diff --git a/apps/server/src/orchestration-v2/ProviderTurnStartService.ts b/apps/server/src/orchestration-v2/ProviderTurnStartService.ts
index a506cc6e122c6..cf52173f45c6a 100644
--- a/apps/server/src/orchestration-v2/ProviderTurnStartService.ts
+++ b/apps/server/src/orchestration-v2/ProviderTurnStartService.ts
@@ -45,6 +45,7 @@ import {
   handoffTokenCapConfig,
   handoffBudget,
   attachmentTokenAllowance,
+  estimateTextTokens,
   contextUsageForHandoff,
   historicalMessage,
   latestNativeContextUsage,
@@ -998,7 +999,7 @@ export const layer: Layer.Layer<
                     : 0;
                 return (
                   sum +
-                  (historical === null ? 0 : Buffer.byteLength(historical.text)) +
+                  (historical === null ? 0 : estimateTextTokens(historical.text)) +
                   nativeAttachments
                 );
               }, 0)
diff --git a/apps/server/src/orchestration-v2/testkit/ProviderSwitch.integration.test.ts b/apps/server/src/orchestration-v2/testkit/ProviderSwitch.integration.test.ts
index f252380bf7f70..306935cdd0975 100644
--- a/apps/server/src/orchestration-v2/testkit/ProviderSwitch.integration.test.ts
+++ b/apps/server/src/orchestration-v2/testkit/ProviderSwitch.integration.test.ts
@@ -510,7 +510,7 @@ describe("orchestration v2 provider switching", () => {
             capabilities: ClaudeProviderCapabilitiesV2,
             modelSelection: CLAUDE_MODEL_SELECTION,
             responseByRunOrdinal: scenario.includes("target-estimate")
-              ? { 2: "x".repeat(10_000) }
+              ? { 2: "x".repeat(64_000) }
               : {},
             capturedTurns,
             ...(scenario.includes("retry") || scenario.includes("unsent") ? { failStartOnce } : {}),
@@ -1230,6 +1230,106 @@ describe("orchestration v2 provider switching", () => {
       ),
   );
 
+  it.live("keeps a native session whose stored prose is far below an unreported window", () =>
+    Effect.scoped(
+      Effect.gen(function* () {
+        const cwd = yield* checkpointWorkspace("handoff-estimate-headroom");
+        const capturedTurns = yield* Ref.make<ReadonlyArray<CapturedTurn>>([]);
+        const generation = yield* Ref.make(0);
+        // About 112 KB of ordinary English is about 25k tokens: a quarter of a 128k window.
+        const prose = "The quick brown fox jumps over the lazy dog. ".repeat(2_500);
+        const layerRegistry = ProviderAdapterRegistry.layerFromAdapters([
+          makeTestAdapter({
+            instanceId: CODEX_MODEL_SELECTION.instanceId,
+            driver: CODEX_DRIVER,
+            capabilities: CodexProviderCapabilitiesV2,
+            modelSelection: CODEX_MODEL_SELECTION,
+            responseByRunOrdinal: { 2: "Source work" },
+            capturedTurns,
+          }),
+          // No window and no usage, like an ACP registry agent that sends no usage_update.
+          makeTestAdapter({
+            instanceId: CLAUDE_MODEL_SELECTION.instanceId,
+            driver: CLAUDE_DRIVER,
+            capabilities: ClaudeProviderCapabilitiesV2,
+            modelSelection: CLAUDE_MODEL_SELECTION,
+            responseByRunOrdinal: { 1: prose },
+            capturedTurns,
+            nativeThreadGeneration: generation,
+          }),
+        ]);
+        yield* Effect.gen(function* () {
+          const orchestrator = yield* Orchestrator.OrchestratorV2;
+          const worker = yield* EffectWorker.OrchestrationEffectWorkerV2;
+          const dispatch = (ordinal: number, text: string, selection: ModelSelection) =>
+            orchestrator.dispatch({
+              type: "message.dispatch",
+              commandId: CommandId.make(`headroom:${ordinal}`),
+              threadId,
+              messageId: MessageId.make(`headroom:${ordinal}`),
+              createdBy: "user",
+              creationSource: "web",
+              text,
+              attachments: [],
+              modelSelection: selection,
+              dispatchMode: { type: "start_immediately" },
+            });
+          const wait = (ordinal: number) =>
+            orchestrator.streamStoredEvents.pipe(
+              Stream.filter(
+                ({ event }) =>
+                  event.type === "run.updated" &&
+                  event.payload.ordinal === ordinal &&
+                  (event.payload.status === "completed" || event.payload.status === "failed"),
+              ),
+              Stream.runHead,
+              Effect.andThen(worker.drain()),
+            );
+          yield* orchestrator.dispatch({
+            type: "thread.create",
+            commandId: CommandId.make("headroom:create"),
+            threadId,
+            projectId,
+            createdBy: "user",
+            creationSource: "web",
+            title: "Handoff estimate headroom",
+            modelSelection: CLAUDE_MODEL_SELECTION,
+            runtimeMode: "full-access",
+            interactionMode: "default",
+            branch: null,
+            worktreePath: null,
+          });
+          yield* dispatch(1, "Write the long report", CLAUDE_MODEL_SELECTION);
+          yield* wait(1);
+          yield* dispatch(2, "Continue elsewhere", CODEX_MODEL_SELECTION);
+          yield* wait(2);
+          yield* dispatch(3, "Back to the first provider", CLAUDE_MODEL_SELECTION);
+          yield* wait(3);
+          assert.equal(
+            (yield* orchestrator.getThreadProjection(threadId)).runs.at(-1)?.status,
+            "completed",
+          );
+          // A second generation means T3 discarded the native session as full.
+          assert.equal(yield* Ref.get(generation), 1);
+        }).pipe(
+          Effect.provide(
+            ProviderReplayHarness.layerWithRegistry(
+              {
+                name: "handoff-estimate-headroom",
+                runtimePolicyOverride: {
+                  cwd,
+                  approvalPolicy: "never",
+                  sandboxPolicy: { type: "readOnly" },
+                },
+              },
+              layerRegistry,
+            ),
+          ),
+        );
+      }),
+    ),
+  );
+
   it.live.each(["native", "legacy"] as const)(
     "starts a replacement native thread as new after the resume fallback with %s attempts",
     (attempts) =>
diff --git a/packages/provider-core/src/server/handoffBudget.ts b/packages/provider-core/src/server/handoffBudget.ts
index 5f9f2f3ae33f6..735647693d06d 100644
--- a/packages/provider-core/src/server/handoffBudget.ts
+++ b/packages/provider-core/src/server/handoffBudget.ts
@@ -95,7 +95,12 @@ export function attachmentTokenAllowance(attachments: ReadonlyArray<ChatAttachme
   );
 }
 
-// One UTF-8 byte per token is a pessimistic estimate, not a guarantee for custom models.
+// Stored history is mostly prose and tool output, about four UTF-8 bytes per token.
+export function estimateTextTokens(text: string): number {
+  return Math.ceil(Buffer.byteLength(text) / 4);
+}
+
+// One UTF-8 byte per token keeps the new turn and imported history pessimistic.
 export function handoffBudget(input: {
   readonly tokenCap: number;
   readonly userText: string;