MODEL_CAPS

open-sse/translator/index.js

@@ -51,11 +51,38 @@ function ensureInitialized() {
   require("./response/ollama-to-openai.js");
 }
 
+// Strip multimodal content blocks (image/audio/video) from messages if model doesn't support them
+function stripUnsupportedMultimodal(body, multimodal = {}) {
+  if (!body.messages || !Array.isArray(body.messages)) return;
+  for (const msg of body.messages) {
+    if (!Array.isArray(msg.content)) continue;
+    msg.content = msg.content.filter(part => {
+      if (part.type === "image_url" || part.type === "image") return multimodal.image === true;
+      if (part.type === "audio_url" || part.type === "input_audio") return multimodal.audio === true;
+      return true; // keep text, tool_use, tool_result, etc.
+    });
+    // If content array becomes empty after filtering, replace with empty string to avoid API errors
+    if (msg.content.length === 0) msg.content = "";
+  }
+}
+
 // Translate request: source -> openai -> target
-export function translateRequest(sourceFormat, targetFormat, model, body, stream = true, credentials = null, provider = null, reqLogger = null) {
+export function translateRequest(sourceFormat, targetFormat, model, body, stream = true, credentials = null, provider = null, reqLogger = null, caps = null) {
   ensureInitialized();
   let result = body;
 
+  // Apply model capability guards before translation
+  if (caps) {
+    // Strip multimodal content if model doesn't support it
+    stripUnsupportedMultimodal(result, caps.multimodal || {});
+
+    // Strip thinking config if model doesn't support thinking
+    if (!caps.thinking) {
+      delete result.thinking;
+      delete result.reasoning_effort;
+    }
+  }
+
   // Normalize thinking config: remove if lastMessage is not user
   normalizeThinkingConfig(result);
 

open-sse/translator/request/openai-to-gemini.js

@@ -428,11 +428,15 @@ function wrapInCloudCodeEnvelopeForClaude(model, claudeRequest, credentials = nu
   return envelope;
 }
 
+// Detect if model should use Claude backend in Antigravity
+// Claude models have specific ID patterns — more reliable than caps at routing level
+function isClaudeModel(model) {
+  return model.toLowerCase().includes("claude");
+}
+
 // OpenAI -> Antigravity (Sandbox Cloud Code with wrapper)
 export function openaiToAntigravityRequest(model, body, stream, credentials = null) {
-  const isClaude = model.toLowerCase().includes("claude");
-
-  if (isClaude) {
+  if (isClaudeModel(model)) {
     const claudeRequest = openaiToClaudeRequestForAntigravity(model, body, stream);
     return wrapInCloudCodeEnvelopeForClaude(model, claudeRequest, credentials);
   }

open-sse/translator/request/openai-to-kiro.js

@@ -20,8 +20,8 @@ function convertMessages(messages, tools, model) {
   let pendingImages = [];
   let currentRole = null;
 
-  // Only Claude models support images in Kiro
-  const supportsImages = model && model.toLowerCase().includes("claude");
+  // Image support is pre-filtered by caps in translateRequest before reaching here
+  const supportsImages = true;
 
   const flushPending = () => {
     if (currentRole === "user") {