From 2e7040240d201f433d0fe42f922dbfbe953c6c8b Mon Sep 17 00:00:00 2001
From: "devin-ai-integration[bot]"
 <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Date: Sat, 1 Aug 2026 17:10:01 -0700
Subject: [PATCH] feat(config): accept STRIX_REASONING_EFFORT=max for providers
 that support it (#956)

Co-authored-by: Ahmed Allam <allam@usestrix.com>
---
 docs/advanced/configuration.mdx |  2 +-
 strix/config/models.py          | 11 +++++++----
 strix/config/settings.py        |  2 +-
 strix/core/inputs.py            | 18 +++++++++++++++++-
 strix/interface/main.py         |  4 ++--
 tests/test_inputs.py            | 11 +++++++++++
 6 files changed, 39 insertions(+), 9 deletions(-)

diff --git a/docs/advanced/configuration.mdx b/docs/advanced/configuration.mdx
index af98b8b8..027f6d47 100644
--- a/docs/advanced/configuration.mdx
+++ b/docs/advanced/configuration.mdx
@@ -36,7 +36,7 @@ Configure Strix using environment variables or a config file.
 </ParamField>
 
 <ParamField path="STRIX_REASONING_EFFORT" default="high" type="string">
-  Control thinking effort for reasoning models. Valid values: `none`, `minimal`, `low`, `medium`, `high`, `xhigh`. Defaults to `medium` for quick scan mode.
+  Control thinking effort for reasoning models. Valid values: `none`, `minimal`, `low`, `medium`, `high`, `xhigh`, `max`. Defaults to `medium` for quick scan mode.
 </ParamField>
 
 <ParamField path="STRIX_MEMORY_COMPRESSOR_TIMEOUT" default="30" type="integer">
diff --git a/strix/config/models.py b/strix/config/models.py
index 1c84ef3c..a3bb481a 100644
--- a/strix/config/models.py
+++ b/strix/config/models.py
@@ -83,10 +83,13 @@ class _CodexResponsesModel(OpenAIResponsesModel):
         effort = self._reasoning_effort
         if effort and effort != "none":
             # Clamp to efforts the backend accepts.
-            if effort == "minimal":
-                effort = "low"
-            elif effort == "xhigh":
-                effort = "high"
+            match effort:
+                case "minimal":
+                    effort = "low"
+                case "xhigh" | "max":
+                    effort = "high"
+                case _:
+                    pass
             overrides = overrides.resolve(ModelSettings(reasoning=Reasoning(effort=effort)))
         return model_settings.resolve(overrides)
 
diff --git a/strix/config/settings.py b/strix/config/settings.py
index e53d125c..0f26cd62 100644
--- a/strix/config/settings.py
+++ b/strix/config/settings.py
@@ -8,7 +8,7 @@ from pydantic import AliasChoices, Field
 from pydantic_settings import BaseSettings, SettingsConfigDict
 
 
-ReasoningEffort = Literal["none", "minimal", "low", "medium", "high", "xhigh"]
+ReasoningEffort = Literal["none", "minimal", "low", "medium", "high", "xhigh", "max"]
 
 _BASE_CONFIG = SettingsConfigDict(
     case_sensitive=False,
diff --git a/strix/core/inputs.py b/strix/core/inputs.py
index 34a2d4b3..81f46076 100644
--- a/strix/core/inputs.py
+++ b/strix/core/inputs.py
@@ -147,7 +147,7 @@ def make_model_settings(
         and model_supports_reasoning(model_name)
     ):
         model_settings = model_settings.resolve(
-            ModelSettings(reasoning=Reasoning(effort=reasoning_effort)),
+            _reasoning_settings(reasoning_effort, model_settings.extra_args),
         )
     if force_required_tool_choice and _accepts_required_tool_choice(model_name):
         model_settings = model_settings.resolve(ModelSettings(tool_choice="required"))
@@ -162,6 +162,22 @@ def make_model_settings(
     return model_settings
 
 
+def _reasoning_settings(
+    effort: ReasoningEffort,
+    extra_args: dict[str, Any] | None,
+) -> ModelSettings:
+    """``max`` is not in the OpenAI SDK's ``Reasoning.effort`` enum, so send it as
+    a raw body field instead — also keeping it clear of LiteLLM's DeepSeek mapping,
+    which collapses every ``reasoning_effort`` level to plain thinking-enabled.
+    Providers that don't support ``max`` reject the request.
+    """
+    if effort != "max":
+        return ModelSettings(reasoning=Reasoning(effort=effort))
+    return ModelSettings(
+        extra_args={**(extra_args or {}), "extra_body": {"reasoning_effort": "max"}},
+    )
+
+
 def _prompt_cache_extra_args(model_name: str) -> dict[str, Any] | None:
     """LiteLLM ``cache_control_injection_points`` for Claude prompt caching.
 
diff --git a/strix/interface/main.py b/strix/interface/main.py
index 4d88beda..abe77fa3 100644
--- a/strix/interface/main.py
+++ b/strix/interface/main.py
@@ -171,8 +171,8 @@ def validate_environment() -> None:
                     error_text.append("• ", style="white")
                     error_text.append("STRIX_REASONING_EFFORT", style="bold cyan")
                     error_text.append(
-                        " - Reasoning effort level: none, minimal, low, medium, high, xhigh "
-                        "(default: high)\n",
+                        " - Reasoning effort level: none, minimal, low, medium, high, xhigh, "
+                        "max (default: high)\n",
                         style="white",
                     )
 
diff --git a/tests/test_inputs.py b/tests/test_inputs.py
index 871ea149..83a431ad 100644
--- a/tests/test_inputs.py
+++ b/tests/test_inputs.py
@@ -129,6 +129,17 @@ def test_prompt_cache_kept_for_non_bedrock_claude_even_if_unmapped(monkeypatch:
         ]
 
 
+def test_max_reasoning_effort_sent_as_raw_body_field() -> None:
+    # "max" is absent from the OpenAI SDK's Reasoning enum, and LiteLLM's DeepSeek
+    # mapping collapses every effort to thinking-enabled, so it has to ride along
+    # as a raw body field to reach the provider.
+    settings = make_model_settings(
+        "max", model_name="deepseek/deepseek-v4-flash", request_timeout=30
+    )
+    assert settings.reasoning is None
+    assert settings.extra_args == {"timeout": 30, "extra_body": {"reasoning_effort": "max"}}
+
+
 def test_conversation_tail_breakpoint_moves_with_appended_transcript() -> None:
     # LiteLLM must place the index=-1 cache_control on the last message however
     # long the transcript grows.