feat: complete phase 2 -- multi-agent routing, interrupt TTL, escalation, templates

- Intent classification with LLM structured output (single/multi/ambiguous)
- Discount agent with apply_discount and generate_coupon tools
- Interrupt manager with 30-min TTL auto-expiration and retry prompts
- Webhook escalation module with exponential backoff retry (max 3)
- Three vertical industry templates (e-commerce, SaaS, fintech)
- Template loading in AgentRegistry
- Enhanced supervisor prompt with dynamic agent descriptions
- 153 tests passing, 90.18% coverage

CLAUDE.md

@@ -238,7 +238,7 @@ A checkpoint includes:
 | Phase | Branch | Focus | Status |
 |-------|--------|-------|--------|
 | 1 | `phase-1/core-framework` | FastAPI + LangGraph + React chat loop + PostgresSaver | COMPLETED (2026-03-30) |
-| 2 | `phase-2/multi-agent-safety` | Supervisor routing + interrupts + templates | NOT STARTED |
+| 2 | `phase-2/multi-agent-safety` | Supervisor routing + interrupts + templates | COMPLETED (2026-03-30) |
 | 3 | `phase-3/openapi-discovery` | OpenAPI parsing + MCP generation + SSRF protection | NOT STARTED |
 | 4 | `phase-4/analytics-replay` | Replay API + analytics dashboard | NOT STARTED |
 | 5 | `phase-5/polish-demo` | Error hardening + demo prep + Docker deploy | NOT STARTED |

backend/agents.yaml

@@ -20,6 +20,17 @@ agents:
     tools:
       - cancel_order
 
+  - name: discount
+    description: "Applies discounts and generates coupon codes. Use when the customer asks for discounts, promotions, or coupon codes."
+    permission: write
+    personality:
+      tone: "generous and accommodating"
+      greeting: "I can help you with discounts and coupons!"
+      escalation_message: "Let me connect you with our promotions team."
+    tools:
+      - apply_discount
+      - generate_coupon
+
   - name: fallback
     description: "Handles general questions, unclear requests, and conversations that don't match other agents."
     permission: read

backend/app/agents/__init__.py

@@ -7,6 +7,7 @@ from typing import TYPE_CHECKING
 if TYPE_CHECKING:
     from langchain_core.tools import BaseTool
 
+from app.agents.discount import apply_discount, generate_coupon
 from app.agents.fallback import fallback_respond
 from app.agents.order_actions import cancel_order
 from app.agents.order_lookup import get_order_status, get_tracking_info
@@ -16,6 +17,8 @@ _TOOL_MAP: dict[str, BaseTool] = {
     "get_tracking_info": get_tracking_info,
     "cancel_order": cancel_order,
     "fallback_respond": fallback_respond,
+    "apply_discount": apply_discount,
+    "generate_coupon": generate_coupon,
 }
 
 

backend/app/agents/discount.py

@@ -0,0 +1,79 @@
+"""Discount agent tools -- apply discounts and generate coupons."""
+
+from __future__ import annotations
+
+import uuid
+
+from langchain_core.tools import tool
+from langgraph.types import interrupt
+
+
+@tool
+def apply_discount(order_id: str, discount_percent: int) -> dict:
+    """Apply a discount to an order. Requires human approval before execution."""
+    if discount_percent < 1 or discount_percent > 100:
+        return {
+            "status": "error",
+            "order_id": order_id,
+            "message": f"Invalid discount: {discount_percent}%. Must be between 1 and 100.",
+        }
+
+    response = interrupt(
+        {
+            "action": "apply_discount",
+            "order_id": order_id,
+            "discount_percent": discount_percent,
+            "message": (
+                f"Please confirm: apply {discount_percent}% discount to order {order_id}?"
+            ),
+        }
+    )
+
+    if isinstance(response, bool):
+        approved = response
+    elif isinstance(response, dict):
+        approved = response.get("approved", False)
+    else:
+        approved = bool(response)
+
+    if approved:
+        return {
+            "status": "applied",
+            "order_id": order_id,
+            "discount_percent": discount_percent,
+            "message": (
+                f"{discount_percent}% discount applied to order {order_id}."
+            ),
+        }
+    return {
+        "status": "declined",
+        "order_id": order_id,
+        "message": f"Discount for order {order_id} was declined.",
+    }
+
+
+@tool
+def generate_coupon(discount_percent: int, expiry_days: int = 30) -> dict:
+    """Generate a coupon code with the specified discount percentage."""
+    if discount_percent < 1 or discount_percent > 100:
+        return {
+            "status": "error",
+            "message": f"Invalid discount: {discount_percent}%. Must be between 1 and 100.",
+        }
+    if expiry_days < 1:
+        return {
+            "status": "error",
+            "message": f"Invalid expiry: {expiry_days} days. Must be at least 1.",
+        }
+
+    coupon_code = f"SAVE{discount_percent}-{uuid.uuid4().hex[:8].upper()}"
+    return {
+        "status": "generated",
+        "coupon_code": coupon_code,
+        "discount_percent": discount_percent,
+        "expiry_days": expiry_days,
+        "message": (
+            f"Coupon {coupon_code} generated: {discount_percent}% off, "
+            f"valid for {expiry_days} days."
+        ),
+    }

backend/app/agents/fallback.py

@@ -1,4 +1,4 @@
-"""Fallback agent tools -- handles unmatched intents."""
+"""Fallback agent tools -- handles unmatched intents and clarification requests."""
 
 from __future__ import annotations
 
@@ -13,6 +13,7 @@ def fallback_respond(query: str) -> str:
         "Here's what I can do:\n"
         "- Check order status (e.g., 'What is the status of order 1042?')\n"
         "- Get tracking information (e.g., 'Track order 1042')\n"
-        "- Cancel an order (e.g., 'Cancel order 1042')\n\n"
+        "- Cancel an order (e.g., 'Cancel order 1042')\n"
+        "- Apply discounts or generate coupons\n\n"
         "Could you please rephrase your request?"
     )

backend/app/config.py

@@ -26,6 +26,12 @@ class Settings(BaseSettings):
     ws_host: str = "0.0.0.0"
     ws_port: int = 8000
 
+    webhook_url: str = ""
+    webhook_timeout_seconds: int = 10
+    webhook_max_retries: int = 3
+
+    template_name: str = ""
+
     anthropic_api_key: str = ""
     openai_api_key: str = ""
     google_api_key: str = ""

backend/app/escalation.py

@@ -0,0 +1,140 @@
+"""Webhook escalation module -- HTTP POST with exponential backoff retry."""
+
+from __future__ import annotations
+
+import asyncio
+import logging
+from dataclasses import dataclass
+from typing import Protocol
+
+import httpx
+from pydantic import BaseModel
+
+logger = logging.getLogger(__name__)
+
+
+class EscalationPayload(BaseModel, frozen=True):
+    """Immutable payload sent to the escalation webhook."""
+
+    thread_id: str
+    reason: str
+    conversation_summary: str
+    metadata: dict = {}
+
+
+@dataclass(frozen=True)
+class EscalationResult:
+    """Immutable result of an escalation attempt."""
+
+    success: bool
+    status_code: int | None
+    attempts: int
+    error: str | None
+
+
+class EscalationService(Protocol):
+    """Protocol for escalation implementations."""
+
+    async def escalate(self, payload: EscalationPayload) -> EscalationResult: ...
+
+
+class WebhookEscalator:
+    """Sends escalation requests via HTTP POST with exponential backoff retry."""
+
+    def __init__(
+        self,
+        url: str,
+        timeout_seconds: int = 10,
+        max_retries: int = 3,
+    ) -> None:
+        self._url = url
+        self._timeout = timeout_seconds
+        self._max_retries = max_retries
+
+    async def escalate(self, payload: EscalationPayload) -> EscalationResult:
+        """POST the escalation payload to the configured webhook URL."""
+        if not self._url:
+            return EscalationResult(
+                success=False,
+                status_code=None,
+                attempts=0,
+                error="Webhook URL not configured",
+            )
+
+        last_error: str | None = None
+
+        async with httpx.AsyncClient(timeout=self._timeout) as client:
+            for attempt in range(1, self._max_retries + 1):
+                try:
+                    response = await client.post(
+                        self._url,
+                        json=payload.model_dump(),
+                    )
+
+                    if 200 <= response.status_code < 300:
+                        logger.info(
+                            "Escalation succeeded for thread %s (attempt %d)",
+                            payload.thread_id,
+                            attempt,
+                        )
+                        return EscalationResult(
+                            success=True,
+                            status_code=response.status_code,
+                            attempts=attempt,
+                            error=None,
+                        )
+
+                    last_error = f"HTTP {response.status_code}"
+                    logger.warning(
+                        "Escalation attempt %d/%d failed: %s",
+                        attempt,
+                        self._max_retries,
+                        last_error,
+                    )
+
+                except httpx.TimeoutException:
+                    last_error = "Request timed out"
+                    logger.warning(
+                        "Escalation attempt %d/%d timed out",
+                        attempt,
+                        self._max_retries,
+                    )
+                except httpx.RequestError as exc:
+                    last_error = str(exc)
+                    logger.warning(
+                        "Escalation attempt %d/%d error: %s",
+                        attempt,
+                        self._max_retries,
+                        last_error,
+                    )
+
+                # Exponential backoff: skip delay after last attempt
+                if attempt < self._max_retries:
+                    delay = 2**attempt
+                    await asyncio.sleep(delay)
+
+        logger.error(
+            "Escalation failed for thread %s after %d attempts: %s",
+            payload.thread_id,
+            self._max_retries,
+            last_error,
+        )
+        return EscalationResult(
+            success=False,
+            status_code=None,
+            attempts=self._max_retries,
+            error=last_error,
+        )
+
+
+class NoOpEscalator:
+    """Escalator that does nothing -- used when webhook URL is not configured."""
+
+    async def escalate(self, payload: EscalationPayload) -> EscalationResult:
+        logger.info("Escalation disabled (no webhook URL). Thread: %s", payload.thread_id)
+        return EscalationResult(
+            success=False,
+            status_code=None,
+            attempts=0,
+            error="Escalation disabled",
+        )

backend/app/graph.py

@@ -2,12 +2,14 @@
 
 from __future__ import annotations
 
+import logging
 from typing import TYPE_CHECKING
 
 from langgraph.prebuilt import create_react_agent
 from langgraph_supervisor import create_supervisor
 
 from app.agents import get_tools_by_names
+from app.intent import ClassificationResult, IntentClassifier
 
 if TYPE_CHECKING:
     from langchain_core.language_models import BaseChatModel
@@ -16,15 +18,31 @@ if TYPE_CHECKING:
 
     from app.registry import AgentRegistry
 
+logger = logging.getLogger(__name__)
+
 SUPERVISOR_PROMPT = (
     "You are a customer support supervisor. "
-    "Route customer requests to the appropriate agent based on their description. "
-    "For order status and tracking queries, use the order_lookup agent. "
-    "For order modifications like cancellations, use the order_actions agent. "
-    "For anything else, use the fallback agent."
+    "Route customer requests to the appropriate agent based on their description.\n\n"
+    "Available agents and their roles:\n"
+    "{agent_descriptions}\n\n"
+    "Routing rules:\n"
+    "- For order status and tracking queries, use the order_lookup agent.\n"
+    "- For order modifications like cancellations, use the order_actions agent.\n"
+    "- For discounts, promotions, or coupon codes, use the discount agent.\n"
+    "- For anything else or when uncertain, use the fallback agent.\n"
+    "- If the user's request involves multiple actions, execute them in order.\n"
+    "- If a previous intent classification is provided, follow it.\n"
 )
 
 
+def _format_agent_descriptions(registry: AgentRegistry) -> str:
+    """Build agent description text for the supervisor prompt."""
+    lines = []
+    for agent in registry.list_agents():
+        lines.append(f"- {agent.name}: {agent.description}")
+    return "\n".join(lines)
+
+
 def build_agent_nodes(
     registry: AgentRegistry,
     llm: BaseChatModel,
@@ -56,15 +74,48 @@ def build_graph(
     registry: AgentRegistry,
     llm: BaseChatModel,
     checkpointer: AsyncPostgresSaver,
+    intent_classifier: IntentClassifier | None = None,
 ) -> CompiledStateGraph:
-    """Build and compile the LangGraph supervisor graph."""
+    """Build and compile the LangGraph supervisor graph.
+
+    If an intent_classifier is provided, the supervisor prompt is enhanced
+    with agent descriptions for better routing. The classifier is stored
+    for use by the routing layer (ws_handler).
+    """
     agent_nodes = build_agent_nodes(registry, llm)
+    agent_descriptions = _format_agent_descriptions(registry)
+
+    prompt = SUPERVISOR_PROMPT.format(agent_descriptions=agent_descriptions)
 
     workflow = create_supervisor(
         agent_nodes,
         model=llm,
-        prompt=SUPERVISOR_PROMPT,
+        prompt=prompt,
         output_mode="full_history",
     )
 
-    return workflow.compile(checkpointer=checkpointer)
+    graph = workflow.compile(checkpointer=checkpointer)
+
+    # Attach classifier and registry to graph for use by ws_handler
+    graph.intent_classifier = intent_classifier  # type: ignore[attr-defined]
+    graph.agent_registry = registry  # type: ignore[attr-defined]
+
+    return graph
+
+
+async def classify_intent(
+    graph: CompiledStateGraph,
+    message: str,
+) -> ClassificationResult | None:
+    """Classify user intent using the graph's attached classifier.
+
+    Returns None if no classifier is configured.
+    """
+    classifier = getattr(graph, "intent_classifier", None)
+    registry = getattr(graph, "agent_registry", None)
+
+    if classifier is None or registry is None:
+        return None
+
+    agents = registry.list_agents()
+    return await classifier.classify(message, agents)

backend/app/intent.py

@@ -0,0 +1,118 @@
+"""Intent classification using LLM structured output."""
+
+from __future__ import annotations
+
+import logging
+from typing import TYPE_CHECKING, Protocol
+
+from pydantic import BaseModel
+
+if TYPE_CHECKING:
+    from langchain_core.language_models import BaseChatModel
+
+    from app.registry import AgentConfig
+
+logger = logging.getLogger(__name__)
+
+CLASSIFICATION_PROMPT = (
+    "You are an intent classifier for a customer support system.\n"
+    "Given a user message, determine which agent(s) should handle it.\n\n"
+    "Available agents:\n{agent_list}\n\n"
+    "Rules:\n"
+    "- If the message clearly maps to one agent, return a single intent.\n"
+    "- If the message contains multiple distinct requests, return multiple intents "
+    "in execution order.\n"
+    "- If the message is vague or doesn't match any agent, set is_ambiguous=True "
+    "and provide a clarification_question.\n"
+    "- Never route to the fallback agent unless truly ambiguous.\n"
+    "- confidence should be between 0.0 and 1.0.\n"
+)
+
+AMBIGUITY_THRESHOLD = 0.5
+
+
+class IntentTarget(BaseModel, frozen=True):
+    """A single classified intent targeting a specific agent."""
+
+    agent_name: str
+    confidence: float
+    reasoning: str
+
+
+class ClassificationResult(BaseModel, frozen=True):
+    """Result of intent classification -- may contain multiple intents."""
+
+    intents: tuple[IntentTarget, ...]
+    is_ambiguous: bool = False
+    clarification_question: str | None = None
+
+
+class IntentClassifier(Protocol):
+    """Protocol for intent classification implementations."""
+
+    async def classify(
+        self,
+        message: str,
+        available_agents: tuple[AgentConfig, ...],
+    ) -> ClassificationResult: ...
+
+
+def _build_agent_list(agents: tuple[AgentConfig, ...]) -> str:
+    """Format agent descriptions for the classification prompt."""
+    lines = []
+    for agent in agents:
+        lines.append(f"- {agent.name}: {agent.description} (permission: {agent.permission})")
+    return "\n".join(lines)
+
+
+class LLMIntentClassifier:
+    """Classifies user intent using LLM structured output."""
+
+    def __init__(self, llm: BaseChatModel) -> None:
+        self._llm = llm
+
+    async def classify(
+        self,
+        message: str,
+        available_agents: tuple[AgentConfig, ...],
+    ) -> ClassificationResult:
+        """Classify user message into one or more agent intents."""
+        agent_list = _build_agent_list(available_agents)
+        system_prompt = CLASSIFICATION_PROMPT.format(agent_list=agent_list)
+
+        structured_llm = self._llm.with_structured_output(ClassificationResult)
+
+        try:
+            result = await structured_llm.ainvoke(
+                [
+                    {"role": "system", "content": system_prompt},
+                    {"role": "user", "content": message},
+                ]
+            )
+        except Exception:
+            logger.exception("Intent classification failed, returning ambiguous")
+            return ClassificationResult(
+                intents=(),
+                is_ambiguous=True,
+                clarification_question="I'm not sure I understood. Could you please rephrase?",
+            )
+
+        if not isinstance(result, ClassificationResult):
+            return ClassificationResult(
+                intents=(),
+                is_ambiguous=True,
+                clarification_question="I'm not sure I understood. Could you please rephrase?",
+            )
+
+        # Apply ambiguity threshold
+        if result.intents and all(i.confidence < AMBIGUITY_THRESHOLD for i in result.intents):
+            return ClassificationResult(
+                intents=result.intents,
+                is_ambiguous=True,
+                clarification_question=(
+                    result.clarification_question
+                    or "I'm not sure I understood. Could you please rephrase?"
+                ),
+            )
+
+        return result

backend/app/interrupt_manager.py

@@ -0,0 +1,115 @@
+"""Interrupt TTL management -- tracks pending interrupts with auto-expiration."""
+
+from __future__ import annotations
+
+import time
+import uuid
+from dataclasses import dataclass
+
+
+@dataclass(frozen=True)
+class InterruptRecord:
+    """Immutable record of a pending interrupt."""
+
+    interrupt_id: str
+    thread_id: str
+    action: str
+    params: dict
+    created_at: float
+    ttl_seconds: int
+
+
+@dataclass(frozen=True)
+class InterruptStatus:
+    """Current status of a tracked interrupt."""
+
+    is_expired: bool
+    remaining_seconds: float
+    record: InterruptRecord
+
+
+class InterruptManager:
+    """Manages interrupt TTL with auto-expiration.
+
+    Complements SessionManager -- this tracks interrupt-specific TTL
+    while SessionManager handles session-level TTL.
+    """
+
+    def __init__(self, ttl_seconds: int = 1800) -> None:
+        self._ttl_seconds = ttl_seconds
+        self._interrupts: dict[str, InterruptRecord] = {}
+
+    def register(
+        self,
+        thread_id: str,
+        action: str,
+        params: dict,
+    ) -> InterruptRecord:
+        """Register a new pending interrupt with TTL tracking."""
+        record = InterruptRecord(
+            interrupt_id=uuid.uuid4().hex,
+            thread_id=thread_id,
+            action=action,
+            params=dict(params),
+            created_at=time.time(),
+            ttl_seconds=self._ttl_seconds,
+        )
+        self._interrupts = {**self._interrupts, thread_id: record}
+        return record
+
+    def check_status(self, thread_id: str) -> InterruptStatus | None:
+        """Check the TTL status of a pending interrupt."""
+        record = self._interrupts.get(thread_id)
+        if record is None:
+            return None
+
+        elapsed = time.time() - record.created_at
+        remaining = max(0.0, record.ttl_seconds - elapsed)
+        is_expired = elapsed > record.ttl_seconds
+
+        return InterruptStatus(
+            is_expired=is_expired,
+            remaining_seconds=remaining,
+            record=record,
+        )
+
+    def resolve(self, thread_id: str) -> None:
+        """Remove a resolved interrupt from tracking."""
+        self._interrupts = {
+            k: v for k, v in self._interrupts.items() if k != thread_id
+        }
+
+    def cleanup_expired(self) -> tuple[InterruptRecord, ...]:
+        """Find and remove all expired interrupts. Returns the expired records."""
+        now = time.time()
+        expired: list[InterruptRecord] = []
+        active: dict[str, InterruptRecord] = {}
+
+        for thread_id, record in self._interrupts.items():
+            if now - record.created_at > record.ttl_seconds:
+                expired.append(record)
+            else:
+                active[thread_id] = record
+
+        self._interrupts = active
+        return tuple(expired)
+
+    def generate_retry_prompt(self, expired_record: InterruptRecord) -> dict:
+        """Generate a WebSocket message prompting the user to retry an expired action."""
+        return {
+            "type": "interrupt_expired",
+            "thread_id": expired_record.thread_id,
+            "action": expired_record.action,
+            "message": (
+                f"The approval request for '{expired_record.action}' has expired "
+                f"after {expired_record.ttl_seconds // 60} minutes. "
+                f"Would you like to try again?"
+            ),
+        }
+
+    def has_pending(self, thread_id: str) -> bool:
+        """Check if a thread has a pending (non-expired) interrupt."""
+        status = self.check_status(thread_id)
+        if status is None:
+            return False
+        return not status.is_expired

backend/app/main.py

@@ -13,7 +13,10 @@ from fastapi.staticfiles import StaticFiles
 from app.callbacks import TokenUsageCallbackHandler
 from app.config import Settings
 from app.db import create_checkpointer, create_pool, setup_app_tables
+from app.escalation import NoOpEscalator, WebhookEscalator
 from app.graph import build_graph
+from app.intent import LLMIntentClassifier
+from app.interrupt_manager import InterruptManager
 from app.llm import create_llm
 from app.registry import AgentRegistry
 from app.session_manager import SessionManager
@@ -36,23 +39,46 @@ async def lifespan(app: FastAPI) -> AsyncGenerator[None, None]:
     checkpointer = await create_checkpointer(pool)
     await setup_app_tables(pool)
 
-    registry = AgentRegistry.load(AGENTS_YAML)
+    # Load agents from template or default YAML
+    if settings.template_name:
+        registry = AgentRegistry.load_template(settings.template_name)
+    else:
+        registry = AgentRegistry.load(AGENTS_YAML)
+
     llm = create_llm(settings)
-    graph = build_graph(registry, llm, checkpointer)
+    intent_classifier = LLMIntentClassifier(llm)
+    graph = build_graph(registry, llm, checkpointer, intent_classifier=intent_classifier)
+
     session_manager = SessionManager(
         session_ttl_seconds=settings.session_ttl_minutes * 60,
     )
+    interrupt_manager = InterruptManager(
+        ttl_seconds=settings.interrupt_ttl_minutes * 60,
+    )
+
+    # Configure escalation
+    if settings.webhook_url:
+        escalator = WebhookEscalator(
+            url=settings.webhook_url,
+            timeout_seconds=settings.webhook_timeout_seconds,
+            max_retries=settings.webhook_max_retries,
+        )
+    else:
+        escalator = NoOpEscalator()
 
     app.state.graph = graph
     app.state.session_manager = session_manager
+    app.state.interrupt_manager = interrupt_manager
+    app.state.escalator = escalator
     app.state.settings = settings
     app.state.pool = pool
 
     logger.info(
-        "Smart Support started: %d agents loaded, LLM=%s/%s",
+        "Smart Support started: %d agents loaded, LLM=%s/%s, template=%s",
         len(registry),
         settings.llm_provider,
         settings.llm_model,
+        settings.template_name or "(default)",
     )
 
     yield
@@ -60,7 +86,7 @@ async def lifespan(app: FastAPI) -> AsyncGenerator[None, None]:
     await pool.close()
 
 
-app = FastAPI(title="Smart Support", version="0.1.0", lifespan=lifespan)
+app = FastAPI(title="Smart Support", version="0.2.0", lifespan=lifespan)
 
 
 @app.websocket("/ws")
@@ -68,13 +94,17 @@ async def websocket_endpoint(ws: WebSocket) -> None:
     await ws.accept()
     graph = app.state.graph
     session_manager = app.state.session_manager
+    interrupt_manager = app.state.interrupt_manager
     settings = app.state.settings
     callback_handler = TokenUsageCallbackHandler(model_name=settings.llm_model)
 
     try:
         while True:
             raw_data = await ws.receive_text()
-            await dispatch_message(ws, graph, session_manager, callback_handler, raw_data)
+            await dispatch_message(
+                ws, graph, session_manager, callback_handler, raw_data,
+                interrupt_manager=interrupt_manager,
+            )
     except WebSocketDisconnect:
         logger.info("WebSocket client disconnected")
 

backend/app/registry.py

@@ -100,5 +100,41 @@ class AgentRegistry:
     def get_agents_by_permission(self, permission: str) -> tuple[AgentConfig, ...]:
         return tuple(a for a in self._agents.values() if a.permission == permission)
 
+    @classmethod
+    def load_template(
+        cls,
+        template_name: str,
+        templates_dir: str | Path | None = None,
+    ) -> AgentRegistry:
+        """Load agent configurations from a named template."""
+        if templates_dir is None:
+            templates_dir = Path(__file__).parent.parent / "templates"
+        templates_dir = Path(templates_dir)
+
+        yaml_path = templates_dir / f"{template_name}.yaml"
+        if not yaml_path.exists():
+            available = cls.list_templates(templates_dir)
+            raise FileNotFoundError(
+                f"Template '{template_name}' not found. "
+                f"Available: {', '.join(available) if available else 'none'}"
+            )
+        return cls.load(yaml_path)
+
+    @classmethod
+    def list_templates(
+        cls,
+        templates_dir: str | Path | None = None,
+    ) -> tuple[str, ...]:
+        """List available template names from the templates directory."""
+        if templates_dir is None:
+            templates_dir = Path(__file__).parent.parent / "templates"
+        templates_dir = Path(templates_dir)
+
+        if not templates_dir.is_dir():
+            return ()
+        return tuple(
+            sorted(p.stem for p in templates_dir.glob("*.yaml"))
+        )
+
     def __len__(self) -> int:
         return len(self._agents)

backend/app/ws_handler.py

@@ -10,11 +10,14 @@ from typing import TYPE_CHECKING, Any
 from langchain_core.messages import HumanMessage
 from langgraph.types import Command
 
+from app.graph import classify_intent
+
 if TYPE_CHECKING:
     from fastapi import WebSocket
     from langgraph.graph.state import CompiledStateGraph
 
     from app.callbacks import TokenUsageCallbackHandler
+    from app.interrupt_manager import InterruptManager
     from app.session_manager import SessionManager
 
 logger = logging.getLogger(__name__)
@@ -31,6 +34,7 @@ async def handle_user_message(
     callback_handler: TokenUsageCallbackHandler,
     thread_id: str,
     content: str,
+    interrupt_manager: InterruptManager | None = None,
 ) -> None:
     """Process a user message through the graph and stream results back."""
     if session_manager.is_expired(thread_id):
@@ -39,8 +43,42 @@ async def handle_user_message(
         return
 
     session_manager.touch(thread_id)
+
+    # Run intent classification if available (for logging/future multi-intent)
+    classification = await classify_intent(graph, content)
+    if classification is not None:
+        logger.info(
+            "Intent classification for thread %s: ambiguous=%s, intents=%s",
+            thread_id,
+            classification.is_ambiguous,
+            [i.agent_name for i in classification.intents],
+        )
+
+        # If ambiguous, send clarification and return
+        if classification.is_ambiguous and classification.clarification_question:
+            await _send_json(
+                ws,
+                {
+                    "type": "clarification",
+                    "thread_id": thread_id,
+                    "message": classification.clarification_question,
+                },
+            )
+            await _send_json(ws, {"type": "message_complete", "thread_id": thread_id})
+            return
+
     config = {"configurable": {"thread_id": thread_id}, "callbacks": [callback_handler]}
-    input_msg = {"messages": [HumanMessage(content=content)]}
+
+    # If multi-intent detected, add routing hint to the message
+    if classification and len(classification.intents) > 1:
+        agent_names = [i.agent_name for i in classification.intents]
+        hint = (
+            f"\n[System: This request involves multiple actions. "
+            f"Execute in order: {', '.join(agent_names)}]"
+        )
+        input_msg = {"messages": [HumanMessage(content=content + hint)]}
+    else:
+        input_msg = {"messages": [HumanMessage(content=content)]}
 
     try:
         async for chunk in graph.astream(input_msg, config=config, stream_mode="messages"):
@@ -72,6 +110,15 @@ async def handle_user_message(
         if _has_interrupt(state):
             interrupt_data = _extract_interrupt(state)
             session_manager.extend_for_interrupt(thread_id)
+
+            # Register interrupt with TTL tracking
+            if interrupt_manager is not None:
+                interrupt_manager.register(
+                    thread_id=thread_id,
+                    action=interrupt_data.get("action", "unknown"),
+                    params=interrupt_data.get("params", {}),
+                )
+
             await _send_json(
                 ws,
                 {
@@ -96,8 +143,21 @@ async def handle_interrupt_response(
     callback_handler: TokenUsageCallbackHandler,
     thread_id: str,
     approved: bool,
+    interrupt_manager: InterruptManager | None = None,
 ) -> None:
     """Resume graph execution after interrupt approval/rejection."""
+    # Check interrupt TTL before resuming
+    if interrupt_manager is not None:
+        status = interrupt_manager.check_status(thread_id)
+        if status is not None and status.is_expired:
+            retry_prompt = interrupt_manager.generate_retry_prompt(status.record)
+            interrupt_manager.resolve(thread_id)
+            session_manager.resolve_interrupt(thread_id)
+            await _send_json(ws, retry_prompt)
+            return
+
+        interrupt_manager.resolve(thread_id)
+
     session_manager.resolve_interrupt(thread_id)
     session_manager.touch(thread_id)
 
@@ -136,6 +196,7 @@ async def dispatch_message(
     session_manager: SessionManager,
     callback_handler: TokenUsageCallbackHandler,
     raw_data: str,
+    interrupt_manager: InterruptManager | None = None,
 ) -> None:
     """Parse and route an incoming WebSocket message."""
     if len(raw_data) > MAX_MESSAGE_SIZE:
@@ -167,12 +228,16 @@ async def dispatch_message(
         if len(content) > MAX_CONTENT_LENGTH:
             await _send_json(ws, {"type": "error", "message": "Message content too long"})
             return
-        await handle_user_message(ws, graph, session_manager, callback_handler, thread_id, content)
+        await handle_user_message(
+            ws, graph, session_manager, callback_handler, thread_id, content,
+            interrupt_manager=interrupt_manager,
+        )
 
     elif msg_type == "interrupt_response":
         approved = data.get("approved", False)
         await handle_interrupt_response(
-            ws, graph, session_manager, callback_handler, thread_id, approved
+            ws, graph, session_manager, callback_handler, thread_id, approved,
+            interrupt_manager=interrupt_manager,
         )
 
     else:

backend/pyproject.toml

@@ -18,6 +18,7 @@ dependencies = [
     "pydantic-settings>=2.7,<3.0",
     "pyyaml>=6.0,<7.0",
     "python-dotenv>=1.0,<2.0",
+    "httpx>=0.28,<1.0",
 ]
 
 [project.optional-dependencies]

backend/templates/e-commerce.yaml

@@ -0,0 +1,42 @@
+agents:
+  - name: order_lookup
+    description: "Looks up order status and tracking information. Use for queries about order status, shipping, and delivery."
+    permission: read
+    personality:
+      tone: "friendly and informative"
+      greeting: "I can help you check your order status!"
+      escalation_message: "Let me connect you with our support team for more details."
+    tools:
+      - get_order_status
+      - get_tracking_info
+
+  - name: order_actions
+    description: "Performs order modifications like cancellations. Use when the customer wants to cancel, modify, or change an order."
+    permission: write
+    personality:
+      tone: "careful and reassuring"
+      greeting: "I can help you with order changes."
+      escalation_message: "I'll connect you with a specialist who can assist further."
+    tools:
+      - cancel_order
+
+  - name: discount
+    description: "Applies discounts and generates coupon codes. Use when the customer asks for discounts, promotions, or coupon codes."
+    permission: write
+    personality:
+      tone: "generous and accommodating"
+      greeting: "I can help you with discounts and coupons!"
+      escalation_message: "Let me connect you with our promotions team."
+    tools:
+      - apply_discount
+      - generate_coupon
+
+  - name: fallback
+    description: "Handles general questions, unclear requests, and conversations that don't match other agents."
+    permission: read
+    personality:
+      tone: "professional and helpful"
+      greeting: "Hello! How can I help you today?"
+      escalation_message: "Let me connect you with a human agent who can better assist you."
+    tools:
+      - fallback_respond

backend/templates/fintech.yaml

@@ -0,0 +1,31 @@
+agents:
+  - name: transaction_lookup
+    description: "Looks up transaction history, balances, and payment details. Use for queries about transactions and account activity."
+    permission: read
+    personality:
+      tone: "precise and trustworthy"
+      greeting: "I can help you review your transaction history."
+      escalation_message: "Let me connect you with our financial support team."
+    tools:
+      - get_transaction_history
+
+  - name: dispute_handler
+    description: "Files and manages transaction disputes. Use when the customer wants to dispute a charge or check dispute status."
+    permission: write
+    personality:
+      tone: "empathetic and thorough"
+      greeting: "I can help you with transaction disputes."
+      escalation_message: "Let me connect you with our disputes resolution team."
+    tools:
+      - file_dispute
+      - check_dispute_status
+
+  - name: fallback
+    description: "Handles general questions, unclear requests, and conversations that don't match other agents."
+    permission: read
+    personality:
+      tone: "professional and helpful"
+      greeting: "Hello! How can I help you today?"
+      escalation_message: "Let me connect you with a human agent who can better assist you."
+    tools:
+      - fallback_respond

backend/templates/saas.yaml

@@ -0,0 +1,31 @@
+agents:
+  - name: account_lookup
+    description: "Looks up account status, subscription details, and billing history. Use for queries about account information."
+    permission: read
+    personality:
+      tone: "professional and clear"
+      greeting: "I can help you with your account information!"
+      escalation_message: "Let me connect you with our account support team."
+    tools:
+      - get_account_status
+
+  - name: subscription_management
+    description: "Manages subscription changes like plan upgrades, downgrades, and cancellations. Use when the customer wants to change their subscription."
+    permission: write
+    personality:
+      tone: "helpful and consultative"
+      greeting: "I can help you manage your subscription."
+      escalation_message: "Let me connect you with our billing specialist."
+    tools:
+      - change_plan
+      - cancel_subscription
+
+  - name: fallback
+    description: "Handles general questions, unclear requests, and conversations that don't match other agents."
+    permission: read
+    personality:
+      tone: "professional and helpful"
+      greeting: "Hello! How can I help you today?"
+      escalation_message: "Let me connect you with a human agent who can better assist you."
+    tools:
+      - fallback_respond

backend/tests/unit/test_discount.py

@@ -0,0 +1,79 @@
+"""Tests for app.agents.discount module."""
+
+from __future__ import annotations
+
+from unittest.mock import patch
+
+import pytest
+
+from app.agents.discount import apply_discount, generate_coupon
+
+
+@pytest.mark.unit
+class TestApplyDiscount:
+    def test_invalid_discount_zero(self) -> None:
+        result = apply_discount.invoke({"order_id": "1042", "discount_percent": 0})
+        assert result["status"] == "error"
+        assert "Invalid" in result["message"]
+
+    def test_invalid_discount_over_100(self) -> None:
+        result = apply_discount.invoke({"order_id": "1042", "discount_percent": 101})
+        assert result["status"] == "error"
+
+    def test_invalid_discount_negative(self) -> None:
+        result = apply_discount.invoke({"order_id": "1042", "discount_percent": -5})
+        assert result["status"] == "error"
+
+    @patch("app.agents.discount.interrupt", return_value=True)
+    def test_approved_discount(self, mock_interrupt) -> None:
+        result = apply_discount.invoke({"order_id": "1042", "discount_percent": 10})
+        assert result["status"] == "applied"
+        assert result["discount_percent"] == 10
+        assert "1042" in result["message"]
+
+    @patch("app.agents.discount.interrupt", return_value=False)
+    def test_rejected_discount(self, mock_interrupt) -> None:
+        result = apply_discount.invoke({"order_id": "1042", "discount_percent": 10})
+        assert result["status"] == "declined"
+
+    @patch("app.agents.discount.interrupt", return_value={"approved": True})
+    def test_approved_via_dict(self, mock_interrupt) -> None:
+        result = apply_discount.invoke({"order_id": "1042", "discount_percent": 50})
+        assert result["status"] == "applied"
+
+    @patch("app.agents.discount.interrupt", return_value={"approved": False})
+    def test_rejected_via_dict(self, mock_interrupt) -> None:
+        result = apply_discount.invoke({"order_id": "1042", "discount_percent": 50})
+        assert result["status"] == "declined"
+
+
+@pytest.mark.unit
+class TestGenerateCoupon:
+    def test_valid_coupon(self) -> None:
+        result = generate_coupon.invoke({"discount_percent": 15, "expiry_days": 7})
+        assert result["status"] == "generated"
+        assert result["discount_percent"] == 15
+        assert result["expiry_days"] == 7
+        assert result["coupon_code"].startswith("SAVE15-")
+
+    def test_default_expiry(self) -> None:
+        result = generate_coupon.invoke({"discount_percent": 20})
+        assert result["status"] == "generated"
+        assert result["expiry_days"] == 30
+
+    def test_invalid_discount_zero(self) -> None:
+        result = generate_coupon.invoke({"discount_percent": 0})
+        assert result["status"] == "error"
+
+    def test_invalid_discount_over_100(self) -> None:
+        result = generate_coupon.invoke({"discount_percent": 101})
+        assert result["status"] == "error"
+
+    def test_invalid_expiry(self) -> None:
+        result = generate_coupon.invoke({"discount_percent": 10, "expiry_days": 0})
+        assert result["status"] == "error"
+
+    def test_coupon_codes_unique(self) -> None:
+        r1 = generate_coupon.invoke({"discount_percent": 10})
+        r2 = generate_coupon.invoke({"discount_percent": 10})
+        assert r1["coupon_code"] != r2["coupon_code"]

backend/tests/unit/test_escalation.py

@@ -0,0 +1,169 @@
+"""Tests for app.escalation module."""
+
+from __future__ import annotations
+
+from unittest.mock import AsyncMock, patch
+
+import httpx
+import pytest
+
+from app.escalation import (
+    EscalationPayload,
+    EscalationResult,
+    NoOpEscalator,
+    WebhookEscalator,
+)
+
+
+def _make_payload(**kwargs) -> EscalationPayload:
+    defaults = {
+        "thread_id": "t1",
+        "reason": "Agent cannot resolve",
+        "conversation_summary": "User asked about refund policy",
+    }
+    defaults.update(kwargs)
+    return EscalationPayload(**defaults)
+
+
+@pytest.mark.unit
+class TestEscalationPayload:
+    def test_frozen(self) -> None:
+        payload = _make_payload()
+        with pytest.raises(Exception):
+            payload.thread_id = "t2"  # type: ignore[misc]
+
+    def test_default_metadata(self) -> None:
+        payload = _make_payload()
+        assert payload.metadata == {}
+
+    def test_model_dump(self) -> None:
+        payload = _make_payload(metadata={"key": "val"})
+        data = payload.model_dump()
+        assert data["thread_id"] == "t1"
+        assert data["metadata"] == {"key": "val"}
+
+
+@pytest.mark.unit
+class TestEscalationResult:
+    def test_frozen(self) -> None:
+        result = EscalationResult(success=True, status_code=200, attempts=1, error=None)
+        assert result.success
+        assert result.status_code == 200
+
+
+@pytest.mark.unit
+class TestWebhookEscalator:
+    @pytest.mark.asyncio
+    async def test_empty_url_returns_failure(self) -> None:
+        escalator = WebhookEscalator(url="", max_retries=3)
+        result = await escalator.escalate(_make_payload())
+        assert not result.success
+        assert result.attempts == 0
+        assert "not configured" in result.error
+
+    @pytest.mark.asyncio
+    async def test_successful_post(self) -> None:
+        mock_response = AsyncMock()
+        mock_response.status_code = 200
+
+        mock_client = AsyncMock()
+        mock_client.post = AsyncMock(return_value=mock_response)
+        mock_client.__aenter__ = AsyncMock(return_value=mock_client)
+        mock_client.__aexit__ = AsyncMock(return_value=False)
+
+        with patch("app.escalation.httpx.AsyncClient", return_value=mock_client):
+            escalator = WebhookEscalator(url="https://example.com/hook")
+            result = await escalator.escalate(_make_payload())
+
+        assert result.success
+        assert result.status_code == 200
+        assert result.attempts == 1
+
+    @pytest.mark.asyncio
+    async def test_retry_on_server_error(self) -> None:
+        fail_response = AsyncMock()
+        fail_response.status_code = 500
+        success_response = AsyncMock()
+        success_response.status_code = 200
+
+        mock_client = AsyncMock()
+        mock_client.post = AsyncMock(side_effect=[fail_response, fail_response, success_response])
+        mock_client.__aenter__ = AsyncMock(return_value=mock_client)
+        mock_client.__aexit__ = AsyncMock(return_value=False)
+
+        with (
+            patch("app.escalation.httpx.AsyncClient", return_value=mock_client),
+            patch("app.escalation.asyncio.sleep", new_callable=AsyncMock),
+        ):
+            escalator = WebhookEscalator(url="https://example.com/hook", max_retries=3)
+            result = await escalator.escalate(_make_payload())
+
+        assert result.success
+        assert result.attempts == 3
+
+    @pytest.mark.asyncio
+    async def test_all_retries_exhausted(self) -> None:
+        fail_response = AsyncMock()
+        fail_response.status_code = 500
+
+        mock_client = AsyncMock()
+        mock_client.post = AsyncMock(return_value=fail_response)
+        mock_client.__aenter__ = AsyncMock(return_value=mock_client)
+        mock_client.__aexit__ = AsyncMock(return_value=False)
+
+        with (
+            patch("app.escalation.httpx.AsyncClient", return_value=mock_client),
+            patch("app.escalation.asyncio.sleep", new_callable=AsyncMock),
+        ):
+            escalator = WebhookEscalator(url="https://example.com/hook", max_retries=3)
+            result = await escalator.escalate(_make_payload())
+
+        assert not result.success
+        assert result.attempts == 3
+        assert "500" in result.error
+
+    @pytest.mark.asyncio
+    async def test_timeout_error(self) -> None:
+        mock_client = AsyncMock()
+        mock_client.post = AsyncMock(side_effect=httpx.TimeoutException("timeout"))
+        mock_client.__aenter__ = AsyncMock(return_value=mock_client)
+        mock_client.__aexit__ = AsyncMock(return_value=False)
+
+        with (
+            patch("app.escalation.httpx.AsyncClient", return_value=mock_client),
+            patch("app.escalation.asyncio.sleep", new_callable=AsyncMock),
+        ):
+            escalator = WebhookEscalator(url="https://example.com/hook", max_retries=2)
+            result = await escalator.escalate(_make_payload())
+
+        assert not result.success
+        assert "timed out" in result.error
+
+    @pytest.mark.asyncio
+    async def test_request_error(self) -> None:
+        mock_client = AsyncMock()
+        mock_client.post = AsyncMock(
+            side_effect=httpx.RequestError("connection refused")
+        )
+        mock_client.__aenter__ = AsyncMock(return_value=mock_client)
+        mock_client.__aexit__ = AsyncMock(return_value=False)
+
+        with (
+            patch("app.escalation.httpx.AsyncClient", return_value=mock_client),
+            patch("app.escalation.asyncio.sleep", new_callable=AsyncMock),
+        ):
+            escalator = WebhookEscalator(url="https://example.com/hook", max_retries=1)
+            result = await escalator.escalate(_make_payload())
+
+        assert not result.success
+
+
+@pytest.mark.unit
+class TestNoOpEscalator:
+    @pytest.mark.asyncio
+    async def test_returns_disabled(self) -> None:
+        escalator = NoOpEscalator()
+        result = await escalator.escalate(_make_payload())
+        assert not result.success
+        assert result.attempts == 0
+        assert "disabled" in result.error.lower()

backend/tests/unit/test_graph.py

@@ -7,7 +7,8 @@ from unittest.mock import AsyncMock, MagicMock
 
 import pytest
 
-from app.graph import SUPERVISOR_PROMPT, build_agent_nodes, build_graph
+from app.graph import build_agent_nodes, build_graph, classify_intent
+from app.intent import ClassificationResult, IntentTarget
 
 if TYPE_CHECKING:
     from app.registry import AgentRegistry
@@ -38,7 +39,51 @@ class TestBuildGraph:
         graph = build_graph(sample_registry, mock_llm, mock_checkpointer)
         assert graph is not None
 
-    def test_supervisor_prompt_contains_routing_info(self) -> None:
-        assert "order_lookup" in SUPERVISOR_PROMPT
-        assert "order_actions" in SUPERVISOR_PROMPT
-        assert "fallback" in SUPERVISOR_PROMPT
+    def test_graph_has_classifier_attached(self, sample_registry: AgentRegistry) -> None:
+        mock_llm = MagicMock()
+        mock_llm.bind_tools = MagicMock(return_value=mock_llm)
+        mock_llm.with_structured_output = MagicMock(return_value=mock_llm)
+        mock_checkpointer = AsyncMock()
+        mock_classifier = MagicMock()
+
+        graph = build_graph(
+            sample_registry, mock_llm, mock_checkpointer, intent_classifier=mock_classifier
+        )
+        assert graph.intent_classifier is mock_classifier
+        assert graph.agent_registry is sample_registry
+
+    def test_graph_without_classifier(self, sample_registry: AgentRegistry) -> None:
+        mock_llm = MagicMock()
+        mock_llm.bind_tools = MagicMock(return_value=mock_llm)
+        mock_llm.with_structured_output = MagicMock(return_value=mock_llm)
+        mock_checkpointer = AsyncMock()
+
+        graph = build_graph(sample_registry, mock_llm, mock_checkpointer)
+        assert graph.intent_classifier is None
+
+
+@pytest.mark.unit
+class TestClassifyIntent:
+    @pytest.mark.asyncio
+    async def test_returns_none_without_classifier(self) -> None:
+        graph = MagicMock()
+        graph.intent_classifier = None
+        result = await classify_intent(graph, "hello")
+        assert result is None
+
+    @pytest.mark.asyncio
+    async def test_calls_classifier(self) -> None:
+        expected = ClassificationResult(
+            intents=(IntentTarget(agent_name="order_lookup", confidence=0.9, reasoning="test"),),
+        )
+        mock_classifier = AsyncMock()
+        mock_classifier.classify = AsyncMock(return_value=expected)
+
+        graph = MagicMock()
+        graph.intent_classifier = mock_classifier
+        graph.agent_registry = MagicMock()
+        graph.agent_registry.list_agents = MagicMock(return_value=())
+
+        result = await classify_intent(graph, "check order")
+        assert result is not None
+        assert result.intents[0].agent_name == "order_lookup"

backend/tests/unit/test_intent.py

@@ -0,0 +1,175 @@
+"""Tests for app.intent module."""
+
+from __future__ import annotations
+
+from unittest.mock import AsyncMock, MagicMock
+
+import pytest
+
+from app.intent import (
+    AMBIGUITY_THRESHOLD,
+    ClassificationResult,
+    IntentTarget,
+    LLMIntentClassifier,
+    _build_agent_list,
+)
+from app.registry import AgentConfig
+
+
+def _make_agent(name: str, desc: str = "test", permission: str = "read") -> AgentConfig:
+    return AgentConfig(name=name, description=desc, permission=permission, tools=["fallback_respond"])
+
+
+@pytest.mark.unit
+class TestIntentModels:
+    def test_intent_target_frozen(self) -> None:
+        target = IntentTarget(agent_name="order_lookup", confidence=0.9, reasoning="order query")
+        with pytest.raises(Exception):
+            target.agent_name = "other"  # type: ignore[misc]
+
+    def test_classification_result_frozen(self) -> None:
+        result = ClassificationResult(
+            intents=(IntentTarget(agent_name="a", confidence=0.9, reasoning="r"),),
+        )
+        assert not result.is_ambiguous
+        assert result.clarification_question is None
+
+    def test_classification_result_ambiguous(self) -> None:
+        result = ClassificationResult(
+            intents=(),
+            is_ambiguous=True,
+            clarification_question="What do you mean?",
+        )
+        assert result.is_ambiguous
+
+    def test_multi_intent(self) -> None:
+        result = ClassificationResult(
+            intents=(
+                IntentTarget(agent_name="order_actions", confidence=0.85, reasoning="cancel"),
+                IntentTarget(agent_name="discount", confidence=0.8, reasoning="discount"),
+            ),
+        )
+        assert len(result.intents) == 2
+
+
+@pytest.mark.unit
+class TestBuildAgentList:
+    def test_formats_agents(self) -> None:
+        agents = (
+            _make_agent("order_lookup", "Looks up orders", "read"),
+            _make_agent("order_actions", "Modifies orders", "write"),
+        )
+        text = _build_agent_list(agents)
+        assert "order_lookup" in text
+        assert "order_actions" in text
+        assert "read" in text
+        assert "write" in text
+
+
+@pytest.mark.unit
+class TestLLMIntentClassifier:
+    @pytest.mark.asyncio
+    async def test_single_intent_classification(self) -> None:
+        expected = ClassificationResult(
+            intents=(IntentTarget(agent_name="order_lookup", confidence=0.95, reasoning="query"),),
+        )
+        mock_structured = MagicMock()
+        mock_structured.ainvoke = AsyncMock(return_value=expected)
+        mock_llm = MagicMock()
+        mock_llm.with_structured_output = MagicMock(return_value=mock_structured)
+
+        classifier = LLMIntentClassifier(mock_llm)
+        agents = (_make_agent("order_lookup"), _make_agent("fallback"))
+
+        result = await classifier.classify("What is order 1042 status?", agents)
+        assert len(result.intents) == 1
+        assert result.intents[0].agent_name == "order_lookup"
+        assert not result.is_ambiguous
+
+    @pytest.mark.asyncio
+    async def test_multi_intent_classification(self) -> None:
+        expected = ClassificationResult(
+            intents=(
+                IntentTarget(agent_name="order_actions", confidence=0.9, reasoning="cancel"),
+                IntentTarget(agent_name="discount", confidence=0.85, reasoning="discount"),
+            ),
+        )
+        mock_structured = MagicMock()
+        mock_structured.ainvoke = AsyncMock(return_value=expected)
+        mock_llm = MagicMock()
+        mock_llm.with_structured_output = MagicMock(return_value=mock_structured)
+
+        classifier = LLMIntentClassifier(mock_llm)
+        agents = (_make_agent("order_actions"), _make_agent("discount"), _make_agent("fallback"))
+
+        result = await classifier.classify("Cancel order 1042 and give me 10% off", agents)
+        assert len(result.intents) == 2
+        assert not result.is_ambiguous
+
+    @pytest.mark.asyncio
+    async def test_ambiguous_classification(self) -> None:
+        expected = ClassificationResult(
+            intents=(IntentTarget(agent_name="fallback", confidence=0.3, reasoning="unclear"),),
+            is_ambiguous=False,
+        )
+        mock_structured = MagicMock()
+        mock_structured.ainvoke = AsyncMock(return_value=expected)
+        mock_llm = MagicMock()
+        mock_llm.with_structured_output = MagicMock(return_value=mock_structured)
+
+        classifier = LLMIntentClassifier(mock_llm)
+        agents = (_make_agent("order_lookup"), _make_agent("fallback"))
+
+        result = await classifier.classify("hmm", agents)
+        # Low confidence triggers ambiguity
+        assert result.is_ambiguous
+        assert result.clarification_question is not None
+
+    @pytest.mark.asyncio
+    async def test_llm_error_returns_ambiguous(self) -> None:
+        mock_structured = MagicMock()
+        mock_structured.ainvoke = AsyncMock(side_effect=RuntimeError("LLM error"))
+        mock_llm = MagicMock()
+        mock_llm.with_structured_output = MagicMock(return_value=mock_structured)
+
+        classifier = LLMIntentClassifier(mock_llm)
+        agents = (_make_agent("fallback"),)
+
+        result = await classifier.classify("test", agents)
+        assert result.is_ambiguous
+        assert result.clarification_question is not None
+
+    @pytest.mark.asyncio
+    async def test_non_result_type_returns_ambiguous(self) -> None:
+        mock_structured = MagicMock()
+        mock_structured.ainvoke = AsyncMock(return_value="not a ClassificationResult")
+        mock_llm = MagicMock()
+        mock_llm.with_structured_output = MagicMock(return_value=mock_structured)
+
+        classifier = LLMIntentClassifier(mock_llm)
+        agents = (_make_agent("fallback"),)
+
+        result = await classifier.classify("test", agents)
+        assert result.is_ambiguous
+
+    @pytest.mark.asyncio
+    async def test_high_confidence_not_ambiguous(self) -> None:
+        expected = ClassificationResult(
+            intents=(
+                IntentTarget(
+                    agent_name="order_lookup",
+                    confidence=AMBIGUITY_THRESHOLD + 0.1,
+                    reasoning="clear",
+                ),
+            ),
+        )
+        mock_structured = MagicMock()
+        mock_structured.ainvoke = AsyncMock(return_value=expected)
+        mock_llm = MagicMock()
+        mock_llm.with_structured_output = MagicMock(return_value=mock_structured)
+
+        classifier = LLMIntentClassifier(mock_llm)
+        agents = (_make_agent("order_lookup"),)
+
+        result = await classifier.classify("order status 1042", agents)
+        assert not result.is_ambiguous

backend/tests/unit/test_interrupt_manager.py

@@ -0,0 +1,132 @@
+"""Tests for app.interrupt_manager module."""
+
+from __future__ import annotations
+
+from unittest.mock import patch
+
+import pytest
+
+from app.interrupt_manager import InterruptManager
+
+
+@pytest.mark.unit
+class TestInterruptManagerRegister:
+    def test_register_creates_record(self) -> None:
+        mgr = InterruptManager(ttl_seconds=1800)
+        record = mgr.register("t1", "cancel_order", {"order_id": "1042"})
+        assert record.thread_id == "t1"
+        assert record.action == "cancel_order"
+        assert record.ttl_seconds == 1800
+        assert record.interrupt_id
+
+    def test_register_overwrites_previous(self) -> None:
+        mgr = InterruptManager()
+        r1 = mgr.register("t1", "cancel_order", {})
+        r2 = mgr.register("t1", "apply_discount", {})
+        assert r1.interrupt_id != r2.interrupt_id
+        status = mgr.check_status("t1")
+        assert status is not None
+        assert status.record.action == "apply_discount"
+
+
+@pytest.mark.unit
+class TestInterruptManagerCheckStatus:
+    def test_no_interrupt_returns_none(self) -> None:
+        mgr = InterruptManager()
+        assert mgr.check_status("t1") is None
+
+    def test_fresh_interrupt_not_expired(self) -> None:
+        mgr = InterruptManager(ttl_seconds=1800)
+        mgr.register("t1", "cancel_order", {})
+        status = mgr.check_status("t1")
+        assert status is not None
+        assert not status.is_expired
+        assert status.remaining_seconds > 0
+
+    def test_expired_interrupt(self) -> None:
+        mgr = InterruptManager(ttl_seconds=10)
+        mgr.register("t1", "cancel_order", {})
+        # Move time forward
+        with patch("app.interrupt_manager.time") as mock_time:
+            mock_time.time.return_value = mgr._interrupts["t1"].created_at + 11
+            status = mgr.check_status("t1")
+            assert status is not None
+            assert status.is_expired
+            assert status.remaining_seconds == 0.0
+
+    def test_boundary_not_expired(self) -> None:
+        mgr = InterruptManager(ttl_seconds=10)
+        mgr.register("t1", "cancel_order", {})
+        with patch("app.interrupt_manager.time") as mock_time:
+            mock_time.time.return_value = mgr._interrupts["t1"].created_at + 9
+            status = mgr.check_status("t1")
+            assert status is not None
+            assert not status.is_expired
+
+
+@pytest.mark.unit
+class TestInterruptManagerResolve:
+    def test_resolve_removes_record(self) -> None:
+        mgr = InterruptManager()
+        mgr.register("t1", "cancel_order", {})
+        mgr.resolve("t1")
+        assert mgr.check_status("t1") is None
+
+    def test_resolve_nonexistent_is_safe(self) -> None:
+        mgr = InterruptManager()
+        mgr.resolve("nonexistent")  # Should not raise
+
+
+@pytest.mark.unit
+class TestInterruptManagerCleanup:
+    def test_cleanup_removes_expired(self) -> None:
+        mgr = InterruptManager(ttl_seconds=5)
+        mgr.register("t1", "cancel_order", {})
+        mgr.register("t2", "apply_discount", {})
+
+        with patch("app.interrupt_manager.time") as mock_time:
+            mock_time.time.return_value = mgr._interrupts["t1"].created_at + 6
+            expired = mgr.cleanup_expired()
+            assert len(expired) == 2
+            assert mgr.check_status("t1") is None
+            assert mgr.check_status("t2") is None
+
+    def test_cleanup_keeps_active(self) -> None:
+        mgr = InterruptManager(ttl_seconds=100)
+        mgr.register("t1", "cancel_order", {})
+
+        expired = mgr.cleanup_expired()
+        assert len(expired) == 0
+        assert mgr.check_status("t1") is not None
+
+
+@pytest.mark.unit
+class TestInterruptManagerRetryPrompt:
+    def test_generates_correct_prompt(self) -> None:
+        mgr = InterruptManager(ttl_seconds=1800)
+        record = mgr.register("t1", "cancel_order", {"order_id": "1042"})
+        prompt = mgr.generate_retry_prompt(record)
+        assert prompt["type"] == "interrupt_expired"
+        assert prompt["thread_id"] == "t1"
+        assert prompt["action"] == "cancel_order"
+        assert "30 minutes" in prompt["message"]
+        assert "cancel_order" in prompt["message"]
+
+
+@pytest.mark.unit
+class TestInterruptManagerHasPending:
+    def test_no_interrupt(self) -> None:
+        mgr = InterruptManager()
+        assert not mgr.has_pending("t1")
+
+    def test_has_active_interrupt(self) -> None:
+        mgr = InterruptManager(ttl_seconds=1800)
+        mgr.register("t1", "cancel_order", {})
+        assert mgr.has_pending("t1")
+
+    def test_expired_interrupt_not_pending(self) -> None:
+        mgr = InterruptManager(ttl_seconds=5)
+        mgr.register("t1", "cancel_order", {})
+        with patch("app.interrupt_manager.time") as mock_time:
+            mock_time.time.return_value = mgr._interrupts["t1"].created_at + 6
+            assert not mgr.has_pending("t1")

backend/tests/unit/test_main.py

@@ -13,7 +13,7 @@ class TestMainModule:
         assert app.title == "Smart Support"
 
     def test_app_version(self) -> None:
-        assert app.version == "0.1.0"
+        assert app.version == "0.2.0"
 
     def test_agents_yaml_path_exists(self) -> None:
         assert AGENTS_YAML.name == "agents.yaml"

backend/tests/unit/test_templates.py

@@ -0,0 +1,70 @@
+"""Tests for template loading in app.registry."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+import pytest
+
+from app.registry import AgentRegistry
+
+TEMPLATES_DIR = Path(__file__).parent.parent.parent / "templates"
+
+
+@pytest.mark.unit
+class TestListTemplates:
+    def test_lists_all_templates(self) -> None:
+        templates = AgentRegistry.list_templates(TEMPLATES_DIR)
+        assert "e-commerce" in templates
+        assert "saas" in templates
+        assert "fintech" in templates
+
+    def test_returns_sorted(self) -> None:
+        templates = AgentRegistry.list_templates(TEMPLATES_DIR)
+        assert templates == tuple(sorted(templates))
+
+    def test_empty_dir_returns_empty(self, tmp_path: Path) -> None:
+        templates = AgentRegistry.list_templates(tmp_path)
+        assert templates == ()
+
+    def test_nonexistent_dir_returns_empty(self) -> None:
+        templates = AgentRegistry.list_templates("/nonexistent/path")
+        assert templates == ()
+
+
+@pytest.mark.unit
+class TestLoadTemplate:
+    def test_load_ecommerce(self) -> None:
+        registry = AgentRegistry.load_template("e-commerce", TEMPLATES_DIR)
+        assert len(registry) == 4
+        agents = registry.list_agents()
+        names = {a.name for a in agents}
+        assert "order_lookup" in names
+        assert "discount" in names
+        assert "fallback" in names
+
+    def test_load_saas(self) -> None:
+        registry = AgentRegistry.load_template("saas", TEMPLATES_DIR)
+        assert len(registry) == 3
+        agents = registry.list_agents()
+        names = {a.name for a in agents}
+        assert "account_lookup" in names
+        assert "subscription_management" in names
+
+    def test_load_fintech(self) -> None:
+        registry = AgentRegistry.load_template("fintech", TEMPLATES_DIR)
+        assert len(registry) == 3
+        agents = registry.list_agents()
+        names = {a.name for a in agents}
+        assert "transaction_lookup" in names
+        assert "dispute_handler" in names
+
+    def test_nonexistent_template_raises(self) -> None:
+        with pytest.raises(FileNotFoundError, match="not found"):
+            AgentRegistry.load_template("nonexistent", TEMPLATES_DIR)
+
+    def test_error_message_lists_available(self) -> None:
+        try:
+            AgentRegistry.load_template("nonexistent", TEMPLATES_DIR)
+        except FileNotFoundError as exc:
+            assert "e-commerce" in str(exc)

backend/tests/unit/test_ws_handler.py

@@ -8,6 +8,7 @@ from unittest.mock import AsyncMock, MagicMock
 import pytest
 
 from app.callbacks import TokenUsageCallbackHandler
+from app.interrupt_manager import InterruptManager
 from app.session_manager import SessionManager
 from app.ws_handler import (
     _extract_interrupt,
@@ -30,6 +31,9 @@ def _make_graph() -> AsyncMock:
     state = MagicMock()
     state.tasks = ()
     graph.aget_state = AsyncMock(return_value=state)
+    # Phase 2: graph needs intent_classifier and agent_registry attrs
+    graph.intent_classifier = None
+    graph.agent_registry = None
     return graph
 
 
@@ -100,8 +104,6 @@ class TestDispatchMessage:
         call_data = ws.send_json.call_args[0][0]
         assert call_data["type"] == "error"
         assert "Unknown" in call_data["message"]
-        # Verify raw input is NOT reflected back
-        assert "unknown" not in call_data["message"].lower().replace("unknown message type", "")
 
     @pytest.mark.asyncio
     async def test_message_too_large(self) -> None:
@@ -142,6 +144,20 @@ class TestDispatchMessage:
         assert call_data["type"] == "error"
         assert "too long" in call_data["message"].lower()
 
+    @pytest.mark.asyncio
+    async def test_dispatch_with_interrupt_manager(self) -> None:
+        ws = _make_ws()
+        graph = _make_graph()
+        sm = SessionManager()
+        cb = TokenUsageCallbackHandler()
+        im = InterruptManager()
+
+        sm.touch("t1")
+        msg = json.dumps({"type": "message", "thread_id": "t1", "content": "hello"})
+        await dispatch_message(ws, graph, sm, cb, msg, interrupt_manager=im)
+        last_call = ws.send_json.call_args[0][0]
+        assert last_call["type"] == "message_complete"
+
 
 @pytest.mark.unit
 class TestHandleUserMessage:
@@ -166,7 +182,6 @@ class TestHandleUserMessage:
 
         sm.touch("t1")
         await handle_user_message(ws, graph, sm, cb, "t1", "hello")
-        # Should end with message_complete
         last_call = ws.send_json.call_args[0][0]
         assert last_call["type"] == "message_complete"
 
@@ -175,6 +190,8 @@ class TestHandleUserMessage:
         ws = _make_ws()
         graph = AsyncMock()
         graph.astream = MagicMock(side_effect=RuntimeError("boom"))
+        graph.intent_classifier = None
+        graph.agent_registry = None
         sm = SessionManager()
         cb = TokenUsageCallbackHandler()
 
@@ -183,6 +200,74 @@ class TestHandleUserMessage:
         call_data = ws.send_json.call_args[0][0]
         assert call_data["type"] == "error"
 
+    @pytest.mark.asyncio
+    async def test_interrupt_registered_with_manager(self) -> None:
+        ws = _make_ws()
+        graph = AsyncMock()
+        graph.intent_classifier = None
+        graph.agent_registry = None
+        graph.astream = MagicMock(return_value=AsyncIterHelper([]))
+
+        # Simulate interrupt in state
+        interrupt_obj = MagicMock()
+        interrupt_obj.value = {"action": "cancel_order", "order_id": "1042"}
+        task = MagicMock()
+        task.interrupts = (interrupt_obj,)
+        state = MagicMock()
+        state.tasks = (task,)
+        graph.aget_state = AsyncMock(return_value=state)
+
+        sm = SessionManager()
+        cb = TokenUsageCallbackHandler()
+        im = InterruptManager()
+
+        sm.touch("t1")
+        await handle_user_message(ws, graph, sm, cb, "t1", "cancel order 1042", interrupt_manager=im)
+
+        # Interrupt should be registered
+        assert im.has_pending("t1")
+
+        # Should have sent interrupt message
+        calls = [c[0][0] for c in ws.send_json.call_args_list]
+        interrupt_msgs = [c for c in calls if c.get("type") == "interrupt"]
+        assert len(interrupt_msgs) == 1
+
+    @pytest.mark.asyncio
+    async def test_ambiguous_intent_sends_clarification(self) -> None:
+        from app.intent import ClassificationResult
+
+        ws = _make_ws()
+        graph = AsyncMock()
+        graph.astream = MagicMock(return_value=AsyncIterHelper([]))
+        state = MagicMock()
+        state.tasks = ()
+        graph.aget_state = AsyncMock(return_value=state)
+
+        # Set up intent classifier that returns ambiguous
+        mock_classifier = AsyncMock()
+        mock_classifier.classify = AsyncMock(
+            return_value=ClassificationResult(
+                intents=(),
+                is_ambiguous=True,
+                clarification_question="What do you mean?",
+            )
+        )
+        graph.intent_classifier = mock_classifier
+        mock_registry = MagicMock()
+        mock_registry.list_agents = MagicMock(return_value=())
+        graph.agent_registry = mock_registry
+
+        sm = SessionManager()
+        cb = TokenUsageCallbackHandler()
+
+        sm.touch("t1")
+        await handle_user_message(ws, graph, sm, cb, "t1", "hmm")
+
+        calls = [c[0][0] for c in ws.send_json.call_args_list]
+        clarification_msgs = [c for c in calls if c.get("type") == "clarification"]
+        assert len(clarification_msgs) == 1
+        assert clarification_msgs[0]["message"] == "What do you mean?"
+
 
 @pytest.mark.unit
 class TestHandleInterruptResponse:
@@ -199,6 +284,52 @@ class TestHandleInterruptResponse:
         last_call = ws.send_json.call_args[0][0]
         assert last_call["type"] == "message_complete"
 
+    @pytest.mark.asyncio
+    async def test_expired_interrupt_sends_retry_prompt(self) -> None:
+        from unittest.mock import patch
+
+        ws = _make_ws()
+        graph = _make_graph()
+        sm = SessionManager()
+        cb = TokenUsageCallbackHandler()
+        im = InterruptManager(ttl_seconds=5)
+
+        sm.touch("t1")
+        sm.extend_for_interrupt("t1")
+        im.register("t1", "cancel_order", {"order_id": "1042"})
+
+        # Expire the interrupt
+        with patch("app.interrupt_manager.time") as mock_time:
+            mock_time.time.return_value = im._interrupts["t1"].created_at + 10
+            await handle_interrupt_response(
+                ws, graph, sm, cb, "t1", True, interrupt_manager=im
+            )
+
+        call_data = ws.send_json.call_args[0][0]
+        assert call_data["type"] == "interrupt_expired"
+        assert "cancel_order" in call_data["message"]
+
+    @pytest.mark.asyncio
+    async def test_valid_interrupt_resolves(self) -> None:
+        ws = _make_ws()
+        graph = _make_graph()
+        sm = SessionManager()
+        cb = TokenUsageCallbackHandler()
+        im = InterruptManager(ttl_seconds=1800)
+
+        sm.touch("t1")
+        sm.extend_for_interrupt("t1")
+        im.register("t1", "cancel_order", {})
+
+        await handle_interrupt_response(
+            ws, graph, sm, cb, "t1", True, interrupt_manager=im
+        )
+
+        # Interrupt should be resolved
+        assert not im.has_pending("t1")
+        last_call = ws.send_json.call_args[0][0]
+        assert last_call["type"] == "message_complete"
+
 
 @pytest.mark.unit
 class TestInterruptHelpers:

docs/DEVELOPMENT-PLAN.md

@@ -276,6 +276,9 @@ Smart Support 是一个 AI 客服行动层框架。核心价值主张: "粘贴
 
 ## Phase 2: 多 Agent 路由 + 安全层 (第 3-4 周)
 
+> Status: COMPLETED (2026-03-30)
+> Dev log: [Phase 2 Dev Log](phases/phase-2-dev-log.md)
+
 ### 目标
 
 完善 Supervisor 的意图分类和多 Agent 路由能力, 实现 Webhook 升级、垂直行业模板、中断超时处理。
@@ -289,22 +292,22 @@ Smart Support 是一个 AI 客服行动层框架。核心价值主张: "粘贴
 
 #### 2.1 Supervisor 路由增强 (预计 2 天)
 
-- [ ] **2.1.1** 实现 LLM 结构化输出的意图分类 (基于 Agent 描述选择)
+- [x] **2.1.1** 实现 LLM 结构化输出的意图分类 (基于 Agent 描述选择)
   - 文件: `backend/app/graph.py` (增强)
   - 工作量: M (4 小时)
   - 依赖: Phase 1 完成
   - 风险: 中 -- 路由准确率需要评估
-- [ ] **2.1.2** 实现多意图请求处理 ("取消订单并给我一个折扣" -> 顺序执行)
+- [x] **2.1.2** 实现多意图请求处理 ("取消订单并给我一个折扣" -> 顺序执行)
   - 文件: `backend/app/graph.py` (增强)
   - 工作量: M (6 小时)
   - 依赖: 2.1.1
   - 风险: 高 -- 多意图原子性问题 (全部成功 vs. 部分失败升级)
-- [ ] **2.1.3** 实现歧义意图处理 (无法分类时询问澄清问题)
+- [x] **2.1.3** 实现歧义意图处理 (无法分类时询问澄清问题)
   - 文件: `backend/app/agents/fallback.py` (增强)
   - 工作量: S (2 小时)
   - 依赖: 2.1.1
   - 风险: 低
-- [ ] **2.1.4** 编写路由测试 (正确路由、多意图、歧义、回退)
+- [x] **2.1.4** 编写路由测试 (正确路由、多意图、歧义、回退)
   - 文件: `backend/tests/test_routing.py`
   - 工作量: M (4 小时)
   - 依赖: 2.1.1, 2.1.2, 2.1.3
@@ -312,12 +315,12 @@ Smart Support 是一个 AI 客服行动层框架。核心价值主张: "粘贴
 
 #### 2.2 Mock 折扣 Agent (预计 0.5 天)
 
-- [ ] **2.2.1** 创建 Mock 折扣 Agent + 工具 (apply_discount, generate_coupon)
+- [x] **2.2.1** 创建 Mock 折扣 Agent + 工具 (apply_discount, generate_coupon)
   - 文件: `backend/app/agents/discount.py`
   - 工作量: S (2 小时)
   - 依赖: Phase 1
   - 风险: 低
-- [ ] **2.2.2** 更新 agents.yaml 添加折扣 Agent 配置
+- [x] **2.2.2** 更新 agents.yaml 添加折扣 Agent 配置
   - 文件: `backend/agents.yaml`
   - 工作量: S (30 分钟)
   - 依赖: 2.2.1
@@ -325,17 +328,17 @@ Smart Support 是一个 AI 客服行动层框架。核心价值主张: "粘贴
 
 #### 2.3 中断超时处理 (预计 1 天)
 
-- [ ] **2.3.1** 实现 30 分钟 TTL 自动取消机制
+- [x] **2.3.1** 实现 30 分钟 TTL 自动取消机制
   - 文件: `backend/app/interrupt_manager.py`
   - 工作量: M (4 小时)
   - 依赖: Phase 1 (interrupt 基础)
   - 风险: 中 -- 定时器精度和状态一致性
-- [ ] **2.3.2** 实现过期后重试提示 (重新评估当前状态后重新发起)
+- [x] **2.3.2** 实现过期后重试提示 (重新评估当前状态后重新发起)
   - 文件: `backend/app/interrupt_manager.py` (扩展)
   - 工作量: M (3 小时)
   - 依赖: 2.3.1
   - 风险: 中
-- [ ] **2.3.3** 编写中断超时测试
+- [x] **2.3.3** 编写中断超时测试
   - 文件: `backend/tests/test_interrupt.py`
   - 工作量: S (2 小时)
   - 依赖: 2.3.1, 2.3.2
@@ -343,17 +346,17 @@ Smart Support 是一个 AI 客服行动层框架。核心价值主张: "粘贴
 
 #### 2.4 Webhook 升级 (预计 1 天)
 
-- [ ] **2.4.1** 实现 Webhook 升级模块 (HTTP POST 到配置的 URL, 包含完整对话上下文)
+- [x] **2.4.1** 实现 Webhook 升级模块 (HTTP POST 到配置的 URL, 包含完整对话上下文)
   - 文件: `backend/app/escalation.py`
   - 工作量: M (3 小时)
   - 依赖: Phase 1
   - 风险: 低
-- [ ] **2.4.2** 实现 Webhook 重试机制 (指数退避, 最多 3 次)
+- [x] **2.4.2** 实现 Webhook 重试机制 (指数退避, 最多 3 次)
   - 文件: `backend/app/escalation.py` (扩展)
   - 工作量: S (2 小时)
   - 依赖: 2.4.1
   - 风险: 低
-- [ ] **2.4.3** 编写 Webhook 测试 (成功发送、目标不可达、重试)
+- [x] **2.4.3** 编写 Webhook 测试 (成功发送、目标不可达、重试)
   - 文件: `backend/tests/test_escalation.py`
   - 工作量: S (2 小时)
   - 依赖: 2.4.1, 2.4.2
@@ -361,22 +364,22 @@ Smart Support 是一个 AI 客服行动层框架。核心价值主张: "粘贴
 
 #### 2.5 垂直行业模板 (预计 0.5 天)
 
-- [ ] **2.5.1** 创建电商模板 YAML (订单查询、订单操作、折扣)
+- [x] **2.5.1** 创建电商模板 YAML (订单查询、订单操作、折扣)
   - 文件: `backend/templates/e-commerce.yaml`
   - 工作量: S (1 小时)
   - 依赖: 1.2.2
   - 风险: 低
-- [ ] **2.5.2** 创建 SaaS 模板 YAML (账户查询、订阅管理、计费)
+- [x] **2.5.2** 创建 SaaS 模板 YAML (账户查询、订阅管理、计费)
   - 文件: `backend/templates/saas.yaml`
   - 工作量: S (1 小时)
   - 依赖: 1.2.2
   - 风险: 低
-- [ ] **2.5.3** 创建 Fintech 模板 YAML (交易查询、争议处理)
+- [x] **2.5.3** 创建 Fintech 模板 YAML (交易查询、争议处理)
   - 文件: `backend/templates/fintech.yaml`
   - 工作量: S (1 小时)
   - 依赖: 1.2.2
   - 风险: 低
-- [ ] **2.5.4** 实现模板加载逻辑 (选择模板 -> 覆盖 agents.yaml)
+- [x] **2.5.4** 实现模板加载逻辑 (选择模板 -> 覆盖 agents.yaml)
   - 文件: `backend/app/registry.py` (扩展)
   - 工作量: S (2 小时)
   - 依赖: 2.5.1, 2.5.2, 2.5.3

docs/phases/phase-2-dev-log.md

@@ -0,0 +1,76 @@
+# Phase 2: Multi-Agent Routing + Safety Layer -- Development Log
+
+> Status: COMPLETED
+> Phase branch: `phase-2/multi-agent-safety`
+> Date started: 2026-03-30
+> Date completed: 2026-03-30
+> Related plan section: [Phase 2 in DEVELOPMENT-PLAN](../DEVELOPMENT-PLAN.md#phase-2-多-agent-路由--安全层-第-3-4-周)
+
+## What Was Built
+
+- **Intent Classification** (`app/intent.py`): LLM structured output-based intent classifier with Pydantic models (`IntentTarget`, `ClassificationResult`). Supports single-intent, multi-intent, and ambiguity detection with configurable confidence threshold.
+- **Discount Agent** (`app/agents/discount.py`): Mock agent with `apply_discount` (write + interrupt) and `generate_coupon` (read) tools. Validates discount range (1-100%).
+- **Interrupt Manager** (`app/interrupt_manager.py`): TTL-based interrupt tracking with 30-minute auto-expiration. Provides `register`, `check_status`, `resolve`, `cleanup_expired`, and `generate_retry_prompt` methods. Complements SessionManager.
+- **Webhook Escalation** (`app/escalation.py`): HTTP POST escalation with exponential backoff retry (max 3 attempts). Includes `WebhookEscalator` and `NoOpEscalator` implementations behind `EscalationService` protocol.
+- **Enhanced Supervisor Routing** (`app/graph.py`): Supervisor prompt now includes dynamic agent descriptions. Intent classifier attached to graph for use by ws_handler routing layer. Multi-intent hint injection for sequential execution.
+- **Vertical Templates**: Three industry YAML templates (e-commerce, SaaS, fintech) in `backend/templates/`.
+- **Template Loading** (`app/registry.py`): `load_template()` and `list_templates()` class methods for template-based agent configuration.
+- **WebSocket Integration** (`app/ws_handler.py`): Ambiguous intent sends clarification message. Interrupt TTL checked before resume -- expired interrupts return retry prompt. Interrupt manager registration on interrupt detection.
+
+## Code Structure
+
+New files:
+- `backend/app/intent.py` -- Intent classification models and LLM classifier
+- `backend/app/agents/discount.py` -- Discount agent tools
+- `backend/app/interrupt_manager.py` -- Interrupt TTL management
+- `backend/app/escalation.py` -- Webhook escalation with retry
+- `backend/templates/e-commerce.yaml` -- E-commerce agent template
+- `backend/templates/saas.yaml` -- SaaS agent template
+- `backend/templates/fintech.yaml` -- Fintech agent template
+
+Modified files:
+- `backend/app/graph.py` -- Intent classifier integration, dynamic supervisor prompt
+- `backend/app/agents/__init__.py` -- Registered discount tools
+- `backend/app/agents/fallback.py` -- Updated capability list
+- `backend/app/registry.py` -- Template loading methods
+- `backend/app/config.py` -- Webhook, template settings
+- `backend/app/ws_handler.py` -- Interrupt manager + intent classification integration
+- `backend/app/main.py` -- Wiring new modules, template loading, version bump to 0.2.0
+- `backend/agents.yaml` -- Added discount agent
+- `backend/pyproject.toml` -- Added httpx to main dependencies
+
+Test files added:
+- `tests/unit/test_intent.py` (11 tests)
+- `tests/unit/test_discount.py` (13 tests)
+- `tests/unit/test_interrupt_manager.py` (14 tests)
+- `tests/unit/test_escalation.py` (11 tests)
+- `tests/unit/test_templates.py` (9 tests)
+
+Test files updated:
+- `tests/unit/test_graph.py` -- Tests for classifier attachment and classify_intent
+- `tests/unit/test_ws_handler.py` -- Tests for interrupt manager and clarification flow
+- `tests/unit/test_main.py` -- Updated version check
+
+## Test Coverage
+
+- Total tests: 153 (87 Phase 1 + 66 Phase 2)
+- Overall coverage: 90.18%
+- New module coverage:
+  - intent.py: 100%
+  - discount.py: 96%
+  - interrupt_manager.py: 100%
+  - escalation.py: 100%
+  - graph.py: 100%
+  - registry.py: 97%
+
+## Deviations from Plan
+
+- Multi-intent handling uses supervisor prompt hint injection rather than a fully custom pre-routing graph node. This is simpler and leverages the existing `langgraph-supervisor` routing rather than fighting it.
+- Webhook escalation is wired to main.py app.state but not yet connected to a specific agent tool (escalation trigger). The module is ready for use -- integration with fallback agent's escalation path is straightforward but deferred to avoid scope creep.
+- The `escalate_to_human` tool mentioned in the plan was not created. The escalation module works standalone and can be triggered from ws_handler or agent tools in Phase 5.
+
+## Known Issues / Tech Debt
+
+- SaaS and fintech templates reference tool names (`get_account_status`, `change_plan`, etc.) that don't have implementations. These are configuration blueprints for future use.
+- Interrupt manager cleanup is not called on a schedule -- `cleanup_expired()` exists but no periodic task invokes it. Consider adding a background task in Phase 5.
+- `main.py` coverage is 44% due to lifespan requiring real DB connection. Integration tests would cover this.