fix(routing): preserve profile and delegation parity

agent/chat_completion_helpers.py

@@ -164,6 +164,25 @@ def _validated_openrouter_provider_sort(raw_sort: Any) -> Optional[str]:
     return None
 
 
+def _provider_preferences_for_agent(agent) -> Dict[str, Any]:
+    """Build the validated provider-routing object shared by request paths."""
+    preferences: Dict[str, Any] = {}
+    if agent.providers_allowed:
+        preferences["only"] = agent.providers_allowed
+    if agent.providers_ignored:
+        preferences["ignore"] = agent.providers_ignored
+    if agent.providers_order:
+        preferences["order"] = agent.providers_order
+    provider_sort = _validated_openrouter_provider_sort(agent.provider_sort)
+    if provider_sort:
+        preferences["sort"] = provider_sort
+    if agent.provider_require_parameters:
+        preferences["require_parameters"] = True
+    if agent.provider_data_collection:
+        preferences["data_collection"] = agent.provider_data_collection
+    return preferences
+
+
 def _env_float(name: str, default: float) -> float:
     try:
         return float(os.getenv(name, str(default)))
@@ -801,21 +820,8 @@ def build_api_kwargs(agent, api_messages: list) -> dict:
         _omit_temp = False
         _fixed_temp = None
 
-    # Provider preferences (OpenRouter-style)
-    _prefs: Dict[str, Any] = {}
-    if agent.providers_allowed:
-        _prefs["only"] = agent.providers_allowed
-    if agent.providers_ignored:
-        _prefs["ignore"] = agent.providers_ignored
-    if agent.providers_order:
-        _prefs["order"] = agent.providers_order
-    _provider_sort = _validated_openrouter_provider_sort(agent.provider_sort)
-    if _provider_sort:
-        _prefs["sort"] = _provider_sort
-    if agent.provider_require_parameters:
-        _prefs["require_parameters"] = True
-    if agent.provider_data_collection:
-        _prefs["data_collection"] = agent.provider_data_collection
+    # Provider preferences (aggregator profile decides whether to emit them).
+    _prefs = _provider_preferences_for_agent(agent)
 
     # Anthropic-compatible max-output fallback (last resort only — applied in
     # build_kwargs *after* ephemeral/user/profile max_tokens, never overriding
@@ -1690,42 +1696,32 @@ def handle_max_iterations(agent, messages: list, api_call_count: int) -> str:
             if _lm_reasoning_effort is not None:
                 summary_kwargs["reasoning_effort"] = _lm_reasoning_effort
 
-            # Include provider routing preferences
-            provider_preferences = {}
-            if agent.providers_allowed:
-                provider_preferences["only"] = agent.providers_allowed
-            if agent.providers_ignored:
-                provider_preferences["ignore"] = agent.providers_ignored
-            if agent.providers_order:
-                provider_preferences["order"] = agent.providers_order
-            _provider_sort = _validated_openrouter_provider_sort(agent.provider_sort)
-            if _provider_sort:
-                provider_preferences["sort"] = _provider_sort
-            if agent.provider_require_parameters:
-                provider_preferences["require_parameters"] = True
-            if agent.provider_data_collection:
-                provider_preferences["data_collection"] = agent.provider_data_collection
-            if provider_preferences:
-                profile_provider_preferences = None
-                try:
-                    from providers import get_provider_profile
+            # Merge the profile's canonical body even when routing is unset:
+            # profiles may always emit required metadata such as Portal tags.
+            provider_preferences = _provider_preferences_for_agent(agent)
+            profile_extra_body = {}
+            try:
+                from providers import get_provider_profile
 
-                    provider_profile = get_provider_profile(agent.provider)
-                    if provider_profile is not None:
-                        profile_extra_body = provider_profile.build_extra_body(
-                            provider_preferences=provider_preferences,
-                        )
-                        profile_provider_preferences = profile_extra_body.get("provider")
-                except Exception:
-                    pass
+                provider_profile = get_provider_profile(agent.provider)
+                if provider_profile is not None:
+                    profile_extra_body = provider_profile.build_extra_body(
+                        session_id=getattr(agent, "session_id", None),
+                        provider_preferences=provider_preferences or None,
+                        model=agent.model,
+                        base_url=agent.base_url,
+                        reasoning_config=agent.reasoning_config,
+                    )
+            except Exception:
+                pass
 
-                if profile_provider_preferences:
-                    summary_extra_body["provider"] = profile_provider_preferences
-                elif (
-                    (agent.provider or "").strip().lower() == "openrouter"
-                    or agent._is_openrouter_url()
-                ):
-                    summary_extra_body["provider"] = provider_preferences
+            if profile_extra_body:
+                summary_extra_body.update(profile_extra_body)
+            if provider_preferences and "provider" not in profile_extra_body and (
+                (agent.provider or "").strip().lower() == "openrouter"
+                or agent._is_openrouter_url()
+            ):
+                summary_extra_body["provider"] = provider_preferences
 
             # Pareto Code router plugin — model-gated. Same shape as
             # the main-loop emission so summary calls on

tests/providers/test_provider_profiles.py

@@ -418,6 +418,7 @@ class TestNousProfile:
         from agent.portal_tags import nous_portal_tags
 
         p = get_provider_profile("nous")
+        assert p is not None
         preferences = {"only": ["deepseek"], "ignore": ["deepinfra"]}
         body = p.build_extra_body(provider_preferences=preferences)
 

tests/run_agent/test_run_agent.py

@@ -3789,7 +3789,7 @@ class TestHandleMaxIterations:
         assert kwargs["extra_body"]["provider"]["only"] == ["Anthropic"]
 
     def test_summary_keeps_provider_preferences_for_nous(self, agent):
-        agent.base_url = "https://inference-api.nousresearch.com/v1"
+        agent.base_url = "https://proxy.example.com/v1"
         agent._base_url_lower = agent.base_url.lower()
         agent.provider = "nous"
         agent.providers_allowed = ["deepseek"]
@@ -3804,6 +3804,9 @@ class TestHandleMaxIterations:
 
         assert result == "Summary"
         kwargs = agent.client.chat.completions.create.call_args.kwargs
+        from agent.portal_tags import nous_portal_tags
+
+        assert kwargs["extra_body"]["tags"] == nous_portal_tags()
         assert kwargs["extra_body"]["provider"] == {
             "only": ["deepseek"],
             "ignore": ["deepinfra"],
@@ -3812,6 +3815,21 @@ class TestHandleMaxIterations:
             "data_collection": "deny",
         }
 
+    def test_summary_keeps_nous_profile_body_without_routing_preferences(self, agent):
+        agent.base_url = "https://proxy.example.com/v1"
+        agent._base_url_lower = agent.base_url.lower()
+        agent.provider = "nous"
+        agent.client.chat.completions.create.return_value = _mock_response(content="Summary")
+        agent._cached_system_prompt = "You are helpful."
+
+        result = agent._handle_max_iterations([{"role": "user", "content": "do stuff"}], 60)
+
+        assert result == "Summary"
+        kwargs = agent.client.chat.completions.create.call_args.kwargs
+        from agent.portal_tags import nous_portal_tags
+
+        assert kwargs["extra_body"] == {"tags": nous_portal_tags()}
+
     def test_summary_drops_invalid_provider_sort(self, agent):
         agent.base_url = "https://openrouter.ai/api/v1"
         agent._base_url_lower = agent.base_url.lower()

tests/tools/test_delegate.py

@@ -1446,6 +1446,8 @@ class TestDelegationProviderIntegration(unittest.TestCase):
         parent.providers_ignored = ["openai/gpt-4o-mini"]
         parent.providers_order = ["google/gemini-2.5-pro"]
         parent.provider_sort = "price"
+        parent.provider_require_parameters = True
+        parent.provider_data_collection = "deny"
 
         with patch("run_agent.AIAgent") as MockAgent:
             mock_child = MagicMock()
@@ -1464,6 +1466,44 @@ class TestDelegationProviderIntegration(unittest.TestCase):
             self.assertIsNone(kwargs["providers_ignored"])
             self.assertIsNone(kwargs["providers_order"])
             self.assertIsNone(kwargs["provider_sort"])
+            self.assertIs(kwargs["provider_require_parameters"], False)
+            self.assertEqual(kwargs["provider_data_collection"], "")
+
+    @patch("tools.delegate_tool._load_config")
+    @patch("tools.delegate_tool._resolve_delegation_credentials")
+    def test_same_provider_inherits_all_routing_preferences(self, mock_creds, mock_cfg):
+        mock_cfg.return_value = {"max_iterations": 45}
+        mock_creds.return_value = {
+            "model": None,
+            "provider": None,
+            "base_url": None,
+            "api_key": None,
+            "api_mode": None,
+        }
+        parent = _make_mock_parent(depth=0)
+        parent.provider = "nous"
+        parent.providers_allowed = ["deepseek"]
+        parent.providers_ignored = ["deepinfra"]
+        parent.providers_order = ["anthropic"]
+        parent.provider_sort = "throughput"
+        parent.provider_require_parameters = True
+        parent.provider_data_collection = "deny"
+
+        with patch("run_agent.AIAgent") as MockAgent:
+            mock_child = MagicMock()
+            mock_child.run_conversation.return_value = {
+                "final_response": "done", "completed": True, "api_calls": 1
+            }
+            MockAgent.return_value = mock_child
+            delegate_task(goal="Keep routing", parent_agent=parent)
+
+        _, kwargs = MockAgent.call_args
+        self.assertEqual(kwargs["providers_allowed"], ["deepseek"])
+        self.assertEqual(kwargs["providers_ignored"], ["deepinfra"])
+        self.assertEqual(kwargs["providers_order"], ["anthropic"])
+        self.assertEqual(kwargs["provider_sort"], "throughput")
+        self.assertIs(kwargs["provider_require_parameters"], True)
+        self.assertEqual(kwargs["provider_data_collection"], "deny")
 
     @patch("tools.delegate_tool._load_config")
     @patch("tools.delegate_tool._resolve_delegation_credentials")

tools/delegate_tool.py

@@ -1288,12 +1288,20 @@ def _build_child_agent(
     child_providers_ignored = getattr(parent_agent, "providers_ignored", None)
     child_providers_order = getattr(parent_agent, "providers_order", None)
     child_provider_sort = getattr(parent_agent, "provider_sort", None)
+    child_provider_require_parameters = getattr(
+        parent_agent, "provider_require_parameters", False
+    )
+    child_provider_data_collection = getattr(
+        parent_agent, "provider_data_collection", None
+    ) or ""
     child_openrouter_min_coding_score = getattr(parent_agent, "openrouter_min_coding_score", None)
     if override_provider:
         child_providers_allowed = None
         child_providers_ignored = None
         child_providers_order = None
         child_provider_sort = None
+        child_provider_require_parameters = False
+        child_provider_data_collection = ""
         # Note: openrouter_min_coding_score is model-gated (only emitted on
         # openrouter/pareto-code), so we keep it inherited even when the
         # provider is overridden — it's a no-op on any other model.
@@ -1326,6 +1334,8 @@ def _build_child_agent(
         providers_ignored=child_providers_ignored,
         providers_order=child_providers_order,
         provider_sort=child_provider_sort,
+        provider_require_parameters=child_provider_require_parameters,
+        provider_data_collection=child_provider_data_collection,
         openrouter_min_coding_score=child_openrouter_min_coding_score,
         tool_progress_callback=child_progress_cb,
         iteration_budget=None,  # fresh budget per subagent

website/docs/user-guide/features/provider-routing.md

@@ -167,7 +167,7 @@ provider_routing:
 
 ## How It Works
 
-Provider routing preferences are passed to OpenRouter or Nous Portal via the `extra_body.provider` field on every API call. (`extra_body` is the OpenAI Python SDK argument; it becomes the top-level `provider` object in the JSON request.) This applies to both:
+Provider routing preferences are passed to OpenRouter or Nous Portal on agent chat requests and iteration-limit summaries via the `extra_body.provider` field. (`extra_body` is the OpenAI Python SDK argument; it becomes the top-level `provider` object in the JSON request.) Auxiliary tasks such as compression and title generation are configured independently under `auxiliary.<task>.extra_body`.
 
 - **CLI mode** — configured in `~/.hermes/config.yaml`, loaded at startup
 - **Gateway mode** — same config file, loaded when the gateway starts