diff --git a/src/server/request-log.ts b/src/server/request-log.ts
index f4c4df6626..430fc01b8f 100644
--- a/src/server/request-log.ts
+++ b/src/server/request-log.ts
@@ -1017,7 +1017,11 @@ function finalizedUsage(
       // ESTIMATE via capEstimateAtContextWindow, and Math.max preserves a real
       // provider-reported count, so it needs no further reduction.
       ? (finalUsage.estimated && contextWindow !== undefined && finalUsage.inputTokens > contextWindow
-          ? { ...finalUsage, inputTokens: contextWindow }
+          ? {
+              ...finalUsage,
+              inputTokens: contextWindow,
+              totalTokens: contextWindow + finalUsage.outputTokens,
+            }
           : finalUsage)
       : usageFallback;
   const totalTokens = usageTotalTokens(loggedUsage);
diff --git a/tests/request-log-estimate-cap.test.ts b/tests/request-log-estimate-cap.test.ts
index b51f7aeb6b..3554dd62b5 100644
--- a/tests/request-log-estimate-cap.test.ts
+++ b/tests/request-log-estimate-cap.test.ts
@@ -31,6 +31,23 @@ describe("request log token-estimate context-window cap (codex-router PR #140)",
     expect(attempt.usage?.inputTokens).toBe(100_000);
   });
 
+  test("capping adapter-estimated input recomputes an explicit total", () => {
+    const attempt = beginRequestAttempt(1, "kiro", "deepseek-3.2", "kiro");
+    finishRequestAttempt(attempt, 200, 10, {
+      inputTokens: 200_000,
+      outputTokens: 4,
+      totalTokens: 200_004,
+      estimated: true,
+    });
+    expect(attempt.usage).toEqual({
+      inputTokens: DEEPSEEK_WINDOW,
+      outputTokens: 4,
+      totalTokens: DEEPSEEK_WINDOW + 4,
+      estimated: true,
+    });
+    expect(attempt.totalTokens).toBe(DEEPSEEK_WINDOW + 4);
+  });
+
   test("a positive provider-reported input count is never reduced by the cap", () => {
     const attempt = beginRequestAttempt(1, "kiro", "deepseek-3.2", "kiro");
     noteAttemptSend(attempt, 200_000); // estimate would exceed the window