diff --git a/apps/server/src/orchestration-v2/Orchestrator.ts b/apps/server/src/orchestration-v2/Orchestrator.ts
index bcec0a13cf20a..5d3c4391b1fe8 100644
--- a/apps/server/src/orchestration-v2/Orchestrator.ts
+++ b/apps/server/src/orchestration-v2/Orchestrator.ts
@@ -6279,6 +6279,9 @@ const makeOrchestrator = Effect.fn("orchestrationV2.Orchestrator.layer")(functio
         ...(command.restartContinuationOfRunId === undefined
           ? {}
           : { restartContinuationOfRunId: command.restartContinuationOfRunId }),
+        ...(command.usageLimitContinuationOfRunId === undefined
+          ? {}
+          : { usageLimitContinuationOfRunId: command.usageLimitContinuationOfRunId }),
         ...(continuedTaskId === undefined ? {} : { delegatedTaskId: continuedTaskId }),
         ...(freshContext ? { freshContext: true } : {}),
       };
@@ -6659,6 +6662,9 @@ const makeOrchestrator = Effect.fn("orchestrationV2.Orchestrator.layer")(functio
         ...(command.restartContinuationOfRunId === undefined
           ? {}
           : { restartContinuationOfRunId: command.restartContinuationOfRunId }),
+        ...(command.usageLimitContinuationOfRunId === undefined
+          ? {}
+          : { usageLimitContinuationOfRunId: command.usageLimitContinuationOfRunId }),
         ...(dispatchMode.type === "defer_start" && dispatchMode.workspaceStrategy !== undefined
           ? { workspacePreparation: dispatchMode.workspaceStrategy }
           : {}),
@@ -7445,6 +7451,9 @@ const makeOrchestrator = Effect.fn("orchestrationV2.Orchestrator.layer")(functio
       ...(command.restartContinuationOfRunId === undefined
         ? {}
         : { restartContinuationOfRunId: command.restartContinuationOfRunId }),
+      ...(command.usageLimitContinuationOfRunId === undefined
+        ? {}
+        : { usageLimitContinuationOfRunId: command.usageLimitContinuationOfRunId }),
       ...(continuedTaskId === undefined ? {} : { delegatedTaskId: continuedTaskId }),
       ...wakeWorkStartedAt(projection.runs, command),
     };
diff --git a/apps/server/src/orchestration-v2/ProviderRuntimeRecoveryService.ts b/apps/server/src/orchestration-v2/ProviderRuntimeRecoveryService.ts
index 222747c413982..89e43e5b1ede4 100644
--- a/apps/server/src/orchestration-v2/ProviderRuntimeRecoveryService.ts
+++ b/apps/server/src/orchestration-v2/ProviderRuntimeRecoveryService.ts
@@ -442,6 +442,64 @@ export const make = Effect.gen(function* () {
           },
         });
       }
+      const unsentResumeRunIds = runs.flatMap((run) =>
+        run.startedAt === null && run.usageLimitContinuationOfRunId !== undefined ? [run.id] : [],
+      );
+      const limitRecovery = projection.thread.limitRecovery;
+      // The unsent resume spent the worker's request id, so a fresh one lets it send again. An
+      // untagged resume from an older server stays cancelled: its copy couldn't be hidden.
+      if (limitRecovery?.autoResume === true && unsentResumeRunIds.length > 0) {
+        events.push({
+          id: yield* allocateEventId(),
+          type: "thread.metadata-updated",
+          threadId: projection.thread.id,
+          occurredAt: now,
+          payload: {
+            ...projection.thread,
+            limitRecovery: { ...limitRecovery, requestId: commandId },
+            updatedAt: now,
+          },
+        });
+      }
+      // The worker sends the auto-resume again, so the unsent copy leaves the transcript the way
+      // cancelled queued input does.
+      const unsentResumePrompts =
+        unsentResumeRunIds.length === 0
+          ? []
+          : (yield* projections
+              .getThreadRecords(projection.thread.id, ["turnItems"], {
+                turnItemTypes: ["user_message"],
+                turnItemRunIds: unsentResumeRunIds,
+              })
+              .pipe(
+                Effect.mapError(
+                  (cause) =>
+                    new ProviderRuntimeRecoveryError({
+                      operation: "read-projections",
+                      threadId: projection.thread.id,
+                      cause,
+                    }),
+                ),
+              )).turnItems;
+      const unsentResumePromptIds = new Set(unsentResumePrompts.map((item) => item.id));
+      for (const item of unsentResumePrompts) {
+        if (item.type !== "user_message") continue;
+        const active =
+          item.status === "pending" || item.status === "running" || item.status === "waiting";
+        events.push({
+          id: yield* allocateEventId(),
+          type: "turn-item.updated",
+          threadId: projection.thread.id,
+          ...(item.runId === null ? {} : { runId: item.runId }),
+          occurredAt: now,
+          payload: {
+            ...item,
+            ...(active ? { status: "cancelled", completedAt: now } : {}),
+            inputIntent: "queued_turn",
+            updatedAt: now,
+          },
+        });
+      }
       for (const question of questions) {
         events.push({
           id: yield* allocateEventId(),
@@ -552,6 +610,7 @@ export const make = Effect.gen(function* () {
         for (const item of projection.turnItems.filter(
           (candidate) =>
             candidate.runId === run.id &&
+            !unsentResumePromptIds.has(candidate.id) &&
             (candidate.nodeId === null || !messageRequestNodeIds.has(candidate.nodeId)) &&
             !isAppOwnedDelegationItem(candidate) &&
             // A link card waits on the user, not the run, and answers by message.
diff --git a/apps/server/src/orchestration-v2/runtimeLayer.test.ts b/apps/server/src/orchestration-v2/runtimeLayer.test.ts
index b90f491dfa94e..170ceee73a331 100644
--- a/apps/server/src/orchestration-v2/runtimeLayer.test.ts
+++ b/apps/server/src/orchestration-v2/runtimeLayer.test.ts
@@ -1,6 +1,7 @@
 import { PullRequestProviderError } from "@t3tools/source-control-core/server/PullRequestProvider";
 import { SourceControlProviderKind } from "@t3tools/contracts";
 import { limitRecoveryCommand } from "./UsageLimitRecoveryWorker.ts";
+import { createOrchestrationV2TurnItemVisibility } from "@t3tools/shared/orchestrationV2Timeline";
 import * as McpProviderSessions from "@t3tools/provider-core/server/McpProviderSessions";
 import * as SourceControlProviderRegistry from "../sourceControl/SourceControlProviderRegistry.ts";
 import * as NodeCrypto from "@effect/platform-node/NodeCrypto";
@@ -8136,6 +8137,8 @@ it.layer(layerTest)("usage-limit recovery", (it) => {
 
   it.effect.each([
     "resume",
+    "resume-started",
+    "resume-untagged",
     "queued-resume",
     "rearm",
     "snooze-race",
@@ -8614,17 +8617,97 @@ it.layer(layerTest)("usage-limit recovery", (it) => {
         yield* orchestrator.dispatch(resume);
       }
       const after = yield* orchestrator.getThreadProjection(threadId);
-      if (scenario === "resume") {
+      if (
+        scenario === "resume" ||
+        scenario === "resume-started" ||
+        scenario === "resume-untagged"
+      ) {
         assert.deepInclude(after.messages.at(-1), {
           text: "Continue where you left off.",
           createdBy: "agent",
           creationSource: "server",
         });
+        const continuation = after.runs.at(-1)!;
+        assert.equal(continuation.status, "starting");
+        if (scenario === "resume-started") {
+          const startedAt = yield* DateTime.now;
+          yield* events.write({
+            events: [
+              {
+                id: EventId.make("recovery:continuation-started:resume-started"),
+                type: "run.updated",
+                threadId,
+                runId: continuation.id,
+                occurredAt: startedAt,
+                payload: { ...continuation, status: "running", startedAt },
+              },
+            ],
+          });
+        }
+        if (scenario === "resume-untagged") {
+          // A server older than the run tag stored the resume run without it.
+          const { usageLimitContinuationOfRunId: _, ...untagged } = continuation;
+          yield* events.write({
+            events: [
+              {
+                id: EventId.make("recovery:continuation-untagged"),
+                type: "run.updated",
+                threadId,
+                runId: continuation.id,
+                occurredAt: yield* DateTime.now,
+                payload: untagged,
+              },
+            ],
+          });
+        }
+        yield* (yield* ProviderRuntimeRecoveryService.ProviderRuntimeRecoveryService).reconcile(
+          "startup",
+        );
+        const restartedShell = (yield* orchestrator.getShellSnapshot()).threads.find(
+          (thread) => thread.id === threadId,
+        )!;
+        const retry = limitRecoveryCommand(
+          restartedShell,
+          true,
+          DateTime.toEpochMillis(yield* DateTime.now),
+        );
+        if (scenario === "resume-started") assert.isNull(retry);
+        else assert(retry?.type === "message.dispatch");
+        if (scenario === "resume") {
+          assert.notEqual(retry!.commandId, resume!.commandId);
+        } else {
+          assert.equal(
+            restartedShell.limitRecovery?.requestId,
+            after.thread.limitRecovery?.requestId,
+          );
+        }
+        if (retry !== null) {
+          yield* orchestrator.dispatch(retry);
+          yield* orchestrator.dispatch(retry);
+        }
+        const resumed = yield* orchestrator.getThreadProjection(threadId);
+        const visible = createOrchestrationV2TurnItemVisibility({
+          runs: resumed.runs,
+          attempts: resumed.attempts,
+          items: resumed.turnItems,
+        });
+        assert.deepEqual(
+          resumed.turnItems.flatMap((item) =>
+            item.type === "user_message" && visible(item) ? [item.runId] : [],
+          ),
+          [run.id, resumed.runs.at(-1)!.id],
+        );
+        assert.deepEqual(
+          resumed.runs.map((row) => row.status),
+          scenario === "resume" ? ["failed", "cancelled", "starting"] : ["failed", "cancelled"],
+        );
       }
       assert.lengthOf(
         after.runs,
         before.runs.length +
           (scenario === "resume" ||
+          scenario === "resume-started" ||
+          scenario === "resume-untagged" ||
           stopBeforeArm ||
           scenario === "queued-resume" ||
           scenario === "independent-patches" ||
@@ -8636,6 +8719,8 @@ it.layer(layerTest)("usage-limit recovery", (it) => {
         after.messages,
         before.messages.length +
           (scenario === "resume" ||
+          scenario === "resume-started" ||
+          scenario === "resume-untagged" ||
           stopBeforeArm ||
           scenario === "queued-resume" ||
           scenario === "independent-patches" ||
diff --git a/packages/contracts/src/orchestrationV2.ts b/packages/contracts/src/orchestrationV2.ts
index 461ec5e4b3f3e..5c430ef75581a 100644
--- a/packages/contracts/src/orchestrationV2.ts
+++ b/packages/contracts/src/orchestrationV2.ts
@@ -692,6 +692,7 @@ export const OrchestrationV2Run = Schema.Struct({
   checkpointId: Schema.NullOr(CheckpointId),
   contextHandoffId: Schema.NullOr(ContextHandoffId),
   restartContinuationOfRunId: Schema.optional(RunId),
+  usageLimitContinuationOfRunId: Schema.optional(RunId),
   delegatedTaskId: Schema.optional(NodeId),
   workStartedAt: Schema.optional(Schema.DateTimeUtc),
   restartCancelledBackgroundWork: Schema.optional(