4.6.0+ with zod ≥ 4.4: every run completion waits a fixed 6 s (cancels 31 s) because FLUSH/CANCEL IPC acks are rejected

[marcus-dk]

What happens

Since 4.6.0, in projects that resolve zod to 4.4 or newer, every run on a managed worker takes ~6.1 s longer between the task finishing and the completion reaching the engine. Compute (usage_duration) is unchanged; execution_duration - usage_duration goes from ~0.1–0.25 s on 4.5.15 to 6.10–6.29 s on 4.6.0 and 4.6.2, for every task, warm or cold. Cancellation is hit harder: the CANCEL ack budget is 30 s + 1 s. Still present on main (c2b7a72) and in 4.6.4: zodIpc.ts is unchanged since #4039.

We see it on every task we run: the gap appears exactly at the SDK upgrade and never drops below ~6.09 s, where it was ~0.1 s before.

Why

TaskRunProcess.cleanup() sends FLUSH to the task run process with sendWithAck(..., timeoutInMs + 1_000) (5 s + 1 s) before completeRunAttempt. FLUSH and CANCEL declare callback: z.void(), so the child acks with { type: "ACK", id, message: undefined }. process.send JSON-serialises the packet and drops the message key. In zodIpc.ts the Packet schema declares message: z.any(); under zod 3 a z.any() key was optional at parse time, under zod 4.4+ it is required (see the version table below). Since #4039, zodIpc.ts imports zod/v4, and @trigger.dev/core usually resolves zod from the user's project (^3.25.56 || ^4.0.0). So any project that resolves zod 4.4 or newer (this repo pins 4.5.4) fails Packet.safeParse on every ack, #handlePacket returns silently, and sendWithAck waits out its full timeout. Nothing is logged because the rejection is swallowed by tryCatch.

Repro (zod alone)

zod 4.6.5: z.discriminatedUnion("type", [...ACK { message: z.any(), id }]).safeParse({ type: "ACK", id: 0 })
  -> message: Invalid input: expected nonoptional, received undefined

The same schema across versions (root zod and zod/v4 behave the same):

zod 3.25.76  accepted
zod 4.3.6    accepted
zod 4.4.0    REJECTED
zod 4.5.4    REJECTED

That is probably why this has gone unreported: projects on zod 3.x or 4.0–4.3 are unaffected, and billed compute (usage_duration) does not change.

Repro (two ZodIpcConnections over a forked child, 3 s ack timeouts)

core 4.6.2:  FLUSH: REJECTED after 3002ms (sendWithAck() timeout)   CANCEL: REJECTED after 3001ms
core 4.5.15: FLUSH: ack in 2ms                                       CANCEL: ack in 0ms

End-to-end confirmation (trigger dev, same server, same one-task project, six runs each)

worker	task compute	gap between task end and completion
stock trigger.dev@4.6.2	14 ms	6.27 – 6.34 s
same install, only message: z.any().optional() on the ACK packet in core's dist	14 ms	0.40 – 0.46 s

The dev controller runs the same TaskRunProcess.cleanup() → FLUSH → completeRunAttempt sequence as the managed controller, and that line is the only change between the two workers.

Fix

Make the ACK packet's message optional again (z.any().optional()), which restores the zod 3 type and runtime behaviour. Only the ACK packet can legitimately carry undefined (the two void callbacks); no message catalog entry has an undefined payload, so the messageSchema.payload: z.unknown() envelope is unaffected. Happy to open a PR with this change and a JSON-round-trip test that fails on main and passes with the fix.