fix(llm-validation): accept stepfun multimodal models and return 422 on capability mismatch

backend/src/services/__tests__/test_llm_prompt_templates.py

@@ -74,6 +74,7 @@ def test_render_prompt_replaces_known_placeholders():
 def test_is_multimodal_model_detects_known_vision_models():
     assert is_multimodal_model("gpt-4o") is True
     assert is_multimodal_model("claude-3-5-sonnet") is True
+    assert is_multimodal_model("stepfun/step-3.5-flash:free", "openrouter") is True
     assert is_multimodal_model("text-only-model") is False
 # [/DEF:test_is_multimodal_model_detects_known_vision_models:Function]
 

backend/src/services/llm_prompt_templates.py

@@ -9,7 +9,7 @@
 from __future__ import annotations
 
 from copy import deepcopy
-from typing import Dict, Any
+from typing import Dict, Any, Optional
 
 
 # [DEF:DEFAULT_LLM_PROMPTS:Constant]
@@ -131,10 +131,21 @@ def normalize_llm_settings(llm_settings: Any) -> Dict[str, Any]:
 # @PURPOSE: Heuristically determine whether model supports image input required for dashboard validation.
 # @PRE: model_name may be empty or mixed-case.
 # @POST: Returns True when model likely supports multimodal input.
-def is_multimodal_model(model_name: str) -> bool:
+def is_multimodal_model(model_name: str, provider_type: Optional[str] = None) -> bool:
     token = (model_name or "").strip().lower()
     if not token:
         return False
+    provider = (provider_type or "").strip().lower()
+    text_only_markers = (
+        "text-only",
+        "embedding",
+        "rerank",
+        "whisper",
+        "tts",
+        "transcribe",
+    )
+    if any(marker in token for marker in text_only_markers):
+        return False
     multimodal_markers = (
         "gpt-4o",
         "gpt-4.1",
@@ -143,8 +154,21 @@ def is_multimodal_model(model_name: str) -> bool:
         "gemini",
         "claude-3",
         "claude-sonnet-4",
+        "omni",
+        "multimodal",
+        "pixtral",
+        "llava",
+        "internvl",
+        "qwen-vl",
+        "qwen2-vl",
+        "stepfun/step-3.5",
     )
-    return any(marker in token for marker in multimodal_markers)
+    if any(marker in token for marker in multimodal_markers):
+        return True
+    # OpenRouter model ids are heterogeneous; keep permissive path for known StepFun family.
+    if provider == "openrouter" and token.startswith("stepfun/step-3.5"):
+        return True
+    return False
 # [/DEF:is_multimodal_model:Function]