Skip to content

The agent that grows with you

fix(agent): a plugin claiming a host never replaces core's reasoning gate there

Tekniumf4a935b47ada
diff --git a/agent/chat_completion_helpers.py b/agent/chat_completion_helpers.py
index b7ee890b31a..d8a4b815ae6 100644
--- a/agent/chat_completion_helpers.py
+++ b/agent/chat_completion_helpers.py
@@ -1442,8 +1442,7 @@ def _build_chat_completions_kwargs(agent, api_messages, tools_for_api, reasoning
_is_or = agent._is_openrouter_url()
_host = agent._base_url_lower
_is_lmstudio = (agent.provider or "").strip().lower() == "lmstudio"
- from providers import profile_for_host
- _profile_endpoint = profile_for_host(str(agent.base_url or "")) is not None
+ _profile_endpoint = agent._reasoning_endpoint_profile() is not None
# _fixed_temperature_for_model may return the OMIT_TEMPERATURE sentinel
# (temperature omitted entirely), a numeric override, or None.
diff --git a/agent/reasoning_params.py b/agent/reasoning_params.py
index 5cc06ce1949..7c6c145b8eb 100644
--- a/agent/reasoning_params.py
+++ b/agent/reasoning_params.py
@@ -17,6 +17,9 @@ _OPENROUTER_REASONING_PREFIXES = (
"qwen/qwen3", "tencent/hy", "xiaomi/",
)
+# Hosts whose reasoning gate core owns: a plugin claiming one in ``hostnames`` never replaces it.
+_CORE_REASONING_GATED_HOSTS = ("nousresearch.com", "ai-gateway.vercel.sh", "ollama.com")
+
# Probe results cache per (model, base_url). Definitive values cache permanently; an
# "unknown" value (empty list / None) caches 60s so a transient failure neither sticks
# for the session nor round-trips every turn.
@@ -92,9 +95,7 @@ class ReasoningParamsMixin:
url = self._base_url_lower
if base_url_host_matches(url, "nousresearch.com") or base_url_host_matches(url, "ai-gateway.vercel.sh"):
return True
- from providers import profile_for_host
-
- if (profile := profile_for_host(str(getattr(self, "base_url", None) or ""))) is not None:
+ if (profile := self._reasoning_endpoint_profile()) is not None:
# A claimed endpoint gates on its own catalog: reasoning goes out iff it declares a payload.
from agent.provider_route_hooks import ask_profile
return ask_profile(profile, "reasoning_payload", None, self.model, None,
@@ -137,13 +138,23 @@ class ReasoningParamsMixin:
return False
return bool(_cached_probe(self, "_ollama_thinking_cache", ollama_model_supports_thinking, None, lambda v: v is not None))
- def _endpoint_reasoning_payload(self) -> dict | None:
- """``reasoning_payload`` of the profile claiming this endpoint for the agent's model and
- ``reasoning_config`` (``providers.profile_for_host``), or None when no profile claims it."""
+ def _reasoning_endpoint_profile(self):
+ """The profile claiming this endpoint (``providers.profile_for_host``) that gates its reasoning, or None.
+ Endpoints core gates itself (Nous Portal, Vercel AI Gateway, LM Studio, Ollama Cloud, OpenRouter) keep
+ core's gate even when a plugin claims their host for metadata or auxiliary routing."""
+ url = self._base_url_lower
+ if ((self.provider or "").strip().lower() == "lmstudio" or self._is_openrouter_url()
+ or any(base_url_host_matches(url, host) for host in _CORE_REASONING_GATED_HOSTS)):
+ return None
from providers import profile_for_host
+ return profile_for_host(str(getattr(self, "base_url", None) or ""))
+
+ def _endpoint_reasoning_payload(self) -> dict | None:
+ """``reasoning_payload`` of the profile gating this endpoint's reasoning for the agent's model and
+ ``reasoning_config`` (``_reasoning_endpoint_profile``), or None when no profile does."""
base_url = getattr(self, "base_url", None)
- profile = profile_for_host(str(base_url or ""))
+ profile = self._reasoning_endpoint_profile()
if profile is None:
return None
from agent.provider_route_hooks import ask_profile
diff --git a/providers/base.py b/providers/base.py
index e166f4b391e..c51c7bdcfbc 100644
--- a/providers/base.py
+++ b/providers/base.py
@@ -596,6 +596,8 @@ class ProviderProfile:
# ── Lane L4: catalog / metadata seams ─────────────────────
# Hosts (exact or any subdomain) that serve THIS provider's API, so a ``custom`` endpoint or an
# auto-routed aux client at one of them is recognised as this provider (``providers.profile_for_host``).
+ # It also makes ``reasoning_payload`` the reasoning gate there, except on hosts core gates itself
+ # (OpenRouter, Ollama Cloud, Nous Portal, Vercel AI Gateway, the ``lmstudio`` provider).
hostnames: tuple = ()
# Hosts that identify this provider's catalog for model metadata (context length, models.dev) but
# take other credentials; matched only by ``profile_for_host(url, metadata=True)``.
diff --git a/tests/agent/test_reasoning_gate_claimed_hosts.py b/tests/agent/test_reasoning_gate_claimed_hosts.py
new file mode 100644
index 00000000000..511ce48356d
--- /dev/null
+++ b/tests/agent/test_reasoning_gate_claimed_hosts.py
@@ -0,0 +1,47 @@
+"""A plugin claiming a host in ``hostnames`` (catalog / aux routing) never replaces the reasoning gate core
+owns for that host (OpenRouter, Ollama Cloud): only hosts no core gate covers follow ``reasoning_payload``."""
+
+from unittest.mock import MagicMock, patch
+
+import pytest
+
+
+def claiming_profile(request):
+ import providers
+ from providers.base import ProviderProfile
+
+ providers.list_providers() # discovery first, so the fake survives it
+ providers.register_provider(ProviderProfile(name="hostclaimer", hostnames=("openrouter.ai", "ollama.com")))
+
+ def _cleanup():
+ providers._REGISTRY.pop("hostclaimer", None)
+ providers._SOURCES.pop("hostclaimer", None)
+ providers._PROVIDER_LIST_CACHE = None
+
+ request.addfinalizer(_cleanup)
+
+
+def _agent(provider, base_url, model):
+ from run_agent import AIAgent
+
+ with patch("agent.process_bootstrap.OpenAI") as mock_openai:
+ mock_openai.return_value = MagicMock()
+ return AIAgent(api_key="k", base_url=base_url, provider=provider, api_mode="chat_completions", model=model,
+ quiet_mode=True, skip_context_files=True, skip_memory=True, enabled_toolsets=[],
+ reasoning_config={"enabled": True, "effort": "high"})
+
+
+def test_openrouter_keeps_its_reasoning_gate_and_body_when_a_plugin_claims_the_host(claiming_profile):
+ agent = _agent("openrouter", "https://openrouter.ai/api/v1", "anthropic/claude-sonnet-4.5")
+ with patch("hermes_cli.models_reasoning_caps.openrouter_model_reasoning_capabilities",
+ return_value={"supports_reasoning": True}):
+ assert agent._supports_reasoning_extra_body() is True
+ extra_body = agent._build_api_kwargs([{"role": "user", "content": "x"}]).get("extra_body") or {}
+ assert extra_body.get("reasoning") == {"enabled": True, "effort": "high"}
+
+
+def test_ollama_cloud_keeps_its_capability_gate_when_a_plugin_claims_the_host(claiming_profile):
+ agent = _agent("custom", "https://ollama.com/v1", "gpt-oss:120b")
+ with patch.object(type(agent), "_ollama_supports_thinking_cached", return_value=True):
+ assert agent._supports_reasoning_extra_body() is True