[WRONG BRANCH] fix(kiro): avoid mixed-script estimator allocations

src/adapters/kiro.ts

@@ -1,5 +1,5 @@
 import { decodeEventStream } from "../lib/eventstream-decoder";
-import { estimateTokens } from "../lib/token-estimate";
+import { estimateTokens, estimateTokensFromCharacterCounts } from "../lib/token-estimate";
 import { debugProviderDiagnostic } from "../lib/debug";
 import { resolveKiroApiRegion, resolveKiroRequestProfile } from "../oauth/kiro";
 import { KIRO_MODEL_CONTEXT_WINDOWS, normalizeKiroModelId } from "../providers/kiro-models";
@@ -209,8 +209,9 @@ function estimateKiroWireTokens(text: string, modelId: string): number {
   if (!text) return 0;
   const cjk = kiroCjkCount(text);
   if (cjk === 0) return Math.ceil(estimateKiroTokens(text, modelId) * KIRO_LATIN_WIRE_EXPANSION);
-  const latinTokens = estimateKiroTokens("x".repeat(text.length - cjk), modelId);
-  const cjkTokens = estimateKiroTokens("\uac00".repeat(cjk), modelId);
+  const prefixedModelId = `kiro/${modelId}`;
+  const latinTokens = estimateTokensFromCharacterCounts(text.length - cjk, 0, prefixedModelId);
+  const cjkTokens = estimateTokensFromCharacterCounts(0, cjk, prefixedModelId);
   return Math.ceil(latinTokens * KIRO_LATIN_WIRE_EXPANSION + cjkTokens);
 }
 

src/lib/token-estimate.ts

@@ -142,12 +142,24 @@ export function estimateTokens(text: string, modelId?: string, contextWindow?: n
   if (!text) return 0;
   const len = text.length;
   if (len === 0) return 0;
-  const latinRatio = charsPerToken(modelId);
   const cjk = countCjk(text);
+  return estimateTokensFromCharacterCounts(len - cjk, cjk, modelId, contextWindow);
+}
+
+/** Estimate tokens from already-counted script buckets without materializing replacement text. */
+export function estimateTokensFromCharacterCounts(
+  latin: number,
+  cjk: number,
+  modelId?: string,
+  contextWindow?: number,
+): number {
+  const len = latin + cjk;
+  if (len === 0) return 0;
+  const latinRatio = charsPerToken(modelId);
   // Continuous in the CJK share: no threshold, so one added Korean character moves the estimate
   // by a fraction of a token instead of switching the whole blob to a different divisor.
   const estimate = cjk === 0
     ? Math.ceil(len / latinRatio)
-    : Math.ceil((len - cjk) / latinRatio + cjk / CJK_CHARS_PER_TOKEN);
+    : Math.ceil(latin / latinRatio + cjk / CJK_CHARS_PER_TOKEN);
   return capEstimateAtContextWindow(Math.max(1, estimate), contextWindow);
 }

tests/lib/token-estimate.test.ts

@@ -1,5 +1,10 @@
 import { describe, expect, test } from "bun:test";
-import { capEstimateAtContextWindow, charsPerToken, estimateTokens } from "../../src/lib/token-estimate";
+import {
+  capEstimateAtContextWindow,
+  charsPerToken,
+  estimateTokens,
+  estimateTokensFromCharacterCounts,
+} from "../../src/lib/token-estimate";
 
 describe("script-segmented ratio", () => {
   const korean = "한국어 텍스트는 토큰 밀도가 높아서 영어 기준 추정이 과소계산됩니다 ".repeat(10);
@@ -10,6 +15,13 @@ describe("script-segmented ratio", () => {
     return Math.ceil((s.length - cjk) / latin + cjk / 1.5);
   };
 
+  test("pre-counted script buckets preserve estimates without replacement strings", () => {
+    expect(estimateTokensFromCharacterCounts(10, 3, "kiro/kiro-auto"))
+      .toBe(estimateTokens("xxxxxxxxxx한한한", "kiro/kiro-auto"));
+    expect(estimateTokensFromCharacterCounts(250_000_000, 1, "kiro/kiro-auto"))
+      .toBe(Math.ceil(250_000_000 / 2.8 + 1 / 1.5));
+  });
+
   test("CJK characters are counted at their own denser ratio, not the model ratio", () => {
     expect(estimateTokens(korean, "gpt-5.6-sol")).toBe(expected(korean, 4));
     expect(estimateTokens(korean, "kiro/claude-opus-5")).toBe(expected(korean, 2.8));