NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
Packagist
Packagist, #3244 most downloaded on Packagist
Embedded durable workflow runtime and orchestration engine for Laravel applications.
Latest stable
2.28.1
Last release 8 days ago
14 Nov 2026
Ships on a steady schedule
a new release about every 8 days
Nearly every release is documented
notes for 60 of the last 60 stable releases
Nothing withdrawn
no release was ever pulled
5 years old
633 releases · first in 2022
Drag across the ruler to compare, or tap a release to open it.
This security patch requires Laravel 12.72.0 or later in 12.x , or 13.41.0 or later in 13.x . Update your application lockfile and run Composer audit
This security patch requires Laravel 12.72.0 or later in 12.x, or 13.41.0 or later in 13.x. Update your application lockfile and run Composer audit when upgrading Workflow.
These minimums include fixes for Concurrency process task-output injection, temporary signed URL expiration bypass, and temporary signed URL expiration bypass through unencoded numeric-keyed route parameters.
Workflow runtime behavior and schema are unchanged from 2.28.0. Current example and qualification lockfiles are refreshed, with the HTTP example preserving its declared PHP 8.3 support. See #1133 and #1141.
Behavior change: an un-awaited select() loser is cancelled when the run closes. To keep a service-mode activity running, schedule it with cancellation
Your coding agent can read these notes before it upgrades. Set up the MCP server →
Behavior change: an un-awaited select() loser is cancelled when the run
closes. To keep a service-mode activity running, schedule it with
cancellation_policy: abandon; its outcome is then never recorded. See the
first entry below. No migration.
A run no longer leaves work running after its own code closes it (#1117).
Before, an activity or timer the workflow never awaited, such as a
select() loser, stayed open when the run completed, failed or continued as
new. The activity still executed, and its ActivityStarted and outcome were
appended to the closed run's history after the terminal event. Now the close
records ActivityCancelled for each open activity and TimerCancelled for
each pending timer in the same transaction, before WorkflowCompleted,
WorkflowFailed or WorkflowContinuedAsNew, and cancels their tasks. Such a
cancellation carries cancellation_reason: run_closed, and replay ignores
it, because the workflow never observed it: it never changes a selection
winner or a parallel group's failure. A
late claim, heartbeat or outcome for such an activity is refused with
run_closed, including one left behind by a run closed before upgrading.
Nothing is appended to the closed run. Exception: a service-mode activity
scheduled with cancellation_policy: abandon keeps running. The terminal
event lists it under the new abandoned_activities payload key, and its
heartbeats and outcome are accepted but not appended. Cancel, terminate,
timeout and cooperative cancellation already settled open work this way.
Children keep their parent-close policy and continue-as-new hand-over.
Behavior change: if a workflow needs a loser's effect, await the loser
before returning.
Replaying a workflow that repeats the same message-stream send no longer
lists every earlier matching row again for each send (#1014). The exactly
matched candidates for a send identity are now loaded once per execution
and extended with the rows that execution writes, so N identical sends in
one replay cost one candidate query instead of N, and O(N) rows instead of
O(N²). Matching is unchanged: exact field comparison, send order and the
redrive chain.
An expired open activity row of a closed run is now closed instead of being
selected and skipped on every timeout sweep (#1117 follow-up). An activity
that its run's close left running under the abandon policy is marked failed
when its own schedule_to_close deadline passes without an outcome. Any
other open row of a closed run, such as one left by a close before
upgrading, is cancelled. Its task is cancelled and nothing is appended to
the closed run's history. Before, such rows stayed open and could fill the
sweep's batch limit.
Closing a child no longer reads every sibling's history under the parent lock ( #892 ). Before, each child close loaded the full history of every othe
Closing a child no longer reads every sibling's history under the parent
lock (#892). Before, each child close loaded the full history of every
other child of its parent to record the outcome, check a parallel group
and re-project the parent, so a wide fan-out did quadratic history reads
while blocking the parent's own workflow task. Now a close reads the
closing child's history and the parent's own history and links. A
sibling's outcome comes from the parent's resolution event, or from the
sibling's terminal event when none is recorded yet. Measured with 50
children, closing all of them read 96,052 sibling history rows on the
embedded path and 26,460 through the service-mode bridge for a parallel
group; both now read none, and history queries per close no longer grow
with the number of children. Outcomes, wake-ups and projections are
unchanged.
A child close now costs the parent the same number of queries however many
siblings it has (#892). The parent re-projection reuses the child runs it
batch-loads with its child links instead of fetching each sibling again for
every wait, lineage and summary lookup, and it reads the parent's lineage
rows in one query. Embedded, queries per close went from 218, 372 and 790
at 2, 10 and 30 children to 172, 158 and 156. Through the service-mode
bridge (parallel group) they went from 153, 333 and 755 to 130, 141 and 143.
The close that completes a parallel group still records every member's
resolution once. Projections are unchanged.
Embedded task repair and execution now stay inside the app's namespaces
(#889). Before, apps sharing one database with the same connection and
queue names could repair, redispatch, time out and run each other's tasks
with the wrong app's code and credentials. The watchdog (including the
queue worker's wake and workflow:v2:repair-pass), existing and missing
task repair, run deadlines, activity timeouts, service-call deadline,
dispatch and linked-call repair, workflow:v2:schedule-tick, and the
RunWorkflowTask,
RunActivityTask, RunTimerTask and RunServiceCall jobs now act only on
the new workflows.v2.namespaces allowlist (DW_V2_NAMESPACES). Unset, it
is exactly workflows.v2.namespace, so a single-namespace app behaves as
before. A comma-separated list serves each listed namespace (default is
the unscoped one); an empty value serves none; * is refused. A queued
job outside the allowlist, including one queued before upgrading, is
skipped before any effect and logged as a warning. Serving every namespace
needs the separate workflows.v2.all_namespaces (DW_V2_ALL_NAMESPACES)
opt-in. Upgrade note: an embedded app that starts, schedules or signals
workflows in namespaces other than DW_V2_NAMESPACE must list them in
DW_V2_NAMESPACES, or their tasks and schedules wait for an app that
serves them. A process that repairs or ticks every namespace, such as the
standalone server, must set DW_V2_ALL_NAMESPACES=true. The allowlist is an
execution-safety boundary, so one app never runs another app's work with the
wrong code or credentials. It is not tenant isolation: apps that share one
database can still read each other's rows.
Work refused by the namespace allowlist is now counted, so a misconfigured
upgrade no longer stalls tasks with a log line as the only sign (#1112).
The operator metrics snapshot has a new namespace_scope_refusals block
with the app's allowlist and refusals per namespace and path over the last
15 minutes: job counts each skipped queued job, and repair, timeout
and schedule_tick count repair candidates, expired activity and run
deadlines, and due schedules that a pass left in a namespace outside the
allowlist after they had waited more than two minutes (sampled at most once
a minute per path). The new advisory namespace_scope_refusals health check
warns while refusals keep arriving and clears 15 minutes after the last one.
The skipped-job warning is now logged once per namespace a minute; the rest
are counted. Counts live in the default cache store in one-minute buckets,
with no table and no migration. Use a shared store (Redis, database or
Memcached) so the web process that serves the health snapshot sees the
workers' counts.
Upgrade: no migration. If an embedded app starts, schedules or signals workflows in namespaces other than DW_V2_NAMESPACE, list them in DW_V2_NAMESPACES. A process that repairs or ticks every namespace sets DW_V2_ALL_NAMESPACES=true.
A child handed to a continued run now gets its retries ( #1105 ). Before, a child that failed after its parent continued as new was never retried, bec
A child handed to a continued run now gets its retries (#1105). Before, a
child that failed after its parent continued as new was never retried,
because the retry needed the parent to be waiting. It is now retried under
the retry policy the parent recorded, with the same child call identity and
attempt count, while a run of the chain still supervises it. The retry run's
WorkflowStarted names the original parent run, and the supervision moves to
the retry run, so the chain's final run applies the parent-close policy to it.
Nothing is appended to the closed predecessor. Once the attempts are spent,
the failure is non-retryable, or the chain has closed, the failure is final.
Embedded and service-mode runs behave the same. No migration.
Failed-run redrive now reuses a completed child workflow call (#864).
Before, any child workflow in the failed run's history refused redrive with
unsupported_history. The successor records the child's
ChildWorkflowScheduled, ChildRunStarted and ChildRunCompleted events
with reused_from_run_id and reused_recorded_at, keeps the original child
instance, run, link id and call id, and does not start the child again or
write a child link of its own. Executor, query and fiber replay resume the
call with the recorded output at the time it first closed, so now() after
the call matches the failed run. Redrive chains carry the reuse forward.
Child steps that replay would not reproduce are refused before any write,
with new reasons: unsupported_open_child, unsupported_child_outcome
(failed, cancelled or terminated children), unsupported_child_workflow
(parallel, scoped, retried or continued children, or mismatched call
identity), external_child_result_not_supported and
unavailable_child_result. Runs a service worker started stay refused with
service_redrive_history_not_qualified when they would reuse a child step.
Runs that are themselves children are still refused. No migration.
Rollback: earlier releases ignore reused_recorded_at on child events, so
replaying such a successor there can give a different workflow time, and
deadlines derived from it, after the reused call; they also refuse to
redrive it again (unsupported_history). Once a successor with a reused
child call exists, keep workers on this release or later, or confirm that
no such history can replay before rolling back.
Workflow code can read the history budget replay-safely with
historyBudgetAtStep() and shouldContinueAsNewAtStep() (#859). They count
only the history recorded before the current workflow step, so the same line
returns the same value on every replay, and a loop that checks the budget
each iteration continues as new once, at the first iteration that crosses the
threshold. The existing historyLength(), historySize(), historyFanOut(),
shouldContinueAsNew() and historyBudgetPressure() still read the whole
current history and are unchanged, so runs that already decided with them
keep replaying. Inside a loop they can answer differently on replay; use the
new methods there.
Upgrade migrations that fill in existing rows no longer lock the table for
the whole backfill on PostgreSQL (#871). The output codec (2026_07_13),
local activity exception (2026_09_24), schedule occurrence (2026_09_29)
and service call deadline (2026_10_11) migrations now commit the new
column first, with a one-second lock wait, and then fill rows in small
batches that each commit on their own. The 2026_09_29 and 2026_10_11
indexes are now built after the backfill without blocking writes
(CREATE INDEX CONCURRENTLY on PostgreSQL, LOCK=NONE on MySQL and
MariaDB). An interrupted run keeps the batches it finished and can be run
again. The finished schema and data are the same as before, and installs
that already ran these migrations are not affected.
A parent that continues as new while children it started are still open
now hands them to the continued run (#868). Embedded and service-mode
continue-as-new record one workflow_child_supervisions row per open child,
in the same transaction. The child link, child call, sequence, deadlines and
parent-close policy stay as recorded. When the chain's final run completes,
fails, is cancelled, is terminated or times out, it applies each child's
request_cancel, request_cancellation or terminate policy and records
the outcome in its own history. abandon stays abandon. Before, these
children kept running and their policy was never applied. Nothing is
appended to the closed predecessor, and the continued run gets no handle on
the children. The unused ChildCallService::transferChildCallsToContinuedRun()
is removed: it re-keyed child calls to the continued run, where their
sequences could collide with its own children. Run the new migration.
An activity outcome no longer locks the attempt row after it holds the task,
execution and run (#1000). It closes the attempt with one guarded conditional
update instead. Lease renewal locks the attempt first, so the old order could
deadlock a renewal against an outcome for the same attempt on PostgreSQL and
MySQL, mainly on the stale, cancelled and terminated paths.
Changing a schedule's action now drops fires buffered under the old action ( #876 ). Each dropped fire is recorded as a ScheduleTriggerSkipped event w
ScheduleTriggerSkippedaction_changed and its occurrence_time. Updates thatStructuralLimitExceededException instead ofInvalidArgumentException. Its new limitName says which limit was hit,structural_limit_name. An embedded upsertmemo_size_bytes andsearch_attribute_size_bytes limits. A configured limit can tighten amemo_size_bytesInvalidArgumentException for theseStructuralLimitExceededException.Heartbeats no longer grow run history with every call ( #870 ). Each heartbeat still renews the lease, moves the heartbeat deadline and stores the lat
ActivityHeartbeatRecorded event is written only when progress changed since the last recorded event and at least max(heartbeat timeout, 60 seconds) has passed. Heartbeats without progress write none, and progress not yet recorded is written once when the attempt closes. Run detail, Waterline and history export read the latest progress from the attempt. Liveness, timeouts and retries are unchanged, and older histories replay and export as before. Expect far fewer heartbeat events.changed_fields lists only real changes, and only a spec or jitter change moves the next fire. Creating a schedule with a taken id, including a deleted or max_runs_exhausted one, raises Workflow\V2\Exceptions\ScheduleAlreadyExistsException, matching the Server's 409 schedule_already_exists.runs.replay_blocked, oldest_replay_blocked_at and max_replay_blocked_age_ms count open runs whose workflow task cannot replay, and a new advisory replay_blocked_runs health check warns while any exist.Run migrations. A new migration adds nullable last_heartbeat_progress, last_heartbeat_progress_at and last_heartbeat_recorded_at columns to activity_attempts. It is additive, with no backfill. Dependency floors are unchanged.
ActivityHeartbeatRecorded event is written only when progress changed
since the last recorded event and at least max(heartbeat timeout, 60
seconds) has passed. Heartbeats without progress write none. Progress that
was not yet recorded is recorded once when the attempt closes. Run detail,
Waterline and history export read the latest progress from the attempt.
Liveness, timeouts and retries are unchanged, and older histories with one
event per heartbeat replay and export as before. This is a behavior change:
expect far fewer heartbeat events.last_heartbeat_progress,
last_heartbeat_progress_at and last_heartbeat_recorded_at columns to
activity_attempts. It is additive, with no backfill.changed_fields lists only fields whose stored value
changed. Only a spec or jitter change moves the next fire, so
metadata-only updates cannot skip a due occurrence (#876, #1089).max_runs_exhausted schedule, raises
Workflow\V2\Exceptions\ScheduleAlreadyExistsException instead of a raw
database error, matching the Server's 409 schedule_already_exists
(#876, #1090).runs.replay_blocked, oldest_replay_blocked_at,
max_replay_blocked_age_ms), and a new advisory replay_blocked_runs
health check warns while any exist (#861, #1091).Service-mode size limits: commands from SDK workers now get the remaining size limits that embedded workflows already apply. Side-effect results, loca
No database migration, history format or dependency floor change.
Saga compensation and cancellation: embedded compensate() now runs under a cancellation shield, as the PHP and Python SDK sagas do. A cooperative canc
compensate() now runs under a cancellation shield, as the PHP and Python SDK sagas do. A cooperative cancellation request that arrives during compensation no longer skips the remaining compensations. Cleanup deadlines, termination and run timeouts still apply (#872).maxAttempts below 1 now fails the run with a clear message, because the budget counts the first attempt. Previously it silently ran once. For unlimited attempts, omit maxAttempts and set the activity class $tries to 0. Histories that recorded maxAttempts: 0 under an earlier release keep replaying after the upgrade (#860).new DateTime(), new DateTimeImmutable(), new Carbon() and new CarbonImmutable() when they read the current time, plus today(), shuffle, array_rand and str_shuffle. Fixed dates are not flagged (#888).No database migration, history format or dependency floor change.
maxAttempts below 1 is refused when a new activity or
local activity is scheduled, and fails the run with a message that the
budget counts the first attempt (#860). Previously it ran one attempt
silently. For unlimited attempts, omit maxAttempts and set the activity
class $tries to 0.maxAttempts: 0 under an earlier release keep
replaying after an upgrade: replay accepts that recorded budget when the
code still passes 0, and it means one attempt. Before this change such runs
failed replay with a retry policy mismatch.compensate() runs under a cancellation shield, as the PHP
and Python SDK sagas do. A cooperative cancellation request that arrives
during compensation no longer skips the remaining compensations. The cleanup
deadline, termination and run timeouts still apply (#872, #1077).new DateTime(), new DateTimeImmutable(), new Carbon() and new CarbonImmutable() when they
read the current time, plus today(), shuffle, array_rand and
str_shuffle. Fixed dates are not flagged (#888, #1080).Laravel Cloud queues: timers, awaitWithTimeout() and delayed v2 tasks longer than 15 minutes failed on Laravel Cloud's managed queues with DelaySecond
awaitWithTimeout() and delayed v2 tasks longer than 15 minutes failed on Laravel Cloud's managed queues with DelaySeconds must be >= 0 and <= 900. The cloud queue driver wraps Amazon SQS, but the 900-second delay cap only applied to the sqs driver. Long delays on cloud now wait in steps of at most 900 seconds, as on plain SQS. Reported and fixed for 1.x by @Joel-Jensen (#1072, released in 1.1.1); carried to 2.x in #1074.inbox()->receive() returned nothing (#1024). Live update handlers now record the message ids each receive returned on their UpdateApplied event, and queries and replays return exactly those messages (#1076). Updates applied by earlier releases keep the previous lookup.Workflow\V2\StateMachine with native Saga integration (#1071, contributed by @chrom). See docs/business-state-machines.md. The API is marked experimental and may change in a minor release.No database migration or dependency floor change. UpdateApplied history gains an optional message_receives field. Older runtimes ignore it and fall back to the previous lookup.
cloud
queue driver wraps Amazon SQS, which rejects delays over 900 seconds, but
the cap only applied to the sqs driver. On Laravel Cloud, timers,
awaitWithTimeout() and delayed v2 tasks longer than 15 minutes failed with
DelaySeconds must be >= 0 and <= 900. They now wait in steps of at most
900 seconds, as on plain SQS (#1074, reported in #1072).inbox()->receive() at the wrong step and saw
nothing (#1024). Live update handlers now record the message ids each
receive returned on their UpdateApplied event, and replays return exactly
those messages. Updates applied by earlier releases keep the previous
lookup (#1076).Workflow\V2\StateMachine with native Saga integration (#1071, see
docs/business-state-machines.md). The API is marked experimental and may
change in a minor release.UpdateApplied history
gains an optional message_receives field. Older runtimes ignore it.Opt-in occurrence deadlines keep the first timer's exact deadline through false-wake reopens. The original timeout budget stays attached to the logica
Opt-in occurrence deadlines keep the first timer's exact deadline through false-wake reopens. The original timeout budget stays attached to the logical chain. Conflicting, timed-out or unprovable chains refuse with whole-completion rollback. Legacy chains retain their per-open behavior.
The embedded worker protocol advertises 1.22 and still accepts 1.20 requests. There is no database migration or support-floor change: PHP ^8.2 and Laravel ^12.69.0 or ^13.30.0.
After a workflow has used this policy, roll back to an older Core version only by restoring a pre-upgrade backup. Older normalizers can discard the policy and create a mixed chain; upgrading again does not repair that history.
Verified on Packagist at commit 4e56debcb59546269d8e1a69f31c5e7a111dd7ae. Published-package verification and five Laravel upgrade cells passed, alongside the full source/main qualification in #1067 and #1068.
This publishes Core support. PHP Server advertisement and the SDK release/installed portable qualification remain separate steps tracked in #1066 and sdk-php#269.
deadline_policy: "occurrence", physical false-wake reopens retain the first
timer's exact deadline and original timeout budget. Conflicting, timed-out
or unprovable chains refuse with whole-completion rollback. Legacy chains
first opened without the policy retain their existing per-open behavior;
an anchored chain must keep its policy on every reopen.Exclude the test tree from Composer distribution archives (#1064), avoiding the tests/Fixtures and tests/fixtures directory collision when vendor file
tests/Fixtures and tests/fixtures directory collision when vendor
files are synchronized to a case-insensitive filesystem. Source checkouts
retain the complete test corpus. Runtime code, history, schema, public APIs
and the 2.21.0 platform support floors are unchanged.Published-package verification passed all six jobs: exact installed package/platform conformance and all five supported Laravel 12/13 upgrade and queue-worker cells. Canonical post-product PHP, private-pip and Composer cleanup passed.
Workflow2.21.0 supports PHP8.2+ with patched Laravel12/13 ( ^12.69.0|^13.30.0 ). Laravel10/11 and PHP8.1 applications should keep their Workflow2.20.x
Workflow2.21.0 supports PHP8.2+ with patched Laravel12/13 (^12.69.0|^13.30.0). Laravel10/11 and PHP8.1 applications should keep their Workflow2.20.x lock while upgrading their platform. Stable-v1 history ownership and coexistence rules are unchanged.
This release adds namespace-bound schedule visibility predicates and optional guarded local-activity retry contract2 helpers. The default bridge does not enable the new retry role. It also includes the MySQL parallel-Activity continuation prevention fix from2.20.5; already-stranded runs are not automatically repaired.
Published package verification passed all6 jobs: the exact installed package/terminal cleanup and all5 supported Laravel12/13 upgrade/queue-worker cells. Immutable source: cb9420019476d400e1df329f063cfb1a3b86ab18.
Workflow 2.21.0 requires PHP ^8.2 and patched Laravel
^12.69.0|^13.30.0. Laravel 10/11 and PHP 8.1 are no longer supported.
Upgrade the application platform first while preserving its existing
Workflow 2.20.x lock, then update Workflow. The stable-v1 authority and
ownership of existing durable history are unchanged.
The embedded source/published upgrade matrices now cover Laravel 12/13 on their supported PHP 8.2–8.4 intersection. Laravel 13 still requires PHP 8.3 or newer. Removed-framework constraints refuse before installation; retired-platform advisory exceptions are not carried into 2.21.
Add namespace-bound schedule visibility predicates for ScheduleId,
Status, WorkflowType, TaskQueue and Note. Equality and AND use
bound values; unsupported fields and operators return typed refusals.
Add guarded prepared local-activity retry contract 2 helpers and an optional task-bridge interface. Preparation snapshots, report validation, retry decisions and lost-claim recovery retain original authority and fences. The default bridge does not implement the new role; these helpers do not enable a runtime capability or change existing v1 bridges.
Include the parallel activity continuation prevention fix described under 2.20.5 below. Existing stranded runs still require separate recovery.
Fix a MySQL repeatable-read race where the last successful activity in a parallel group could leave its workflow waiting. The activity barrier now rea
Fix a MySQL repeatable-read race where the last successful activity in a parallel group could leave its workflow waiting. The activity barrier now reads current history while holding the parent-run lock, including named storage and nested groups. Duplicate completion still wakes once. Fixes #1052.
This patch changes no schema, payload, public API or platform support. It prevents new missed continuations; it does not automatically resume runs already waiting without a continuation. Rolling back restores the defect.
Qualified with independent-connection MySQL predecessor/candidate controls and the normal database, unit, static-analysis and coverage suites. Published-package verification and all twelve supported upgrade cells passed.
Fix an embedded MySQL/InnoDB REPEATABLE READ deadlock between continueAsNew() and signal admission for an unrelated workflow instance.
Fix an embedded MySQL/InnoDB REPEATABLE READ deadlock between continueAsNew() and signal admission for an unrelated workflow instance.
The correction preserves caller-owned transaction read views and avoids replaying user workflow code to recover from a deadlock. It also preserves pending default and named signal handoffs across continuation.
The published package passed the installed MySQL regression and scope checks. Affected applications retain their own upgrade verification tracked in #1043.
Contributor: @Loosand.
Rollback, embedded Workflow package: 2.20.3 changes no schema, history event or public API. Rolling back restores the shared-cursor, empty receive rep
Rollback, embedded Workflow package: 2.20.3 changes no schema, history event or public API. Rolling back restores the shared-cursor, empty receive replay and fake task-drain precision bugs. A kept stream now records receives at the current step. Earlier versions look them up at the opening step and can replay those runs differently after rollback.
A message on one stream stays receivable until it is consumed or expires
(#1011). The run's message cursor is shared by signals, updates and every
stream, so a signal, an update or a receive on another stream used to move
it past older pending messages, and inbox()->receive(), peek(),
pendingCount() and hasPending() no longer returned them. They now read
every pending message of the stream, in sequence order.
WorkflowStub::runReadyTasks() now drains fake-workflow tasks at their
exact due time, including fractional seconds (#1026). It compares the
sampled clock using the configured task model's persisted precision.
SQLite previously skipped tasks due at that instant because it compared
their fractional timestamps with a whole-second query value. Future
fractional deadlines remain pending until they are due.
Known limitation (#1024): a query can return nothing for a message an update handler received, because the query replays the handler at the wrong step. The query changes no message rows. The workflow's own replay and completed output still return what the handler received. This issue predates 2.20.3 and remains open.
Full changes: https://github.com/durable-workflow/workflow/compare/2.20.2...2.20.3
Rollback, embedded Workflow package: 2.20.2 changes no schema, history event or API. Rolling back to 2.20.1 or earlier brings back the stranded pendin
Rollback, embedded Workflow package: 2.20.2 changes no schema, history event or API. Rolling back to 2.20.1 or earlier brings back the stranded pending stream messages at redrive (#1013) and at continue-as-new (#1018). Messages that 2.20.2 already moved to a successor or continued run stay there.
Continue-as-new moves the run's pending inbound stream messages to the
continued run (#1018). Before, only the message cursor carried over, so a
stream message still pending when a run continued as new stayed on the
closed run and was never received. This covers embedded execution and
external workers that complete a task with continue_as_new. Signals were
not affected.
A redrive moves the failed run's pending inbound messages to the successor (#1013). Before, only the message cursor carried over, so a message that arrived while the run was failed stayed on the failed run and the successor never received it. A replayed receive whose original call found nothing can now return one of the moved messages.
Known limitation, not fixed in 2.20.2 (#1011): stream receives still start
after the run's shared message cursor, which signals and receives on other
streams also advance. A stream message that is still pending behind that
cursor stays hidden from inbox()->receive() and receiveOne(). A replayed
receive whose original call found nothing can also read the stream again
and return a message. The fix (#1017) follows in 2.20.3.
Full changes: https://github.com/durable-workflow/workflow/compare/2.20.1...2.20.2
Rollback, embedded Workflow package: 2.20.1 changes no schema, history event or API. Rolling back to 2.20.0 or earlier brings back the duplicate sends
MessageStream.outbox()->sendReference() used to write a new outbound and inbound
message each time. A replay now returns the row the original send wrote,
including a row written by the run a successor was redriven from.inbox()->receive() and receiveOne() now return the messages that call
consumed at that step, by this run or by the run it was redriven from.
Before, a replay could return different messages or none, and could
consume a newer message under the earlier step.idempotency_key. A run with such rows
matches its sends to the earliest rows and writes only new sends.Full changes: https://github.com/durable-workflow/workflow/compare/2.20.0...2.20.1
Rollback, embedded Workflow package:
reused_recorded_at on reused signal, memo and search-attribute events,
so a redrive successor with those events can replay with a different
workflow time and deadlines there. Keep 2.19.0 or later workers, or
confirm no such history can replay, before rolling back below 2.19.0.unsupported_history) or a reused signal wait
(mismatched_signal_receipt) before any write. A request for a run that
already has its successor returns that successor.service_redrive_history_not_qualified fence.reused_from_run_id and reused_recorded_at, ahead of the step it
ran before, and replay runs the update handler again from its recorded
arguments so the workflow state matches. The update is not delivered again.
A rejected update, and one the failure closed before it ran, are not reused.
Only three update shapes are accepted, in history order: a rejection, an
acceptance closed by the failure, or an acceptance, application and
successful completion with identical arguments. Every event in each shape
must name the same update, command and handler and belong to the source run.
An acceptance closed by the failure must record a terminal reason and a
failure, with no result and no step. An update the source
accepted but never resolved is refused with unsupported_pending_update,
any other shape or binding with mismatched_update_history or
incomplete_update_step, an externally stored update payload with
external_update_payload_not_supported, and an inline payload that can't be
decoded with unavailable_update_payload. A run a service worker started
that applied an update is refused with service_redrive_history_not_qualified
(#864).reused_from_run_id names
the immediate source, so a chain of redrives reuses each step once.
reused_from_run_id and reused_recorded_at come as a pair on every reused
update and signal event. A reused event keeps its original time, and a
history where one key is missing, blank or names another source is refused.
Any external_storage key on a signal payload counts as external, even a
null one. Before,
redriving such a successor was refused with mismatched_signal_receipt
(#864).parents and the failed run's continuedWorkflows show
link_type: redrive with history_authority: typed_history and
diagnostic_only: false, because the successor's WorkflowStarted records
recovery_kind: redrive, continued_from_run_id and workflow_link_id.
Both sides take the step (resume_step_sequence) and time from that event,
not from the link row. The start must name the successor's own run and
instance, an earlier run of the same instance as the source, a nonblank link
id and a positive resume step, and exactly one redrive link row must join
that source to the successor under that link id. Before, both sides showed
a diagnostic-only fallback. A link the successor's history does not confirm stays a single
diagnostic fallback (#864).Full changes: https://github.com/durable-workflow/workflow/compare/2.19.0...2.20.0
Failed-run redrive recovers more workflows: workflow-code memo and search-attribute upserts, applied signal waits, and failures in workflow code, incl
Failed-run redrive recovers more workflows: workflow-code memo and search-attribute upserts, applied signal waits, and failures in workflow code, including onto a corrected definition when you ask for it.
MemoUpserted and SearchAttributesUpserted steps in the completed prefix are reused at their original time. Memo and search-attribute rows carry to the successor unchanged, so a successor that fails again can be redriven again.await('name')) that received and applied its signal is reused with its value and original time. The message cursor carries over, so a consumed signal isn't delivered again.WorkflowFailed now records failed_step_sequence with failed_step_kind workflow_code or structural_limit, and the successor resumes at that step.
allow_definition_change: true.
definition_changed_from on the successor's WorkflowStarted. Run detail shows it as workflow_definition_changed_from, and describe as run.definition_changed_from. A later redrive on the same definition records none.Rollback. Releases before 2.19.0 ignore reused_recorded_at on the signal, memo and search-attribute events that a 2.19.0 redrive successor reuses. Replaying such a successor on an older release can change its workflow time and the deadlines derived from it. Once a 2.19.0 redrive successor with reused signal waits or upserts exists, keep 2.19.0 or later workers, or confirm that no such history can replay before rolling back.
New failed_step_kind values. workflow_code and structural_limit join activity. Readers that switch on failed_step_kind should accept them. Failures written by earlier releases have no boundary and still refuse redrive with unrecorded_failure_boundary.
allow_definition_change is an explicit opt-in. Without it, a changed definition is refused as before. Runs started by a service worker are always refused with definition_change_requires_replay_check, even if a local class is registered for the type later.
New typed refusals:
workflow_code_failure_requires_corrected_definitiondefinition_incompatible_with_prefix, with definition_mismatch {sequence, kind}definition_check_over_budgetdefinition_change_requires_replay_checkunsupported_signal_wait, unsupported_pending_signal, incomplete_signal_step, mismatched_signal_receipt, external_signal_payload_not_supportedservice_redrive_history_not_qualifiedunsupported_visibility_metadata now refuses only a workflow-code row with no matching reused upsert step.
Service mode (#1006):
service_redrive_history_not_qualified when its completed prefix has a memo upsert, search-attribute upsert or signal wait to reuse. The refusal happens before any write. The fence lifts only through a replay-capability check once durable-workflow/sdk-python#235, durable-workflow/sdk-php#244 and durable-workflow/sdk-rust#172 ship installed-qualified.Main CI was green at 28893f97 (merge of #1005). Release Verification runs on this tag.
reused_recorded_at on the signal,
memo and search-attribute events that a 2.19.0 redrive successor reuses.
Replaying such a successor on an older release can change its workflow time
and the deadlines derived from it. Once a 2.19.0 redrive successor with
reused signal waits or upserts exists, keep 2.19.0 or later workers, or
confirm that no such history can replay before rolling back (#864).MemoUpserted or SearchAttributesUpserted event
refused redrive with unsupported_history. The successor reuses those
events in step order, carrying reused_from_run_id and
reused_recorded_at, and replay uses the original time. Memo and search
attribute rows now carry to the successor unchanged, keeping the step that
wrote them and whether they were inherited, so a successor that fails again
can be redriven again. A row written by workflow code with no matching
reused upsert step is still refused with unsupported_visibility_metadata
(#864).await('name')) that received and applied its signal. The successor
records the wait, receipt and application with reused_from_run_id and
reused_recorded_at, and replay uses the original time, so the workflow
gets the same value without a new signal. Redrive also carries the source's
message cursor, so a consumed message is not delivered again. A wait with a
timeout, cancellation scope or parallel group is refused with
unsupported_signal_wait, a signal the source received but never applied
with unsupported_pending_signal, and a wait without its receipt and
application with incomplete_signal_step. A receipt that names another
run, instance, command or wait is refused with mismatched_signal_receipt,
and an externally stored signal payload with
external_signal_payload_not_supported. Updates in the completed prefix
are still refused (#864).WorkflowFailed now records failed_step_sequence with failed_step_kind
workflow_code or structural_limit, the step the code never took. The
successor reuses every earlier step and resumes there. A workflow-code
failure on the same definition is refused with
workflow_code_failure_requires_corrected_definition, because the same code
would throw again. Pass allow_definition_change: true to redrive onto a
corrected definition: inside the redrive transaction, the new code replays
the reused prefix in memory, and redrive refuses with
definition_incompatible_with_prefix when it can't (with
definition_mismatch.sequence and .kind naming the first step it
disagrees with), or with
definition_check_over_budget when the prefix exceeds the history budget,
before writing anything. An accepted change records
definition_changed_from on the successor's WorkflowStarted, and run
detail (workflow_definition_changed_from) and describe
(run.definition_changed_from) show it. A later redrive on the same
definition records no change. A run started by a service worker is refused
with definition_change_requires_replay_check, even when a local class is
registered for its type later. Failures
recorded by earlier releases have no boundary and still refuse with
unrecorded_failure_boundary (#864).service_redrive_history_not_qualified, before writing anything. Embedded
runs, and service runs whose prefix holds only activities, side effects,
version markers and fired timers, redrive as before. The refusal lifts only
through a replay-capability check once SDK replay of those reused events
ships (durable-workflow/sdk-python#235, durable-workflow/sdk-php#244,
durable-workflow/sdk-rust#172) (#864).Failed-run redrive recovers more ordinary workflows, keyed signals and updates keep their admissions, and V2 explains the Eloquent models it rejects.
Failed-run redrive recovers more ordinary workflows, keyed signals and updates keep their admissions, and V2 explains the Eloquent models it rejects.
can_redrive and redrive_blocked_reason, using the same decision redrive() applies (#990).reused_from_run_id and reused_recorded_at, and replay uses the original time, so now() after a reused step matches the source run.A retry with the same caller key returns the retained admission. Reusing a key for a different name, typed arguments or explicit run target is refused without durable writes. Caller keys must be valid UTF-8, 1 to 255 code points.
Passing an Eloquent model, ModelIdentifier or Eloquent collection across a V2 payload boundary fails with an unsupported_value_type message that names the model class and says to pass the key and load the model inside the activity. Before, the message named Laravel's internal ModelIdentifier type.
The first refused completion or failure from an expired or superseded attempt appends one ActivityOutcomeRejected event. It records the attempt identity, lease owner, attempted outcome, refusal reason and observation time. It is payload-free: the result, failure, stack and request snapshot are never stored. It grants no business outcome, renews no lease and changes no retry.
workflow_commands.intent_digest column, with no backfill. Run migrations, then drain and upgrade all embedded database writers together.ActivityOutcomeRejected since 2.13.0. After this release writes one, do not roll back below 2.13.0. Rolling back within 2.13.0 and later is fine.Main CI was green at 36dc304d (merge of #1003). Release Verification runs on this tag.
Redrive a failed run that has start-time memos or search attributes. Before,
any memo or search attribute row refused redrive with
unsupported_visibility_metadata. The successor now inherits the start
values, as continue-as-new does. A value upserted by workflow code is still
refused until redrive replays that part of the history (#864).
Run detail and describe report whether redrive would accept a run.
can_redrive and redrive_blocked_reason come from the same decision
redrive() applies, so the action an operator sees matches what redrive
does. Runs that are not failed report run_not_failed, and a source that
was already redriven reports already_redriven (#864).
Redrive a failed run whose completed steps include side effects or version
markers. Before, any SideEffectRecorded or VersionMarkerRecorded event
refused redrive with unsupported_history. The successor reuses them in step
order without running the side effect again. Like reused activity
completions, they carry reused_from_run_id and reused_recorded_at, and
replay uses reused_recorded_at so workflow time after the step matches the
source run. New refusal reasons: invalid_step_sequence,
duplicate_step_sequence and external_side_effect_result_not_supported
(#864).
Redrive a failed run whose completed steps include a fired timer(). The
successor records the timer as already fired, under a new timer id, with no
timer row or task, so it does not wait again. TimerFired carries
reused_from_run_id and reused_recorded_at, and replay uses the original
fire time. Condition and signal timeouts, cancellation cleanup timers and
timers in a cancellation scope or parallel group are refused with
unsupported_timer. A timer step without both its scheduled and fired events
is refused with incomplete_timer_step (#864).
Retain keyed signal admissions and compare canonical typed intent for keyed signals and updates. Conflicting key reuse refuses without durable writes; undecodable legacy receipts remain explicitly unverified. This strengthens the earlier same-key update receipt contract (#877).
Require valid UTF-8 caller keys of 1..255 code points before admission.
Explicit empty or non-string keys refuse; null clears an inherited key.
Apply the nullable intent_digest migration, drain and upgrade all embedded
database writers together, and allocate new keys for different operations.
Native Server and published SDK/CLI qualification remain separate.
V2 payloads reject an Eloquent model, ModelIdentifier or Eloquent
collection with an unsupported_value_type message that names the model
class and says to pass the key and load the model inside the activity.
Before, the message named Laravel's internal ModelIdentifier type. The
Avro argument path no longer runs serializeModels, and failure and
throwable normalization are unchanged (#997).
Record a bounded diagnostic when a late activity outcome is refused. The
first refused completion or failure from an expired or superseded attempt
appends one ActivityOutcomeRejected event with the attempt identity, lease
owner, attempted outcome, refusal reason and observation time. It is
payload-free: the result, failure, stack and request snapshot are never
stored. It grants no business outcome, renews no lease and changes no retry.
Readers have understood this event since 2.13.0, so after an upgraded
release writes one, do not roll back below 2.13.0 (#878, #901).
workflow:v2:replay-verify now accepts a failed run whose replay reproduces its recorded failure. A matching exception class and message reports replay
workflow:v2:replay-verify now accepts a failed run whose replay reproduces its recorded failure. A matching exception class and message reports replayed with replay.recorded_failure; a redacted bundle compares by class only. A different exception still reports replay_error and adds error.recorded_failure. The reason set is unchanged, so Server's replay verification contract is unaffected.
Fixes #983. Main CI green at 5998e6ae. Release Verification runs on this tag.
ReplayDiff and workflow:v2:replay-verify reported
replay_error for any run that failed because workflow code threw, even when
replay reproduced the recorded WorkflowFailed exception. A matching
exception class and message now report replayed with
replay.recorded_failure. A different exception stays replay_error and
names the recorded failure in error.recorded_failure (#983).Laravel 9 has no fixed release for the TokenGuard advisory GHSA-f9x3-7vq9-cph6 (CVE-2026-114509). Composer keeps Laravel 9 applications on 2.16.x, whi…
Retention-pruned runs no longer export as complete history. A run whose details were pruned exports history_complete: false, with workflow.details_state: "pruned" and details_pruned_at. Its dedupe key gains a :details-pruned: suffix, so a post-prune export cannot collide with the full export taken earlier. workflow:v2:replay-verify fails such a bundle with history.details_pruned, and import refuses it as terminal_history_incomplete. Unpruned bundles are unchanged apart from the additive details_state: "retained" field.
Laravel 9 is no longer supported: the constraint is now ^10.50.4|^11.56.1|^12.69.0|^13.30.0. Laravel 9 has no fixed release for the TokenGuard advisory GHSA-f9x3-7vq9-cph6 (CVE-2026-114509). Composer keeps Laravel 9 applications on 2.16.x, which stays installable.
Fixes #856 and #922. Main CI green at 0897a58f (run 41545363761, attempt 2 after a test-harness race was diagnosed and filed separately). Release Verification runs on this tag.
WorkflowRunRetentionCleanup::pruneRun(), HistoryExport::forRun() sets
history_complete to false, adds workflow.details_state: "pruned" and
workflow.details_pruned_at, and appends :details-pruned:{time} to the
dedupe_key, so the post-prune export can't be mistaken for the full export
taken before cleanup. Every bundle now carries details_state; runs that
were never pruned keep their existing dedupe key. Bundle verification reports
pruned bundles as history.details_pruned, so replay verification fails
them, and history import rejects them (#856).laravel/framework constraint, which is
now ^10.50.4|^11.56.1|^12.69.0|^13.30.0. Laravel 9 has no fix for the
TokenGuard advisory GHSA-f9x3-7vq9-cph6. Composer keeps Laravel 9
applications on Workflow 2.16.x, so updating does not break them; to move
past 2.16.x, upgrade the framework to a patched Laravel 12 or 13 release
first. The embedded upgrade contract no longer lists Laravel 9 cells, and the
dependency security policy records the removal. The dependency security audit
now requires Composer to refuse this package with Laravel 9 (#922).Embedded schedule evaluation now accepts workflow:v2:schedule-tick --limit=N. Set the default through workflows.v2.schedules.tick_limit / DW_V2_SCHEDU
Embedded schedule evaluation now accepts workflow:v2:schedule-tick --limit=N. Set the default through workflows.v2.schedules.tick_limit / DW_V2_SCHEDULE_TICK_LIMIT; it remains 100. An invalid value fails without evaluating schedules.
The command retains the existing namespace fairness and missed-fire behavior. Each pass bounds due starts and buffered actions separately. Finite-burst bounds assume successfully startable candidates and on-cadence ticks; failed starts remain due and can extend the wait. Applications with an older published config/workflows.php must add the new nested schedules setting or refresh their configuration before the environment setting takes effect.
Source: protected merge 637ea0b5a25af29159116b3976d602c4811768a9, PR #979. Full main quality, unit coverage, database features and upgrade qualification passed in run 41519147789. The committed public-platform mirror contract was verified before tagging. Published-package and installed consumer verification follow the tag.
workflow:v2:schedule-tick accepts --limit=N, and the new
workflows.v2.schedules.tick_limit config (DW_V2_SCHEDULE_TICK_LIMIT)
sets the default. The default stays 100, so existing behavior is
unchanged. Each pass starts at most that many due schedules and drains at
most that many buffered actions, shared round-robin across namespaces;
the rest fire on later passes. A limit that is not a positive integer is
refused, and the command fails without evaluating schedules (#875).Scheduled starts now use namespace-specific instance IDs. Schedules with the same ID in different namespaces can backfill, tick or trigger the same oc
Scheduled starts now use namespace-specific instance IDs. Schedules with the same ID in different namespaces can backfill, tick or trigger the same occurrence independently.
Existing instance IDs remain readable, and recorded occurrences remain deduplicated.
Fix: #977. Source and full main qualification passed at a8a552b569f7a21fead0d8a2e9abe26fa7b37b15. Published-package verification and installed acceptance follow publication.
schedule.ns:{namespace hash}:{length}:{schedule_id}:{suffix}, so the same
occurrence no longer collides across namespaces or after the configured
default namespace changes. Instances started under the earlier
schedule:{schedule_id}:{suffix} form stay readable, and occurrences that
already fired stay deduplicated through schedule history (#874).Embedded workflow starts and all subsequent tasks now preserve requested priority and fairness. Timer/signal resumes, continue-as-new and child retry
Embedded workflow starts and all subsequent tasks now preserve requested priority and fairness. Timer/signal resumes, continue-as-new and child retry retain the run scheduling class. Explicit task/activity overrides keep precedence, and default-class operation remains compatible with older configured tables.
Qualified through #976 and full source build41443622443. Published-package verification and installed lifecycle acceptance are tracked in #971.
StartOptions::withPriority() and withFairness() on runs started
through the embedded WorkflowStub; they were accepted but not stored, so
embedded runs always used the default priority and fairness class (#971).Fix repeated daylight-saving schedule occurrences in the embedded Workflow engine.
Fix repeated daylight-saving schedule occurrences in the embedded Workflow engine.
Source fix and qualification cover New York and Kyiv transitions, three database engines, unit/combined coverage and the supported Laravel upgrade matrix. The product-train metadata matches 2.15.13. A fresh published 2.15.12 consumer reproduces the three original daily-schedule failures and provides the installed upgrade comparison.
Follow #882 for published-package verification and installed after-results. The eight-language schedule guide follows qualified publication and installed acceptance.
30 1 * * * schedule in America/New_York no
longer starts a second run at 01:30 EST on the day clocks go back, and
30 2 * * * no longer fires early inside the repeated hour. A time skipped
when clocks go forward still fires once, moved forward by the gap (02:30
runs at 03:30). Schedules whose hour field selects every hour (*,
0-23 or a step of one) keep running on elapsed time, so hourly and
sub-hourly schedules fire through both passes of the repeated hour (#882).Namespaced activity polling now checks task, run and canonical instance ownership before applying the batch limit. This prevents inconsistent legacy r
Namespaced activity polling now checks task, run and canonical instance ownership before applying the batch limit. This prevents inconsistent legacy rows from hiding eligible activities.
Fix and qualification. Server adoption is tracked in Server #765.
Fix embedded history imports and previously imported rows resolving or changing state outside their namespace. Validate copied references, require sto
Fix embedded history imports and previously imported rows resolving or changing state outside their namespace. Validate copied references, require stored relationship evidence before recovering lineage, and keep current-run pointers, retention, command effects and projected metadata scoped to their canonical owners. Preserve valid imports in either endpoint order and idempotent repeats.
Namespace comparisons now remain exact under case-insensitive database collations. Imports and lineage repairs use the configured storage connection. PostgreSQL write imports own a serializable transaction with up to three complete conflict attempts; callers must finish an existing storage transaction before importing. Preview remains available inside a caller transaction.
Correct bounded history byte counts for NUL-bearing JSON strings and keys on older SQLite JSON1 versions without hydrating history or changing stored payloads.
See #929, #928, and #968. Applications and pinned downstream images need a dependency update or rebuild to receive this package.
Core 2.15.10 restores parent delivery when a child resolves after the parent worker has claimed its history and the worker completes into another wait
Core 2.15.10 restores parent delivery when a child resolves after the parent worker has claimed its history and the worker completes into another wait. The bridge offers that outcome one follow-up replay task, preserving task coalescing, terminal parents and parallel child barriers.
The release also clarifies that Laravel 9 support is limited to existing lockfiles that need to upgrade or drain. Its TokenGuard advisory remains unfixed; keep the documented exposure restriction and migration policy. Laravel constraints and Composer security blocking are unchanged.
No database migration, worker-protocol change or SDK release is required. Server images pinned to an earlier Core package need their own qualified dependency update and publication.
Fix workflow storage correctness when DW_STORAGE_CONNECTION or custom V2 models select a connection other than the application's default.
Fix workflow storage correctness when DW_STORAGE_CONNECTION or custom V2 models select a connection other than the application's default.
The fresh Composer distribution matches all 1,007 committed runtime files. All 22 storage cases pass on PostgreSQL 17 and all 22 pass on MySQL 8, including competing claimants, actual row locks, rollback and physical database queue publication. Four independent-application-rollback probes also pass.
Main qualification passes, and published verification passes the package contract plus all 16 Laravel/PHP upgrade journeys. Local unit qualification passes 5,279 tests / 74,821 assertions (45 skips, 15 PHPUnit notices); unit coverage passes its gate at 77.60%.
Application changes requiring atomicity with workflow storage must use the same Laravel connection and transaction. Different connections commit independently; queue transports retain their own transaction and delivery behavior.
The complete local quality cycle also passes, including 2,551 PostgreSQL feature tests / 25,150 assertions (18 skips). The updated connection and transaction guide is deployed and verified in all eight languages.
Laravel 9 retains the existing upgrade-or-drain requirement documented in #922.
Keep engine transactions, task claims and deferred callbacks on the workflow
storage connection selected by DW_STORAGE_CONNECTION or custom V2 models.
Hold claim locks until commit and roll back failed starts without partial
durable state. Lifecycle events wait for the outer transaction to commit and
retain their execution trace context (#893).
Require Laravel 10.50.4 or later on the Laravel 10 line so TokenGuard rejects non-string tokens in both user lookup and credential validation. Laravel 9 retains the documented upgrade or drain requirement (#922).
Back off between SQLite host-query preparation and dispatch-claim retries after a proven root rollback, so a brief competing writer does not exhaust all attempts immediately. Preserve caller-owned transactions and durable query claim fencing (#947).
Add bounded SQLite backoff for competing host-query preparation and dispatch-claim writes, preserving transaction ownership and the existing retry lim
Laravel 9 remains affected by the TokenGuard issue; follow the documented upgrade or drain guidance. Issue #922 remains open for that line.
All five repository quality commands pass, including 5,255 unit tests and 2,550 PostgreSQL feature tests. The published package/source/platform contract and all 16 published Laravel/PHP upgrade journeys pass in Release Verification.
Fixes cancellation and termination exceeding PHP resource limits for large, valid pending inventories. See #958 and #959.
Fixes cancellation and termination exceeding PHP resource limits for large, valid pending inventories. See #958 and #959.
Two separate synthetic SQLite HTTP fixtures each admitted 2,000 timers and 2,000 activities normally, then closed with concurrent maintenance. Cancellation used 29.0744 seconds of CPU and termination 28.8921 seconds, each peaking at 108 MiB under unchanged PHP limits of 128 MiB and 30 seconds. Canonical history, projections, original-lease fencing and repeat closure were verified. Qualification details and contention caveat.
SQLite poll-mode maintenance now pauses 200 ms outside transactions between selected candidates. A fully saturated pass at default phase limits can add up to 35 seconds of wall waits, including synchronous repair API and embedded queue-loop callers. Empty or throttled passes add no pause.
Installing this library release does not update an existing pinned Server image. Server consumers need an image containing this Core version; follow #958 for downstream delivery.
Schedule an accepted signal's delivery turn when another leased replay finishes waiting. Received named waits remain eligible with locally available d
Schedule an accepted signal's delivery turn when another leased replay finishes waiting. Received named waits remain eligible with locally available definitions, while the existing attempted-signal and command-order fences are preserved. Fresh replay applies both signals once and leaves no extra workflow task while the timer remains pending.
Includes four timing/definition regressions and buffered/ordering controls. All five local quality commands pass, including PostgreSQL features and 77.24% unit coverage. Main qualification covers the SQL-engine, coverage and upgrade gates. The committed source-release metadata declares 2.15.6.
Refs #956, which remains open for normal Server integration and exact published Server/SDK acceptance. Core change: #957.
Fix retained local and cancellation-scope checkpoint receipts ending in complete_update or fail_update. The returned next_sequence now admits the foll
Fix retained local and cancellation-scope checkpoint receipts ending in complete_update or fail_update. The returned next_sequence now admits the following checkpoint or scope-open under the same retained claim. Retrying an older receipt corrects its cursor without replaying commands or changing history, leases or task state. Update outcomes keep their separate identity from yielded workflow steps.
Includes focused before/after coverage and an official cold-replay regression fixture. Full main qualification passed its quality, coverage and SQL-engine gates. The committed source-release metadata matches this tag.
Refs #953, which remains open for exact published Server and affected-consumer acceptance.
Direct workflow-bound service updates now reject a zero completion timeout with a typed validation error before storing a call or evaluating admission
Direct workflow-bound service updates now reject a zero completion timeout with a typed validation error before storing a call or evaluating admission. Blank and padded completion modes use the shared wait policy.
Accepted-only calls, keyed submissions and existing transactions retain their bounded behavior. Replays return the original call. Query zero waits and unresolved-catalog auditing keep their existing contracts.
Includes regression coverage and documentation for the completion/accepted-only distinction. See PR #951 and #947.
Admitted workflow-query service calls can use the existing durable dispatcher without holding a database transaction while a worker answers. The trust
Admitted workflow-query service calls can use the existing durable dispatcher without holding a database transaction while a worker answers. The trusted host capability persists one accepted query intent, fences expired or replaced claims, and retries short result writes independently of business attempts.
The repair includes file-backed SQLite concurrent-replay and independent-writer controls, claim recovery, immutable query arguments/target, authorization, cancellation/deadline and per-attempt retry coverage. Ordinary published-package and supported Laravel upgrade verification passed.
The Server integration and exact published-image acceptance remain in Server #794. Existing Server images retain their bundled Core until a new image is released.
Source: #950.
Host bridge signals and updates now deduplicate inside the instance writer lock. A bounded identity token preserves the complete adapter, action, opaq
Host bridge signals and updates now deduplicate inside the instance writer lock. A bounded identity token preserves the complete adapter, action, opaque provider key and target tuple, while accepted legacy receipts remain reusable. Conflicting retained metadata fails closed.
Ordinary signal request IDs and service-call identities keep their existing behavior. Host-created identity objects cannot be forged by JSON context or serialized into application payloads. This patch adds no schema migration.
Validation: full ECS/PHPStan; 5,119 unit tests / 72,539 assertions; PostgreSQL feature suite 2,513 cases / 24,675 assertions; 77.21% unit line coverage. Normal published-package/platform and supported Laravel upgrade verification passed in run 40481078834. Merge 15ac2f1ccd5a2e4944bf39c31562d66a62a9d333 has the exact qualified PR #949 tree; Packagist source/dist resolve to that immutable release.
Server consumers require a separate dependency update, image publication and installed acceptance. See durable-workflow/server#777 for that follow-through.
Workflow-bound service calls now preserve JSON and encoded arguments/results, reconcile linked timeout and update outcomes, follow canonical continue-
Workflow-bound service calls now preserve JSON and encoded arguments/results, reconcile linked timeout and update outcomes, follow canonical continue-as-new successors in bounded passes, and preserve the winning terminal/cancellation receipt.
HTTP replays authorize the selected call's owner and current catalog policy. Hosts can prepare a canonical admission audit before charging policy, then finalize only their own uncommitted pending row.
This patch adds no schema migration or worker-protocol change. Server images pin their Core dependency and need a separate image release to carry these changes.
Validation: full ECS/PHPStan; 5,107 unit tests / 72,526 assertions; PostgreSQL feature suite 2,513 cases / 24,675 assertions; 77.19% V2 unit line coverage; all main checks green; normal published-package/platform and supported Laravel upgrade verification passed. See PR #948 and issue #947 for scope and the remaining installed Nexus/recovery qualification.
Embedded Laravel applications can now listen for Workflow\V2\Events\WorkflowCancelled and Workflow\V2\Events\WorkflowTerminated when a run closes.
Embedded Laravel applications can now listen for Workflow\V2\Events\WorkflowCancelled and Workflow\V2\Events\WorkflowTerminated when a run closes.
reason and the persisted UTC closedAt timestamp.WorkflowFailed. Termination still does not execute workflow cleanup code.This is an additive PHP event API; no database migration or wire-protocol change is required. These are in-process Laravel notifications, without a durable external-delivery guarantee. Broader transaction/storage work remains tracked in #893.
Implementation and qualification: #946. Delivery and installed-consumer/documentation acceptance: #896.
Fix direct service execution with caller workflow context returning scheduler admission instead of its requested terminal result or typed failure (#94
Fix direct service execution with caller workflow context returning scheduler admission instead of its requested terminal result or typed failure (#942, #943).
External workflow completion now validates a supported canonical worker protocol before parsing commands or making changes. Protocol 1.20 completion remains inline; protocol 1.21 completion uses durable scheduling. The existing two-argument WorkflowTaskBridge::complete() contract is preserved. Hosts passing an authenticated worker version use the additive ProtocolAwareWorkflowTaskBridge::completeForProtocol() role.
Reduce unnecessary retained-wait projection writes while preserving workflow progress (#899, #941).
Upgrading from 2.14.1 requires no new schema migration. Server adoption and installed consumer qualification are tracked in the linked owning issues.
Workflow 2.14.1 applies pending-count structural limits consistently to external-worker task completions. Scheduling an activity, timer or child workf
Workflow 2.14.1 applies pending-count structural limits consistently to external-worker task completions. Scheduling an activity, timer or child workflow now checks its limit before any command writes. An over-limit batch rolls back completely and retains the workflow task lease without durable changes.
Fixes #933 via #939. The exact release source is af9dbadb41bf96f51ad78047022c392887e3c7a6, qualified by the full main build, including all database shards, coverage, Laravel upgrades and Sonar, plus exact-source boundary, dependency-audit and CodeQL checks.
The normal published-package verification runs for this tag. PHP Server images and pinned applications require their own adoption and installed acceptance.
Workflow 2.14.0 makes workflow service-operation admission and execution predictable under concurrency and crash recovery. Trusted host authorizers no
Workflow 2.14.0 makes workflow service-operation admission and execution predictable under concurrency and crash recovery. Trusted host authorizers now own principal and namespace authority; replay and keyed creation are reauthorized against locked canonical state. Service effects dispatch outside completion locks, cap synchronous waits at 60 seconds, and preserve one opaque effect identity through recovery.
Workflow task completions now enforce hard structural limits for history events, history bytes, metered rows and external payload bytes. Lifecycle events dispatch only after commit, predictable external payloads are staged only after admission, and local-activity timing values are bounded to storable ranges.
Changes: #934 and #937. The exact release source is b814c6c485b1c072f4092d3f478c7deba9c8c241, qualified by the full main build, including all database shards, coverage, Laravel upgrades and Sonar.
The normal published-package verification runs for this tag. PHP Server images and pinned applications require their own adoption and installed acceptance.
WorkflowServiceOperationAuthorizer before
upgrading. The default refuses service operations; explicitly trusted embedded
hosts can opt into RunServiceOperationAuthorizer for same-namespace calls
with an unprivileged workflow principal. Authenticated servers supply their
own current principal and authorize requested target scope.authorizeReplay hook.record_local_activity timing fields (duration_ms, heartbeat
elapsed_ms, backoff_seconds) and refuse a report whose walk back from
completion leaves the storable timestamp range (#930).history_transaction_size, history_transaction_bytes,
history_transaction_rows, external_payload_transaction_bytes). Metered
rows are history events and local activity records for now; other command
writes are not counted yet. A refusal rolls the whole completion back and
keeps the task's lease. A run's deadline close is never metered. The limits
appear in the structural limit snapshot.PredictableExternalPayloadStorageDriver interface. A
driver that implements it has payloads staged and stored only after the
completion passes every limit. The built-in local driver implements it, and
now stores objects atomically after checking their length and hash. Other
drivers, including the filesystem driver, keep storing payloads as before.Cooperative cancellation now selects the live task for a retried activity, preserves completed tasks, and allows successful and duplicate stop acknowl
Cooperative cancellation now selects the live task for a retried activity, preserves completed tasks, and allows successful and duplicate stop acknowledgements. Application heartbeats retry rolled-back SQLite metadata contention while preserving attempt authority, deadlines and one progress event per accepted call.
Changes: #924 and #925. The exact release source is b2a3369110b8cbdf060802ff91e475110e66e333, qualified by the full main build, including all database shards, coverage, Laravel upgrades and Sonar.
The normal published-package verification runs for this tag. PHP Server images and pinned applications require their own adoption and installed acceptance.
Recognize ActivityOutcomeRejected diagnostics in history pages, timeline inspection, exports and cold replay while preserving business command positio
Recognize ActivityOutcomeRejected diagnostics in history pages, timeline inspection, exports and cold replay while preserving business command positions. Diagnostic summaries use the existing UTF-8 storage bound and tolerate malformed fields without rewriting history payloads.
This is the reader prerequisite for #878 / #901. It does not produce the new event or accept late outcomes. Update and qualify participating applications, Server roles, inspection tools and rollback artifacts, then normal SDK and Waterline readers, before enabling a separate producer. The current portable schema/catalog and closed-run terminal-history decisions remain prerequisites for that producer.
Source: ae18295693273907d0f750b479eb9b8a593e5803; tree c75f48065008711a43238943d8e0dea0b3d1e152. PR #919.
ActivityOutcomeRejected diagnostics in history pages,
timeline inspection, exports and cold replay. The reader preserves business
replay positions and does not produce the event or accept a refused outcome.Preserve canonical activity failure messages beyond MySQL TEXT capacity and when messages contain NUL bytes. Diagnostic lists remain bounded, while mo
Preserve canonical activity failure messages beyond MySQL TEXT capacity and when messages contain NUL bytes. Diagnostic lists remain bounded, while model, detail and history reads retain the complete accepted message.
This patch also prevents expired activity callbacks from racing retained-state pruning and fixes PostgreSQL recent-failure and history-budget reads for NUL-bearing history.
Upgrade with the normal package update and php artisan migrate. The additive migration retains existing rows and failure evidence on package rollback. MySQL/MariaDB require an instant column addition; unsupported layouts fail rather than copying a live table. See failure message storage for migration and rollback behavior.
Laravel 11 now requires 11.56.1 or later for the upstream TokenGuard correction. Laravel 9/10 compatibility remains limited to the documented upgrade/drain path and must not expose TokenGuard while migrating. The legacy advisory remains tracked in #922.
Server images and deployed applications require their own dependency update, migration and acceptance. The rejected-outcome history emitter remains separate.
Changes: #918, #920 and #921. Full changelog: https://github.com/durable-workflow/workflow/compare/2.12.0...2.12.2
Read recent failure links and rebuild history-budget counters on PostgreSQL when retained messages contain NUL bytes. Preserve canonical history and exact byte counts without loading full payloads into diagnostic readers.
Refuse expired or settled activity callbacks under the original run lock before acquiring task or execution locks. Preserve committed renewals and the normal live-outcome path. Concurrent retention can finish without a reverse lock dependency or restored details.
Acquire SQLite writer admission before reading a retained run for pruning, within the existing transaction policy and original publication budget. This correction emits no new history event and needs no new event reader.
Require the upstream TokenGuard string-validation fix on Laravel 11, raising its minimum to 11.56.1. Keep the maintained Laravel 12/13 minimums unchanged.
Record the newly published Composer TokenGuard advisory against the retained Laravel 9–11 upgrade/drain compatibility paths, with an explicit review date. Maintained Laravel 12/13 minimums already include the upstream fixes.
Preserve canonical failure messages beyond MySQL TEXT capacity and across NUL-bearing PostgreSQL text parameters. Add nullable encoded overflow storage without backfilling existing rows, retain the exact public message and keep compatibility previews bounded. Metadata-lock waits are bounded and a package rollback retains the added storage and failure evidence.
Datetime search attributes now preserve UTC instants and microseconds, including non-UTC input offsets. Embedded visibility predicates use those exact
Datetime search attributes now preserve UTC instants and microseconds, including non-UTC input offsets. Embedded visibility predicates use those exact instants.
The new workflow:v2:repair-search-datetimes command previews or explicitly repairs selected datetime projections from matching retained, typed history. It refuses incomplete or conflicting evidence.
Derived timeline summaries are bounded by UTF-8 bytes, preserving complete canonical failure evidence and callback-authority recovery notices.
Implemented in #914, #916, and #917. Exact main 789f3b64 passed the full database, upgrade, quality and coverage checks, CodeQL, public boundary and Sonar. Normal published-package and affected-consumer verification follows publication; source checks do not claim that acceptance.
workflow:v2:repair-search-datetimes to preview or explicitly repair
selected datetime projections from matching retained, typed history.CLI history exports now preserve whole-valued floats and signed zero in compact and pretty JSON, including nested memo projections. The unchanged cano
CLI history exports now preserve whole-valued floats and signed zero in compact and pretty JSON, including nested memo projections. The unchanged canonical checksum and Avro authority verify correctly after export.
Previously generated lossy bundles remain invalid. Export retained history again after upgrading.
Fixes #912 in #913. Exact candidate and independent main checks pass at ae6bf6c8a7c4b9597ee4eecb15ecd786e78b0430.
Prevent Laravel dependency injection from silently replacing or discarding a recorded activity argument after an activity method changes.
Prevent Laravel dependency injection from silently replacing or discarding a recorded activity argument after an activity method changes.
Incompatible binding now records ActivityEntryArgumentsMismatchException before the activity body runs. Both queued and local activities use the guard. Legitimate injected services, scalar arguments, trailing defaults and variadic handlers retain their behavior.
The activity's recorded retry policy and deadlines still apply. A configured retry can complete the same run and activity with the original input after compatible code is restored. A single-attempt policy can close the run normally; this patch does not reopen terminal runs or grant extra attempts.
Exact source: 71213ae1e0338cda5967e0840782a981530ed854. All candidate and independent main checks pass, including supported SQL backends, coverage, static analysis, regression corpus and Laravel/PHP upgrade tests. Published-package verification and affected-consumer results are recorded on #910 and #911.
This is a compatible patch with no schema or worker-protocol change. Published Server images need a separately qualified package update and rebuild.
Fix activity delivery during rolling deployments with multiple worker builds.
Fix activity delivery during rolling deployments with multiple worker builds.
Versioned workers can poll unpinned activity tasks as well as tasks pinned to their own build. Namespace, queue, connection, activity type, readiness and availability filters remain enforced, and existing claim authority is unchanged.
The complete bridge suite passes on SQLite, PostgreSQL 17.11 and MySQL 8.0.46 (68 tests / 412 assertions on each). Test-first regressions reproduce the old null/empty-pin exclusion. See #908 and #909.
This is a compatible patch with no schema or worker protocol changes. Pinned Server images need a separate package update and qualified rebuild; this library release does not update an already published Server image.
Fix activity lease repair at fractional-second deadlines. Repair selectors now preserve the task model's full timestamp precision for lease expiry, av
Fix activity lease repair at fractional-second deadlines. Repair selectors now preserve the task model's full timestamp precision for lease expiry, availability, dispatch backoff and claim backoff. An expired lease no longer waits for the next whole-second cutoff to become eligible.
This compatible patch is published from commit 30974a3c07932e8d7e026b7fdd4bb3a27eda082b. Published-package verification passes all 17 jobs across the supported Laravel/PHP upgrade matrix. Server images pinned to an earlier Core version require a separate qualified update.
Legacy Laravel workflows can now set a delay when a queued workflow, activity, signal or child-completion job cannot acquire its overlap lock. This ad
Legacy Laravel workflows can now set a delay when a queued workflow, activity, signal or child-completion job cannot acquire its overlap lock. This addresses rapid database-queue retries reported in #897.
use Workflow\WorkflowOptions;
$workflow->start('argument', new WorkflowOptions(overlapReleaseAfter: 5));
The delay is a nonnegative integer in seconds. The default remains zero. Existing serialized options, routing, lock ownership and expiry remain compatible. Children and continue-as-new inherit the stored setting unless explicit options replace it. Existing runs keep their stored options.
Laravel counts released queue reservations toward attempt limits. Choose a suitable delay and retry budget; the option does not make an arbitrarily low attempt limit safe.
This setting applies to Workflow\Workflow. V2 uses durable task claims and ignores the legacy setting. No portable protocol or database migration changes are included.
Applications staying on 1.x can install 1.1.0, which contains the scoped backport.
Implementation and qualification: #903. Documentation: durable-workflow/durable-workflow.github.io#356.
WorkflowOptions::overlapReleaseAfter delay for legacy
workflow, activity, signal and child-completion jobs contending on overlap
locks. Preserve the zero default, stored options, lock ownership and expiry.Preserve the canonical lease-loss classification and reason in activity attempt views. A later task cancellation no longer hides an already recorded e
Source: #900. Release preparation: #902. This patch changes no lease authority, schema or protocol and does not include the proposed rejected-outcome audit.
The published package/platform and all supported Laravel upgrade checks pass in Release Verification. PHP Server image and Waterline consumer follow-through remain recorded on #878.
Keeps eligible embedded activity leases alive through a fresh metadata-only PHP helper while the original callback retains its process, injected servi
Keeps eligible embedded activity leases alive through a fresh metadata-only PHP helper while the original callback retains its process, injected services, captured state and application connections. The helper preserves original task, worker, session and business deadlines, stops when its original parent exits, and is joined before recording the outcome. It does not fabricate heartbeat progress or revive lost authority.
SQLite maintenance reopens the actual current journal mode instead of replaying stale startup configuration. Environments without the required process capabilities retain the original grant behavior.
Exact source: 1a5ec33f83a68021a163bc78d06e990c1eedc5da, whole c2ef09bd137f5a7b307d1b14bf402c2acada1882, identical to the qualified release candidate. Candidate checks and the preceding implementation's full main checks pass; fresh-main checks and published-package verification are separate.
The service-mode Server/SDK integration and actual six-minute plus killed-worker acceptance remain tracked under #878. This package publication alone does not complete that cross-repository issue.
Add the optional ActivityLeaseRenewal bridge contract and configurable activity-task leases, retaining the 300-second default.
ActivityLeaseRenewal bridge contract and configurable activity-task leases, retaining the 300-second default.Automatic inline callback maintenance requires the separately qualified Server/SDK or embedded integration. This release's optional renewal contract alone does not keep an inline callback alive, and does not prove that a stale callback has physically stopped. Workflow #878 remains open for that follow-through.
Source: cafcd20519877ed640733481ecbc6b97cd2b56ce, identical release tree 3de4f2a76e315709d592b0d2f4270bd8a520a91b. Full main build 39576379949 and release-candidate checks pass before tagging.
ActivityLeaseRenewal bridge role and configurable activity
task leases, with the existing 300-second default. Runtime renewal records no
application progress and cannot extend authored activity or workflow timeouts.Adds optional execution tracing and Laravel log correlation for embedded Durable Workflow. Applications can carry W3C context across independent queue
Adds optional execution tracing and Laravel log correlation for embedded Durable Workflow. Applications can carry W3C context across independent queue workers, activity attempts, retries and recovery while preserving distinct durable, request and trace identities.
Tracing is disabled by default. Applications own the optional OpenTelemetry API (1.10.1 or later in the 1.x line), providers, propagators, sampling, exporters and shutdown. Default installations require no OpenTelemetry package. Baggage is bounded and propagates only through an explicit allowlist, empty by default. Default attributes exclude business arguments, results and HTTP bodies.
Run the package migrations against the application's configured workflow models and storage connection, enable the tracing flag and Laravel log tap together, and restart workers. Nullable carrier columns preserve existing rows. Tracing-enabled workflow execution uses cold replay and clears retained continuation frames. Telemetry does not confer authority or change retry, lease, receipt, cancellation or business outcome rules.
The embedded guide documents relationships, bounded exporter operation, migration, legacy executions and limitations. Implementation #788 and report #881 are merged. Immutable release source is 2bc1ba0b57df34b74de02427d2225bd8e4dc6322, with the same production files as the qualified implementation.
Qualification includes the full source and normal protected suites, canonical coverage gate, source-specific Sonar review, installed migration controls, a default no-API consumer, PHP 8.1 / Monolog 2 compatibility, and all 16 published Laravel 9–13 / PHP 8.1–8.4 upgrades in the normal registry matrix. Installed package identity checks cover all 966 production source/resource files.
The original two-CPU fan-out observer failures remain failed. A corrected test-only runner on four CPUs passes both 64-member installed fan-out cases, with the original deadlines, complete histories/results, one attempt per activity, drained queues and no failed jobs. It narrowly classifies a recovered SQLite queue-producer insertion. #883 records that runner correction and its separate normal gates; it changes no release artifact. These functional results do not establish a two-CPU capacity guarantee.
The fixed study against normal published Core 2.8.0 passes three of four predeclared CPU budgets. Disabled fan-out fails the original 5% target: mean +1.41%, 95% interval [-3.61%, +6.43%], with +4.53% baseline drift. The maintainer explicitly accepts a post-study component release allowance of 7.5% disabled / 15% enabled. This preserves the failed 5% result and uncertainty; it is not a predeclared four-budget pass or a performance SLA. See the complete report.
This release delivers the embedded Laravel component. Workflow #773 remains open for the actual published carrier-enabled Server tuple, cross-SDK acceptance and downstream documentation/consumer delivery. No Cloud deployment, native wire capability or frozen-catalog adoption is claimed.
This minor release adds optional bounded live execution for persistent embedded PHP database-queue workers. Enable selected workflows with an explicit
This minor release adds optional bounded live execution for persistent embedded PHP database-queue workers. Enable selected workflows with an explicit allowlist and immutable application build ID. Sequential remote activities and positive ordinary timers can resume after committed outcomes; unsupported features and object-capable entry signatures use cold replay.
Each worker retains at most four frames for a nonrenewing 30 seconds, subject to a 128 MiB worker-RSS threshold. Fresh task generation, connection/session identity, configuration and committed history checks fence reuse. Rollback, worker loss and uncertain broker acknowledgement discard retained state. Normal worker supervision and durable recovery remain required.
The release preserves 2.7.3's entry-binding refusal, supported repair of the same original, and genuine application-body failures. There are no #748 database migrations or portable protocol changes.
See the configuration and operating limits and measurement reports. Results vary by workload; retention does not establish higher safe SQLite capacity. The reported four-worker 1,000-event failures remain failed and explicit.
Source qualification passes 71 focused continuation/process/cache/profile controls and the full 4,301-test unit suite, all database feature shards, all five supported Laravel source-upgrade journeys, regression corpus, style/static analysis and the 93.75% combined coverage ratchet (no exclusions). The exact PR head also passes dependency-security, public-boundary and Sonar checks. Source build and protected PR build retain the original results.
Normal Packagist installation is verified against the immutable release: source/dist references and all 955 product files match, product classes/functions load from vendor, and 71 installed-package controls pass 2,606 assertions. The published-package verifier and all 16 PHP/Laravel upgrade cells pass across the five supported Laravel families.
Sample App #264 and Server #552 are merged and qualified with normal published Core 2.8.0. Sample's native AMD64/ARM64 installed publication passes. Its subsequent #265 security delivery retains the Core pins and passes its own installed publication. Server 2.9.11 / chart 0.2.1 publication passes, including published SQLite first-run, source-free MySQL Compose and anonymous chart clean install. Exact-image acceptance also matches all 955 Core files and passes 18 PostgreSQL/MySQL/MariaDB TLS probes. Native/shared/frozen migration adoption and Cloud rollout remain separate.
Normally installed Workflow 2.7.3 is available from Packagist at immutable source dda951d93f58b82447348f1b3b4e4e17dba149fb. A fresh Laravel applicatio
Normally installed Workflow 2.7.3 is available from Packagist at immutable source dda951d93f58b82447348f1b3b4e4e17dba149fb. A fresh Laravel application resolved that exact package normally, with no path repository, source autoloader override or mounted product checkout.
Five genuine cold-process controls pass: added required entry argument and scalar-to-DTO incompatibility produce inspectable nonterminal workflow_entry_arguments_mismatch; restoring compatible code plus supported repair completes those same original runs/tasks, retaining exact recorded input, history prefix and original microsecond 30-second run / 60-second execution deadlines. The cold typed result preserves int64 maximum/null/Unicode. Genuine workflow-body TypeError and an application deliberately throwing the internal mismatch exception both remain ordinary business failures. An added optional parameter remains compatible.
Stopped read-only SQL independently confirms 5 runs (3 completed / 2 intentional business failures), 15 tasks (13 completed / 2 failed), 27 events, 2 deliberate repairs, 22 cold workers exiting 0, no remaining queue jobs or failed queue deliveries. Package runtime files independently match the immutable tagged source.
The initial audit used SQLite read-only SQL on a writable directory and failed its file-inventory guard when SQLite created empty WAL/shared-memory files. That failed audit is retained as a reader failure. A separate readback with the whole source mounted kernel read-only passes every row/prefix/deadline/repair check and leaves database/WAL hashes unchanged. No business original, deadline or assertion was altered. The earlier tool-container bootstrap failed before package install or business admission; its idle shell later required 137/no OOM after all actual workers had exited 0.
The repository's independent published-package/Laravel upgrade workflow 39434329508 passes every job, including normal registry resolution and the Laravel 9–13 / PHP 8.1–8.4 upgrade matrix. Source main CI and the applicable quality/security/boundary gates also pass on exact dda951d93f58b82447348f1b3b4e4e17dba149fb / tree f4045b09b42ab9dbc5a4d5fbec1dd9dca2cf028c. Workflow 2.7.3 is qualified and published; downstream consumers must qualify their own exact package/image adoption. This embedded runtime release does not qualify a different Server image, service SDK or native runtime.
Keep an update's caller key independent of proxy request-correlation headers. Preserve opaque case, long Unicode keys and retained legacy receipts; re
New update rows use bounded update-v1: tokens in the internal indexed
workflow_commands.request_id column. Read logical keys through
WorkflowCommand::requestId() or command context. Existing rows are preserved.
Drain and upgrade all processes writing directly to the database before admitting new updates. After the new format is written, an old writer is an unsafe rollback target. Remote SDK clients keep the same HTTP contract.
Self-hosted Server adoption is a separate coordinated release. This library tag does not update a pinned Server image; keep its supported package pin.
Related: Server #505, Server #506, Core implementation #862.
Fix long durable task delays on standard SQS queues.
Fix long durable task delays on standard SQS queues.
Source and qualification: all candidate and independent main feature, coverage/ratchet, supported Laravel upgrade, quality, security and corpus gates pass. Actual Laravel/AWS standard queue serialization and redelivery are exercised with the official SDK mock HTTP transport. This does not qualify a live AWS service or FIFO queue.
Published-package verification and normal installed-consumer results are tracked in #851. Upgrading from before2.7.0 also requires the 2.7.0 reader-upgrade precautions.
Give an eligible declared signal an ordered delivery turn before a remote workflow task completes or continues as new.
run_closed_before_signal_applied, a closure time and a linked SignalRejected event. Original acceptance receipts remain immutable.SignalRejected extends the persisted history enum. Stop and upgrade all embedded/direct-database readers, including Waterline when it shares the database, before enabling the writer. Older Core versions cannot decode this event. Mixed-version reading and rolling back old readers after it is stored are unqualified; repair forward or restore the complete pre-upgrade recovery point.
No database schema or SDK protocol-floor change is required. Service-mode SDKs retain their existing task-completion mechanics.
The exact source tree passed full MySQL/PostgreSQL/MariaDB features, unit coverage/ratchet, five Laravel upgrade configurations, replay-corpus counterfactuals, quality/security/boundary checks and Sonar (1.0% new duplication, zero open/confirmed issues under unchanged settings). Twelve real normally installed Python SDK cases cover complete/CAN, declared/missing declarations and default/named namespaces. Declared signals apply once; unresolved legacy signals close explicitly with no false application. Original deadlines and database bytes are preserved.
Implementation: #858. Follow-through: #857. Published-package verification and the affected PHP Server image/consumer are recorded there separately; publishing this Core package does not update a pinned image or managed deployment.
SignalRejected.
Older Core history enums cannot parse the new event; drain old readers
during the upgrade and do not roll them back onto affected histories.One column per quarter.