fix(anthropic): strip top_p when temperature is set to avoid 400 (#3691)

Integrated into release/v3.8.23

open-sse/handlers/chatCore.ts

@@ -1986,12 +1986,13 @@ export async function handleChatCore({
   // Use credentials.connectionId as a fallback so that requests without an
   // explicit session-level connectionId still register in the pendingRequests map.
   const pendingConnId = connectionId || credentials?.connectionId || null;
-  const pendingRequestId = trackPendingRequest(model, provider, pendingConnId, true, {
-    clientEndpoint: clientRawRequest?.endpoint || "/v1/chat/completions",
-    clientRequest: clientRawRequest?.body ?? body,
-    providerRequest: initialProviderRequest,
-    stage: "registered",
-  }) || generateRequestId();
+  const pendingRequestId =
+    trackPendingRequest(model, provider, pendingConnId, true, {
+      clientEndpoint: clientRawRequest?.endpoint || "/v1/chat/completions",
+      clientRequest: clientRawRequest?.body ?? body,
+      providerRequest: initialProviderRequest,
+      stage: "registered",
+    }) || generateRequestId();
 
   // Initialize rate limit settings from persisted DB (once, lazy)
   await initializeRateLimits();
@@ -2769,7 +2770,10 @@ export async function handleChatCore({
           comboTargetLimits,
         });
         contextLimit = resolved.limit;
-        log?.info?.("CONTEXT", `Combo context limit: ${resolved.limit} (source=${resolved.source})`);
+        log?.info?.(
+          "CONTEXT",
+          `Combo context limit: ${resolved.limit} (source=${resolved.source})`
+        );
       } catch (err) {
         log?.warn?.("CONTEXT", "Failed to resolve combo limits for compression: " + err);
       }
@@ -3104,6 +3108,12 @@ export async function handleChatCore({
           translatedBody.messages,
           DEFAULT_THINKING_CLAUDE_SIGNATURE
         ) as typeof translatedBody.messages;
+
+        // Anthropic API rejects requests with both temperature and top_p.
+        // VS Code Claude extension and similar clients send both; strip top_p.
+        if (translatedBody.temperature !== undefined && translatedBody.top_p !== undefined) {
+          delete translatedBody.top_p;
+        }
       }
 
       // Fix #2468: always extract role:"system" → top-level system.
@@ -5442,17 +5452,14 @@ export async function handleChatCore({
   }
 
   const responseHeaders: Record<string, string> = {
-    ...buildStreamingResponseHeaders(
-      providerResponse.headers,
-      {
-        provider,
-        model,
-        cacheHit: false,
-        latencyMs: 0,
-        usage: null,
-        costUsd: 0,
-      }
-    ),
+    ...buildStreamingResponseHeaders(providerResponse.headers, {
+      provider,
+      model,
+      cacheHit: false,
+      latencyMs: 0,
+      usage: null,
+      costUsd: 0,
+    }),
     "x-omniroute-request-id": pendingRequestId,
   };
 
@@ -5560,7 +5567,9 @@ export async function handleChatCore({
       });
     } catch (e) {
       // Best-effort — don't break the stream completion path if this fails
-      try { console.warn("finalizeMostRecentPendingRequest failed:", e && (e.message || e)); } catch {}
+      try {
+        console.warn("finalizeMostRecentPendingRequest failed:", e && (e.message || e));
+      } catch {}
     }
 
     if (apiKeyInfo?.id && streamUsage) {

open-sse/translator/request/openai-to-claude.ts

@@ -212,7 +212,7 @@ export function openaiToClaudeRequest(model, body, stream) {
   if (body.temperature !== undefined) {
     result.temperature = body.temperature;
   }
-  if (body.top_p !== undefined) {
+  if (body.temperature === undefined && body.top_p !== undefined) {
     result.top_p = body.top_p;
   }
   if (body.stop !== undefined) {

tests/unit/translator-openai-to-claude.test.ts

@@ -84,7 +84,8 @@ test("OpenAI -> Claude maps system messages, parameters and assistant cache mark
   assert.equal(result.stream, true);
   assert.equal(result.max_tokens, 33);
   assert.equal(result.temperature, 0.25);
-  assert.equal(result.top_p, 0.8);
+  // top_p is stripped when temperature is also present (Anthropic rejects both).
+  assert.equal(result.top_p, undefined);
   assert.deepEqual(result.stop_sequences, ["DONE"]);
   assert.equal(result.system[0].text, "Rule A\nRule B\nRule C");
   assert.equal(result.messages[0].role, "user");
@@ -94,6 +95,21 @@ test("OpenAI -> Claude maps system messages, parameters and assistant cache mark
   assert.deepEqual(result.messages[1].content[0].cache_control, { type: "ephemeral" });
 });
 
+test("OpenAI -> Claude strips top_p when temperature is also present", () => {
+  const result = openaiToClaudeRequest(
+    "claude-4-sonnet",
+    {
+      messages: [{ role: "user", content: "Hello" }],
+      temperature: 0.25,
+      top_p: 0.8,
+    },
+    false
+  );
+
+  assert.equal(result.temperature, 0.25);
+  assert.equal(result.top_p, undefined);
+});
+
 test("OpenAI -> Claude converts multimodal content, tool declarations, tool calls and tool results", () => {
   const result = openaiToClaudeRequest(
     "claude-4-sonnet",