[owl] Support disabling reasoning for GPT-5.1 (#845)

Backend - owl (API server)

- Support disabling reasoning for GPT-5.1
- Fix tool call step formatting: Add double new lines

Author: jiahuei

clients/python/src/jamaibase/types/lm.py

@@ -1136,29 +1136,19 @@ class ChatRequestBase(BaseModel):
         ),
         examples=[{}],
     )
-    reasoning_effort: Literal["disable", "minimal", "low", "medium", "high"] | None = Field(
-        "minimal",
-        description=(
-            "Constrains effort on reasoning for reasoning models. "
-            "Currently supported values are `disable`, `minimal`, `low`, `medium`, and `high`. "
-            "Reducing reasoning effort can result in faster responses and fewer tokens used on reasoning in a response. "
-            "For non-OpenAI models, `low` ~ 1024 tokens, `medium` ~ 2048 tokens, `high` ~ 4096 tokens. "
-            "Note that this parameter will be ignored when using models that do not support it, "
-            "such as non-reasoning models."
-        ),
-        examples=["low"],
-    )
-    reasoning_effort: Literal["disable", "minimal", "low", "medium", "high"] | None = Field(
-        None,
-        description=(
-            "Constrains effort on reasoning for reasoning models. "
-            "Currently supported values are `disable`, `minimal`, `low`, `medium`, and `high`. "
-            "Reducing reasoning effort can result in faster responses and fewer tokens used on reasoning in a response. "
-            "For non-OpenAI models, `low` ~ 1024 tokens, `medium` ~ 4096 tokens, `high` ~ 8192 tokens. "
-            "Note that this parameter will be ignored when using models that do not support it, "
-            "such as non-reasoning models."
-        ),
-        examples=["low"],
+    reasoning_effort: Literal["disable", "minimal", "none", "low", "medium", "high"] | None = (
+        Field(
+            None,
+            description=(
+                "Constrains effort on reasoning for reasoning models. "
+                'Currently supported values are `None`, "disable", "none", "minimal", "low", "medium", and "high". '
+                "Reducing reasoning effort can result in faster responses and fewer tokens used on reasoning in a response. "
+                "For non-OpenAI models, `low` ~ 1024 tokens, `medium` ~ 4096 tokens, `high` ~ 8192 tokens. "
+                "Note that this parameter will be ignored when using models that do not support it, "
+                "such as non-reasoning models."
+            ),
+            examples=["low"],
+        )
     )
     thinking_budget: int | None = Field(
         None,

services/api/src/owl/utils/lm.py

@@ -519,7 +519,7 @@ def _prepare_hyperparams(
             return
         # Disable reasoning if requested
         if (
-            reasoning_effort in ("disable", "minimal")
+            reasoning_effort in ("disable", "minimal", "none")
             or thinking_budget == 0
             or (reasoning_effort is None and thinking_budget is None)
         ):
@@ -540,7 +540,14 @@ def _prepare_hyperparams(
                 hyperparams["thinking"] = {"type": "disabled"}
                 return
             elif ctx.inference_provider == CloudProvider.OPENAI:
-                if "gpt-5" in ctx.routing_id:
+                if "gpt-5.1" in ctx.routing_id:
+                    # gpt-5.1: Supported values are: 'none', 'low', 'medium', and 'high'.
+                    hyperparams["reasoning"] = {
+                        "effort": "none",
+                        "summary": reasoning_summary,
+                    }
+                    return
+                elif "gpt-5" in ctx.routing_id:
                     hyperparams["reasoning"] = {
                         "effort": "minimal",
                         "summary": reasoning_summary,
@@ -668,20 +675,22 @@ async def _openai_responses_stream(
                         and hasattr(chunk.item.action, "query")
                         and chunk.item.action.query
                     ):
+                        yield self._stream_delta(Delta(role="assistant", reasoning_content="\n\n"))
                         yield self._stream_delta(
                             Delta(
                                 role="assistant",
-                                reasoning_content=f'\n\nSearched the web for "{chunk.item.action.query}".',
+                                reasoning_content=f'Searched the web for "{chunk.item.action.query}".',
                             )
                         )
                         yield self._stream_delta(Delta(role="assistant", reasoning_content="\n\n"))
                 elif isinstance(chunk.item, ResponseCodeInterpreterToolCall):
                     usage_stats["code_interpreter_calls"] += 1
                     code_snippet = chunk.item.code
+                    yield self._stream_delta(Delta(role="assistant", reasoning_content="\n\n"))
                     yield self._stream_delta(
                         Delta(
                             role="assistant",
-                            reasoning_content=f"\n\nRan Python code:\n\n```python\n{code_snippet}\n```",
+                            reasoning_content=f"Ran Python code:\n\n```python\n{code_snippet}\n```",
                         )
                     )
                     yield self._stream_delta(Delta(role="assistant", reasoning_content="\n\n"))
@@ -744,11 +753,13 @@ async def _openai_responses(
             elif isinstance(item, ResponseFunctionWebSearch) and item.status == "completed":
                 usage_stats["web_search_calls"] += 1
                 if item.action and hasattr(item.action, "query") and item.action.query:
-                    reasoning_parts.append(f'Searched the web for "{item.action.query}".')
+                    reasoning_parts.append(f'\n\nSearched the web for "{item.action.query}".\n\n')
             elif isinstance(item, ResponseCodeInterpreterToolCall) and item.status == "completed":
                 usage_stats["code_interpreter_calls"] += 1
                 code_snippet = item.code
-                reasoning_parts.append(f"Ran Python code:\n\n```python\n{code_snippet}\n```")
+                reasoning_parts.append(
+                    f"\n\nRan Python code:\n\n```python\n{code_snippet}\n```\n\n"
+                )
             elif isinstance(item, ResponseOutputMessage) and item.status == "completed":
                 text_content = item.content[0].text if item.content else ""
                 result_parts.append(text_content)