fix(cascade): give OVH/gpt-oss reasoning headroom so Tier-2 isn't silently dead

gpt-oss-120b is a reasoning model: it spends output tokens on chain-of-thought
before the answer. deep_check called _call_ovh with max_tokens=400, which
length-capped it mid-reasoning -> content=null -> the OVH tier returned nothing
and the cascade always skipped Tier-2. Floor the OVH budget to >=2000, fall back
to reasoning_content when content is null, and raise the client timeout to 90s
for the slower reasoning path.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

backend-compliance/compliance/services/llm_cascade.py

@@ -142,19 +142,26 @@ async def _call_ovh(system: str, user: str, max_tokens: int = 6000) -> str:
     headers = {"Content-Type": "application/json"}
     if key:
         headers["Authorization"] = f"Bearer {key}"
+    # gpt-oss-120b is a REASONING model: it spends output tokens on
+    # chain-of-thought before emitting the answer. A low cap (e.g. deep_check's
+    # max_tokens=400) makes it hit the length limit mid-reasoning and return
+    # content=null — the whole tier then silently yields nothing. Floor the
+    # budget so the reasoning AND the JSON answer fit.
     payload = {
-        "model": model, "temperature": 0.05, "max_tokens": max_tokens,
+        "model": model, "temperature": 0.05, "max_tokens": max(max_tokens, 2000),
         "messages": [{"role": "system", "content": system},
                      {"role": "user", "content": user}],
         "response_format": {"type": "json_object"},
     }
     try:
-        async with httpx.AsyncClient(timeout=45.0) as c:
+        async with httpx.AsyncClient(timeout=90.0) as c:
             r = await c.post(f"{base.rstrip('/')}/v1/chat/completions",
                              json=payload, headers=headers)
             r.raise_for_status()
-        choice = (r.json().get("choices") or [{}])[0]
-        return (choice.get("message") or {}).get("content", "") or ""
+        msg = (r.json().get("choices") or [{}])[0].get("message") or {}
+        # Answer is normally in content; if the model was length-capped the
+        # JSON can land in reasoning_content instead — fall back to it.
+        return (msg.get("content") or "") or (msg.get("reasoning_content") or "")
     except Exception as e:
         logger.warning("ovh cascade tier 2 failed: %s", e)
         return ""

backend-compliance/tests/test_llm_cascade_ovh.py

@@ -0,0 +1,77 @@
+"""Regression tests for the OVH (gpt-oss-120b) tier of the LLM cascade.
+
+gpt-oss-120b is a reasoning model: it spends output tokens on chain-of-thought
+before the answer. Two bugs this pins:
+  1. A small max_tokens (deep_check passed 400) length-caps it mid-reasoning →
+     content=null → the tier silently returns nothing. _call_ovh must floor the
+     budget so reasoning + the JSON answer fit.
+  2. When length-capped, the JSON can land in reasoning_content, not content →
+     _call_ovh must fall back to reasoning_content.
+"""
+import pytest
+from unittest.mock import AsyncMock, MagicMock, patch
+
+from compliance.services import llm_cascade
+
+
+def _resp(data):
+    r = MagicMock()
+    r.raise_for_status = MagicMock()
+    r.json = MagicMock(return_value=data)
+    return r
+
+
+def _client(resp):
+    inst = AsyncMock()
+    inst.post.return_value = resp
+    inst.__aenter__ = AsyncMock(return_value=inst)
+    inst.__aexit__ = AsyncMock(return_value=False)
+    return inst
+
+
+class TestCallOvhReasoning:
+    @pytest.mark.asyncio
+    async def test_reasoning_content_used_when_content_null(self, monkeypatch):
+        monkeypatch.setenv("OVH_LLM_URL", "https://llm.example.com")
+        monkeypatch.setenv("OVH_LLM_MODEL", "gpt-oss-120b")
+        monkeypatch.setenv("OVH_LLM_KEY", "k")
+        resp = _resp({"choices": [{"message": {
+            "content": None,
+            "reasoning_content": '{"erfuellt": true, "confidence": 0.9}'}}]})
+        with patch("httpx.AsyncClient", return_value=_client(resp)):
+            out = await llm_cascade._call_ovh("sys", "user", max_tokens=400)
+        assert '"erfuellt": true' in out
+
+    @pytest.mark.asyncio
+    async def test_small_budget_is_floored(self, monkeypatch):
+        monkeypatch.setenv("OVH_LLM_URL", "https://llm.example.com")
+        monkeypatch.setenv("OVH_LLM_MODEL", "gpt-oss-120b")
+        inst = _client(_resp({"choices": [{"message": {"content": "{}"}}]}))
+        with patch("httpx.AsyncClient", return_value=inst):
+            await llm_cascade._call_ovh("sys", "user", max_tokens=400)
+        assert inst.post.call_args.kwargs["json"]["max_tokens"] >= 2000
+
+    @pytest.mark.asyncio
+    async def test_large_budget_is_preserved(self, monkeypatch):
+        monkeypatch.setenv("OVH_LLM_URL", "https://llm.example.com")
+        monkeypatch.setenv("OVH_LLM_MODEL", "gpt-oss-120b")
+        inst = _client(_resp({"choices": [{"message": {"content": "{}"}}]}))
+        with patch("httpx.AsyncClient", return_value=inst):
+            await llm_cascade._call_ovh("sys", "user", max_tokens=6000)
+        assert inst.post.call_args.kwargs["json"]["max_tokens"] == 6000
+
+    @pytest.mark.asyncio
+    async def test_content_preferred_when_present(self, monkeypatch):
+        monkeypatch.setenv("OVH_LLM_URL", "https://llm.example.com")
+        monkeypatch.setenv("OVH_LLM_MODEL", "gpt-oss-120b")
+        resp = _resp({"choices": [{"message": {
+            "content": '{"erfuellt": false}', "reasoning_content": "noise"}}]})
+        with patch("httpx.AsyncClient", return_value=_client(resp)):
+            out = await llm_cascade._call_ovh("sys", "user")
+        assert out == '{"erfuellt": false}'
+
+    @pytest.mark.asyncio
+    async def test_unconfigured_returns_empty(self, monkeypatch):
+        monkeypatch.delenv("OVH_LLM_URL", raising=False)
+        monkeypatch.delenv("OVH_LLM_MODEL", raising=False)
+        assert await llm_cascade._call_ovh("sys", "user") == ""