Temporal and Inngest rethink how child work and sandboxes get scoped

Temporal reshapes integrations and child-workflow control

The headline change of the week comes from Temporal's Go SDK, which pulled the Google ADK integration out of contrib entirely — it now lives in the separate temporalio/ai-integrations repo under the module path go.temporal.io/googleadk, so anyone importing the old contrib path needs to update their imports before upgrading. The SDK also gained per-child versioning overrides, letting you pin, auto-upgrade, or one-time route the versioning behavior of a child workflow independently of its parent (requires Temporal Server 1.32+). Alongside that, a new experimental transfer type converter enables transfer type conversion from the default data converter, and server poller groups picked up autoscaling support. One regression to watch for: plugin cleanup that synchronously calls the same worker's Stop will now deadlock.

On the server side, Temporal's dependency bumps addressed CVEs and made fair-task garbage collection conditional on range ID, with the internal callback namespace binding now validating both token readings — security hardening that self-hosters should prioritize.

Inngest makes sandboxes disposable

Inngest's JS SDK shipped a handful of updates that mostly concern sandboxes. The disposable sandbox scope lets you declare a sandbox with await using, so it's destroyed automatically when the block exits, and durations now accept compound strings like "4m30s". An experimental group.sandbox() helper runs a callback against a fresh sandbox, tears it down afterwards, and keeps it alive across retried attempts. Separately, CI's steps are now grouped into trace spans — one per job, command, and extra machine — and named for what they do so traces distinguish CI work from your own. That package now requires inngest 4.23.1 or later; older SDKs are unsupported.

Trigger.dev tunes replay and trace performance

Trigger.dev's release rounded out its replay and observability story with per-key concurrency overrides, much faster run traces for large and live runs, and replay options that let you change the payload and trigger settings on a replay. The same improvements shipped through the Helm chart for self-hosted deployments, so platform teams running the chart get the new trace speed and replay controls without upgrading the app image separately.

n8n's steady stream of fixes

n8n cut nine patch releases across its lines, and the story is consistency rather than novelty: webhooks now get their own XML parser per request body, fixing cross-request contention, a fix mirrored into the newer line with the same isolation. Flows that rely on deferred tools saw skill-loaded tool calls resume correctly, including queue-mode execution. The main line added Gemini models shown by name and prioritized first in model pickers, plus a sweep of execution-retry and crash-journal fixes and a size validation guard for the chunk() expression extension. Rounds of maintenance released too — Observer detail lines are kept instead of discarded, with the 1.123 line getting its own housekeeping and Chat Hub segments now handled once per segment in queue mode.