fix: Claude passthrough (#1359), kimi-k2 reasoning (#1360), thinking leak (#1361), Ollama redirect (#1381)

- Eliminate lossy Claude→OpenAI→Claude round-trip for Claude-format providers
- Expand isReasoner to include kimi-k2 and opencode-go provider models
- Block thinking param leak to non-Claude antigravity models (gemini, gpt-oss)
- Allow redirects for Ollama Cloud /v1/models endpoint (301)

open-sse/handlers/chatCore.ts

@@ -1164,52 +1164,16 @@ export async function handleChatCore({
         preserveCacheControl,
       });
       log?.debug?.("FORMAT", "claude-code-compatible bridge enabled");
-    } else if (isClaudePassthrough && preserveCacheControl) {
-      // Pure passthrough: when preserveCacheControl is true, forward the body
-      // as-is without prior normalization. The OpenAI round-trip would strip
-      // cache_control markers; even prepareClaudeRequest can alter structure.
-      // Claude Code sends well-formed Messages API payloads — trust it.
+    } else if (isClaudePassthrough) {
+      // Pure passthrough: forward the body as-is without OpenAI round-trip.
+      // The Claude→OpenAI→Claude double translation was lossy and corrupted
+      // payloads at high context (150+ msgs, 100+ tools). Fix: #1359.
+      // Claude Code sends well-formed Messages API payloads — trust them
+      // regardless of combo strategy or cache_control settings.
       translatedBody = { ...body };
       translatedBody._disableToolPrefix = true;
 
-      log?.debug?.("FORMAT", "claude passthrough with cache_control preservation");
-    } else if (isClaudePassthrough) {
-      // Claude OAuth expects the same Claude Code prompt + structural normalization
-      // as the OpenAI-compatible chat path. Round-trip through OpenAI to reuse the
-      // working Claude translator instead of forwarding raw Messages payloads.
-      const normalizeToolCallId = getModelNormalizeToolCallId(
-        provider || "",
-        model || "",
-        sourceFormat
-      );
-      const preserveDeveloperRole = getModelPreserveOpenAIDeveloperRole(
-        provider || "",
-        model || "",
-        sourceFormat
-      );
-      translatedBody = translateRequest(
-        FORMATS.CLAUDE,
-        FORMATS.OPENAI,
-        model,
-        { ...body },
-        stream,
-        credentials,
-        provider,
-        reqLogger,
-        { normalizeToolCallId, preserveDeveloperRole, preserveCacheControl }
-      );
-      translatedBody = translateRequest(
-        FORMATS.OPENAI,
-        FORMATS.CLAUDE,
-        model,
-        { ...translatedBody, _disableToolPrefix: true },
-        stream,
-        credentials,
-        provider,
-        reqLogger,
-        { normalizeToolCallId, preserveDeveloperRole, preserveCacheControl }
-      );
-      log?.debug?.("FORMAT", "claude->openai->claude normalized passthrough");
+      log?.debug?.("FORMAT", `claude passthrough (preserveCache=${preserveCacheControl})`);
     } else {
       translatedBody = { ...body };
 

open-sse/translator/index.ts

@@ -207,7 +207,9 @@ export function translateRequest(
   // Inject reasoning_content = "" for DeepSeek/Reasoning models assistant messages with tool_calls
   // if omitted by the client, to avoid upstream 400 errors (e.g. "Messages with role 'assistant' that contain tool_calls must also include reasoning_content")
   const isReasoner =
-    provider === "deepseek" || (typeof model === "string" && /r1|reason/i.test(model));
+    provider === "deepseek" ||
+    provider === "opencode-go" ||
+    (typeof model === "string" && /r1|reason|kimi-k2/i.test(model));
   if (isReasoner && result.messages && Array.isArray(result.messages)) {
     for (const msg of result.messages) {
       if (

src/app/api/providers/[id]/models/route.ts

@@ -892,6 +892,8 @@ export async function GET(
         ...SAFE_OUTBOUND_FETCH_PRESETS.modelsPagination,
         guard: getProviderOutboundGuard(),
         proxyConfig: proxy,
+        // Ollama Cloud /v1/models returns 301 redirects (#1381)
+        ...(provider === "ollama-cloud" ? { allowRedirect: true } : {}),
         ...fetchOptions,
       });
 

src/lib/modelCapabilities.ts

@@ -11,6 +11,11 @@ const REASONING_UNSUPPORTED_PATTERNS = [
   "antigravity/claude-sonnet-4-6",
   "antigravity/claude-sonnet-4-5",
   "antigravity/claude-sonnet-4",
+  // Non-Claude antigravity models don't support thinking params (#1361)
+  "antigravity/gemini-",
+  "antigravity/gpt-oss-",
+  "antigravity/gemini-3",
+  "antigravity/tab_",
 ];
 
 type CapabilityInput =