fix(core): preserve new.target in the deterministic Date override so Date subclasses work

[ar-tama]

Fixes #3371

Problem

createContext() replaces the VM's global Date with a plain function. A plain function has no [[Construct]] behavior that forwards new.target, so when user code does class X extends Date, super() returns a fresh plain Date object which becomes this — the subclass instance loses its identity, methods, and fields.

This silently breaks every library that models a date by subclassing Date, e.g. TZDate from @date-fns/tz (the officially recommended way to do time-zone-aware math with date-fns v4). Inside a workflow body, a TZDate degrades to a plain Date, so every zone-aware getter falls back to the container's time zone. Since the epoch value stays correct, this presents as a silent, production-only off-by-one-day bug. Full analysis in #3371.

Fix

Replace the plain-function override with a class so new.target is preserved:

(g as any).Date = class Date extends Date_ {
  constructor(...args: any[]) {
    if (args.length === 0) {
      super(fixedTimestamp);
    } else {
      // @ts-expect-error - Args is `Date` constructor arguments
      super(...args);
    }
  }
};
g.Date.now = () => fixedTimestamp;

Both of the previous fix-ups become unnecessary, because extends already sets up the whole prototype chain:

• (g as any).Date.prototype = Date_.prototype — with extends the real chain is in place (and the assignment would be illegal on a class, whose prototype is non-writable).
• Object.setPrototypeOf(g.Date, Date_) — extends already makes the statics (Date.parse, Date.UTC) inherited; only the Date.now override stays as an own property.

Determinism is unchanged: zero-arg construction still returns the fixed timestamp, and Date.now() is still overridden. Covered by the existing determinism tests plus two new ones (subclassing, statics).

One behavioral note: calling Date() without new now throws (classes are not callable), where the old override returned a Date object — itself already a deviation from the spec, which returns a string. If callable Date() needs to keep working, the alternative is a plain function that branches on new.target and uses Reflect.construct(Date_, args, new.target); happy to switch to that if preferred.

Tests

• should support subclassing \Date` — subclass keeps identity (instanceof`), methods, fields; constructor args are forwarded; zero-arg subclass construction still gets the fixed timestamp
• should preserve \Date` static methods—Date.parse/Date.UTC` still inherited

All 36 tests in packages/core/src/vm/index.test.ts pass, pnpm typecheck is clean. (The runtime.test.ts failure re-routes a misrouted lazy hook resume also fails on an unmodified checkout of main, so it is unrelated.)

🤖 Generated with Claude Code

[changeset-bot]

🦋 Changeset detected

Latest commit: 4c7cd3c

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 16 packages

Name	Type
@workflow/core	Patch
@workflow/builders	Patch
@workflow/cli	Patch
@workflow/next	Patch
@workflow/nitro	Patch
@workflow/vitest	Patch
@workflow/web-shared	Patch
@workflow/web	Patch
workflow	Patch
@workflow/world-testing	Patch
@workflow/astro	Patch
@workflow/nest	Patch
@workflow/rollup	Patch
@workflow/sveltekit	Patch
@workflow/vite	Patch
@workflow/nuxt	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[vercel]

@ar-tama is attempting to deploy a commit to the Vercel Labs Team on Vercel.

A member of the Team first needs to authorize it.

[TooTallNate]

Reviewed at 922287c (base is current main, 0 behind). Build green; all 37 vm tests pass; full core suite green except the one pre-existing failure you correctly identified (more on that below).

Verified:

• The final Reflect.construct(Date_, args, new.target) design is the right call over the class approach from the PR description: a plain function keeps [[Call]], so bare Date() stays callable — and now returns the spec-correct string instead of the old non-spec object. Construction semantics all check out: a plain function retains [[Construct]], super() from a subclass forwards new.target through to Date_, and the constructed object (with the subclass's prototype and a real Date internal slot) becomes this — identity, methods, and fields preserved, which is exactly what TZDate needs.
• Determinism is fully preserved: zero-arg construction and zero-arg super() both pin to fixedTimestamp, Date.now stays overridden (and is inherited by subclass statics via the retained setPrototypeOf), and fixedTimestamp is a live let binding that updateTimestamp advances during replay — the closure reads the current value at each construction, same as before.
• Cross-engine alignment bonus: the QuickJS engine (WORKFLOW_VM=quickjs) never wraps Date at all — it intercepts the WASI clock_time_get syscall, so its Date is fully native and spec-correct. That means QuickJS never had this subclassing bug, and your bare-Date() string change brings node:vm into agreement with QuickJS where the old object-return diverged. No QuickJS-side work needed.
• One honest behavioral note for the record: a workflow that called bare Date() and used the old non-spec object (e.g. .getTime() on it) would diverge when an in-flight run is replayed across this upgrade. That pattern is vanishingly rare, was already broken outside the sandbox, and the new behavior is what both the spec and the other engine do — the right trade.

Two asks before merge:

1. DCO check is failing — the commits need sign-off (rebase with --signoff, or follow the DCO bot's remediation instructions).
2. The changeset text is stale: it describes the override as "now a class that preserves new.target", but the final implementation is a plain function using Reflect.construct (your head commit changed the approach to keep callability). Changesets become release notes — please reword, and consider mentioning that bare Date() now returns the spec time string.

On the unrelated runtime.test.ts failure: confirmed — re-routes a misrouted lazy hook resume with its payload intact fails on an unmodified checkout of main. It's an interaction between two recently merged PRs (#2960's deployment-affinity guard and #3345's lazy-hook fast path); I've reported it with analysis on #3345. Thank you for isolating it precisely instead of ignoring it — that note made the triage immediate.

Clean, well-reasoned fix with exactly the right tests (subclass identity + zero-arg determinism, bare-call string, static inheritance). Approving.

[ar-tama]

Both asks addressed: rebased with --signoff (DCO now passing) and reworded the changeset to describe the final Reflect.construct implementation, including the bare Date() spec-string behavior. Thanks for the review!