feat(sdk): add NVIDIA NIM app-origin attribution

libs/deepagents/deepagents/_models.py

@@ -39,9 +39,10 @@ def resolve_model(model: str | BaseChatModel) -> BaseChatModel:
 
     String models are resolved via `init_chat_model`, composed with any
     provider-specific initialization behavior registered in the
-    `ProviderProfile` registry. Built-in registrations supply the OpenAI
-    Responses API default and OpenRouter app attribution headers; users can
-    layer additional providers or overrides via `register_provider_profile`.
+    `ProviderProfile` registry. Built-in registrations supply NVIDIA NIM and
+    OpenRouter app attribution headers plus the OpenAI Responses API default;
+    users can layer additional providers or overrides via
+    `register_provider_profile`.
 
     Args:
         model: Model string (e.g. `"openai:gpt-5.4"`) or pre-configured

libs/deepagents/deepagents/profiles/_builtin_profiles.py

@@ -33,7 +33,7 @@
     _openai_codex,
 )
 from deepagents.profiles.harness.harness_profiles import _HARNESS_PROFILES
-from deepagents.profiles.provider import _openai, _openrouter
+from deepagents.profiles.provider import _nvidia, _openai, _openrouter
 from deepagents.profiles.provider.provider_profiles import _PROVIDER_PROFILES
 
 logger = logging.getLogger(__name__)
@@ -145,6 +145,7 @@ def _ensure_builtin_profiles_loaded() -> None:
     saved_harness_profiles = dict(_HARNESS_PROFILES)
     saved_bootstrap_harness_keys = _BOOTSTRAP_HARNESS_KEYS
     try:
+        _nvidia.register()
         _openai.register()
         _openrouter.register()
         _anthropic_opus_4_7.register()

libs/deepagents/deepagents/profiles/provider/_nvidia.py

@@ -0,0 +1,50 @@
+"""Built-in NVIDIA provider profile and helpers.
+
+Injects Deep Agents app-origin attribution into NVIDIA NIM requests via the
+header supported by `langchain-nvidia-ai-endpoints`.
+
+Registered directly by `_ensure_builtin_profiles_loaded` during the first
+profile-registry access. Not exposed as an `importlib.metadata` entry point —
+built-ins ship with the SDK and should not depend on install-time metadata to
+activate.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+from deepagents.profiles.provider.provider_profiles import (
+    ProviderProfile,
+    _register_provider_profile_impl,
+)
+
+_NVIDIA_BILLING_ORIGIN_HEADER = "X-BILLING-INVOKE-ORIGIN"
+"""NVIDIA NIM header used to attribute requests to an originating app."""
+
+_NVIDIA_APP_ORIGIN = "DeepAgents"
+"""Deep Agents identity reported to NVIDIA NIM."""
+
+
+def _nvidia_attribution_kwargs() -> dict[str, Any]:
+    """Build default NVIDIA NIM app-attribution kwargs.
+
+    `ChatNVIDIA` accepts `default_headers` and merges them into sync, async,
+    and streaming requests. Returning a new nested mapping on every call keeps
+    profile resolution isolated between model instances.
+
+    Returns:
+        Dictionary of kwargs to spread into `init_chat_model`.
+    """
+    return {
+        "default_headers": {
+            _NVIDIA_BILLING_ORIGIN_HEADER: _NVIDIA_APP_ORIGIN,
+        }
+    }
+
+
+def register() -> None:
+    """Register the built-in NVIDIA provider profile."""
+    _register_provider_profile_impl(
+        "nvidia",
+        ProviderProfile(init_kwargs_factory=_nvidia_attribution_kwargs),
+    )

libs/deepagents/deepagents/profiles/provider/provider_profiles.py

@@ -13,8 +13,9 @@
 runtime state (e.g. environment variables).
 """
 # Built-in profiles are registered lazily on first registry access for
-# `"openai"` (enables the Responses API by default) and `"openrouter"`
-# (enforces a minimum version and injects app-attribution headers).
+# `"nvidia"` (injects NIM app-attribution headers), `"openai"` (enables the
+# Responses API by default), and `"openrouter"` (enforces a minimum version and
+# injects app-attribution headers).
 # Additional providers or per-model overrides can be registered with
 # `register_provider_profile`.
 

libs/deepagents/tests/unit_tests/test_models.py

@@ -31,6 +31,11 @@
     _merge_middleware,
     _merge_profiles,
 )
+from deepagents.profiles.provider._nvidia import (
+    _NVIDIA_APP_ORIGIN,
+    _NVIDIA_BILLING_ORIGIN_HEADER,
+    _nvidia_attribution_kwargs,
+)
 from deepagents.profiles.provider._openrouter import (
     _OPENROUTER_ALLOW_AZURE_ENV,
     _OPENROUTER_APP_TITLE,
@@ -104,6 +109,19 @@ def test_openrouter_prefix_sets_attribution(self) -> None:
         )
         assert result is mock.return_value
 
+    def test_nvidia_prefix_sets_billing_origin(self) -> None:
+        with patch("deepagents._models.init_chat_model") as mock:
+            mock.return_value = MagicMock(spec=BaseChatModel)
+            result = resolve_model("nvidia:nvidia/nemotron-3-super-120b-a12b")
+
+        mock.assert_called_once_with(
+            "nvidia:nvidia/nemotron-3-super-120b-a12b",
+            default_headers={
+                _NVIDIA_BILLING_ORIGIN_HEADER: _NVIDIA_APP_ORIGIN,
+            },
+        )
+        assert result is mock.return_value
+
     def test_openrouter_env_var_overrides_app_url(self) -> None:
         env = {"OPENROUTER_APP_URL": "https://custom.app"}
         with (
@@ -1148,6 +1166,10 @@ def test_openrouter_provider_profile_has_pre_init_and_factory(self) -> None:
         assert profile.pre_init is not None
         assert profile.init_kwargs_factory is not None
 
+    def test_nvidia_provider_profile_has_attribution_factory(self) -> None:
+        profile = get_provider_profile("nvidia:nvidia/nemotron-3-super-120b-a12b")
+        assert profile.init_kwargs_factory is _nvidia_attribution_kwargs
+
     def test_openai_has_no_built_in_harness_profile(self) -> None:
         assert _get_harness_profile("openai:gpt-5") is None
 
@@ -1787,6 +1809,26 @@ def test_openrouter_runs_pre_init_and_factory(self) -> None:
         _, kwargs = mock.call_args
         assert "app_url" in kwargs or "app_title" in kwargs
 
+    def test_nvidia_runs_attribution_factory(self) -> None:
+        with patch("deepagents._models.init_chat_model") as mock:
+            mock.return_value = MagicMock(spec=BaseChatModel)
+            resolve_model("nvidia:nvidia/nemotron-3-super-120b-a12b")
+
+        _, kwargs = mock.call_args
+        assert kwargs["default_headers"] == {
+            _NVIDIA_BILLING_ORIGIN_HEADER: _NVIDIA_APP_ORIGIN,
+        }
+
+    def test_nvidia_caller_headers_override_attribution_default(self) -> None:
+        custom_headers = {_NVIDIA_BILLING_ORIGIN_HEADER: "CustomHarness"}
+
+        kwargs = apply_provider_profile(
+            "nvidia:nvidia/nemotron-3-super-120b-a12b",
+            {"default_headers": custom_headers},
+        )
+
+        assert kwargs["default_headers"] is custom_headers
+
     def test_unknown_provider_passes_no_extra_kwargs(self) -> None:
         with patch("deepagents._models.init_chat_model") as mock:
             mock.return_value = MagicMock(spec=BaseChatModel)