From 8662254ab2fe019fffa4d82cf5263b235c914b5e Mon Sep 17 00:00:00 2001
From: Epoxidex <skosarevivan@yandex.ru>
Date: Thu, 21 May 2026 15:01:14 +0300
Subject: [PATCH] fix(ollama): emit top-level reasoning_effort=none on
 /v1/chat/completions (#25758)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Ollama's /v1/chat/completions silently ignores extra_body.think (it only
honours it on /api/chat — ollama/ollama#14820), so agent.reasoning_effort:
none never actually disabled thinking on OpenAI-compatible Ollama routes.
Emit the top-level reasoning_effort='none' field (which Ollama respects)
alongside think=False (kept for proxies and the native /api/chat path).

The PR's second half (propagating reasoning_config to the background-review
fork) already landed on main via agent/background_review.py, so only the
provider-profile change is salvaged here, resolved onto the current
GLM/effort-aware profile.

Salvaged from PR #29820 by @Epoxidex.
---
 plugins/model-providers/custom/__init__.py         | 11 ++++++++++-
 .../plugins/model_providers/test_custom_profile.py | 14 ++++++++++----
 2 files changed, 20 insertions(+), 5 deletions(-)

diff --git a/plugins/model-providers/custom/__init__.py b/plugins/model-providers/custom/__init__.py
index 2847d161adf0..de744d50914e 100644
--- a/plugins/model-providers/custom/__init__.py
+++ b/plugins/model-providers/custom/__init__.py
@@ -4,7 +4,9 @@ Covers any endpoint registered as provider="custom", including local
 Ollama instances and OpenAI-compatible reasoning endpoints (GLM-5.2 on
 Volcengine ARK, vLLM, llama.cpp). Key quirks:
   - ollama_num_ctx → extra_body.options.num_ctx (local context window)
-  - reasoning_config disabled → extra_body.think = False
+  - reasoning_config disabled → top-level reasoning_effort="none"
+    (Ollama /v1/chat/completions ignores think=False — ollama#14820)
+    + extra_body.think = False for /api/chat and proxies
   - reasoning_config enabled + effort → top-level reasoning_effort
     (the native OpenAI-compatible format GLM/ARK expect; unset omits it
     so the endpoint's server default applies)
@@ -53,6 +55,13 @@ class CustomProfile(ProviderProfile):
             _effort = (reasoning_config.get("effort") or "").strip().lower()
             _enabled = reasoning_config.get("enabled", True)
             if _effort == "none" or _enabled is False:
+                # Ollama's /v1/chat/completions silently ignores
+                # extra_body.think (only /api/chat honours it — ollama#14820)
+                # but respects the top-level reasoning_effort field, so both
+                # are needed to actually stop a thinking-capable model from
+                # reasoning (#25758). Endpoints that recognize neither simply
+                # ignore them.
+                top_level["reasoning_effort"] = "none"
                 extra_body["think"] = False
             elif _effort:
                 top_level["reasoning_effort"] = _effort
diff --git a/tests/plugins/model_providers/test_custom_profile.py b/tests/plugins/model_providers/test_custom_profile.py
index e20ee57b8ceb..152359ed2f30 100644
--- a/tests/plugins/model_providers/test_custom_profile.py
+++ b/tests/plugins/model_providers/test_custom_profile.py
@@ -49,20 +49,26 @@ class TestCustomReasoningWireShape:
         assert tl == {}
 
     def test_disabled_sends_think_false(self, custom_profile):
-        """enabled=False → extra_body.think = False (Ollama thinking-off flag)."""
+        """enabled=False → reasoning_effort='none' top-level + think=False.
+
+        Both fields are required: Ollama's /v1/chat/completions silently
+        ignores extra_body.think (only /api/chat honours it — ollama#14820)
+        but respects top-level reasoning_effort (#25758). think=False stays
+        for proxies and the native /api/chat path.
+        """
         eb, tl = custom_profile.build_api_kwargs_extras(
             reasoning_config={"enabled": False}, model="glm-5.2"
         )
         assert eb == {"think": False}
-        assert tl == {}
+        assert tl == {"reasoning_effort": "none"}
 
     def test_effort_none_sends_think_false(self, custom_profile):
-        """effort='none' is the disable alias → think=False, no effort."""
+        """effort='none' is the disable alias → same dual emission."""
         eb, tl = custom_profile.build_api_kwargs_extras(
             reasoning_config={"enabled": True, "effort": "none"}, model="glm-5.2"
         )
         assert eb == {"think": False}
-        assert tl == {}
+        assert tl == {"reasoning_effort": "none"}
 
     @pytest.mark.parametrize(
         "effort", ["minimal", "low", "medium", "high", "xhigh", "max"]