Improve GPT-OSS writing speed and reliability

services/cowriter_service.py

@@ -72,7 +72,9 @@ def _suggest_llm(
             system_prompt=system_prompt,
             user_prompt=user_prompt,
             temperature=0.78,
-            max_tokens=max_tokens * n + 180,
+            # GPT-OSS counts reasoning and visible text in this budget. Keep a
+            # floor so all requested insert-ready options reach the response.
+            max_tokens=max(1024, max_tokens * n + 220),
         )
         model_used = "standard"
     else:
@@ -81,7 +83,7 @@ def _suggest_llm(
             user_prompt=user_prompt,
             model=model,
             temperature=0.78,
-            max_tokens=max_tokens * n + 180,
+            max_tokens=max(1024, max_tokens * n + 220),
         )
 
     return _parse_suggestions(raw, text, n), model_used

services/llm_client.py

@@ -227,6 +227,11 @@ def _call_provider(
         "temperature": temperature,
         "max_tokens": max_tokens,
     }
+    if model.startswith("openai/gpt-oss-"):
+        # These models default to medium reasoning, which spends extra tokens
+        # and latency on routine writing tasks. Low effort keeps their quality
+        # while returning visible copy much faster and more reliably.
+        payload["reasoning_effort"] = "low"
 
     try:
         resp = httpx.post(url, json=payload, headers=headers, timeout=timeout)

tests/test_llm_client.py

@@ -1,8 +1,10 @@
 """Tests for the current writing-model registry and legacy migrations."""
 
+from unittest.mock import Mock, patch
+
 import pytest
 
-from services.llm_client import resolve_premium_model
+from services.llm_client import _call_provider, resolve_premium_model
 
 
 @pytest.mark.parametrize(
@@ -25,3 +27,23 @@ def test_resolve_premium_model(selector, expected):
 def test_resolve_premium_model_rejects_arbitrary_provider_ids():
     with pytest.raises(ValueError, match="Unsupported writing model"):
         resolve_premium_model("unknown/provider-model")
+
+
+def test_gpt_oss_requests_low_reasoning_for_fast_writing():
+    response = Mock()
+    response.raise_for_status.return_value = None
+    response.json.return_value = {"choices": [{"message": {"content": "Ready"}}]}
+
+    with patch("services.llm_client.httpx.post", return_value=response) as post:
+        result = _call_provider(
+            url="https://provider.test/chat",
+            api_key="secret",
+            model="openai/gpt-oss-20b",
+            messages=[{"role": "user", "content": "Continue this draft"}],
+            temperature=0.7,
+            max_tokens=1024,
+            timeout=30.0,
+        )
+
+    assert result == "Ready"
+    assert post.call_args.kwargs["json"]["reasoning_effort"] == "low"