diff --git a/engine/codeflow_engine/actions/quality_engine/ai/ai_analyzer.py b/engine/codeflow_engine/actions/quality_engine/ai/ai_analyzer.py
index b4c2368..5a970d8 100644
--- a/engine/codeflow_engine/actions/quality_engine/ai/ai_analyzer.py
+++ b/engine/codeflow_engine/actions/quality_engine/ai/ai_analyzer.py
@@ -135,8 +135,6 @@ async def analyze_code(
             ],
             "temperature": 0.3,  # Lower temperature for more deterministic responses
             "max_tokens": 2000,
-            # ✅ Validated and enforced JSON output per provider
-            "response_format": {"type": "json"},
         }
         if provider_name:
             request["provider"] = provider_name
@@ -149,7 +147,7 @@ async def analyze_code(
                 file_path=file_path,
                 provider=provider_name or "default",
             )
-            response = self.llm_manager.complete(request)
+            response = await asyncio.to_thread(self.llm_manager.complete, request)
 
             # LLMResponse is a dataclass and so always truthy; a failed call carries
             # `error` with empty content. Checking truthiness alone would send "" on
diff --git a/engine/codeflow_engine/actions/quality_engine/ai/ai_handler.py b/engine/codeflow_engine/actions/quality_engine/ai/ai_handler.py
index 5aa6323..aac64ac 100644
--- a/engine/codeflow_engine/actions/quality_engine/ai/ai_handler.py
+++ b/engine/codeflow_engine/actions/quality_engine/ai/ai_handler.py
@@ -5,12 +5,16 @@
 """
 
 import time
-from typing import Any
+from typing import TYPE_CHECKING, Any
 
 import structlog
 
 from codeflow_engine.actions.quality_engine.models import ToolResult
 
+
+if TYPE_CHECKING:
+    from codeflow_engine.actions.llm.manager import ActionLLMProviderManager
+
 logger = structlog.get_logger(__name__)
 
 
@@ -206,7 +210,7 @@ async def run_ai_analysis(
         return None
 
 
-async def initialize_llm_manager() -> Any | None:
+async def initialize_llm_manager() -> "ActionLLMProviderManager | None":
     """Initialize the LLM manager for AI analysis.
 
     Delegates to the quality engine's single initializer rather than building a
diff --git a/engine/codeflow_engine/actions/quality_engine/ai/ai_modes.py b/engine/codeflow_engine/actions/quality_engine/ai/ai_modes.py
index 8ebc039..21eefc2 100644
--- a/engine/codeflow_engine/actions/quality_engine/ai/ai_modes.py
+++ b/engine/codeflow_engine/actions/quality_engine/ai/ai_modes.py
@@ -5,8 +5,9 @@
 integrating with the CodeFlow LLM provider system.
 """
 
-import json
+import asyncio
 from enum import StrEnum
+import json
 from pathlib import Path
 from typing import Any
 
@@ -18,6 +19,7 @@
 from codeflow_engine.actions.quality_engine.models import ToolResult
 from codeflow_engine.agents.models import CodeIssue
 
+
 logger = structlog.get_logger(__name__)
 
 # System prompt templates
@@ -167,17 +169,19 @@ async def run_ai_analysis(
                 {"role": "user", "content": analysis_prompt},
             ],
             "temperature": 0.1,
-            "response_format": {"type": "json_object"},
         }
         if provider_name:
             request["provider"] = provider_name
         if model:
             request["model"] = model
 
-        response = llm_manager.complete(request)
+        response = await asyncio.to_thread(llm_manager.complete, request)
 
         if not response or not response.content:
-            logger.warning("No response received from LLM")
+            logger.warning(
+                "No response received from LLM",
+                error=getattr(response, "error", None),
+            )
             return None
 
         # Report what was actually used, not what was asked for — when either was
diff --git a/engine/tests/test_sluice_request_metadata.py b/engine/tests/test_sluice_request_metadata.py
index 725ac57..e7f4c36 100644
--- a/engine/tests/test_sluice_request_metadata.py
+++ b/engine/tests/test_sluice_request_metadata.py
@@ -1,6 +1,6 @@
 """Guards the Sluice request-metadata contract (ADR 10, raised to MUST by ADR 17).
 
-These tests assert the *contract* — what the gateway requires — not what the code
+These tests assert the *contract* — what the gateway requires — not what the code
 currently happens to send. That distinction is the reason this file exists.
 
 The same contract was silently violated for months in house-of-veritas: the call
@@ -12,7 +12,7 @@
 So: every assertion below is written against
 `phoenixvc/sluice` docs/architecture/17-mandatory-request-metadata.md. If an
 assertion here ever disagrees with the implementation, the implementation is what
-moves — unless the ADR itself changed, in which case update the citation too.
+moves — unless the ADR itself changed, in which case update the citation too.
 
 codeflow-engine's virtual key carries no `use:` in the gateway's keys.yaml, which
 classifies it as a *service*: enforced, not exempt.
@@ -40,7 +40,7 @@
 from codeflow_engine.core.llm.sluice_provider import SluiceProvider
 
 # ADR 10 naming rule, unchanged by ADR 17: `app` and `agent` become Prometheus label
-# values, so they must be lowercase kebab-case — stable, one token per dimension.
+# values, so they must be lowercase kebab-case — stable, one token per dimension.
 KEBAB_CASE = re.compile(r"^[a-z0-9]+(-[a-z0-9]+)*$")
 
 SLUICE_BASE_URL = "https://litellm.sluice.example"
@@ -68,7 +68,7 @@ def recorder(monkeypatch: pytest.MonkeyPatch) -> _CompletionsRecorder:
     """Install a fake `openai` module so tests assert on the request, not the network.
 
     Stubbing at the module boundary rather than patching the provider's `client`
-    attribute keeps `_initialize_client` in the code path — that is where the
+    attribute keeps `_initialize_client` in the code path — that is where the
     base_url and api_key actually get bound.
     """
     completions = _CompletionsRecorder()
@@ -144,7 +144,7 @@ def test_does_not_send_superseded_field_names(
 
 
 class TestNamingRules:
-    """ADR 10 naming rules — violations are counted, not rejected, so nothing else catches them."""
+    """ADR 10 naming rules — violations are counted, not rejected, so nothing else catches them."""
 
     def test_prometheus_labelled_fields_are_kebab_case(
         self, recorder: _CompletionsRecorder, sluice_env: None
@@ -199,7 +199,7 @@ def test_known_literal_strings_still_resolve(self) -> None:
 
 
 class TestUntaggedTrafficIsRefused:
-    """An untagged request is accepted by the gateway today — local failure is the only signal."""
+    """An untagged request is accepted by the gateway today — local failure is the only signal."""
 
     def test_sluice_route_without_an_agent_refuses_to_send(
         self, recorder: _CompletionsRecorder, sluice_env: None
@@ -300,7 +300,7 @@ def test_guard_ignores_path_and_trailing_slash_differences(
     ) -> None:
         from codeflow_engine.actions.llm.providers import OpenAIProvider
 
-        # Same host, cosmetically different URL — must not slip past the guard.
+        # Same host, cosmetically different URL — must not slip past the guard.
         with pytest.raises(SluiceMetadataError):
             OpenAIProvider(
                 {"api_key": "sk-test", "base_url": f"{SLUICE_BASE_URL}/v1/"}
@@ -848,6 +848,7 @@ async def test_analysis_requests_are_tagged(
             "app": "codeflow-engine",
             "agent": "quality-analyzer",
         }
+        assert "response_format" not in recorder.calls[0]
 
     @pytest.mark.asyncio
     async def test_does_not_ask_the_gateway_for_a_vendor_model(
@@ -970,3 +971,4 @@ async def test_the_code_analyzer_shares_the_same_tag(
         await AICodeAnalyzer(manager).analyze_code("sample.py", "x = 1\n")
 
         assert sent_metadata(recorder)["agent"] == "quality-analyzer"
+        assert "response_format" not in recorder.calls[0]