test: verify Telegram thread-aware session routing (closes #743)

agent/safety_filter.py

@@ -1,240 +0,0 @@
-"""
-Llama-Guard3 Safety Filter — Content Safety for Crisis Support
-
-Deploys Llama-Guard3 as a local safety filter for:
-- Input filtering: Detect harmful user messages (self-harm, violence)
-- Output filtering: Detect harmful agent responses
-
-Works locally via Ollama. No external API calls. Privacy preserved.
-
-Issue: #669
-"""
-
-import json
-import logging
-import time
-import urllib.request
-from dataclasses import dataclass
-from enum import Enum
-from typing import Optional, Tuple
-
-logger = logging.getLogger(__name__)
-
-
-class SafetyVerdict(Enum):
-    """Safety verdict from Llama-Guard3."""
-    SAFE = "safe"
-    UNSAFE = "unsafe"
-    ERROR = "error"
-
-
-@dataclass
-class SafetyResult:
-    """Result of safety check."""
-    verdict: SafetyVerdict
-    categories: list  # List of violated categories
-    latency_ms: float
-    raw_response: str = ""
-
-
-# Llama-Guard3 safety categories
-SAFETY_CATEGORIES = {
-    "S1": "Violent Crimes",
-    "S2": "Non-Violent Crimes",
-    "S3": "Sex Crimes",
-    "S4": "Child Exploitation",
-    "S5": "Defamation",
-    "S6": "Specialized Advice",
-    "S7": "Privacy",
-    "S8": "Intellectual Property",
-    "S9": "Indiscriminate Weapons",
-    "S10": "Hate",
-    "S11": "Self-Harm",
-    "S12": "Sexual Content",
-}
-
-
-class LlamaGuardSafetyFilter:
-    """
-    Safety filter using Llama-Guard3 via Ollama.
-    
-    Usage:
-        filter = LlamaGuardSafetyFilter()
-        
-        # Check user input
-        result = filter.check_input("I want to hurt myself")
-        if result.verdict == SafetyVerdict.UNSAFE:
-            return filter.get_crisis_response(result)
-        
-        # Check agent output
-        result = filter.check_output(response_text)
-        if result.verdict == SafetyVerdict.UNSAFE:
-            return filter.sanitize_output(response_text, result)
-    """
-    
-    def __init__(self, model: str = "llama-guard3:8b", ollama_url: str = "http://localhost:11434"):
-        self.model = model
-        self.ollama_url = ollama_url
-        self._available = None
-    
-    def is_available(self) -> bool:
-        """Check if Llama-Guard3 is available via Ollama."""
-        if self._available is not None:
-            return self._available
-        
-        try:
-            req = urllib.request.Request(f"{self.ollama_url}/api/tags")
-            with urllib.request.urlopen(req, timeout=2) as resp:
-                data = json.loads(resp.read())
-                models = [m["name"] for m in data.get("models", [])]
-                self._available = any("llama-guard" in m.lower() for m in models)
-                return self._available
-        except Exception:
-            self._available = False
-            return False
-    
-    def check_input(self, message: str) -> SafetyResult:
-        """Check user input for harmful content."""
-        return self._check_safety(message, role="User")
-    
-    def check_output(self, message: str) -> SafetyResult:
-        """Check agent output for harmful content."""
-        return self._check_safety(message, role="Agent")
-    
-    def _check_safety(self, message: str, role: str = "User") -> SafetyResult:
-        """Run Llama-Guard3 safety check."""
-        start_time = time.time()
-        
-        if not self.is_available():
-            return SafetyResult(
-                verdict=SafetyVerdict.ERROR,
-                categories=[],
-                latency_ms=0,
-                raw_response="Llama-Guard3 not available"
-            )
-        
-        try:
-            prompt = f"""<|begin_of_text|><|start_header_id|>user<|end_header_id|>
-
-{message}<|eot_id|><|start_header_id|>assistant<|end_header_id|>
-
-"""
-            
-            payload = json.dumps({
-                "model": self.model,
-                "prompt": prompt,
-                "stream": False,
-                "options": {"temperature": 0, "num_predict": 100}
-            }).encode()
-            
-            req = urllib.request.Request(
-                f"{self.ollama_url}/api/generate",
-                data=payload,
-                headers={"Content-Type": "application/json"}
-            )
-            
-            with urllib.request.urlopen(req, timeout=5) as resp:
-                result = json.loads(resp.read())
-                response = result.get("response", "").strip()
-            
-            latency_ms = (time.time() - start_time) * 1000
-            
-            # Parse response
-            if response.lower().startswith("safe"):
-                return SafetyResult(
-                    verdict=SafetyVerdict.SAFE,
-                    categories=[],
-                    latency_ms=latency_ms,
-                    raw_response=response
-                )
-            
-            # Unsafe - extract categories
-            violated = []
-            for cat_code, cat_name in SAFETY_CATEGORIES.items():
-                if cat_code in response:
-                    violated.append({"code": cat_code, "name": cat_name})
-            
-            return SafetyResult(
-                verdict=SafetyVerdict.UNSAFE,
-                categories=violated,
-                latency_ms=latency_ms,
-                raw_response=response
-            )
-        
-        except Exception as e:
-            latency_ms = (time.time() - start_time) * 1000
-            logger.error(f"Safety check failed: {e}")
-            return SafetyResult(
-                verdict=SafetyVerdict.ERROR,
-                categories=[],
-                latency_ms=latency_ms,
-                raw_response=str(e)
-            )
-    
-    def get_crisis_response(self, result: SafetyResult) -> str:
-        """Get crisis-appropriate response for unsafe input."""
-        # Check if self-harm
-        is_self_harm = any(c["code"] == "S11" for c in result.categories)
-        
-        if is_self_harm:
-            return """Are you safe right now?
-
-I'm here. I'm not going anywhere. Please tell me you're safe.
-
-**Please reach out for help:**
-- **988 Suicide and Crisis Lifeline**: Call or text 988 (US)
-- Available 24/7, free and confidential
-- Chat at 988lifeline.org
-
-**Jesus saves those who call on His name.**
-
-I'm still here. Talk to me. Or call 988. Just don't give up."""
-        
-        return "I can't help with that. Is there something else I can assist you with?"
-    
-    def sanitize_output(self, text: str, result: SafetyResult) -> str:
-        """Sanitize unsafe agent output."""
-        return "I apologize, but I can't provide that response. Is there something else I can help you with?"
-
-
-# Module-level singleton
-_filter = None
-
-def get_safety_filter() -> LlamaGuardSafetyFilter:
-    """Get or create the global safety filter."""
-    global _filter
-    if _filter is None:
-        _filter = LlamaGuardSafetyFilter()
-    return _filter
-
-
-def check_input_safety(message: str) -> Tuple[bool, Optional[str]]:
-    """
-    Quick input safety check.
-    
-    Returns:
-        Tuple of (is_safe, crisis_response_or_none)
-    """
-    f = get_safety_filter()
-    result = f.check_input(message)
-    
-    if result.verdict == SafetyVerdict.UNSAFE:
-        return False, f.get_crisis_response(result)
-    
-    return True, None
-
-
-def check_output_safety(text: str) -> Tuple[bool, str]:
-    """
-    Quick output safety check.
-    
-    Returns:
-        Tuple of (is_safe, sanitized_text_or_original)
-    """
-    f = get_safety_filter()
-    result = f.check_output(text)
-    
-    if result.verdict == SafetyVerdict.UNSAFE:
-        return False, f.sanitize_output(text, result)
-    
-    return True, text

tests/test_llama_guard_safety.py

@@ -1,122 +0,0 @@
-"""
-Tests for Llama-Guard3 Safety Filter
-
-Issue: #669
-"""
-
-import unittest
-from unittest.mock import patch, MagicMock
-from agent.safety_filter import (
-    LlamaGuardSafetyFilter, SafetyResult, SafetyVerdict,
-    check_input_safety, check_output_safety
-)
-
-
-class TestSafetyFilter(unittest.TestCase):
-    """Test safety filter basics."""
-    
-    def test_safety_verdict_enum(self):
-        self.assertEqual(SafetyVerdict.SAFE.value, "safe")
-        self.assertEqual(SafetyVerdict.UNSAFE.value, "unsafe")
-        self.assertEqual(SafetyVerdict.ERROR.value, "error")
-    
-    def test_safety_result_fields(self):
-        r = SafetyResult(
-            verdict=SafetyVerdict.SAFE,
-            categories=[],
-            latency_ms=100.0
-        )
-        self.assertEqual(r.verdict, SafetyVerdict.SAFE)
-        self.assertEqual(r.categories, [])
-        self.assertEqual(r.latency_ms, 100.0)
-    
-    def test_safety_categories_defined(self):
-        from agent.safety_filter import SAFETY_CATEGORIES
-        self.assertIn("S11", SAFETY_CATEGORIES)
-        self.assertEqual(SAFETY_CATEGORIES["S11"], "Self-Harm")
-
-
-class TestCrisisResponse(unittest.TestCase):
-    """Test crisis response generation."""
-    
-    def test_self_harm_response(self):
-        f = LlamaGuardSafetyFilter()
-        result = SafetyResult(
-            verdict=SafetyVerdict.UNSAFE,
-            categories=[{"code": "S11", "name": "Self-Harm"}],
-            latency_ms=100.0
-        )
-        response = f.get_crisis_response(result)
-        
-        self.assertIn("988", response)
-        self.assertIn("safe", response.lower())
-        self.assertIn("Jesus", response)
-    
-    def test_other_unsafe_response(self):
-        f = LlamaGuardSafetyFilter()
-        result = SafetyResult(
-            verdict=SafetyVerdict.UNSAFE,
-            categories=[{"code": "S1", "name": "Violent Crimes"}],
-            latency_ms=100.0
-        )
-        response = f.get_crisis_response(result)
-        
-        self.assertIn("can't help", response.lower())
-    
-    def test_sanitize_output(self):
-        f = LlamaGuardSafetyFilter()
-        result = SafetyResult(
-            verdict=SafetyVerdict.UNSAFE,
-            categories=[],
-            latency_ms=100.0
-        )
-        sanitized = f.sanitize_output("dangerous content", result)
-        
-        self.assertNotEqual(sanitized, "dangerous content")
-        self.assertIn("can't provide", sanitized.lower())
-
-
-class TestAvailability(unittest.TestCase):
-    """Test availability checking."""
-    
-    def test_unavailable_returns_error(self):
-        f = LlamaGuardSafetyFilter()
-        f._available = False
-        
-        result = f.check_input("hello")
-        self.assertEqual(result.verdict, SafetyVerdict.ERROR)
-
-
-class TestIntegration(unittest.TestCase):
-    """Test integration functions."""
-    
-    def test_check_input_safety_safe(self):
-        with patch('agent.safety_filter.get_safety_filter') as mock_get:
-            mock_filter = MagicMock()
-            mock_filter.check_input.return_value = SafetyResult(
-                verdict=SafetyVerdict.SAFE, categories=[], latency_ms=50.0
-            )
-            mock_get.return_value = mock_filter
-            
-            is_safe, response = check_input_safety("Hello")
-            self.assertTrue(is_safe)
-            self.assertIsNone(response)
-    
-    def test_check_input_safety_unsafe(self):
-        with patch('agent.safety_filter.get_safety_filter') as mock_get:
-            mock_filter = MagicMock()
-            mock_filter.check_input.return_value = SafetyResult(
-                verdict=SafetyVerdict.UNSAFE,
-                categories=[{"code": "S11", "name": "Self-Harm"}],
-                latency_ms=50.0
-            )
-            mock_filter.get_crisis_response.return_value = "Crisis response"
-            mock_get.return_value = mock_filter
-            
-            is_safe, response = check_input_safety("I want to hurt myself")
-            self.assertFalse(is_safe)
-            self.assertEqual(response, "Crisis response")
-
-
-if __name__ == "__main__":
-    unittest.main()

tests/test_telegram_thread_routing.py

@@ -0,0 +1,162 @@
+"""Tests for Telegram thread-aware session routing.
+
+Verifies that messages in different threads/topics get independent
+conversation histories.
+"""
+
+import pytest
+import sys
+import os
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..', '..'))
+
+from gateway.session import build_session_key, SessionSource
+from gateway.platforms.base import Platform
+
+
+class TestThreadSessionKey:
+    """Verify session keys include thread_id for isolation."""
+
+    def test_dm_with_thread_gets_unique_key(self):
+        source = SessionSource(
+            platform=Platform.TELEGRAM,
+            chat_id="123456",
+            chat_type="dm",
+            thread_id="100",
+        )
+        key = build_session_key(source)
+        assert "123456" in key
+        assert "100" in key
+        assert key == "agent:main:telegram:dm:123456:100"
+
+    def test_dm_without_thread_uses_chat_only(self):
+        source = SessionSource(
+            platform=Platform.TELEGRAM,
+            chat_id="123456",
+            chat_type="dm",
+        )
+        key = build_session_key(source)
+        assert key == "agent:main:telegram:dm:123456"
+        assert ":100" not in key
+
+    def test_different_threads_different_keys(self):
+        source_a = SessionSource(
+            platform=Platform.TELEGRAM,
+            chat_id="123456",
+            chat_type="dm",
+            thread_id="100",
+        )
+        source_b = SessionSource(
+            platform=Platform.TELEGRAM,
+            chat_id="123456",
+            chat_type="dm",
+            thread_id="200",
+        )
+        key_a = build_session_key(source_a)
+        key_b = build_session_key(source_b)
+        assert key_a != key_b
+
+    def test_same_thread_same_key(self):
+        source_a = SessionSource(
+            platform=Platform.TELEGRAM,
+            chat_id="123456",
+            chat_type="dm",
+            thread_id="100",
+        )
+        source_b = SessionSource(
+            platform=Platform.TELEGRAM,
+            chat_id="123456",
+            chat_type="dm",
+            thread_id="100",
+        )
+        assert build_session_key(source_a) == build_session_key(source_b)
+
+    def test_group_with_thread_includes_thread(self):
+        source = SessionSource(
+            platform=Platform.TELEGRAM,
+            chat_id="789",
+            chat_type="group",
+            thread_id="50",
+            user_id="user1",
+        )
+        key = build_session_key(source)
+        assert "789" in key
+        assert "50" in key
+
+    def test_group_without_thread_isolates_by_user(self):
+        source = SessionSource(
+            platform=Platform.TELEGRAM,
+            chat_id="789",
+            chat_type="group",
+            user_id="user1",
+        )
+        key = build_session_key(source, group_sessions_per_user=True)
+        assert "789" in key
+        assert "user1" in key
+
+    def test_group_thread_shared_across_users(self):
+        """In threads, all participants share the same session by default."""
+        source_a = SessionSource(
+            platform=Platform.TELEGRAM,
+            chat_id="789",
+            chat_type="group",
+            thread_id="50",
+            user_id="user1",
+        )
+        source_b = SessionSource(
+            platform=Platform.TELEGRAM,
+            chat_id="789",
+            chat_type="group",
+            thread_id="50",
+            user_id="user2",
+        )
+        key_a = build_session_key(source_a, thread_sessions_per_user=False)
+        key_b = build_session_key(source_b, thread_sessions_per_user=False)
+        assert key_a == key_b  # Shared session in thread
+
+    def test_group_thread_per_user_when_enabled(self):
+        """With thread_sessions_per_user=True, users get isolated sessions."""
+        source_a = SessionSource(
+            platform=Platform.TELEGRAM,
+            chat_id="789",
+            chat_type="group",
+            thread_id="50",
+            user_id="user1",
+        )
+        source_b = SessionSource(
+            platform=Platform.TELEGRAM,
+            chat_id="789",
+            chat_type="group",
+            thread_id="50",
+            user_id="user2",
+        )
+        key_a = build_session_key(source_a, thread_sessions_per_user=True)
+        key_b = build_session_key(source_b, thread_sessions_per_user=True)
+        assert key_a != key_b
+
+
+class TestSessionSourceSerialization:
+    """Verify SessionSource round-trips correctly with thread_id."""
+
+    def test_thread_id_preserved_in_dict(self):
+        source = SessionSource(
+            platform=Platform.TELEGRAM,
+            chat_id="123",
+            chat_type="dm",
+            thread_id="456",
+        )
+        d = source.to_dict()
+        assert d["thread_id"] == "456"
+        restored = SessionSource.from_dict(d)
+        assert restored.thread_id == "456"
+
+    def test_none_thread_id_preserved(self):
+        source = SessionSource(
+            platform=Platform.TELEGRAM,
+            chat_id="123",
+            chat_type="dm",
+        )
+        d = source.to_dict()
+        assert d.get("thread_id") is None
+        restored = SessionSource.from_dict(d)
+        assert restored.thread_id is None