fix(vision): providers that reject images in tool results no longer get the multimodal envelope — meta-ai/Muse Spark 400 (#101668, salvages #89987)

plugins/model-providers/meta-ai/__init__.py

@@ -121,6 +121,9 @@ def _base_url() -> str:
     api_mode="codex_responses",
     # Muse Spark is natively multimodal (image/video/pdf/audio in, text out).
     supports_vision=True,
+    # ...but only on user turns: an image envelope inside a role:tool message
+    # 400s "messages[N].content did not match any supported type" (#101668).
+    supports_vision_tool_messages=False,
     # Cheap contributor tier is a good default for auxiliary tasks
     # (compaction, title generation, vision) when this is the main provider.
     default_aux_model="muse-spark-1.2-contributor",

tests/providers/test_meta_ai_profile.py

@@ -28,6 +28,8 @@ def test_profile_registered(self):
         assert p.api_mode == "codex_responses"
         assert "MODEL_API_KEY" in p.env_vars
         assert p.supports_vision is True
+        # Images are accepted on user turns only; tool-result envelopes 400 (#101668).
+        assert p.supports_vision_tool_messages is False
         assert p.default_aux_model == "muse-spark-1.2-contributor"
         assert p.default_max_tokens == 16384
         assert p.fallback_models == ("muse-spark-1.2",)

tests/tools/test_vision_native_fast_path.py

@@ -90,6 +90,30 @@ def test_empty_provider_no(self):
         assert _supports_media_in_tool_results("", "anything") is False
         assert _supports_media_in_tool_results(None, "anything") is False  # type: ignore[arg-type]
 
+    def test_profile_tool_message_veto_overrides_supports_vision(self):
+        """supports_vision_tool_messages=False is a hard veto even when the
+        profile declares supports_vision=True (xiaomi/MiMo 400s on list-type
+        tool-result content, #89981)."""
+        assert _supports_media_in_tool_results("xiaomi", "mimo-v2.5") is False
+
+    def test_profile_veto_applies_even_when_vision_capable_lookup_agrees(self):
+        """A capability source marking the model vision-capable must not
+        re-open the native fast path for a provider that rejects it."""
+        from tools.vision_tools import _should_use_native_vision_fast_path
+        from agent.auxiliary_client import set_runtime_main, clear_runtime_main
+        from agent import image_routing
+
+        set_runtime_main("xiaomi", "mimo-v2.5")
+        try:
+            with patch.object(
+                image_routing, "decide_image_input_mode", return_value="native"
+            ), patch.object(
+                image_routing, "_lookup_supports_vision", return_value=True
+            ):
+                assert _should_use_native_vision_fast_path() is False
+        finally:
+            clear_runtime_main()
+
 
 # ─── _build_native_vision_tool_result ────────────────────────────────────────
 

tools/vision_tools.py

@@ -1077,6 +1077,22 @@ def _dims_ok(w: int, h: int) -> bool:
 # ---------------------------------------------------------------------------
 
 
+def _profile_rejects_tool_media(provider: str) -> bool:
+    """Hard veto: the provider's ``ProviderProfile`` declares
+    ``supports_vision_tool_messages=False`` — images are accepted in user
+    messages but list-type tool-result content is rejected with 400
+    (xiaomi/MiMo "text is not set"). ``supports_vision`` alone must not
+    override this, or the multimodal tool-result envelope 400s every turn
+    and the image never enters context (#89981).
+    """
+    try:
+        from providers import get_provider_profile
+        profile = get_provider_profile(str(provider or "").strip().lower())
+        return profile is not None and profile.supports_vision_tool_messages is False
+    except Exception:
+        return False
+
+
 def _supports_media_in_tool_results(provider: str, model: str) -> bool:
     """Whether the given provider+model combination accepts image content
     inside a tool-result message.
@@ -1100,7 +1116,7 @@ def _supports_media_in_tool_results(provider: str, model: str) -> bool:
     if not isinstance(provider, str):
         return False
     p = provider.strip().lower()
-    if not p:
+    if not p or _profile_rejects_tool_media(p):
         return False
 
     # Aggregators that route to multiple vendors — assume support since
@@ -1170,6 +1186,11 @@ def _should_use_native_vision_fast_path() -> bool:
         cfg = load_config()
         if decide_image_input_mode(provider, model, cfg) != "native":
             return False
+        # The profile veto applies ahead of the capability lookup too: a
+        # model marked vision-capable by models.dev / custom_providers must
+        # not re-open the multimodal-envelope route the profile rejects.
+        if _profile_rejects_tool_media(provider):
+            return False
         return (
             _supports_media_in_tool_results(provider, model)
             or _lookup_supports_vision(provider, model, cfg) is True