Distillation

Date: Sep 17, 2026

• 🔴 (10:02) User directed: “Replace local ownership with durable multi-node ownership when clustered.”
• 🟡 (10:02) User supplied the session-input implementation for read-only review. Visible APIs and fields include admit, isRecoveryInput, find, SessionInputTable.wake_pending, SessionRecoveryTable.continuation_message_id, and SessionMessage.ID.
• 🟡 (10:02) Session-input admission rejects an input already recognized as recovery work by dying with new LifecycleConflict({ id: input.id }); this check appears in multiple admission paths.
• 🟡 (10:02) Session-input logic returns admission results shaped as { admitted: stored, shouldWake }, treats an existing wakePending value as immediately wakeable, and validates recovery projection identity, including input.session_id !== recovery.session_id.
• 🟡 (10:02) User supplied packages/core/src/session/runner/index.ts for read-only review. It exports SessionRunner, defines RunError, Interface, and Service, and documents the runner as executing “one local continuation from already-recorded Session history.”
• 🟡 (10:02) SessionRunner.RunError is the union LLMError | SessionRunnerModel.Error | MessageDecodeError | ContextSnapshotDecodeError | SystemContext.InitializationBlocked | ToolOutputStore.Error.
• 🟡 (10:02) SessionRunner.Interface.run accepts { sessionID: SessionSchema.ID, force: boolean, executionID: string, ownerID: string, recoveryInputID?: SessionMessage.ID } and returns Effect.Effect<void, RunError>.
• 🟡 (10:02) SessionRunner.Service is registered as Context.Service<Service, Interface>()("@opencode/v2/SessionRunner").
• 🟡 (10:02) User supplied the session-runner implementation for read-only review. Its design comment says to retain orchestration over smaller collaborators rather than rebuilding the legacy implementation.
• 🟡 (10:02) Runner orchestration starts each recorded local call eagerly and awaits all settlements before continuation; visible execution uses concurrency: "unbounded".
• 🟡 (10:02) Runner ownership checks match both SessionExecutionTable.owner_id and the active claim’s claim.ownerID; the implementation still marks clustered durable multi-node ownership as unfinished.
• 🟡 (10:02) Provider requests include the header "x-session-affinity": session.id.
• 🟡 (10:02) Runner captures a starting snapshot with snapshots.capture(), restores the provider stream through restore(providerStream).pipe(Effect.exit), publishes overflow failures, and can continue through runAfterOverflowCompaction.
• 🟡 (10:02) Runner ignores interrupted failed settlements when applying ordinary failure handling via settled._tag === "Failure" && !Cause.hasInterrupts(settled.cause).
• 🟡 (10:02) When the provider fails to return a tool result, runner publication invokes publisher.failUnsettledTools("Provider did not return a tool result", true).
• 🟡 (10:02) Runner exits without a provider turn when !recovering && !input.force && !hasSteer && !hasQueue; otherwise continuation processing can repeat under while (needsContinuation).
• 🟡 (10:02) Runner design notes state that durable continuation recovery remains a separate future slice requiring an explicit retry policy.