"""Tests for the unified restaurant agent: language detection, tool filtering.""" from __future__ import annotations from unittest.mock import MagicMock import pytest from pydantic_ai import ToolDefinition from app.agents.deps import AgentDeps, CallerIdentity from app.agents.restaurant import ( _TOOL_CAPABILITIES, _filter_tools, detect_language, restaurant_agent, ) # --------------------------------------------------------------------------- # Language detection # --------------------------------------------------------------------------- class TestDetectLanguage: def test_dutch_message(self): assert detect_language("Ik wil een tafel reserveren") == "nl" def test_english_message(self): assert detect_language("What are your opening hours?") == "en" def test_dutch_greeting(self): assert detect_language("Hallo, goedemiddag") == "nl" def test_english_greeting(self): assert detect_language("Hello, how are you?") == "en" def test_empty_message_defaults_nl(self): assert detect_language("") == "nl" def test_punctuation_stripped(self): assert detect_language("Morgen, 4 personen.") == "nl" def test_mixed_defaults_nl(self): # Equal hits → nl wins assert detect_language("is") == "nl" # --------------------------------------------------------------------------- # Tool capability mapping # --------------------------------------------------------------------------- class TestToolCapabilities: def test_all_tools_mapped(self): """Every tool registered on the agent has a capability mapping.""" for tool in restaurant_agent._function_toolset.tools.values(): assert tool.name in _TOOL_CAPABILITIES, ( f"Tool {tool.name!r} missing from _TOOL_CAPABILITIES" ) def test_faq_tools(self): faq_tools = {k for k, v in _TOOL_CAPABILITIES.items() if v == "faq"} assert faq_tools == { "search_knowledge_base", "is_open_now", "get_opening_hours", "get_restaurant_policies", } def test_reservation_tools(self): res_tools = {k for k, v in _TOOL_CAPABILITIES.items() if v == "reservation"} assert res_tools == { "check_availability", "create_reservation", "modify_reservation", "cancel_reservation", "find_available_slots", "find_reservation", "get_restaurant_info", } def test_takeaway_tools(self): tk_tools = {k for k, v in _TOOL_CAPABILITIES.items() if v == "takeaway"} assert tk_tools == { "search_menu", "create_order", "get_payment_link", } # --------------------------------------------------------------------------- # Dynamic tool filtering via prepare_tools # --------------------------------------------------------------------------- def _make_tool_def(name: str) -> ToolDefinition: return ToolDefinition(name=name, description=f"Tool {name}", parameters_json_schema={}) def _make_deps(capabilities: set[str]) -> AgentDeps: return AgentDeps( session=MagicMock(), restaurant_id="rest-1", http_client=MagicMock(), caller=CallerIdentity( channel="dashboard", verified=True, identity_key="test-user", ), enabled_capabilities=capabilities, ) class TestFilterTools: @pytest.mark.anyio async def test_faq_only(self): ctx = MagicMock() ctx.deps = _make_deps({"faq"}) defs = [ _make_tool_def("search_knowledge_base"), _make_tool_def("check_availability"), _make_tool_def("search_menu"), ] result = await _filter_tools(ctx, defs) assert result is not None names = {td.name for td in result} assert names == {"search_knowledge_base"} @pytest.mark.anyio async def test_faq_and_reservation(self): ctx = MagicMock() ctx.deps = _make_deps({"faq", "reservation"}) defs = [ _make_tool_def("search_knowledge_base"), _make_tool_def("check_availability"), _make_tool_def("create_reservation"), _make_tool_def("search_menu"), ] result = await _filter_tools(ctx, defs) assert result is not None names = {td.name for td in result} assert names == {"search_knowledge_base", "check_availability", "create_reservation"} @pytest.mark.anyio async def test_all_capabilities(self): ctx = MagicMock() ctx.deps = _make_deps({"faq", "reservation", "takeaway"}) defs = [_make_tool_def(name) for name in _TOOL_CAPABILITIES] result = await _filter_tools(ctx, defs) assert result is not None # Dashboard/verified caller: verification tools are excluded non_verification = {n for n, c in _TOOL_CAPABILITIES.items() if c != "verification"} assert len(result) == len(non_verification) @pytest.mark.anyio async def test_empty_capabilities(self): ctx = MagicMock() ctx.deps = _make_deps(set()) defs = [_make_tool_def("search_knowledge_base")] result = await _filter_tools(ctx, defs) assert result is not None assert len(result) == 0 # --------------------------------------------------------------------------- # Reservation system prompt: verification block is channel/verification-aware # --------------------------------------------------------------------------- class TestReservationPromptVerificationBlock: """Regression: the prompt used to instruct the model to call ``send_verification_code`` unconditionally. On the dashboard the tool is filtered out by ``_filter_tools``, so the model would hallucinate "I don't have the tools to verify you" mid-reply. The prompt must mirror the same gate as ``_filter_tools``.""" def _build_ctx(self, *, channel: str, verified: bool) -> MagicMock: deps = AgentDeps( session=MagicMock(), restaurant_id="rest-1", http_client=MagicMock(), caller=CallerIdentity( channel=channel, verified=verified, identity_key="anchor", conversation_id="conv-1", ), enabled_capabilities={"reservation", "faq"}, ) ctx = MagicMock() ctx.deps = deps return ctx @pytest.mark.anyio async def test_dashboard_prompt_omits_verification_instructions(self): from app.agents.restaurant import _reservation_prompt prompt = await _reservation_prompt(self._build_ctx(channel="dashboard", verified=True)) assert prompt is not None assert "send_verification_code" not in prompt assert "verify_code" not in prompt assert "verification_required" not in prompt assert "direct access" in prompt @pytest.mark.anyio async def test_unverified_website_prompt_includes_verification_flow(self): from app.agents.restaurant import _reservation_prompt prompt = await _reservation_prompt(self._build_ctx(channel="website", verified=False)) assert prompt is not None assert "send_verification_code" in prompt assert "verify_code" in prompt assert "verification_required" in prompt # Operator-style "direct access" wording must not leak into guest prompts. assert "direct access" not in prompt @pytest.mark.anyio async def test_verified_whatsapp_prompt_omits_verification_instructions(self): from app.agents.restaurant import _reservation_prompt prompt = await _reservation_prompt(self._build_ctx(channel="whatsapp", verified=True)) assert prompt is not None assert "send_verification_code" not in prompt @pytest.mark.anyio async def test_unverified_website_prompt_has_explicit_post_verify_steer(self): """Regression: with only a vague "retry the reservation operation" instruction, the model would emit a generic greeting after verify_code and forget the original cancel/modify intent.""" from app.agents.restaurant import _reservation_prompt prompt = await _reservation_prompt(self._build_ctx(channel="website", verified=False)) assert prompt is not None # Explicit post-verify steer that bans a fresh greeting. assert "POST-VERIFY" in prompt assert "DO NOT greet" in prompt assert "ORIGINAL" in prompt or "original" in prompt # --------------------------------------------------------------------------- # History window must span a verification round-trip # --------------------------------------------------------------------------- class TestHistoryWindow: """Regression: the previous limit of 10 messages caused the agent to evict the original user intent ("I want to cancel my reservation") by the time it generated the post-verify reply. A verification round-trip routinely produces 10+ messages (user msg → tool call → tool return, repeated for find_reservation, send_verification_code, verify_code).""" def test_history_processor_limit_spans_verification_flow(self): import inspect from app.agents import restaurant as agent_module source = inspect.getsource(agent_module) # The agent constructor uses partial(keep_recent_messages, limit=N). # N must be large enough to keep the first user message through a # full verification round-trip. assert "keep_recent_messages, limit=30" in source, ( "history limit must be ≥30 to survive a verification round-trip; " "lowering it re-introduces the post-verify context-loss bug" )