Order Logic Skill

4. Order Suggestion — message that suggests a menu item type not already in the order

• for example, if the order has a main item, then suggest a side

5. Order Confirmation — summary text of the order items along with a request for confirmation of its correctness

• presented if the order contains at least one of each menu item type or if the user has confirmed nothing else is wanted

6. Next Step — defines what the supervisor should do next; maps directly to a tool call:

• take_order — record the parsed item and quantity
• answer_menu_question — return an answer from menu data
• get_error_response — return a predefined error message (off-topic or invalid order)
• get_non_error_response — return a predefined suggestion or confirmation message
• summarize_order — present the final order summary

7. Non Error Messages — predefined messages selected by get_non_error_response, mapped to situations such as Order Suggestions or Order Confirmations; the LLM is not called to generate these
8. Off Topic Messages — predefined messages selected by get_error_response, mapped to specific off-topic scenarios; see Off-Topic Handling below

How Order Logic Connects to the Supervisor

Every user message is first classified by the supervisor agent (src/agents/supervisor.py) into one of three intents via SupervisorDecision.intent:

Order logic only applies when intent is order_entry. The other intents bypass order logic entirely.

Order Item Validation (Error Cases)

The take_order tool enforces two hard rules. Both return an error string instead of recording the item:

1. Item not on menu — the item name (and aliases) must match an entry in menu.json. If not found, an error response is returned listing available items.
2. Quantity out of range — quantity must be between 1 and 5 (inclusive). A quantity of 0 or greater than 5 is rejected with an error.

These errors surface back through the supervisor's response. The LLM does not generate error text — errors are returned directly from the tool as strings.

Order Step Decision Tree

After take_order succeeds, the response type depends on the current state of the order:

User input classified as order_entry
    └─ take_order(item, quantity)
          ├─ ERROR: item not on menu       → get_error_response
          ├─ ERROR: quantity out of range  → get_error_response
          └─ SUCCESS: item recorded
                ├─ Order has ONLY a Main (no Side, no Drink)
                │     └─ get_non_error_response("next-step-only-main-ordered")
                ├─ Order has a Main AND a Side (no Drink)
                │     └─ get_non_error_response("next-step-main-and-side-ordered")
                ├─ Any other successful combination
                │     └─ get_non_error_response("next-step-generic")
                └─ User says nothing else is wanted
                      └─ summarize_complete_order → get_non_error_response("ending-comment")

Key rule: Predefined responses are always selected from resource files — the LLM is never called to generate the text of a final response.

Off-Topic Handling

When the supervisor classifies input as off_topic:

1. The LLM generates a response prefixed with [OFF-TOPIC] (enforced by the system prompt).
2. The track_off_topic middleware detects the prefix and increments off_topic_count in state.
3. get_error_response is called with an error_type and a level. The level is determined by how many off-topic turns have occurred in the session (counted from [OFF-TOPIC] markers in conversation history).
4. If off_topic_count reaches OFF_TOPIC_LIMIT (3), track_off_topic terminates the session and routes to __end__.

Error Level Escalation

Each off-topic type (sexual-content, prompt-engineering, not-understandable, simply-unrelated) has messages at levels 1 and 2. The third off-topic message — regardless of type — uses the universal "any" entry at level 3:

The "any" level 3 message is a final warning delivered before track_off_topic ends the session. The LLM is responsible for counting prior [OFF-TOPIC] responses in the conversation history to determine the correct level to pass.

The logged off_topic_count is always the post-increment value. See langchain-middleware-skill for middleware implementation details.

Additional Resources

full menu non-error messages project goals examples