fix(translator): make K3 reasoning preservation model-driven

changelog.d/fixes/9496-kimi-k3-responses-replay.md

@@ -1 +1 @@
-- fix(translator): preserve Kimi K3 Responses reasoning through Kimi Coding Claude-format and native Moonshot requests, keeping it on the matching assistant tool call or completed turn instead of dropping it or carrying it across a user boundary (#9496)
+- fix(translator): preserve authentic K3 Responses reasoning by model across providers, projecting it onto the matching assistant tool call or completed turn instead of dropping it or carrying it across a user boundary (#9496)

open-sse/translator/helpers/responsesApiHelper.ts

@@ -2,7 +2,7 @@
  * Convert OpenAI Responses API format to standard chat completions format.
  * Delegates to the canonical translator to avoid logic duplication.
  */
-import { shouldPreserveResponsesReasoningContent } from "../../utils/reasoningContentInjector.ts";
+import { requiresAuthenticReasoningContent } from "../../utils/reasoningContentInjector.ts";
 import { openaiResponsesToOpenAIRequest } from "../request/openai-responses.ts";
 import { toRecord } from "../request/openai-responses/helpers.ts";
 
@@ -23,7 +23,7 @@ export function convertResponsesApiFormat(
     credentials && typeof credentials === "object" && !Array.isArray(credentials)
       ? (credentials as Record<string, unknown>)
       : {};
-  const translationCredentials = shouldPreserveResponsesReasoningContent(provider, model)
+  const translationCredentials = requiresAuthenticReasoningContent(provider, model)
     ? { ...credentialRecord, _preserveReasoningContent: true }
     : credentials;
   const converted = openaiResponsesToOpenAIRequest(

open-sse/translator/index.ts

@@ -13,10 +13,7 @@ import {
   providerHonorsOpenAIFormatCacheControl,
   resolveConnectionCacheOverride,
 } from "../utils/cacheControlPolicy.ts";
-import {
-  requiresAuthenticReasoningContent,
-  shouldPreserveResponsesReasoningContent,
-} from "../utils/reasoningContentInjector.ts";
+import { requiresAuthenticReasoningContent } from "../utils/reasoningContentInjector.ts";
 import { isInternalReasoningPlaceholder } from "../utils/reasoningPlaceholder.ts";
 import {
   coerceToolSchemas,
@@ -245,8 +242,7 @@ export function translateRequest(
     normalizedModel
   );
   const preserveResponsesReasoning =
-    sourceFormat === FORMATS.OPENAI_RESPONSES &&
-    shouldPreserveResponsesReasoningContent(normalizedProvider, normalizedModel);
+    sourceFormat === FORMATS.OPENAI_RESPONSES && requiresAuthenticReasoning;
 
   // Phase 2: Apply thinking budget control before normalization
   result = applyThinkingBudget(result);

open-sse/utils/reasoningContentInjector.ts

@@ -30,35 +30,24 @@ const THINKING_MODEL_PATTERNS: RegExp[] = [
   /\bmimo\b/i, // xiaomi-tokenplan mimo family (e.g. xiaomi-tokenplan/mimo-v2.5-pro)
 ];
 
-const AUTHENTIC_REASONING_MODEL_PATTERN = /(?:^|\/)kimi-k(?:3|2\.7-code)(?:$|-)/i;
+const K3_AUTHENTIC_REASONING_PATTERN = /(?:^|\/)(?:kimi-)?k3(?:$|-)/i;
+const NATIVE_K27_AUTHENTIC_REASONING_PATTERN = /(?:^|\/)kimi-k2\.7-code(?:$|-)/i;
 
 /**
- * Native Moonshot K3/K2.7 replay must use the original reasoning content.
- * A fabricated placeholder changes preserved-thinking history and is not a
- * valid substitute when the client and reasoning cache both lack the field.
+ * K3 requires authentic reasoning regardless of which provider serves it.
+ * Native Moonshot K2.7 retains the same preserved-thinking contract. Empty
+ * protocol markers remain valid only after client content and replay miss.
  */
 export function requiresAuthenticReasoningContent(provider: unknown, model: unknown): boolean {
+  const normalizedModel = String(model ?? "").trim();
+  if (K3_AUTHENTIC_REASONING_PATTERN.test(normalizedModel)) return true;
+
   const normalizedProvider = String(provider ?? "")
     .trim()
     .toLowerCase();
-  const normalizedModel = String(model ?? "").trim();
   return (
     (normalizedProvider === "moonshot" || normalizedProvider === "kimi") &&
-    AUTHENTIC_REASONING_MODEL_PATTERN.test(normalizedModel)
-  );
-}
-
-export function shouldPreserveResponsesReasoningContent(
-  provider: unknown,
-  model: unknown
-): boolean {
-  const normalizedProvider = String(provider ?? "")
-    .trim()
-    .toLowerCase();
-  return (
-    normalizedProvider === "kimi-coding" ||
-    normalizedProvider === "kimi-coding-apikey" ||
-    requiresAuthenticReasoningContent(normalizedProvider, model)
+    NATIVE_K27_AUTHENTIC_REASONING_PATTERN.test(normalizedModel)
   );
 }
 

tests/unit/responses-translation-fixes.test.ts

@@ -57,9 +57,13 @@ test("production Responses conversion preserves Kimi K3 reasoning history", () =
   };
 
   for (const { provider, model } of [
+    { provider: "kimi-coding", model: "k3" },
     { provider: "kimi-coding-apikey", model: "k3-256k" },
     { provider: "moonshot", model: "kimi-k3" },
     { provider: "kimi", model: "kimi-k3" },
+    { provider: "some-other", model: "k3" },
+    { provider: "some-other", model: "k3-256k" },
+    { provider: "some-other", model: "kimi-k3" },
   ]) {
     const converted = convertResponsesApiFormat(body, {}, provider, model) as {
       messages: Array<Record<string, unknown>>;
@@ -67,10 +71,12 @@ test("production Responses conversion preserves Kimi K3 reasoning history", () =
     assert.equal(converted.messages[1].reasoning_content, "I should search first.");
   }
 
-  const generic = convertResponsesApiFormat(body, {}, "openai", "gpt-5") as {
-    messages: Array<Record<string, unknown>>;
-  };
-  assert.equal(Object.hasOwn(generic.messages[1], "reasoning_content"), false);
+  for (const provider of ["kimi-coding", "kimi-coding-apikey"]) {
+    const generic = convertResponsesApiFormat(body, {}, provider, "kimi-k2.6") as {
+      messages: Array<Record<string, unknown>>;
+    };
+    assert.equal(Object.hasOwn(generic.messages[1], "reasoning_content"), false, provider);
+  }
 });
 
 test("Responses→Chat: input_image converted to image_url with detail", () => {