number.

Supersedes the tart.mjs portion of #25134 (drops that branch's unrelated FreeBSD support and azure-import removal so this stays scoped).

Test plan

• node --check scripts/{tart,machine,utils}.mjs
• bun -e 'import {tart} from "./scripts/tart.mjs"' — module loads, no unresolved refs
• tart.getImage({os:"darwin",arch:"aarch64",release:"14"}) → ghcr.io/cirruslabs/macos-sonoma-base
• tart.toMachine("x") returns {snapshot, upload, spawn, spawnSafe, attach, close, ...}
• import {spawnScp} from "./scripts/utils.mjs" resolves
• On a Mac with tart installed: node scripts/machine.mjs create-image --cloud=tart --os=darwin --arch=aarch64 --release=14 boots a VM and runs bootstrap (requires ci: split downloadCacheDir from cacheDir, add BUN_DEPS_CACHE_PATH #29314's bootstrap.sh darwin fixes — separate follow-up)

[robobun]

^{Updated 3:05 AM PT - Apr 17th, 2026}

❌ @robobun, your commit bdf4cca has 4 failures in Build #46017 (All Failures):

• test/js/web/broadcastchannel/broadcast-channel-worker-gc.test.ts - code 1 on 🐧 13 x64-asan
• test/js/bun/cron/in-process-cron.test.ts - crash reported on 🐧 13 x64-asan
• test/bake/dev/esm.test.ts - code 1 on 🪟 2019 x64
• test/bundler/bundler_defer.test.ts - code 1 on 🪟 2019 x64

---

🧪   To try this PR locally:

bunx bun-pr 29315

That installs a local version of the PR into your bun-29315 executable, so you can run:

bun-29315 --bun

[github-actions]

This PR may be a duplicate of:

1. ci: unregress tart image gen for macos #25134 - Both fix tart.mjs for macOS CI image generation (unresolved refs, getVm bug, spawnScp relocation); this PR explicitly supersedes the tart.mjs portion of ci: unregress tart image gen for macos #25134

🤖 Generated with Claude Code

[claude]

---

Additional findings (outside current diff — PR may have been updated during review):

• 🔴 scripts/tart.mjs:285-291 — The PR removes the rdp() function from tart.toMachine()'s return object, but machine.mjs unconditionally calls machine.rdp() when options.rdp is truthy — which happens with both --rdp and --vnc flags. Since tart uses the rdp option to enable vnc-experimental in runVm(), passing --vnc --cloud=tart is a legitimate usage that will now throw TypeError: machine.rdp is not a function. Fix: guard the call site in machine.mjs with if (typeof machine.rdp === "function"), or add a descriptive stub to tart.toMachine().

  Extended reasoning...

  What the bug is and how it manifests

  Before this PR, tart.toMachine() returned an object that included an rdp function (though it called the non-existent spawnRdp, causing a ReferenceError at call time). This PR explicitly drops it with the comment "macOS doesn't speak RDP anyway." The returned machine object now contains only: spawn, spawnSafe, attach, upload, snapshot, close, and [Symbol.asyncDispose]. There is no rdp property, so machine.rdp is undefined.

  The specific code path that triggers it

  machine.mjs parses CLI flags and sets:

  rdp: \!\!args["rdp"] || \!\!args["vnc"],

  Then, unconditionally at the call site:

  if (options.rdp) {
    await startGroup("Connecting with RDP...", async () => {
      const { hostname, username, password } = await machine.rdp();
      ...
    });
  }

  There is no guard checking typeof machine.rdp === "function" before invoking it.

  Why existing code doesn't prevent it

  The Machine typedef marks rdp as optional (@property {() => Promise<RdpCredentials>} [rdp]), correctly signaling that not all cloud implementations need to provide it. However, the call site only checks options.rdp (the CLI flag), not whether the cloud-specific machine object actually exposes the method.

  What the impact would be

  For tart, the --vnc flag is a meaningful option: tart.createMachine() passes 'vnc-experimental': rdp to runVm(), which correctly enables VNC on the VM. A user running node scripts/machine.mjs create-image --cloud=tart --vnc would have the VM start with VNC enabled, but the process would immediately throw TypeError: machine.rdp is not a function when machine.mjs tries to call machine.rdp(). The VM remains running (orphaned) while the script crashes.

  How to fix it

  Option A — guard the call site in machine.mjs:

  if (options.rdp && typeof machine.rdp === "function") {
    await startGroup("Connecting with RDP...", ...);
  }

  Option B — add a descriptive stub in tart.toMachine():

  const rdp = async () => {
    throw new Error("macOS/tart does not support RDP; use --vnc to enable VNC-only access");
  };

  Step-by-step proof

  1. User runs: node scripts/machine.mjs create-image --cloud=tart --os=darwin --arch=aarch64 --release=14 --vnc
  2. options.rdp is set to true (line: rdp: \!\!args["rdp"] || \!\!args["vnc"])
  3. tart.createMachine() is called; it passes 'vnc-experimental': rdp (i.e., true) to runVm() — VNC is legitimately enabled on the VM.
  4. tart.toMachine(machineId) returns an object with no rdp property.
  5. machine.mjs checks if (options.rdp) — this is true.
  6. machine.rdp() is called — but machine.rdp is undefined.
  7. TypeError: machine.rdp is not a function is thrown. The VM is left running and orphaned.

• 🔴 scripts/tart.mjs:264-277 — The PR removes the rdp() stub from tart.toMachine(), but machine.mjs unconditionally calls machine.rdp() when options.rdp is truthy — which happens for both --rdp and --vnc flags. Because --vnc --cloud=tart is a valid and meaningful flag combination (it enables vnc-experimental on the VM), users who pass it will immediately crash with TypeError: machine.rdp is not a function. Add a typeof machine.rdp === 'function' guard before the call site in machine.mjs, or add a tart-specific stub that throws a descriptive error.

  Extended reasoning...

  What the bug is and how it manifests

  In machine.mjs, the main() function builds options.rdp as \!\!args['rdp'] || \!\!args['vnc'] — so passing either --rdp or --vnc makes options.rdp truthy. Later, the code has an unconditional call:

  if (options.rdp) {
    await startGroup('Connecting with RDP...', async () => {
      const { hostname, username, password } = await machine.rdp();

  After this PR, tart.toMachine() returns an object without any rdp property (the old stub was removed; the new snapshot function replaced it). So machine.rdp is undefined, and calling machine.rdp() throws TypeError: machine.rdp is not a function.

  The specific code path that triggers it

  1. User runs: node scripts/machine.mjs create-image --cloud=tart --os=darwin --arch=aarch64 --release=14 --vnc
  2. options.rdp is set to true (because \!\!args['vnc'] is true).
  3. tart.createMachine() correctly passes 'vnc-experimental': rdp to runVm(), enabling VNC on the VM. This is intentional and useful.
  4. main() reaches if (options.rdp) — condition is true.
  5. machine.rdp is undefined on the tart machine object.
  6. await machine.rdp() throws TypeError: machine.rdp is not a function.

  Why existing code does not prevent it

  The Machine typedef marks rdp as optional (@property {() => Promise<RdpCredentials>} [rdp]), signaling that not all cloud implementations need to provide it. However, the call site in machine.mjs has no guard like typeof machine.rdp === 'function' — it only checks options.rdp (the CLI flag), not whether machine.rdp is actually callable. Before this PR, tart's rdp() stub existed (though it called the nonexistent spawnRdp, producing a ReferenceError); after this PR, the property is absent entirely, changing the failure mode to a TypeError at the call site rather than inside the function.

  What the impact would be

  Any operator who passes --vnc --cloud=tart (a legitimate option for obtaining graphical access to a macOS VM) will hit the crash before any VNC connection can be made. The VNC flag is wired up correctly on the tart side — the VM would start with vnc-experimental enabled — but machine.mjs aborts with a TypeError before doing anything useful.

  How to fix it

  Option A — guard the call site in machine.mjs:

  if (options.rdp && typeof machine.rdp === 'function') {
    await startGroup('Connecting with RDP...', async () => {
      const { hostname, username, password } = await machine.rdp();
      ...
    });
  }

  Option B — add a descriptive stub in tart.toMachine():

  const rdp = async () => {
    throw new Error('RDP is not supported on tart (macOS) VMs; use --vnc for VNC access');
  };

  Option A is safer for the general case (other future cloud implementations may also omit rdp); option B gives a better error message.

  Step-by-step proof

  1. args['vnc'] is truthy so options.rdp = true.
  2. tart.createMachine() calls tart.runVm(machineId, { ..., 'vnc-experimental': true }) — the VM starts with VNC enabled. Correct.
  3. tart.toMachine(machineId) returns { cloud: 'tart', id, spawn, spawnSafe, attach, upload, snapshot, close, [Symbol.asyncDispose] } — no rdp field.
  4. Back in main(): if (options.rdp) evaluates to true — enter block.
  5. machine.rdp evaluates to undefined.
  6. await machine.rdp() throws TypeError: machine.rdp is not a function. Process crashes.

[claude]

LGTM — all previously flagged 🔴 issues have been addressed in this revision.

Extended reasoning...

Overview

This PR fixes which was completely broken since it landed: SSH/SCP helpers were called without being imported (causing ReferenceError), getVm() returned a truthy object even for non-existent VMs (preventing image pulls), and snapshot() was absent. The changes also move spawnScp from machine.mjs to utils.mjs to break a circular import dependency, switch cirruslabs base images from -xcode to -base variants, and add macOS 26 (tahoe) support.

Security risks

No security-sensitive code is touched. Auth credentials for the registry (TART_REGISTRY_USERNAME/TART_REGISTRY_PASSWORD) are read directly by the tart CLI from the environment — no new credential-handling code is introduced. SSH still uses password auth via sshpass (hardcoded admin/admin from cirruslabs base images), but this is the existing approach and is limited to local tart VMs.

Level of scrutiny

This is CI tooling (image-build scripts), not production user-facing code. The blast radius is confined to the macOS CI image-creation pipeline. The changes are well-scoped: bug fixes with clear before/after intent, no novel algorithms or protocol changes.

Other factors

All six rounds of inline comments (flagging unhandled rejection in runVm, VM orphan on missing sshpass, silent push output, orphaned VM on toMachine throw) are marked resolved and the fixes are visible in the current diff. Remaining open issues are pre-existing minor cosmetics (undefined username in SCP error message, stdio:inherit nullifying stderr capture) inherited verbatim from machine.mjs — not regressions introduced here.

[alii]

@robobun adopt

[robobun]

✅ Adopted — up to date with main + base, diff is exactly scripts/{tart,machine,utils}.mjs.

Since picking up:

• 6761fa6 — guard machine.rdp() call site (tart doesn't provide it; --vnc --cloud=tart would throw)
• e465ec1 — cloneVm(): inherit stdio on the now-reachable remote-pull path so Buildkite sees progress during the ~25GB pull
• d2bbd54 — createMachine() fills imageId/instanceType/region so the "Created machine" table isn't undefined×3

All review threads resolved.

[claude]

LGTM — all the previously flagged blockers have been addressed across the iteration commits.

Extended reasoning...

Overview

This PR fixes a series of longstanding bugs in scripts/tart.mjs that had prevented machine.mjs --cloud=tart from ever completing: missing SSH/SCP imports (ReferenceError), getVm() always-truthy bug suppressing remote image pulls, the dead rdp() referencing non-existent spawnRdp, and a missing snapshot() implementation. It also moves spawnScp to utils.mjs to avoid circular deps, switches from -xcode to -base cirruslabs images, and adds macOS 26 (Tahoe) support.

Security risks

No security-sensitive code paths are touched. The hardcoded admin:admin credentials in connect() are the cirruslabs default for local tart VMs and are expected; the sshpass pre-flight check is gated before any VM resources are allocated. No auth logic, crypto, or permission code is modified.

Level of scrutiny

Medium. This is CI infrastructure / scripting code, not production bun runtime code. The changes are incremental fixes to a module that was previously broken, so the risk of regression to the broader codebase is low. The test plan is necessarily limited (full macOS VM tests require a Mac with tart installed), but the static correctness checks (--check, module import smoke test) pass.

Other factors

The PR went through thorough review with multiple rounds of bug reports and fixes. All previously flagged red bugs were addressed: sshpass guard moved before resource allocation, throwOnError:false on the unawaited tart run promise, { stdio: "inherit" } on long-running clone and push operations (preventing Buildkite no-output timeouts), and the mimalloc hash sync in process.test.js. The one remaining nit (missing imageId/instanceType/region in toMachine()) is purely cosmetic — it only affects a diagnostic console.table printout.

[claude]

LGTM — all critical issues addressed.

Extended reasoning...

Overview

This PR fixes long-standing breakage in scripts/tart.mjs: missing imports for SSH/SCP helpers, an always-truthy getVm() that prevented remote image pulls, a missing snapshot() implementation, and switches from -xcode to lighter -base cirruslabs images. It also relocates spawnScp from machine.mjs to utils.mjs to avoid a circular dependency, and adds a guard for the machine.rdp call site.

Security risks

No security-sensitive code touched. Auth for registry pushes relies on env vars (TART_REGISTRY_USERNAME/TART_REGISTRY_PASSWORD) read directly by tart — no credentials are handled in JS. The hardcoded admin/admin SSH credentials are standard for cirruslabs base images on a local VM network.

Level of scrutiny

This is CI tooling — scripts that build and publish macOS VM images. It is not in any production or runtime path. The changes are mechanical: import fixes, a null-check, stdio inheritance for long-running spawns, and a new snapshot() method. Low-risk CI scripting changes that warranted review but not extended scrutiny once the critical bugs were fixed.

Other factors

All five red/yellow inline bugs from prior review rounds were resolved by robobun (commits c28e1ad, e465ec1, d2bbd54, 6761fa6, 7aa892d, and others). All review threads are marked resolved. The one remaining new finding (JSDoc typedef label? should be label) is cosmetic — no current call site omits the label — and is called out in an inline comment.