fix(translator): replay cached Kimi reasoning before fallback

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

@@ -1 +1 @@
-- 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)
+- fix(translator): preserve authentic K3 Responses reasoning by model across providers, keep it on the matching assistant turn, and make Kimi Coding prefer client reasoning then cached replay before its empty-marker fallback (#9496)

open-sse/translator/index.ts

@@ -501,7 +501,7 @@ export function translateRequest(
   // isReasoner / normalizedProvider / normalizedModel / resolvedCapabilities were
   // resolved up-front (before the OpenAI-format filter) so the #4849 reasoning strip
   // could honor reasoning-replay providers.
-  if (isReasoner && !isKimiCoding && result.messages && Array.isArray(result.messages)) {
+  if (isReasoner && result.messages && Array.isArray(result.messages)) {
     const canReplayReasoningOnly = isReasoningOnlyReplayTarget(normalizedProvider, normalizedModel);
 
     for (const [messageIndex, msg] of result.messages.entries()) {
@@ -550,29 +550,51 @@ export function translateRequest(
         // Has tool_use blocks but no thinking block yet.
         // Reasoning models (Kimi K2, etc.) require a thinking block before tool_use
         // on multi-turn or they regenerate the same tool call infinitely.
-        const hasThinkingBlock = msg.content.some(
+        const thinkingBlock = msg.content.find(
           (b) => b?.type === "thinking" || b?.type === "redacted_thinking"
         );
-        if (hasThinkingBlock) continue;
+        const hasNonEmptyClientThinking =
+          thinkingBlock?.type === "thinking" &&
+          typeof thinkingBlock.thinking === "string" &&
+          thinkingBlock.thinking.trim().length > 0;
+        if (thinkingBlock && (!isKimiCoding || hasNonEmptyClientThinking)) continue;
 
         const toolUseBlocks = msg.content.filter((b) => b?.type === "tool_use");
         const firstToolUseId = toolUseBlocks[0]?.id;
         const firstToolUseIdx = msg.content.findIndex((b) => b?.type === "tool_use");
 
-        // Try reasoning cache first
+        // Client reasoning wins above. Otherwise try authentic replay before
+        // retaining Kimi Code's empty protocol marker as the final fallback.
         if (firstToolUseId) {
           const cached = lookupReasoning(firstToolUseId);
           if (cached) {
-            msg.content.splice(firstToolUseIdx, 0, {
-              type: "thinking",
-              thinking: cached,
-            });
+            if (thinkingBlock) {
+              thinkingBlock.type = "thinking";
+              thinkingBlock.thinking = cached;
+              delete thinkingBlock.data;
+              delete thinkingBlock.signature;
+            } else {
+              msg.content.splice(firstToolUseIdx, 0, {
+                type: "thinking",
+                thinking: cached,
+              });
+            }
             recordReplay();
             continue;
           }
         }
+        if (isKimiCoding) {
+          if (thinkingBlock) {
+            thinkingBlock.type = "thinking";
+            thinkingBlock.thinking = "";
+            delete thinkingBlock.data;
+            delete thinkingBlock.signature;
+          } else {
+            msg.content.splice(firstToolUseIdx, 0, { type: "thinking", thinking: "" });
+          }
+          continue;
+        }
         if (requiresAuthenticReasoning) continue;
-        // Fallback: inject placeholder (must be non-empty for kimi-coding)
         msg.content.splice(firstToolUseIdx, 0, {
           type: "thinking",
           thinking: NON_ANTHROPIC_THINKING_PLACEHOLDER,

tests/unit/translator-helper-branches.test.ts

@@ -698,12 +698,12 @@ test("translateRequest does not replay reasoning-only messages for non-DeepSeek
   clearReasoningCacheAll();
 });
 
-test("translateRequest uses Kimi Coding's empty thinking marker instead of cached replay", () => {
+test("translateRequest replays cached reasoning before Kimi Coding's empty fallback", () => {
   clearReasoningCacheAll();
   cacheReasoningByKey(
     "toolu_kimi_claude",
-    "kimi-coding",
-    "kimi-for-coding",
+    "kimi-coding-apikey",
+    "k3-256k",
     "cached thinking for Kimi tool call"
   );
 
@@ -712,7 +712,7 @@ test("translateRequest uses Kimi Coding's empty thinking marker instead of cache
   const result = translateRequest(
     FORMATS.OPENAI,
     FORMATS.CLAUDE,
-    "kimi-for-coding",
+    "k3-256k",
     {
       reasoning_effort: "high",
       messages: [
@@ -733,24 +733,23 @@ test("translateRequest uses Kimi Coding's empty thinking marker instead of cache
     },
     false,
     null,
-    "kimi-coding"
+    "kimi-coding-apikey"
   );
 
   const assistantMsg = result.messages.find((m) => m.role === "assistant");
   assert.ok(assistantMsg, "assistant message should exist");
   assert.ok(Array.isArray(assistantMsg.content), "content should be array");
 
-  // Kimi Code CLI 0.26 sends an explicit empty thinking marker before tool_use.
   const thinkingBlock = assistantMsg.content.find((b) => b?.type === "thinking");
   assert.ok(thinkingBlock, "thinking block should be injected");
-  assert.equal(thinkingBlock.thinking, "");
+  assert.equal(thinkingBlock.thinking, "cached thinking for Kimi tool call");
 
   // Thinking block should appear before tool_use
   const thinkingIdx = assistantMsg.content.indexOf(thinkingBlock);
   const toolUseIdx = assistantMsg.content.findIndex((b) => b?.type === "tool_use");
   assert.ok(thinkingIdx < toolUseIdx, "thinking block should be before tool_use");
 
-  assert.equal(getReasoningCacheServiceStats().replays, 0);
+  assert.equal(getReasoningCacheServiceStats().replays, 1);
   clearReasoningCacheAll();
 });
 
@@ -760,7 +759,7 @@ test("translateRequest uses an empty Kimi Coding thinking marker on cache miss",
   const result = translateRequest(
     FORMATS.OPENAI,
     FORMATS.CLAUDE,
-    "kimi-for-coding",
+    "k3-256k",
     {
       reasoning_effort: "high",
       messages: [
@@ -774,7 +773,7 @@ test("translateRequest uses an empty Kimi Coding thinking marker on cache miss",
     },
     false,
     null,
-    "kimi-coding"
+    "kimi-coding-apikey"
   );
 
   const assistantMsg = result.messages.find((m) => m.role === "assistant");
@@ -790,11 +789,17 @@ test("translateRequest uses an empty Kimi Coding thinking marker on cache miss",
 
 test("translateRequest does NOT inject duplicate thinking for Claude-format messages with existing thinking block", () => {
   clearReasoningCacheAll();
+  cacheReasoningByKey(
+    "toolu_existing",
+    "kimi-coding-apikey",
+    "k3-256k",
+    "cached thinking must not replace client thinking"
+  );
 
   const result = translateRequest(
     FORMATS.OPENAI,
     FORMATS.CLAUDE,
-    "kimi-for-coding",
+    "k3-256k",
     {
       messages: [
         { role: "user", content: "hi" },
@@ -810,7 +815,7 @@ test("translateRequest does NOT inject duplicate thinking for Claude-format mess
     },
     false,
     null,
-    "kimi-coding"
+    "kimi-coding-apikey"
   );
 
   const assistantMsg = result.messages.find((m) => m.role === "assistant");
@@ -823,6 +828,7 @@ test("translateRequest does NOT inject duplicate thinking for Claude-format mess
     "I already have this",
     "original thinking should be preserved"
   );
+  assert.equal(getReasoningCacheServiceStats().replays, 0);
 
   clearReasoningCacheAll();
 });