Add rewind support

[YunchuWang]

Summary

Adds rewind support to the durabletask-js SDK, reaching parity with the Python reference implementation in microsoft/durabletask-python#121. Rewind resets a failed orchestration to a runnable state so that failed activities re-run while previously successful results are preserved.

Before this change, durabletask-js had only the client entry point (client.rewindInstance(instanceId, reason)); the worker-side history rewrite and the in-memory backend support were missing, so rewind was a no-op end-to-end.

Architecture

Rewind spans three layers, split exactly as in the Python reference:

1. Client (already existed): rewindInstance(instanceId, reason) sends a RewindInstanceRequest over gRPC.
2. Backend (in-memory, new here): validates the instance is FAILED, appends an ExecutionRewoundEvent, flips status to RUNNING, and re-dispatches the work item. On terminal completion it now appends an executionCompleted "bookend" event — this is the precondition the worker uses to detect a rewind.
3. Worker (new here): when it sees executionRewound in the new events and executionCompleted in the committed history, it short-circuits normal replay and calls buildRewindResult(), which rewrites history — dropping taskFailed, the matching failed taskScheduled, subOrchestrationInstanceFailed, and executionCompleted; assigning a fresh execution id to executionStarted; keeping successful results and the executionRewound event — wrapped in a RewindOrchestrationAction. The backend then applies the new history and recursively rewinds failed sub-orchestrations.

Note: The history rewrite in buildRewindResult is a backend contract — the in-memory and DTS backends rely on it. Any change to how history is rewritten must be mirrored in those backends.

Changes

• Proto: re-synced orchestrator_service.proto from durabletask-protobuf main and regenerated the JS/TS stubs. Adds RewindOrchestrationAction and OrchestratorAction.rewindOrchestration = 9. The re-sync also pulls in benign, purely-additive upstream drift: ActivityRequest.tags, PurgeInstanceFilter.timeout, and [deprecated=true] annotations on OrchestratorResponse.isPartial/chunkIndex. No fields were renumbered or removed.
• Worker: new worker/rewind.ts (buildRewindResult); orchestration-executor.ts gains the rewind short-circuit and an informational EXECUTIONREWOUND no-op case.
• In-memory backend: rewindInstance, prepareRewind, and processRewindOrchestrationAction (including recursive and purged sub-orchestration handling), the terminal executionCompleted bookend, and a failureDetails presence guard. New helpers newExecutionCompletedEvent / newExecutionRewoundEvent; new rewindOrchestration() test-client method.

Tests

• 14 new unit cases for buildRewindResult (test/build-rewind-result.spec.ts).
• 8 new in-memory e2e cases (test/rewind-e2e.spec.ts): failed activity, preserves successful results, not-found, non-failed instance, sub-orchestration, purged sub-orchestration, no reason, and rewind-twice.
• Full suite: 1119 passing / 63 suites. build:core and lint are green.

Known limitation

Consistent with the Core/.NET/Python implementations: timer events emitted between retry attempts of an activity scheduled with a RetryPolicy are not removed during the history rewrite, so rewinding a retry-policy activity is not currently supported (documented in worker/rewind.ts).

Reference

Ports andystaples/durabletask-python@d5052d7 (worker._build_rewind_result, testing/in_memory_backend.py, internal/helpers.new_execution_completed_event, and the corresponding unit + e2e tests).

[Copilot]

Pull request overview

Adds end-to-end rewind support to the Durable Task JS SDK (worker + in-memory backend) so failed orchestrations can be reset to runnable state while preserving prior successful results, matching the durabletask-python reference behavior.

Changes:

• Added worker-side history rewrite logic (buildRewindResult) and executor short-circuiting when a rewind is detected.
• Implemented in-memory backend rewind handling, including terminal executionCompleted bookend and processing of RewindOrchestrationAction (with recursive sub-orchestration rewind).
• Updated protocol/proto stubs for RewindOrchestrationAction and related upstream additive proto changes; added unit + e2e tests for rewind.

Reviewed changes

Copilot reviewed 10 out of 11 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
packages/durabletask-js/src/worker/rewind.ts	Implements history rewrite to drop failed activity/sub-orchestration results and produce a RewindOrchestrationAction.
packages/durabletask-js/src/worker/orchestration-executor.ts	Detects rewind work items and short-circuits to buildRewindResult; adds no-op handling for executionRewound events.
packages/durabletask-js/src/testing/in-memory-backend.ts	Adds in-memory backend rewind entrypoint, applies rewind action, appends terminal executionCompleted bookend, and supports recursive sub-orchestration rewind.
packages/durabletask-js/src/testing/test-client.ts	Exposes rewind operation from the test client to the in-memory backend.
packages/durabletask-js/src/utils/pb-helper.util.ts	Adds helpers for creating executionCompleted and executionRewound history events.
packages/durabletask-js/test/build-rewind-result.spec.ts	Unit tests for the worker-side history rewrite behavior.
packages/durabletask-js/test/rewind-e2e.spec.ts	In-memory backend e2e tests covering rewind scenarios, including sub-orchestrations and purged sub-orchestrations.
internal/protocol/protos/orchestrator_service.proto	Proto updates including RewindOrchestrationAction and other additive upstream drift.
packages/durabletask-js/src/proto/orchestrator_service_pb.js	Regenerated JS proto stubs reflecting proto updates.
packages/durabletask-js/src/proto/orchestrator_service_pb.d.ts	Regenerated TS declarations reflecting proto updates.
internal/protocol/SOURCE_COMMIT	Updates the recorded upstream protocol source commit.