Merge pull request #530 from MaEscalanteHe/feature/update-anthropic-models

fix: update Anthropic models and handle thinking/refusal responses

AppImage/config/verified_ai_models.json

@@ -63,18 +63,21 @@
   },
   "anthropic": {
     "models": [
+      "claude-haiku-5-5",
       "claude-haiku-4-5-20251001",
-      "claude-opus-4-8",
-      "claude-sonnet-4-5-20250929",
-      "claude-opus-4-5-20251101",
-      "claude-opus-4-7",
+      "claude-sonnet-5-5",
+      "claude-sonnet-5",
       "claude-sonnet-4-6",
+      "claude-opus-5-5",
+      "claude-opus-4-8",
+      "claude-opus-4-7",
+      "claude-opus-4-6",
+      "claude-opus-4-5-20251101",
       "claude-fable-5",
-      "claude-fable-5-1",
-      "claude-opus-4-6"
+      "claude-fable-5-1"
     ],
-    "recommended": "claude-haiku-4-5-20251001",
-    "_note": "Verified 2026-09-02 — 9 of 11 tested pass. The `claude-haiku-4-5` alias no longer resolves upstream; the working snapshot is `claude-haiku-4-5-20251001` — used as recommended (2.89s). Full family surfaced: haiku/sonnet/opus/fable across generations 4.5/4.6/4.7/4.8/5."
+    "recommended": "claude-haiku-5-5",
+    "_note": "Verified 2026-10-26 against the live API. Added the 5.x family, recommended `claude-haiku-5-5`. Dropped `claude-sonnet-4-5` (retires 2026-11-30). `claude-opus-5` left out: false refusals on plain notifications."
   },
   "openrouter": {
     "models": [

AppImage/scripts/ai_providers/anthropic_provider.py

@@ -1,7 +1,7 @@
 """Anthropic (Claude) provider implementation.
 
 Anthropic's Claude models are excellent for text generation and translation.
-Models use "-latest" aliases that auto-update to newest versions.
+Models are addressed by their pinned IDs (the "-latest" aliases were retired).
 """
 from typing import Optional, List
 from .base import AIProvider, AIProviderError
@@ -15,29 +15,35 @@ class AnthropicProvider(AIProvider):
     API_URL = "https://api.anthropic.com/v1/messages"
     API_VERSION = "2023-06-01"
     
-    # Anthropic model aliases that resolve to pinned snapshots (per docs
-    # at platform.claude.com/docs/en/docs/about-claude/models). Kept as a
-    # runtime fallback for the `/api/notifications/provider-models`
-    # intersection with `verified_ai_models.json`; the JSON is the
-    # authoritative UI-facing list. Refreshed 2026-07-13 to drop the
-    # deprecated claude-3-5-* aliases (retired upstream, no longer in the
-    # Models API response) and add the current-generation IDs. Ordered
-    # cheapest → most capable so the first entry is a safe default.
+    # Current Claude API model IDs (per docs at
+    # platform.claude.com/docs/en/docs/about-claude/models). Kept as a
+    # runtime fallback for `/api/notifications/provider-models` when
+    # `verified_ai_models.json` has no Anthropic entry; the JSON is the
+    # authoritative UI-facing list. Refreshed 2026-10-26 against the live
+    # Models API. Ordered cheapest → most capable so the first entry is a
+    # safe default.
     KNOWN_MODELS = [
-        "claude-haiku-4-5",
+        "claude-haiku-5-5",
+        "claude-haiku-4-5-20251001",
+        "claude-sonnet-5-5",
         "claude-sonnet-5",
         "claude-sonnet-4-6",
+        "claude-opus-5-5",
         "claude-opus-4-8",
+        "claude-opus-4-7",
         "claude-opus-4-6",
+        "claude-opus-4-5-20251101",
         "claude-fable-5",
+        "claude-fable-5-1",
     ]
+
+    # Newer models (Opus 5.x, Fable) may think before answering and the
+    # thinking counts against max_tokens, so a small cap can end the turn
+    # with no text at all. max_tokens is a ceiling, not a cost.
+    MIN_MAX_TOKENS = 1024
     
     def list_models(self) -> List[str]:
-        """Return known Anthropic model aliases.
-        
-        Anthropic doesn't have a public models list API, but their "-latest"
-        aliases auto-update to the newest versions, making them reliable choices.
-        """
+        """Return known Anthropic model IDs."""
         return self.KNOWN_MODELS
     
     def generate(self, system_prompt: str, user_message: str,
@@ -68,7 +74,7 @@ class AnthropicProvider(AIProvider):
             'messages': [
                 {'role': 'user', 'content': user_message},
             ],
-            'max_tokens': max_tokens,
+            'max_tokens': max(max_tokens, self.MIN_MAX_TOKENS),
         }
         
         headers = {
@@ -79,11 +85,20 @@ class AnthropicProvider(AIProvider):
         
         result = self._make_request(self.API_URL, payload, headers)
         
-        try:
-            # Anthropic returns content as array of content blocks
-            content = result['content']
-            if isinstance(content, list) and len(content) > 0:
-                return content[0].get('text', '').strip()
-            return str(content).strip()
-        except (KeyError, IndexError) as e:
-            raise AIProviderError(f"Unexpected response format: {e}")
+        # A safety classifier can decline even benign notifications
+        # (HTTP 200, stop_reason "refusal", empty content). Raising lets the
+        # caller fall back to the original, untranslated notification.
+        if result.get('stop_reason') == 'refusal':
+            category = (result.get('stop_details') or {}).get('category')
+            raise AIProviderError(f"Request declined by Anthropic (category: {category})")
+        
+        # Content is a list of blocks; thinking blocks may come before the
+        # text, so take the first text block rather than content[0].
+        content = result.get('content')
+        if not isinstance(content, list):
+            raise AIProviderError("Unexpected response format: missing content")
+        for block in content:
+            if isinstance(block, dict) and block.get('type') == 'text':
+                return block.get('text', '').strip()
+        raise AIProviderError(
+            f"No text in response (stop_reason: {result.get('stop_reason')})")

AppImage/scripts/flask_notification_routes.py

@@ -514,13 +514,10 @@ def get_provider_models():
                     'message': f'Found {len(models)} local models'
                 })
         
-        # Handle Anthropic - no models list API, return verified models directly
+        # Handle Anthropic - return verified models directly
         if provider == 'anthropic':
-            models = list(verified_models) if verified_models else [
-                'claude-3-5-haiku-latest',
-                'claude-3-5-sonnet-latest',
-                'claude-3-opus-latest',
-            ]
+            from ai_providers.anthropic_provider import AnthropicProvider
+            models = list(verified_models) if verified_models else list(AnthropicProvider.KNOWN_MODELS)
             return jsonify({
                 'success': True,
                 'models': sorted(models),

AppImage/scripts/tests/test_anthropic_provider.py

@@ -0,0 +1,65 @@
+import sys
+import unittest
+from pathlib import Path
+
+
+SCRIPTS_DIR = Path(__file__).resolve().parents[1]
+if str(SCRIPTS_DIR) not in sys.path:
+    sys.path.insert(0, str(SCRIPTS_DIR))
+
+from ai_providers.anthropic_provider import AnthropicProvider
+from ai_providers.base import AIProviderError
+
+
+class CannedAnthropicProvider(AnthropicProvider):
+    def __init__(self, response):
+        super().__init__(api_key="test-key", model="claude-opus-5-5")
+        self.response = response
+        self.captured_payload = None
+
+    def _make_request(self, url, payload, headers):
+        self.captured_payload = payload
+        return self.response
+
+
+class AnthropicProviderTests(unittest.TestCase):
+    def test_skips_thinking_block_before_text(self):
+        provider = CannedAnthropicProvider({
+            "stop_reason": "end_turn",
+            "content": [
+                {"type": "thinking", "thinking": "", "signature": "x"},
+                {"type": "text", "text": "  Hola  "},
+            ],
+        })
+        self.assertEqual(provider.generate("system", "user"), "Hola")
+
+    def test_thinking_only_response_raises(self):
+        provider = CannedAnthropicProvider({
+            "stop_reason": "max_tokens",
+            "content": [{"type": "thinking", "thinking": "", "signature": "x"}],
+        })
+        with self.assertRaises(AIProviderError):
+            provider.generate("system", "user")
+
+    def test_refusal_with_empty_content_raises(self):
+        provider = CannedAnthropicProvider({
+            "stop_reason": "refusal",
+            "stop_details": {"type": "refusal", "category": "cyber"},
+            "content": [],
+        })
+        with self.assertRaisesRegex(AIProviderError, "cyber"):
+            provider.generate("system", "user")
+
+    def test_small_max_tokens_is_raised_to_floor(self):
+        provider = CannedAnthropicProvider({
+            "stop_reason": "end_turn",
+            "content": [{"type": "text", "text": "CONNECTION_OK"}],
+        })
+        provider.generate("system", "user", max_tokens=50)
+        self.assertEqual(provider.captured_payload["max_tokens"], AnthropicProvider.MIN_MAX_TOKENS)
+        provider.generate("system", "user", max_tokens=4000)
+        self.assertEqual(provider.captured_payload["max_tokens"], 4000)
+
+
+if __name__ == "__main__":
+    unittest.main()