A model's safety-filter refusal is an error that says so, not an empty reply

Opus's provider filter declined one of nisse-review's ask_expert
questions (which repositories a Gitea token can list) every time. It came
back as a 200 with no text, and the expert reported "unparseable reply:
empty reply". Both LLM clients now raise ModelRefused for
finish_reason content_filter (chat route) and stop_reason refusal
(passthrough). The expert's handback says what happened and what to do
instead.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01J5Utyfzw7gF7UhjRDEtjTk

nisse-core/src/nisse_core/loop/expert.py

@@ -32,7 +32,7 @@ from nisse_core.log.canonical import (
     StoredEvent,
     UserMessage,
 )
-from nisse_core.models.llm import LLMClient, LLMError, RoleMap
+from nisse_core.models.llm import LLMClient, LLMError, ModelRefused, RoleMap
 from nisse_core.state.store import Transcript
 
 EXPERT_ROLE = "expert"
@@ -168,6 +168,9 @@ class ExpertDesk:
                     model=model,
                     system=system,
                 )
+            except ModelRefused as exc:
+                # Not retried: the same question is declined again.
+                return None, f"{exc}. Ask it differently, or work from the repository"
             except LLMError as exc:
                 return None, str(exc)
             text = "\n".join(

nisse-core/src/nisse_core/models/llm.py

@@ -102,6 +102,24 @@ class TruncatedResponse(LLMError):
     """
 
 
+class ModelRefused(LLMError):
+    """The provider's safety filter declined to answer.
+
+    It comes back as a successful reply with no text (``finish_reason:
+    content_filter`` on the chat route, ``stop_reason: refusal`` on the
+    passthrough), so unchecked it reads as an empty answer. It can be a false
+    positive: Opus declined, every time, a question about which repositories
+    a Gitea token can list. Asking the same thing again gets the same answer.
+    """
+
+
+def _refused(model: str) -> ModelRefused:
+    return ModelRefused(
+        f"{model} declined to answer: its provider's safety filter stopped the "
+        "reply. That says nothing about the subject itself"
+    )
+
+
 class LLMClient(Protocol):
     def complete(
         self,
@@ -216,6 +234,8 @@ class LiteLLMClient:
                 f"{requested_model} hit its output limit mid-answer; raise "
                 "llm.max_output_tokens, or the model's max_output_tokens in LiteLLM"
             )
+        if choice.get("finish_reason") == "content_filter":
+            raise _refused(requested_model)
         produced_by = data.get("model") or requested_model
         # What actually answered, which is not always what was asked for: a
         # proxy fallback answers under another model's name.
@@ -338,6 +358,8 @@ class AnthropicPassthroughClient:
                 f"{requested_model} hit its output limit mid-answer; raise "
                 "llm.max_output_tokens, or the model's max_output_tokens in LiteLLM"
             )
+        if data.get("stop_reason") == "refusal":
+            raise _refused(requested_model)
         content = data.get("content")
         if not isinstance(content, list):
             raise LLMError(f"malformed passthrough response: {str(data)[:300]}")