feat(caps): user-registered models trust upstream vision instead of stripping media

open-sse/providers/capabilities.js

@@ -317,6 +317,11 @@ export const PATTERN_CAPABILITIES = [
   { pattern: "*ling-*",         caps: { reasoning: true, contextWindow: 128000 } },
 ];
 
+// OpenRouter-style gateways validate modalities upstream — a text-only model
+// sent an image gets a clear upstream error instead of silent corruption. So for
+// unknown models on these providers, trust vision instead of stripping images.
+const TRUST_UPSTREAM_VISION = new Set(["openrouter"]);
+
 /**
  * Resolve capabilities for a model using the 4-step fallback chain,
  * merged over DEFAULT_CAPABILITIES so the result is always complete.
@@ -349,6 +354,9 @@ export function getCapabilitiesForModel(provider, model) {
     }
   }
 
-  // 4. Floor
+  // 4. Floor (upstream-validated gateways keep vision on for unknown models)
+  if (provider && TRUST_UPSTREAM_VISION.has(provider)) {
+    return { ...DEFAULT_CAPABILITIES, vision: true };
+  }
   return { ...DEFAULT_CAPABILITIES };
 }