Make tone detection selectable as a category in the Content Filter UI

- Add categories/tone_detection.yaml with metadata so it appears in the
  ContentCategoryConfiguration dropdown alongside other categories
- Wire _load_categories() to detect tone_detection category and enable the
  ToneChecker engine (regex-based, response-only) instead of loading keywords
- Add tests: category-based enablement, blocking, and disabled state

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/categories/tone_detection.yaml

@@ -0,0 +1,24 @@
+# Tone Detection — CPU-only regex detection of inappropriate chatbot tone
+#
+# Detects 6 categories of inappropriate tone in customer-facing chatbot responses:
+#   - Dismissive language ("that's not my problem", "just read the FAQ")
+#   - Blaming the customer ("you should have read the terms", "that's your fault")
+#   - Refusal to help ("I can't help you", "figure it out yourself")
+#   - Sarcasm / condescension ("if you'd been paying attention", "let me spell it out")
+#   - Impatience / frustration ("I've already told you", "are you even listening")
+#   - Unprofessional casual language ("bruh", "lol", "sounds like a you problem")
+#
+# This category uses regex-based pattern matching (not keyword lists).
+# When selected, it enables the ToneChecker engine which runs on LLM responses only.
+#
+# Advanced configuration (safe_phrases, blocked_phrases) can be set via
+# tone_detection_config in litellm_params for programmatic use.
+
+category_name: "tone_detection"
+display_name: "Tone Detection"
+description: "Detects dismissive, blaming, condescending, impatient, or unprofessional tone in customer-facing chatbot responses"
+default_action: "BLOCK"
+
+# No keywords — tone detection uses the ToneChecker regex engine
+keywords: []
+exceptions: []

litellm/proxy/guardrails/guardrail_hooks/litellm_content_filter/content_filter.py

@@ -456,6 +456,16 @@ class ContentFilterGuardrail(CustomGuardrail):
                 )
                 continue
 
+            # Tone detection is a special category — it enables the ToneChecker
+            # regex engine rather than loading keyword lists.
+            if category_name == "tone_detection":
+                if self._tone_checker is None:
+                    self._init_tone_checker({})
+                    verbose_proxy_logger.info(
+                        "Loaded tone_detection category (ToneChecker enabled)"
+                    )
+                continue
+
             # Load category file (custom or default)
             if custom_file:
                 category_file_path = self._resolve_category_file_path(custom_file)

tests/test_litellm/proxy/guardrails/guardrail_hooks/test_tone_detector.py

@@ -425,6 +425,36 @@ class TestInit:
         g = ContentFilterGuardrail(guardrail_name="test-no-tone")
         assert g._tone_checker is None
 
+    def test_tone_checker_enabled_via_category(self):
+        """Selecting tone_detection as a category should enable the ToneChecker."""
+        g = ContentFilterGuardrail(
+            guardrail_name="test-category-tone",
+            categories=[{"category": "tone_detection", "enabled": True, "action": "BLOCK"}],
+        )
+        assert g._tone_checker is not None
+
+    @pytest.mark.asyncio
+    async def test_tone_detection_via_category_blocks(self):
+        """Tone detection enabled via category selection should block bad tone."""
+        g = ContentFilterGuardrail(
+            guardrail_name="test-category-tone",
+            categories=[{"category": "tone_detection", "enabled": True, "action": "BLOCK"}],
+        )
+        with pytest.raises(HTTPException):
+            await g.apply_guardrail(
+                _inputs("That's not my problem."),
+                {},
+                "response",
+            )
+
+    def test_tone_detection_category_disabled(self):
+        """Disabling the tone_detection category should NOT enable ToneChecker."""
+        g = ContentFilterGuardrail(
+            guardrail_name="test-category-disabled",
+            categories=[{"category": "tone_detection", "enabled": False, "action": "BLOCK"}],
+        )
+        assert g._tone_checker is None
+
     def test_invalid_regex_degrades_gracefully(self):
         """Invalid regex in blocked_phrases should degrade gracefully (tone checker disabled)."""
         g = _make_guardrail(blocked_phrases=[r"(unclosed"])