fix(vision): image routing honours ProviderProfile.supports_vision

`decide_image_input_mode` consulted config overrides, the managed runtime,
models.dev and Ollama, but never the registered profile's `supports_vision`
— the field the tool-result media path already trusts. A plugin that
declared vision was therefore native for tool results and text-only for
user-attached images. The declaration is now the last probe in
`_VISION_PROBES`: only an explicit True is a verdict, and per-model catalog
entries still win.

Part of #116408

agent/image_routing.py

@@ -310,12 +310,25 @@ def _probe_ollama(provider: str, model: str, cfg: Optional[Dict[str, Any]]) -> O
     return query_ollama_supports_vision(model, base_url, api_key=api_key)
 
 
+def _probe_provider_profile(provider: str, model: str, cfg: Optional[Dict[str, Any]]) -> Optional[bool]:
+    """Registered ``ProviderProfile.supports_vision`` — the same declaration the tool-result media
+    path (``tools.vision_tools._supports_media_in_tool_results``) already trusts, so a plugin is not
+    vision-capable on one surface and text-only on another. Provider-wide and opt-in: only an
+    explicit ``True`` is a verdict (the field defaults to False, which means "not declared").
+    Runs after the per-model catalogs so a catalog that knows the exact model still wins."""
+    from providers import get_provider_profile
+
+    profile = get_provider_profile(provider)
+    return True if profile is not None and profile.supports_vision is True else None
+
+
 # Capability probes after the config override, in priority order; each returns
 # True/False or None (unknown → next probe). Exceptions are logged and treated as None.
 _VISION_PROBES: Tuple[Tuple[str, Callable[..., Optional[bool]]], ...] = (
     ("managed-runtime caps lookup", _probe_managed_runtime),
     ("caps lookup", _probe_models_dev),
     ("ollama vision probe", _probe_ollama),
+    ("provider profile declaration", _probe_provider_profile),
 )
 
 
@@ -329,7 +342,8 @@ def _lookup_supports_vision(
     """Return True/False if vision capability can be resolved, None if unknown.
 
     Order: config ``supports_vision`` override → :data:`_VISION_PROBES`
-    (managed local runtime → models.dev catalog → Ollama probe).
+    (managed local runtime → models.dev catalog → Ollama probe → registered
+    ``ProviderProfile.supports_vision`` declaration).
     """
     # Named custom providers are canonicalized to ``provider="custom"``; the
     # original name lives in the context-local main runtime. Borrow it only on an

website/docs/developer-guide/model-provider-plugin.md

@@ -100,10 +100,16 @@ Full definition in `providers/base.py`. The most useful ones:
 | `base_url` | str | Default inference endpoint |
 | `models_url` | str | Explicit catalog URL (falls back to `{base_url}/models`) |
 | `auth_type` | str | `api_key` \| `oauth_device_code` \| `oauth_external` \| `copilot` \| `aws_sdk` \| `external_process` |
+<<<<<<< HEAD
 | `auth_handler` | `Callable \| None` | Provider-owned `hermes auth add/status/logout/refresh <name>` — see [Provider-owned auth](#provider-owned-auth-auth_handler-refresh_credential) |
 | `refresh_credential` | `Callable \| None` | Provider-owned rotation of a pooled OAuth row — same section |
 | `fallback_models` | `tuple[str, ...]` | Curated list shown when live catalog fetch fails |
 | `default_headers` | `dict[str, str]` | Sent on every request (e.g. Copilot's `Editor-Version`) |
+=======
+| `fallback_models` | `tuple[str, ...]` | Curated list shown when live catalog fetch fails — in the `/model` picker AND the first-time `hermes setup` / `hermes model` API-key flow, which resolve the catalog the same way (`fetch_models()` merged curated-first with `fallback_models`; `fallback_models` alone when the fetch returns `None` or raises) |
+| `supports_vision` | bool | Declares the endpoint accepts image input. Consumed by the tool-result media path and by the per-turn image-routing probe (`agent/image_routing.decide_image_input_mode`): an image attached to a model of a `supports_vision=True` plugin goes native instead of through `vision_analyze` text, unless a config override or a models.dev per-model entry says otherwise |
+| `default_headers` | `dict[str, str]` | Sent on every request (e.g. Copilot's `Editor-Version`); also forwarded by the default `fetch_models()` catalog request |
+>>>>>>> cac616e8159 (fix(vision): image routing honours ProviderProfile.supports_vision)
 | `fixed_temperature` | Any | `None` = use caller's value; `OMIT_TEMPERATURE` sentinel = don't send temperature at all (Kimi) |
 | `default_max_tokens` | `int \| None` | Provider-level max_tokens cap (Nvidia: 16384) |
 | `unsupported_response_formats` | `tuple` | `response_format` types the API rejects outright; auxiliary requests omit them instead of paying a guaranteed 400 (DeepSeek: `("json_schema",)`) |