workflow

@workflow/core@5.2.3

• security

  @pranaygp - Security: remove the nanoid dependency (pinned at 5.1.6, affected by GHSA-28wg-ghj8-5hjv). Default hook tokens now come from a built-in copy of the same generator, so tokens and seeded random values stay identical for runs recorded on earlier versions. Upgrading nanoid itself would have changed them, because nanoid@5.1.16 draws a different number of random bytes per id.

• @pranaygp - Resolve a run's encryption key from metadata only, so Run#getReadable() and resuming a hook stored without a resume context no longer make the server resolve the run's whole input and output first.

• @VaguelySerious - Fix cross-deployment start() to 4.x deployments being misread as a run invocation, and keep replayed runs in the original run's region.

• @TooTallNate - Stamp the default snapshot threshold into QuickJS runs at start() so they qualify for the event limit exemption, warn when a run approaches the event limit, and recommend WORKFLOW_VM=quickjs in MAX_EVENTS_EXCEEDED and REPLAY_TIMEOUT errors

• @VaguelySerious - Read event logs where a step's creation is recorded on its step_started alone, with no separate step_created.

• @TooTallNate - Preserve workflowVm and snapshotThreshold from executionContext in hydrateResourceIO for observability UIs

• @karthikscale3 - Avoid redundant history reads when independent hooks register concurrently. Keep existing single-event requests parallel and reconcile their committed events together before continuing, with the normal history reload when the responses do not cover a complete consecutive suffix.

• @VaguelySerious - Ignore events a step or hook writes after its outcome during replay instead of failing the run

@workflow/cli@5.0.6

• @pranaygp - Resolve a run's encryption key from metadata only when decrypting in workflow inspect --decrypt and the web UI, instead of making the server resolve the run's whole input and output for each lookup.

@workflow/world@5.0.5

• @VaguelySerious - Fix cross-deployment start() to 4.x deployments being misread as a run invocation, and keep replayed runs in the original run's region.

• @TooTallNate - Exempt QuickJS runs with a positive snapshot threshold from the default per-run event limit, matching the Vercel World

• @karthikscale3 - Avoid redundant history reads when independent hooks register concurrently. Keep existing single-event requests parallel and reconcile their committed events together before continuing, with the normal history reload when the responses do not cover a complete consecutive suffix.

@workflow/world-local@5.1.2

• @NishitaSikka - Restore missing run_created events for interrupted local run creation.

• @TooTallNate - Exempt QuickJS runs with a positive snapshot threshold from the default per-run event limit, matching the Vercel World

• @AndrewBarba - Fix an edge case in createHook({ experimental_force: true }) where a racing resumeHook could be dropped during the handover

@workflow/world-postgres@5.0.5

• @VaguelySerious - Write run_started, run_completed, run_failed, run_cancelled, attr_set, step_retrying, wait_created and wait_completed in one transaction with the entity change they record. A process that died between the two used to leave an entity in a state its event log did not explain; a completed wait with no wait_completed made the run retry the wait forever. A wait left in that state by an earlier version now gets its missing event on the next attempt.

• @dretanacr - Write a step's step_completed or step_failed row update and its event in one transaction, as step_started already does. A process that died between the two used to leave a terminal step with no terminal event, so the replayed step_started was refused as terminal and the run never advanced.

• @pranaygp - Keep a checked-out client in the World's own pg pool from ending the process when the database drops its connection.

• @pranaygp - Store timestamps as timestamp with time zone so createdAt is correct when the Postgres server or session is not in UTC. Migration 0027 converts the columns and repairs created_at values written by earlier versions.

@workflow/world-vercel@5.2.3

• @VaguelySerious - Fix cross-deployment start() to 4.x deployments being misread as a run invocation, and keep replayed runs in the original run's region.

• @karthikscale3 - Avoid redundant history reads when independent hooks register concurrently. Keep existing single-event requests parallel and reconcile their committed events together before continuing, with the normal history reload when the responses do not cover a complete consecutive suffix.

@workflow/web@5.0.6

• @pranaygp - Resolve a run's encryption key from metadata only when decrypting in workflow inspect --decrypt and the web UI, instead of making the server resolve the run's whole input and output for each lookup.

• @TooTallNate - Recommend the QuickJS VM with snapshotting on runs that failed with MAX_EVENTS_EXCEEDED or REPLAY_TIMEOUT

@workflow/web-shared@5.0.6

• @VaguelySerious - Read event logs where a step's creation is recorded on its step_started alone, with no separate step_created.

• @TooTallNate - Add getRunLimitGuidance() for runs that failed with MAX_EVENTS_EXCEEDED or REPLAY_TIMEOUT, and show the run's workflow VM in the sidebar

• @VaguelySerious - Ignore events a step or hook writes after its outcome during replay instead of failing the run

@workflow/errors@5.0.5

• @TooTallNate - Add a QuickJS snapshotting hint and docs link to MaxEventsExceededError

@workflow/core@5.2.2

• @pranaygp - Order setAttributes() resolution with the replay delivery barriers so a branch awaiting it can no longer resume out of event-log order, which could fail the run with REPLAY_DIVERGENCE or hand concurrent branches each other's step results

• @http-samc - Upgrade devalue to 6.0.3, which escapes large strings with native JSON.stringify instead of a per-character loop. Strings containing unpaired surrogates now round-trip exactly instead of being replaced with U+FFFD.

• @VaguelySerious - Remove the experimental WORKFLOW_RESILIENT_STEP_DISPATCH option

@workflow/world-vercel@5.2.2

• @mmarufov - Fix UND_ERR_REQ_CONTENT_LENGTH_MISMATCH by retrying HTTP stream appends with fresh bodies after confirmed 429s, using the shared throttle wait policy. Retry only connection-refused, DNS, and connect-timeout failures; surface other transport errors without replay.

@workflow/core@5.2.1

• @http-samc - Compress workflow arguments on the libuv threadpool when starting a run, so large inputs no longer block the caller's event loop during zstd compression.

• @http-samc - Add a compressArguments option to start(); pass false to store the run's arguments uncompressed when the inputs are known to be incompressible.

• @pranaygp - Record the queue namespace on a run's executionContext.queueNamespace in start(), so restart recovery re-enqueues the run to its own namespace.

@workflow/cli@5.0.4

• @pranaygp - Improve workflow inspect for long runs and scoped lookups: add --all for complete steps, events, and sleeps listings, print reusable cursor hints, and keep pagination on the read path that issued the cursor. Support event ID and short workflow-name lookups, require --runId for individual steps, events, and streams, fix st to mean streams, include run IDs in stream hints, and exit non-zero when a stream read fails.

@workflow/world@5.0.4

• @VaguelySerious - Add InBandSupersededError and optional inBand/expectedSeqInBand event-create params for the in-band writer fence. @workflow/world-vercel forwards them, does not retry a fenced write in-process, and records the fence on its event spans. The runtime does not use them yet.

• @pranaygp - Add optional Queue.registerRoute() and Queue.registerHandler(), implemented by @workflow/world-local and @workflow/world-postgres, so a queue consumer other than the Workflow SDK receives its own prefix's messages on its own route or in process. reenqueueActiveRuns() re-enqueues a run to the queue namespace it recorded, and @workflow/world-postgres keeps a job's queue prefix.

• @VaguelySerious - Implement the optional in-band writer fence and list snapshot in @workflow/world-local and @workflow/world-postgres, declared with the new inBandFence capability.

• @VaguelySerious - Report deliveryCount and createdAt in queue handler meta on @workflow/world-local and @workflow/world-postgres.

• @pranaygp - @workflow/world-vercel declares the inBandFence capability (its backend refuses a stale in-band write on every run with an event sequencer, and answers an in-band write on an older run with a 400), and passes on the in-band counts its backend reports: snapshot on events.list and allocated on a single events.create. The inBandFence contract lets a World list no snapshot when its refusals carry both counters, and @workflow/world-testing takes the counters from a refusal for such a World.

@workflow/world-local@5.1.1

• @pranaygp - Add optional Queue.registerRoute() and Queue.registerHandler(), implemented by @workflow/world-local and @workflow/world-postgres, so a queue consumer other than the Workflow SDK receives its own prefix's messages on its own route or in process. reenqueueActiveRuns() re-enqueues a run to the queue namespace it recorded, and @workflow/world-postgres keeps a job's queue prefix.

• @VaguelySerious - Implement the optional in-band writer fence and list snapshot in @workflow/world-local and @workflow/world-postgres, declared with the new inBandFence capability.

• @fazxes - Fix initDataDir() crashing on Bun for a new data directory: Bun's opendir() reports a missing directory on the first read, not when it is opened.

• @pranaygp - Fix in-band event count conflicts when multiple processes write events, and fix count not surviving world restarts

• @pranaygp - Register a run's stream names under a lock so streams.list no longer drops a stream when several of the run's streams first write at once. A stream written again after clear() is listed again.

• @VaguelySerious - Report a message's cumulative deliveryCount across { timeoutSeconds } redeliveries on @workflow/world-postgres.

• @VaguelySerious - Report deliveryCount and createdAt in queue handler meta on @workflow/world-local and @workflow/world-postgres.

@workflow/world-postgres@5.0.4

• @VaguelySerious - Fix a process crash when the database closes an idle connection

• @pranaygp - Add optional Queue.registerRoute() and Queue.registerHandler(), implemented by @workflow/world-local and @workflow/world-postgres, so a queue consumer other than the Workflow SDK receives its own prefix's messages on its own route or in process. reenqueueActiveRuns() re-enqueues a run to the queue namespace it recorded, and @workflow/world-postgres keeps a job's queue prefix.

• @VaguelySerious - Implement the optional in-band writer fence and list snapshot in @workflow/world-local and @workflow/world-postgres, declared with the new inBandFence capability.

• @VaguelySerious - Report a message's cumulative deliveryCount across { timeoutSeconds } redeliveries on @workflow/world-postgres.

• @pranaygp - Honor pagination.sortOrder in runs.list and steps.list, which always returned newest first.

• @VaguelySerious - Report deliveryCount and createdAt in queue handler meta on @workflow/world-local and @workflow/world-postgres.

@workflow/world-vercel@5.2.1

• @VaguelySerious - Add InBandSupersededError and optional inBand/expectedSeqInBand event-create params for the in-band writer fence. @workflow/world-vercel forwards them, does not retry a fenced write in-process, and records the fence on its event spans. The runtime does not use them yet.

• @pranaygp - @workflow/world-vercel declares the inBandFence capability (its backend refuses a stale in-band write on every run with an event sequencer, and answers an in-band write on an older run with a 400), and passes on the in-band counts its backend reports: snapshot on events.list and allocated on a single events.create. The inBandFence contract lets a World list no snapshot when its refusals carry both counters, and @workflow/world-testing takes the counters from a refusal for such a World.

@workflow/world-testing@5.0.4

• @VaguelySerious - Implement the optional in-band writer fence and list snapshot in @workflow/world-local and @workflow/world-postgres, declared with the new inBandFence capability.

• @pranaygp - @workflow/world-vercel declares the inBandFence capability (its backend refuses a stale in-band write on every run with an event sequencer, and answers an in-band write on an older run with a 400), and passes on the in-band counts its backend reports: snapshot on events.list and allocated on a single events.create. The inBandFence contract lets a World list no snapshot when its refusals carry both counters, and @workflow/world-testing takes the counters from a refusal for such a World. - workflow@5.2.1

@workflow/errors@5.0.4

• @VaguelySerious - Add InBandSupersededError and optional inBand/expectedSeqInBand event-create params for the in-band writer fence. @workflow/world-vercel forwards them, does not retry a fenced write in-process, and records the fence on its event spans. The runtime does not use them yet.

@workflow/utils@5.0.2

• @pranaygp - Add optional Queue.registerRoute() and Queue.registerHandler(), implemented by @workflow/world-local and @workflow/world-postgres, so a queue consumer other than the Workflow SDK receives its own prefix's messages on its own route or in process. reenqueueActiveRuns() re-enqueues a run to the queue namespace it recorded, and @workflow/world-postgres keeps a job's queue prefix.

@workflow/core@5.2.0

• @pranaygp - Start the queued branches of a wide Promise.all fan-out sooner. The batched fan-out now commits its writes in createBatch calls of at most 16 events instead of 32. Each queued branch's message waits on its chunk's commit and then its chunk's batched publish, and both are faster for smaller chunks. In production sweeps, the last branch of a 64-branch fan-out started about 90 ms sooner at p50, and of a 128-branch fan-out about 200 ms sooner, with no change to the first branch at those widths.

• @VaguelySerious - Deliver a hook payload to the pending await when an earlier Promise.race over the same hook was lost

• @TooTallNate - Reject lifecycle registration inside steps, preserve reporting diagnostics for unreadable errors, and document executor registration and stream-lifetime constraints.

• @TooTallNate - Enable VM-memory snapshotting by default for the QuickJS engine, with a threshold of 1000 events. Set WORKFLOW_SNAPSHOT_THRESHOLD=0 to opt out. Runs without an encryption key (for example on world-local and world-postgres) are still not snapshotted unless WORKFLOW_SNAPSHOT_ALLOW_UNENCRYPTED=1 is set.

• @VaguelySerious - Record on the invocation trace whether out-of-band events could affect each inline step boundary

• @pranaygp - Hold a turbo step's awaited step_started until the backgrounded run_started lands, so an explicit WORKFLOW_OPTIMISTIC_INLINE_START=0 no longer gets its first step rejected for a run that has not started yet.

@workflow/cli@5.0.3

• @withkarann - Ship an oclif command manifest so the CLI loads only the command being run, and skip loading the runtime on exit when no World was created. workflow --version starts in about 50 ms instead of 650 ms.

@workflow/world@5.0.3

• @karthikscale3 - Reduce redundant schema compilation by compiling final event schemas only and reusing compiled v4 event-response and paginated response schemas.

@workflow/world-local@5.1.0

• @ruiconti - Store event and step files in one directory per run for faster reads. Local run data written by earlier releases is deleted on upgrade.

• @pranaygp - Make a tagged clear() delete only its own tag's lock files, so it no longer reopens another tag's disposed hooks mid-run.

@workflow/world-vercel@5.2.0

• @alangenfeld - Stream writes now upgrade to WebSocket by default; set WORKFLOW_STREAMS_TRANSPORT=http to opt out.

• @karthikscale3 - Reduce redundant schema compilation by compiling final event schemas only and reusing compiled v4 event-response and paginated response schemas.

@workflow/web@5.0.3

• security

  @TooTallNate - Upgrade react-router and @react-router/* to 7.18.4 to address CVE-2026-34077 and GHSA-qwww-vcr4-c8h2

@workflow/swc-plugin@5.0.1

• @karthikscale3 - Stabilize nested step IDs across scopes and SWC transform modes. Self-hosted runs that already recorded one of the previously colliding IDs may report a replay divergence if they continue on upgraded code; restart those in-flight runs after upgrading.

@workflow/core@5.1.0

• @wycats - Export the existing Serializable type from @workflow/core and workflow.

• @TooTallNate - Remove the unused, stale vm-serde-bundle.generated.ts left over from the retired in-VM QuickJS serializer.

• @alangenfeld - Wait for run creation before writing to writable streams passed as arguments to turbo steps.

• @alangenfeld - Mark step_failed writes that record a failed step body with afterStepBody, so a World can keep waiting out a throttled write instead of re-running the body.

• @alangenfeld - Release drained stream writer sockets without closing the shared stream, preserve handle reuse over HTTP, and dispose transports when public writable streams abort. Propagate source failures to readers of flushable stream pipes.

• @TooTallNate - Serialize only the viewed bytes of a Float16Array (no longer the whole Node Buffer pool behind it), and support Float16Array in the QuickJS engine with the same wire format as the node:vm engine.

• @VaguelySerious - Fix cross-deployment hook resumes and starts writing zstd payloads to runs on Node.js versions that cannot decode them. Required for upgrading to a newer Node.js version while existing runs on older Node.js versions are still receiving hooks.

@workflow/world@5.0.2

• @alangenfeld - Mark step_failed writes that record a failed step body with afterStepBody, so a World can keep waiting out a throttled write instead of re-running the body.

• @alangenfeld - Release drained stream writer sockets without closing the shared stream, preserve handle reuse over HTTP, and dispose transports when public writable streams abort. Propagate source failures to readers of flushable stream pipes.

• @VaguelySerious - Fix cross-deployment hook resumes and starts writing zstd payloads to runs on Node.js versions that cannot decode them. Required for upgrading to a newer Node.js version while existing runs on older Node.js versions are still receiving hooks.

@workflow/world-local@5.0.2

• @alangenfeld - Prevent duplicate hook creation events when a retry publishes before the original token-claim owner.

@workflow/world-postgres@5.0.2

• @VaguelySerious - Reconnect the stream and run-status LISTEN connection after the database drops it, instead of crashing the process.

@workflow/world-vercel@5.1.0

• @VaguelySerious - Default the events transport to WebSockets. Set WORKFLOW_EVENTS_TRANSPORT=http to opt out.

• @pranaygp - Upgrade @vercel/queue to 0.6.0 so queue callbacks for messages that are already claimed or processed (409 / 410) respond 200 instead of logging Queue callback error and returning 500

• @alangenfeld - Wait out a throttled (429) write recording a step body's outcome until the invocation's deadline instead of 30 seconds, and resend a throttled event-log read from its cursor instead of restarting it.

• @alangenfeld - Release drained stream writer sockets without closing the shared stream, preserve handle reuse over HTTP, and dispose transports when public writable streams abort. Propagate source failures to readers of flushable stream pipes.

• @VaguelySerious - Upgrade @vercel/queue to 0.8.0 so throttled queue sends honor the server's Retry-After and queue errors carry clean messages

@workflow/web@5.0.2

• @mitul-s - De-duplicate separate shadcn button implementations between @workflow/web-shared and @workflow/web

@workflow/web-shared@5.0.2

• @mitul-s - De-duplicate separate shadcn button implementations between @workflow/web-shared and @workflow/web

@workflow/builders@5.0.2

• @VaguelySerious - Fail the build when the workflow bundle still contains require(). The workflow sandbox has no require, so an import esbuild left external (most often a Node.js builtin reached through a transitive dependency or a re-export) or an unresolved dynamic require() used to produce a bundle that threw ReferenceError: require is not defined on its first load. The build now reports the specifier, the module that imported it and the import chain back to user code. A require() inside a try/catch block, or behind a typeof require check, is left alone, since that is how packages probe for an optional dependency and the call either fails harmlessly or never runs. Set WORKFLOW_ALLOW_UNSAFE_FLOW_BUNDLE=1 to downgrade the failure to a warning.

@workflow/next@5.0.2

• @VaguelySerious - Fail the build when the workflow bundle still contains require(). The workflow sandbox has no require, so an import esbuild left external (most often a Node.js builtin reached through a transitive dependency or a re-export) or an unresolved dynamic require() used to produce a bundle that threw ReferenceError: require is not defined on its first load. The build now reports the specifier, the module that imported it and the import chain back to user code. A require() inside a try/catch block, or behind a typeof require check, is left alone, since that is how packages probe for an optional dependency and the call either fails harmlessly or never runs. Set WORKFLOW_ALLOW_UNSAFE_FLOW_BUNDLE=1 to downgrade the failure to a warning.

@workflow/utils@5.0.1

• @pranaygp - sleep(), RetryableError's retryAfter, and a hook's experimental_minRetention now throw on an Invalid Date, or on a duration that ends past the latest time a Date can hold, instead of sending the Invalid Date to the World.

@workflow/nest@5.1.0

• @pranaygp - Harden the NestJS integration against the framework features a real application uses. Fastify applications now boot correctly, queue deliveries are no longer capped at 100, app.enableVersioning() no longer moves the workflow routes, setGlobalPrefix(prefix, { exclude }) is honoured, isWorkflowRequest(context) export added, the Vercel app function is bundled for the runtime it is deployed on, --external <pkgs> is honored, and the builder warns when a second copy of @nestjs/core is present

@workflow/core@5.0.1

• @alangenfeld - Dynamic start() refusals (opt-in, validation, same-deployment, World capability, execution-context budget) are now fatal, so a start() inside a step fails fast instead of retrying.

• @alangenfeld - Dynamic start() now publishes the run only after run_created confirms its code was stored, so a refused or failed create can no longer leave an executing run behind.

• @alangenfeld - Dynamic start() validation: cap exportName at 64 characters, detect "use step" only as a real directive, and name non-async, generator, and duplicate workflow declarations.

• @VaguelySerious - Stop re-executing a running lazy or pre-claimed inline step when its message is redelivered.

• @VaguelySerious - Fix nuxt dev and other Nitro v2 dev servers failing with a Rollup "already an existing non-external module id" error, and with "Workflow world runtime was not initialized" after it.

• @TooTallNate - Add experimental threshold-based VM-memory snapshotting to the QuickJS engine via WORKFLOW_SNAPSHOT_THRESHOLD (or per-run executionContext.snapshotThreshold). Once the configured number of events has been processed since the last snapshot, suspensions persist a compressed, encrypted VM snapshot, which is later used for resumption. Runs without an encryption key are only snapshotted with WORKFLOW_SNAPSHOT_ALLOW_UNENCRYPTED=1

• @VaguelySerious - Fix webhooks returning 404 on Nitro apps running Node 24, where serializing the request's headers failed.

• @VaguelySerious - Errors with a read-only stack, such as postgres.js query errors, are now reported as thrown instead of being replaced by Cannot assign to read only property 'stack'.

@workflow/world@5.0.1

• @TooTallNate - Add experimental threshold-based VM-memory snapshotting to the QuickJS engine via WORKFLOW_SNAPSHOT_THRESHOLD (or per-run executionContext.snapshotThreshold). Once the configured number of events has been processed since the last snapshot, suspensions persist a compressed, encrypted VM snapshot, which is later used for resumption. Runs without an encryption key are only snapshotted with WORKFLOW_SNAPSHOT_ALLOW_UNENCRYPTED=1

• @TooTallNate - Add an optional, experimental experimental_snapshots storage interface for the QuickJS engine's VM-memory snapshotting (save/load/delete plus SnapshotMetadata, and encodeSnapshotEnvelope/decodeSnapshotEnvelope.

@workflow/world-local@5.0.1

• @TooTallNate - Add an optional, experimental experimental_snapshots storage interface for the QuickJS engine's VM-memory snapshotting (save/load/delete plus SnapshotMetadata, and encodeSnapshotEnvelope/decodeSnapshotEnvelope.

@workflow/world-postgres@5.0.1

• @cwdx - Add pollInterval option and WORKFLOW_POSTGRES_POLL_INTERVAL_MS env var for the Graphile Worker poll interval (default 500).

• @TooTallNate - Add an optional, experimental experimental_snapshots storage interface for the QuickJS engine's VM-memory snapshotting (save/load/delete plus SnapshotMetadata, and encodeSnapshotEnvelope/decodeSnapshotEnvelope.

@workflow/world-vercel@5.0.1

• @SandboxRD - Fix createQueueDispatcher() and createEventsDispatcher() crashing under Bun with compose is not a function. Bun resolves undici to its built-in module, whose dispatchers can't compose interceptors and are ignored by Bun's fetch, so the plain dispatcher is used instead.

• @alangenfeld - The dynamic execution-context budget error now says that step IDs and exportName count, not only aliases.

• @TooTallNate - Add an optional, experimental experimental_snapshots storage interface for the QuickJS engine's VM-memory snapshotting (save/load/delete plus SnapshotMetadata, and encodeSnapshotEnvelope/decodeSnapshotEnvelope.

• @shalabhc - Add WORKFLOW_EVENTS_TRANSPORT_WS_OVERRIDE_WORKFLOWS, a comma-separated list of workflows whose runs use the WebSocket events transport even when WORKFLOW_EVENTS_TRANSPORT is http or unset.

• @shalabhc - The WebSocket events transport now limits the size of each message. Any frame over WORKFLOW_WS_MAX_MESSAGE_BYTES (default 12 MiB, at most 16 MiB) is sent as several messages and rebuilt by the receiver. The client offers frame-parts in the x-workflow-ws-flags upgrade header so the backend can split large replies too.

• @alangenfeld - Split stream WebSocket write groups by encoded message bytes as well as chunk count, bounded by WORKFLOW_WS_MAX_MESSAGE_BYTES (default 12 MiB, at most 16 MiB).

• @alangenfeld - Retry throttled (429) stream WebSocket writes and closes after Retry-After, moving to HTTP when the socket closes or the wait budget runs out, and retry a close 5xx over HTTP instead of failing the writer.

@workflow/ai@5.0.1

• @VaguelySerious - Publish the 5.x release of @workflow/ai. The 5.0.0 on npm is an old pre-release build that requires workflow 4.x; use 5.0.1 or later with workflow 5.

@workflow/builders@5.0.1

• @withkarann - Fix directive discovery missing "use step" and "use workflow" when a backtick appears earlier in a comment or regex literal

• @VaguelySerious - Keep JSON import attributes in generated bundles, and stop bundling the builder's own serde checker into step and workflow bundles

• @VaguelySerious - Fix next dev not bundling a workflow that is created after its import, or deleted and restored

@workflow/next@5.0.1

• @VaguelySerious - Fix next dev not bundling a workflow that is created after its import, or deleted and restored

• @pranaygp - Scope the Next dev watcher to the framework module graph and stop opening a file descriptor per source file, which fixes file watch limits being reached on large projects

@workflow/nitro@5.0.1

• @VaguelySerious - Fix nuxt dev and other Nitro v2 dev servers failing with a Rollup "already an existing non-external module id" error, and with "Workflow world runtime was not initialized" after it.

Major Changes

• breaking

  ** hook.metadata on hooks returned by getHookByToken() and resumeHook() is now a lazy getter that returns a Promise, like run.returnValue, and needs to be awaited. Looking a hook up by token no longer pays for hydrating metadata that is never read.

• breaking

  Initial v5 beta release

• breaking

  Remove the deprecated experimental_setAttributes and ExperimentalSetAttributesOptions aliases; use setAttributes and SetAttributesOptions instead.

• breaking

  Remove @workflow/core/private and workflow/internal/private public subpath exports. The SWC compiler plugin no longer generates imports from these paths.

• breaking

  Remove the retired step route runtime and make queue health checks target the combined flow handler.

• breaking

  Deterministic sandbox hardening: crypto.subtle.digest in workflow functions now computes synchronously via node:crypto (byte-identical values, deterministic timing under replay); Atomics.waitAsync, async WebAssembly compilation, WeakRef, and FinalizationRegistry are no longer exposed (wall-clock timing and GC observation cannot be replayed).

Minor Changes

• Add opt-in wire-level framing for byte streams (type: 'bytes') so consumers can identify chunk boundaries — a prerequisite for transparent auto-reconnect. The framing decision is gated on a new framedByteStreams capability and recorded per-stream in the serialized ref (framing: 'framed-v1'); legacy raw streams continue to work unchanged.

• start() now delegates run ID generation to world.createRunId(options) when the world provides it, falling back to a monotonic ULID otherwise. The full options bag is passed through so worlds can read whichever fields they recognise. Adds a new region option that worlds may consume — when set, it is also forwarded onto the queue options so the initial workflow message is routed to the matching region.

• Add experimental dynamic workflows: start() accepts workflow source as a string, compiles and stores it with the run through the run-payload serialization pipeline, and replays from that stored code. Steps are exposed to the source through an explicit experimental_dynamic.steps map, which is not a security boundary: dynamic source runs with the deployment's full privileges. Off by default; a deployment opts in with WORKFLOW_EXPERIMENTAL_DYNAMIC_WORKFLOWS=1, and dynamic runs can only start on the current deployment.

• Strengthen the event-creation precondition guard: replay-context writes now also send the number of loaded events, so a snapshot that is missing an event is rejected instead of committing a divergent event log, and a rejection restarts the replay in-process (consuming the events a world may attach to the rejection) rather than re-committing the rejected payload or re-invoking over the queue. @workflow/world-local and @workflow/world-postgres do not implement the check.

• Add an opt-in optimistic-concurrency guard for event creation (WORKFLOW_PRECONDITION_GUARD=1): replay-context event creations send a stateUpdatedAt snapshot timestamp, and the runtime reloads the event log and retries (then falls back to a queue re-invocation) when the backend reports a newer out-of-band event with a 412 PreconditionFailedError.

• Surface workflowCoreVersion from the responding deployment in healthCheck() results.

• Replace hook.hasConflict (a Promise<boolean> property) with hook.getConflict(), a method returning a promise that suspends the workflow to commit hook registration and resolves with the conflicting Run when another active hook owns the token (or null once the hook is registered), without waiting for hook payload data. Code using await hook.hasConflict should migrate to const conflict = await hook.getConflict() and branch on conflict !== null.

• Lazy hook resumption: on a fast path, resumeHook() writes the hook_received event and dispatches the workflow queue message concurrently instead of sequentially, cutting a round trip off resume latency. A (runId, resumeId) dedup constraint keeps the two writers converging on exactly one event; the runtime falls back to the sequential path when dedup support is unavailable or when WORKFLOW_DISABLE_LAZY_HOOK_RESUME=1. resumeHook() resolves to a ResumedHook (a Hook plus an optional resilientResume: true flag, set only when the direct write failed transiently and the resume was recovered via the queue consumer's re-ensure) — preserving the contract introduced alongside the resilient-resume work.

• Pin correlation-ID draw order to event-log order (Node.js VM engine), so two concurrent replays of the same run assign the same IDs even when one loaded a shorter event-log prefix. Set WORKFLOW_LOG_ORDER_DRAWS=0 to opt back into arrival-order delivery resolution.

• Allow passing initial run attributes through start(), and speed up workflow-level setAttribute calls by using native events for recording attributes.

• Use custom class serialization for Run across runtime and workflow VM contexts, and add e2e coverage for Run instance boundary roundtrips

• Add WORKFLOW_TRACE_MODE with a new linked default: each workflow/step invocation span is now its own trace root with span links to the delivery and run-origin contexts, instead of one trace spanning the entire run. world-vercel now explicitly injects W3C traceparent/tracestate/baggage headers on outgoing workflow-server requests.

  Span names are also friendlier: workflow and step spans now use the short function name (e.g. workflow.execute processOrder, step.execute chargeCard, workflow.start processOrder) instead of the uppercase prefixes and full machine names (WORKFLOW_V2 workflow//./src/jobs/order//processOrder). The full name remains available in the workflow.name / step.name span attributes, and new workflowDisplayName / stepDisplayName helpers are exported from @workflow/utils.

  Behavioral changes to telemetry under the new default (set WORKFLOW_TRACE_MODE=continuous to restore the previous trace shape exactly; the span-name change applies in both modes):

  • A run no longer shares one trace ID: the trace of the request that called start() no longer contains the workflow's execution spans — navigate via span links or the workflow.run.id attribute instead.

  • Sampling decisions are made independently per invocation root (previously one parent-based decision covered the whole run), and the number of root spans/traces increases to one per invocation.

  • workflow.execute/step.execute invocation spans (formerly WORKFLOW_V2/STEP) become parentless roots, which changes parent/child-based queries and service-map edges.

  • Re-enqueued queue messages forward the original run-origin trace carrier unchanged, rather than each invocation's current context.

  • Queries or dashboards matching the old WORKFLOW_V2 .../STEP ... span names must switch to the new names.

  • The queue-delivered workflow.execute span kind changed from internal to consumer, matching the queue-delivered step.execute span (this applies in both modes).

  Existing attributes and baggage keys are unchanged, and everything remains a no-op when no OpenTelemetry SDK is registered.

• Inline execution now runs up to WORKFLOW_MAX_INLINE_STEPS (default 3) steps in parallel per suspension, each lazily created. An opt-in WORKFLOW_OPTIMISTIC_INLINE_START (default off) additionally starts step bodies before step_started is confirmed, reconciling the in-flight start before the terminal write so a lost create-claim is discarded; it is off by default because under contention a step body can run more than once (e.g. two runs writing to the workflow stream can corrupt it), so only enable it for idempotent steps.

• The WORKFLOW_PRECONDITION_GUARD event-creation guard is now on by default; opt out with WORKFLOW_PRECONDITION_GUARD=0.

• Add an experimental QuickJS WASM VM engine for workflow execution, opt-in via WORKFLOW_VM=quickjs (or per-run executionContext.workflowVm). The engine performs the same full event replay as the default node:vm engine but runs workflow code in a QuickJS VM compiled to WebAssembly, enabling platforms without node:vm support and laying the groundwork for VM-memory snapshotting.

• Rename experimental_setAttributes to setAttributes — the attributes feature is no longer experimental. The old name remains available as a deprecated alias.

• Make resumeHook() resilient to transient hook_received event write failures (429/5xx) by carrying the payload on the queue message for the runtime to materialize. Returned Hook gets a new resilientResume: true flag when this fallback path is taken.

• Retain workflow execution across inline steps within one invocation; WORKFLOW_RETAINED_VM=0 disables retention. Boundaries whose step arguments are not primitive values fall back to ordinary replay.

• Retained workflow VMs now keep the fast path when step arguments are plain data or standard built-ins (Map, Set, Date, typed arrays, URL, Headers, …), not just primitives. A boundary falls back to a normal replay only when serializing its arguments ran code the workflow controls — a getter, a proxy, a custom serializer — or computed an Error's stack trace.

• Add Run#getWritable() and run.writable to append to an existing run's stream.

• Add an allowReservedAttributes option to start() so framework-level callers can seed reserved $-prefixed run attributes at creation, matching the existing experimental_setAttributes option. The flag is carried through the resilient-start queue input so lazy run creation validates identically.

• Allow start() to be called directly inside workflow functions

• Add a catchable StreamError and classify Workflow stream infrastructure failures as STREAM_ERROR instead of USER_ERROR.

• A WritableStream from a workflow's getWritable() can now be passed as an argument to a child workflow via start(); the child's writes land on the parent run's stream directly for the full lifetime of the child run.

• Add turbo mode (on by default, disable with WORKFLOW_TURBO=0): on the first delivery of a run's first invocation the runtime backgrounds run_started, skips the initial event-log load, and forces optimistic inline start so the run reaches its first steps with no preceding network round-trips. It is safe there because the first delivery has no concurrent handler to race; turbo mode deactivates once a hook or sleep is encountered.

• Add registerLifecycleHooks from workflow/api for best-effort completion and failure reporting with a read-free workflow name, lazy Run instance, and failure cause hydrated from the persisted payload. Run metadata getters avoid resolving input/output payloads.

Patch Changes

• Add experimental_setAttributes() workflow-level helper for attaching string key/value metadata to a workflow run, surfaced as run.attributes

• Log ignored duplicate events at debug instead of info/error, so a straggler no longer prints on every replay of the run.

• Speed up Next.js workflow dev rebuilds, ignore commented imports during HMR discovery, and avoid Turbopack resolving custom-world dynamic imports.

• Embed source content in published sourcemaps.

• Fix spurious "Event cursor missing after initial load" warning

• Fix workflow/next ESM compatibility for next.config.mjs files

• Route Workflow AI examples through AI Gateway and recommend WorkflowAgent for Workflow 5.

• Fix getWritable() returning a new TransformStream per call, which caused racing pipes to reorder chunks when callers acquired a writer per write. Repeat calls within the same step now share a single pipe per (runId, namespace).

• Simplify hardened serialization: intrinsic captures assert at import instead of degrading per use, URLSearchParams now serializes natively on Node 18, and bound-function getters are reported as guest code (previously a getter defined via fn.bind() executed with an empty report).

• Serializing values built inside the node:vm workflow VM no longer executes workflow code, using engine brand checks and host intrinsics captured at boot.

• Skip the per-step incremental events.list round-trip in the inline sequential loop by consuming an event-log delta returned from the step's terminal write (gated to the single-step case with no open hooks or waits).

  Add the opt-in CreateEventParams.sinceCursor contract so a step-terminal events.create can return the event-log delta since that cursor (via EventResult.events/cursor/hasMore).

  Return the inline delta from a step-terminal write when sinceCursor is supplied, computed identically to events.list so the consumed prefix cannot skew from the server log.

  Forward sinceCursor over the v4 wire in @workflow/world-vercel so the server can return the delta on a step-terminal response; older servers ignore it and the runtime falls back to events.list.

• Lazy inline step start: the owned-inline runtime path now sends a single step_started carrying the step input, letting the world create the step on the fly and saving one round-trip per inline step.

  @workflow/world: step_started event data accepts an optional input, and EventResult gains a stepCreated ownership signal.

  @workflow/world-local: step_started with input atomically creates the step plus a synthetic step_created event; a lazy step_started for an already-existing step throws EntityConflictError so concurrent losers skip (exactly-once).

  @workflow/world-postgres: same lazy-create + exactly-once create-claim for the Postgres backend.

  @workflow/world-vercel: sends the step input on step_started over the v4 wire and threads the server's stepCreated signal into EventResult.

• Make run.input and step.input .optional() on the World snapshot schemas so consumers no longer fail validation when the service externalizes payloads as RemoteRef blobs.

• Fix a replay-determinism gap where branch wake order — and therefore step correlation ids — could depend on how much of the event log an invocation had loaded, corrupting runs under concurrent replays (CORRUPTED_EVENT_LOG).

• Memoize hydrated step return values across inline replay iterations, turning the per-invocation step-result decrypt+parse cost from O(N²) to O(N) for sequential workflows. Only primitive results are cached, so deterministic replay is preserved.

• Prepare encrypted replay payloads concurrently and cache the decrypted, decompressed representation for reuse across inline replay iterations.

• QuickJS engine performance: cache compiled WebAssembly modules process-wide, and execute steps inline in a live-VM continuation loop (no queue round-trip per step, cheap events fed before step bodies, delayed wait-continuation dispatch for racing timers).

• Rename useworkflow.dev URLs to workflow-sdk.dev

• Revert the static workflow world target injection: the world package is again resolved at runtime from WORKFLOW_TARGET_WORLD instead of being aliased into host bundles at build time.

• Add serializable AbortController and AbortSignal support across workflow and step boundaries. Workflow code can now construct an AbortController, pass signal to steps, and call abort().

  Behavior change: AbortError thrown from inside a step is now wrapped as FatalError and skips retry semantics. As a result, custom timeouts on fetch inside steps are no longer re-tried by default, and now need to be wrapped in RetryableError to preserve the old behavior.

• Fix stream IDs minted while serializing a step's arguments latching the host wall clock into a run's ID sequence, which gave every entity created afterwards a different correlation ID on each replay

• Allow experimental_setAttributes() to be called from step functions.

• security

  Bump devalue to 5.8.1 to address published security advisories.

• Merge flow and step routes into a single combined handler that executes steps inline when possible, reducing function invocations and queue overhead.

• Use inline sourcemaps for all workspace packages; published packages no longer ship external .js.map files.

• Standardize first-party World packages on createWorld(), support relative target World modules consistently, and align the Postgres World DATABASE_URL fallback with bootstrap.

• Fix world.ts being tree-shaken out of the bundle and unavailable at runtime

• Upgrade runtime validation to Zod 4.5 and enable compilation on SDK-owned Zod schemas.

@workflow/core@4.8.14

• security

  @pranaygp - Security: remove the nanoid dependency (pinned at 5.1.6, affected by GHSA-28wg-ghj8-5hjv). Default hook tokens now come from a built-in copy of the same generator, so tokens and seeded random values stay identical for runs recorded on earlier versions. Upgrading nanoid itself would have changed them, because nanoid@5.1.16 draws a different number of random bytes per id.

@workflow/cli@4.3.18

• @pranaygp - workflow inspect stream exits 1 when it cannot read or decode a stream, and workflow inspect st lists streams instead of steps.

@workflow/world-local@4.4.4

• @pranaygp - Register a run's stream names under a lock so listStreamsByRunId no longer drops a stream when several of the run's streams first write at once. A stream written again after clear() is listed again.

@workflow/world-postgres@4.3.11

• @VaguelySerious - Write run_started, run_completed, run_failed, run_cancelled, step_retrying, wait_created and wait_completed in one transaction with the entity change they record. A process that died between the two used to leave an entity in a state its event log did not explain; a completed wait with no wait_completed made the run retry the wait forever. A wait left in that state by an earlier version now gets its missing event on the next attempt.

• @dretanacr - Write a step's step_completed or step_failed row update and its event in one transaction, as step_started already does. A process that died between the two used to leave a terminal step with no terminal event, so the replayed step_started was refused as terminal and the run never advanced.

• @VaguelySerious - Fix a process crash when the database closes an idle connection

• @pranaygp - Honor pagination.sortOrder in runs.list and steps.list, which always returned newest first.

• @pranaygp - Keep a checked-out client in the World's own pg pool from ending the process when the database drops its connection.

• @pranaygp - Store timestamps as timestamp with time zone so createdAt is correct when the Postgres server or session is not in UTC. Migration 0027 converts the columns and repairs created_at values written by earlier versions.

@workflow/world-vercel@4.7.8

• @mmarufov - Fix UND_ERR_REQ_CONTENT_LENGTH_MISMATCH by retrying HTTP stream appends with fresh bodies after confirmed 429s, with a bounded Retry-After wait. Retry only connection-refused, DNS, and connect-timeout failures; surface other transport errors without replay.

@workflow/core@4.8.13

• @shalabhc - Accept lazy completed and failed runs whose payload is represented by a remote reference.

@workflow/world@4.5.1

• @shalabhc - Accept lazy completed and failed runs whose payload is represented by a remote reference.

@workflow/world-local@4.4.3

• @pranaygp - Make a tagged clear() delete only its own tag's lock files, so it no longer reopens another tag's disposed hooks mid-run.

@workflow/web@4.1.29

• security

  @TooTallNate - Upgrade react-router and @react-router/* to 7.18.4 to address CVE-2026-34077 and GHSA-qwww-vcr4-c8h2

@workflow/core@4.8.12

• @VaguelySerious - Fix an issue with the workflow's deterministic clock tracking advancement on consumption, not on write, which could lead to a determinism issue when concurrent replays called Date.now with different amounts of events read from the log

• @gaojude - Order hook registration and conflict settlements with earlier workflow deliveries so replay preserves concurrent step correlation IDs.

@workflow/world-postgres@4.3.9

• @AndrewBarba - Allow same-run queue wakes to reach a workflow while an inline step is pending, so hook-driven cancellation and steering do not wait for the step to finish.

• @pranaygp - Document that this is a reference implementation whose workflow HTTP routes are unauthenticated, and how to restrict them.

@workflow/world-vercel@4.7.6

• @SandboxRD - Fix createQueueDispatcher() and createEventsDispatcher() crashing under Bun with compose is not a function. Bun resolves undici to its built-in module, whose dispatchers can't compose interceptors and are ignored by Bun's fetch, so the plain dispatcher is used instead.

• @alangenfeld - Honor server-provided Retry-After delays when scheduling queue redeliveries.

@workflow/builders@4.1.16

• @withkarann - Fix directive discovery missing "use step" and "use workflow" when a backtick appears earlier in a comment or regex literal

• @VaguelySerious - Keep JSON import attributes in generated bundles, and stop bundling the builder's own serde checker into step and workflow bundles

@workflow/vite@4.0.26

• @alangenfeld - Rebuild workflows once per file change instead of once per Vite environment.

@workflow/core@4.8.10

• @pranaygp - Fix an unhandled rejection that could exit the process when the encryption-key lookup for a forwarded writable stream failed (for example a run metadata read that timed out) before anything was written to that stream. The lookup now starts on the first write, and a failure rejects that stream instead.

• @TooTallNate - Serialize a DataView as the bytes it views. It previously fell through to devalue's built-in encoding, which persists the whole backing ArrayBuffer — for a view onto Node's pooled Buffer allocator, unrelated process memory.

• @TooTallNate - Retry the max-deliveries run_failed/step_failed write and the step handler's workflow re-queue through queue redelivery when they fail transiently (429, 5xx, transport) instead of acking and leaving the run stuck running.

• @pranaygp - start() with an explicit deploymentId (and so recreateRunFromExisting, i.e. Replay Run) no longer fails in a process that is not itself a deployment; it takes the cross-deployment path instead.

• @pranaygp - Mark WorkflowRunFailedError and WorkflowRunCancelledError as non-retryable, and make FatalError.is() honor the fatal marker, so a step that reads a terminal run's returnValue fails on its first attempt with the error intact instead of exhausting its retry budget first.

• @karthikscale3 - Route unrecognized backend connection and stream failures through existing retry policies, rebuilding shared event connections after repeated HTTP/2 failures. Keep invalid backend URLs, blocked ports, and unsupported request headers out of those retries. Include error cause chains in run-failure logs to expose underlying socket, DNS, and TLS failures.

@workflow/cli@4.3.14

• @TooTallNate - Add a DataView reviver so run payloads containing one still hydrate in the observability UI and CLI. </content>

@workflow/world-local@4.4.2

• @pranaygp - Upgrade undici to 7.30.0, which stops a failed HTTP/2 stream from leaving a phantom in-flight request on its connection.

@workflow/world-postgres@4.3.8

• @pranaygp - Upgrade undici to 7.30.0, which stops a failed HTTP/2 stream from leaving a phantom in-flight request on its connection.

@workflow/world-vercel@4.7.5

• @pranaygp - Send event requests with bodies too large to re-buffer over HTTP/1.1 so they no longer stall behind in-flight HTTP/2 streams, and retry event-log reads whose HTTP/2 stream the peer reset.

• @karthikscale3 - Route unrecognized backend connection and stream failures through existing retry policies, rebuilding shared event connections after repeated HTTP/2 failures. Keep invalid backend URLs, blocked ports, and unsupported request headers out of those retries. Include error cause chains in run-failure logs to expose underlying socket, DNS, and TLS failures.

• @pranaygp - Upgrade undici to 7.30.0, which stops a failed HTTP/2 stream from leaving a phantom in-flight request on its connection.

@workflow/web-shared@4.1.26

• @TooTallNate - Add a DataView reviver so run payloads containing one still hydrate in the observability UI and CLI. </content>

@workflow/next@4.1.14

• @pranaygp - Fix next dev sometimes running the workflow build (and starting a second watcher) twice on startup when Next.js resets process.env between next.config evaluations

@workflow/errors@4.2.2

• @pranaygp - Mark WorkflowRunFailedError and WorkflowRunCancelledError as non-retryable, and make FatalError.is() honor the fatal marker, so a step that reads a terminal run's returnValue fails on its first attempt with the error intact instead of exhausting its retry budget first.

@workflow/core@4.8.9

• @ctgowrie - Stop framed stream reconnects after the consumer cancels, including while completion checks or reconnect acquisition are still pending.

• security

  @Rich-Harris - Bump devalue to 5.9.2 to address published security advisory.

• @torsello - Allow parentheses and square brackets in workflow names

@workflow/world-postgres@4.3.7

• @komly - Fix run creation and slot allocation not happening transactionally

• @komly - Queue deliveries no longer inherit fetch's 300s headers/body deadlines, which redelivered healthy long-running inline work while it was still executing. Deadlines can be set with WORKFLOW_POSTGRES_HEADERS_TIMEOUT_MS and WORKFLOW_POSTGRES_BODY_TIMEOUT_MS.

• @komly - Preserve error names, messages, stacks, and nested causes in Graphile Worker log metadata.

• @komly - Fix stream cursors not being advanced for skipped chunks, so repeated notifications cannot consume the requested start offset twice

• @VaguelySerious - Fix stream readers leaking EventEmitter listeners on EOF, initial query failure, and World close, and fail pending readers when the World is closed.

• @himself65 - Fix readFromStream erroring when rows were written after the stream's first EOF marker. Rows past the first EOF are now ignored consistently across readFromStream(), getStreamChunks(), and getStreamInfo().

@workflow/builders@4.1.14

• @TooTallNate - Deduplicate identical pnpm peer-variant copies of a package instead of failing the build with a duplicate step/workflow ID error. Different implementations that map to the same ID still fail.

@workflow/world-local@4.4.1

• @AndrewBarba - Abort active local queue deliveries when the World closes, including when transport timeouts are disabled.

@workflow/world-vercel@4.7.2

• @karthikscale3 - Recycle the events HTTP/2 connection pool when a streamed response body fails after its headers arrive.

• @VaguelySerious - Authenticate deploymentId: "latest" with the deployment's own OIDC identity instead of an ambient VERCEL_TOKEN, and scope the request to the configured team, fixing spurious 404s when resolving the latest deployment

@workflow/web-shared@4.1.22

• @karthikscale3 - Prevent event payload flicker during decryption.

@workflow/swc-plugin@4.1.3

• @TooTallNate - Register class expressions through an IIFE that closes over the class instead of module-level code that references it by name, fixing the unresolvable AnonymousClass reference emitted for shapes such as var Foo = class { ... } in pre-bundled packages.

@workflow/next@4.1.10

• @NathanColosimo - Discover workflows imported by every Next.js file convention entrypoint.

@workflow/core@4.8.5

• @TooTallNate - Step-argument serialization failures now fail the step with a catchable error (via a step_failed event, like a step-body failure) instead of failing the run from outside the workflow, and when uncaught they fail the run immediately as a USER_ERROR rather than retrying until max queue deliveries.

• @AndrewBarba - Preserve event-log order when adjacent hook payloads resume independent workflow branches.

• @VaguelySerious - Fix a replay-determinism gap where branch wake order — and therefore step correlation ids — could depend on how much of the event log an invocation had loaded, corrupting runs under concurrent replays (CORRUPTED_EVENT_LOG).

• @alangenfeld - Verify stream tail when the reconnecting framed reader receives an EOF to ensure it is due to stream completion (and not other world-side behavior)

@workflow/world@4.5.0

• @karthikscale3 - Accept and send an optional runId when listing events by correlation ID

• @ctgowrie - Use the active queue namespace when re-enqueuing workflow runs during world startup recovery.

@workflow/world-local@4.4.0

• @karthikscale3 - Add WORKFLOW_LOCAL_RECOVER_ACTIVE_RUNS env var as a fallback for the recoverActiveRuns option, so re-enqueueing of pending/running runs on startup can be disabled without a custom world module.

• @ctgowrie - Use the active queue namespace when re-enqueuing workflow runs during world startup recovery.

• @pranaygp - Reduce local sequential-step replay I/O with bounded recent-event and storage-directory caches.

• @karthikscale3 - Accept and send an optional runId when listing events by correlation ID

@workflow/world-postgres@4.3.5

• @shin4141 - Commit step entities and their step_created events atomically.

• @ctgowrie - Use the active queue namespace when re-enqueuing workflow runs during world startup recovery.

• @joeyhotz - Throw EntityConflictError when a run_created event targets a run that already exists, instead of resolving with no run. This matches world-local and world-vercel, and stops start() from throwing Missing 'run' in server response for 'run_created' event when the resilient start path wins the race.

• @karthikscale3 - Accept and send an optional runId when listing events by correlation ID

@workflow/world-vercel@4.7.1

• @karthikscale3 - Accept and send an optional runId when listing events by correlation ID

@workflow/web@4.1.21

• @karthikscale3 - Accept and send an optional runId when listing events by correlation ID

@workflow/ai@4.2.1

• @VaguelySerious - Reduce DurableAgent step boundary payload by reconstructing StepResult outside the step instead of inside.

@workflow/nitro@4.1.11

• @NathanColosimo - Clean up temporary Nitro Vite servers and Workflow build contexts after builds.

@workflow/world@4.4.0

• @VaguelySerious - Add WORKFLOW_NODE_HTTP to run the Vercel and Local World HTTP layers on Node's built-in node:http and node:https modules instead of their usual HTTP client library.

@workflow/world-local@4.3.0

• @VaguelySerious - Add WORKFLOW_NODE_HTTP to run the Vercel and Local World HTTP layers on Node's built-in node:http and node:https modules instead of their usual HTTP client library.

@workflow/world-postgres@4.3.4

• @VaguelySerious - Reject a hook resume that races the hook's disposal, instead of journaling it after hook_disposed and corrupting the owning run's event log.

@workflow/world-vercel@4.7.0

• @VaguelySerious - Add WORKFLOW_NODE_HTTP to run the Vercel and Local World HTTP layers on Node's built-in node:http and node:https modules instead of their usual HTTP client library.

@workflow/core@4.8.0

• @shalabhc - Enforce a server-supplied per-run event limit (default 25K)

@workflow/world@4.3.1

• @shalabhc - Enforce a server-supplied per-run event limit (default 25K)

@workflow/world-local@4.2.4

• @AndrewBarba - Retry transient EPERM unlink failures on Windows when deleting hook and entity JSON files. A concurrent reader briefly holding a file open made deleteJSON throw a share-violation EPERM, which surfaced as a failed operation — for example a failed run.cancel() while deleteAllHooksForRun raced hook polling.

@workflow/world-vercel@4.6.1

• @VaguelySerious - Cancel the v4 event frame stream when a reader stops early, so the response body's undici connection returns to the pool instead of leaking.

• @shalabhc - Enforce a server-supplied per-run event limit (default 25K)

• @VaguelySerious - Event-log requests now multiplex over a single HTTP/2 connection instead of opening one connection per in-flight request

• @VaguelySerious - Raise HTTP/2 receive windows on the events client, slightly decreasing read latency

• @VaguelySerious - Retry stream close on retriable 5xx. Close is idempotent on the server (unlike chunk appends, which keep their no-5xx retry policy), and the server may return retriable 503s expecting the writer to close again.

• @VaguelySerious - Enable HTTP/2 for the events API and stream write requests.

• @VaguelySerious - Unify HTTP request handling into a shared core, extending OTEL client spans, trace-context propagation, and DEBUG logging to the v4 events, stream, and Vercel API request paths

@workflow/web@4.1.16

• @VaguelySerious - Fix run and event observability pages hanging (~16s) and showing no data in the bundled server build, caused by HTTP/2 requests failing to reach node:http2.

@workflow/nitro@4.1.6

• @VaguelySerious - Fix HTTP/2 requests failing in production builds (Vite/Nitro, TanStack Start) where undici's bundled node:http2 could not load and fell back to a stub.

@workflow/sveltekit@4.0.15

• @VaguelySerious - Fix HTTP/2 requests failing in SvelteKit production builds (undici's bundled node:http2 could not load, falling back to a stub).

@workflow/errors@4.2.1

• @shalabhc - Enforce a server-supplied per-run event limit (default 25K)

@workflow/core@4.7.0

• @VaguelySerious - Add an optimistic-concurrency guard for event creation (on by default; opt out with WORKFLOW_PRECONDITION_GUARD=0): replay-context event creations send a stateUpdatedAt snapshot timestamp, and the runtime reloads the event log and retries (then falls back to a fresh re-invocation) when the backend reports a newer out-of-band event with a 412 PreconditionFailedError. Backends without guard support ignore the snapshot, so this is backward-compatible and fails open.

@workflow/world@4.3.0

• @VaguelySerious - Add an optimistic-concurrency guard for event creation (on by default; opt out with WORKFLOW_PRECONDITION_GUARD=0): replay-context event creations send a stateUpdatedAt snapshot timestamp, and the runtime reloads the event log and retries (then falls back to a fresh re-invocation) when the backend reports a newer out-of-band event with a 412 PreconditionFailedError. Backends without guard support ignore the snapshot, so this is backward-compatible and fails open.

@workflow/world-vercel@4.6.0

• @VaguelySerious - Add an optimistic-concurrency guard for event creation (on by default; opt out with WORKFLOW_PRECONDITION_GUARD=0): replay-context event creations send a stateUpdatedAt snapshot timestamp, and the runtime reloads the event log and retries (then falls back to a fresh re-invocation) when the backend reports a newer out-of-band event with a 412 PreconditionFailedError. Backends without guard support ignore the snapshot, so this is backward-compatible and fails open.

• @VaguelySerious - Bound HTTP requests with a 30s headers/body timeout instead of undici's 5-minute default, and retry transport timeouts and other transient failures in-process, so a stalled connection no longer leaves a run silently paused for ~5 minutes until the queue redelivers it. Override with WORKFLOW_VERCEL_HEADERS_TIMEOUT_MS / WORKFLOW_VERCEL_BODY_TIMEOUT_MS.

@workflow/world-testing@4.1.14

• @VaguelySerious - Fix flaky Local World tests caused by concurrent test servers sharing one data directory and re-enqueuing each other's in-flight runs on startup

@workflow/sveltekit@4.0.14

• @TooTallNate - The generated webhook route now passes the incoming request directly to its handler instead of copying it into a new Request, so the request body is no longer read before the handler runs.

@workflow/astro@4.0.14

• @TooTallNate - The generated webhook route now passes the incoming request directly to its handler instead of copying it into a new Request, so the request body is no longer read before the handler runs.

@workflow/errors@4.2.0

• @VaguelySerious - Add an optimistic-concurrency guard for event creation (on by default; opt out with WORKFLOW_PRECONDITION_GUARD=0): replay-context event creations send a stateUpdatedAt snapshot timestamp, and the runtime reloads the event log and retries (then falls back to a fresh re-invocation) when the backend reports a newer out-of-band event with a 412 PreconditionFailedError. Backends without guard support ignore the snapshot, so this is backward-compatible and fails open.