fix(agent): guard multimodal tool content behind provider profile flag

plugins/model-providers/anthropic/__init__.py

@@ -47,6 +47,7 @@ def fetch_models(
     base_url="https://api.anthropic.com",
     auth_type="api_key",
     default_aux_model="claude-haiku-4-5-20251001",
+    supports_multimodal_tool_content=True,
 )
 
 register_provider(anthropic)

plugins/model-providers/openai-codex/__init__.py

@@ -10,6 +10,7 @@
     env_vars=(),  # OAuth external — no API key
     base_url="https://chatgpt.com/backend-api/codex",
     auth_type="oauth_external",
+    supports_multimodal_tool_content=True,
 )
 
 register_provider(openai_codex)

plugins/model-providers/openrouter/__init__.py

@@ -103,6 +103,7 @@ def build_api_kwargs_extras(
     signup_url="https://openrouter.ai/keys",
     base_url="https://openrouter.ai/api/v1",
     models_url="https://openrouter.ai/api/v1/models",
+    supports_multimodal_tool_content=True,
     fallback_models=(
         "anthropic/claude-sonnet-4.6",
         "openai/gpt-5.4",

plugins/model-providers/xai/__init__.py

@@ -10,6 +10,7 @@
     env_vars=("XAI_API_KEY",),
     base_url="https://api.x.ai/v1",
     auth_type="api_key",
+    supports_multimodal_tool_content=True,
 )
 
 register_provider(xai)

providers/base.py

@@ -68,6 +68,13 @@ class ProviderProfile:
     # ── Client-level quirks (set once at client construction) ─
     default_headers: dict[str, str] = field(default_factory=dict)
 
+    # Whether the provider accepts multimodal content (list of text + image_url
+    # parts) inside ``role: "tool"`` messages.  Most OpenAI-compatible providers
+    # only accept plain-string content for tool messages even when they support
+    # images in user messages.  Default False — only providers known to support
+    # it should set True (e.g. Anthropic, OpenAI GPT-4o).
+    supports_multimodal_tool_content: bool = False
+
     # ── Request-level quirks ─────────────────────────────────
     # Temperature: None = use caller's default, OMIT_TEMPERATURE = don't send
     fixed_temperature: Any = None

run_agent.py

@@ -3182,6 +3182,24 @@ def _model_supports_vision(self) -> bool:
         except Exception:
             return False
 
+    def _provider_supports_multimodal_tool_content(self) -> bool:
+        """Return True if the provider accepts multimodal content in tool messages.
+
+        Many providers support images in user messages but require string-only
+        content in ``role: "tool"`` messages.  This checks the provider profile
+        flag; defaults to True for known multimodal-friendly providers
+        (OpenAI, Anthropic) and False otherwise.  See #27344.
+        """
+        try:
+            from providers import get_provider_profile
+            profile = get_provider_profile(self.provider)
+            if profile is not None:
+                return bool(profile.supports_multimodal_tool_content)
+        except Exception:
+            pass
+        # Conservative default: assume no multimodal tool support unless declared.
+        return False
+
     def _preprocess_anthropic_content(self, content: Any, role: str) -> Any:
         if not self._content_has_image_parts(content):
             return content
@@ -3321,7 +3339,12 @@ def _tool_result_content_for_active_model(self, tool_name: str, result: Any) ->
             return content
 
         if self._model_supports_vision():
-            return content
+            # Even if the model supports images in user messages, the provider
+            # may not accept multimodal content (list of text+image parts) inside
+            # tool-role messages.  Check the provider profile flag; fall back to
+            # text summary when not supported.  (#27344)
+            if self._provider_supports_multimodal_tool_content():
+                return content
 
         summary = _multimodal_text_summary(result)
         if tool_name == "computer_use":

tests/tools/test_computer_use.py

@@ -618,6 +618,7 @@ def test_computer_use_image_result_preserved_for_vision_model(self):
         from run_agent import AIAgent
 
         agent = object.__new__(AIAgent)
+        agent.provider = "openrouter"
         result = {
             "_multimodal": True,
             "content": [
@@ -626,12 +627,43 @@ def test_computer_use_image_result_preserved_for_vision_model(self):
             ],
         }
 
-        with patch.object(agent, "_model_supports_vision", return_value=True):
+        with patch.object(agent, "_model_supports_vision", return_value=True), \
+             patch.object(agent, "_provider_supports_multimodal_tool_content", return_value=True):
             content = agent._tool_result_content_for_active_model("computer_use", result)
 
         assert content is result["content"]
         assert any(part.get("type") == "image_url" for part in content)
 
+    def test_computer_use_image_result_downgraded_for_vision_model_without_tool_support(self):
+        """Vision-capable model whose provider doesn't support multimodal tool content.
+
+        Regression test for #27344: MiMo supports images in user messages but
+        requires string content in tool messages.  The multimodal image should
+        be downgraded to text_summary, not passed through as a list.
+        """
+        from run_agent import AIAgent
+
+        agent = object.__new__(AIAgent)
+        agent.provider = "xiaomi"
+        agent.model = "mimo-v2.5"
+        result = {
+            "_multimodal": True,
+            "content": [
+                {"type": "text", "text": "screen captured"},
+                {"type": "image_url", "image_url": {"url": "data:image/png;base64,x"}},
+            ],
+            "text_summary": "screen captured summary",
+        }
+
+        with patch.object(agent, "_model_supports_vision", return_value=True), \
+             patch.object(agent, "_provider_supports_multimodal_tool_content", return_value=False):
+            content = agent._tool_result_content_for_active_model("computer_use", result)
+
+        # Should NOT contain image_url — must be downgraded to error + text_summary
+        parsed = json.loads(content)
+        assert "computer_use returned screenshot/image content" in parsed["error"]
+        assert "image_url" not in str(content)
+
     def test_other_multimodal_tool_uses_text_summary_for_text_only_model(self):
         from run_agent import AIAgent