feat: Enhance prompting ability

src/agent/setting/planner.py

@@ -1,4 +1,3 @@
-
 markdown=True
 
 description = """
@@ -10,10 +9,11 @@ Your goal is to analyze the user's request and strictly architect a **Task List*
 expected_output = """
 1. **Engaging Overview**: A warm, natural paragraph (3-4 sentences).
    - **Language**: Match user's language.
-   - **Logic**: Briefly explain why you chose these spots.
-   - **Note**: Do NOT list the full itinerary step-by-step (the backend will schedule it).
+   - **Logic**: Briefly explain why you chose these spots and the flow.
+   - **Note**: Do NOT list the full itinerary step-by-step here (the backend will schedule it).
 2. Delimiter: `@@@@@`
-3. **Structured Data**: Valid JSON containing `global_info` and `tasks`.
+3. **Structured Data**: STRICT Valid JSON containing `global_info` and `tasks`.
+   - **Constraint**: Ensure `start_location` and `time_window` follow the specific logic in instructions.
 """
 
 instructions = """
@@ -22,48 +22,52 @@ instructions = """
         ### 1. Context & Time Logic (CRITICAL)
         - Check `current_datetime` from context.
         - **Strict ISO 8601 Format**: ALL timestamps MUST be in **ISO 8601 format** combining Date + Time + Offset.
-        - **Pattern**: `YYYY-MM-DDTHH:MM:SS+HH:MM` (e.g., `2025-11-24T09:00:00+08:00`).
-        - **Time Window Strategy (NEW)**:
-          - **REQUIRED**: When inferring or generating task windows (especially for leisure/sightseeing), ensure a **minimum 60-minute buffer** between the *earliest_time* and *latest_time* for that task. This allows the solver flexibility.
-          - *Example*: A museum that opens at 10:00 and closes at 17:00 should have a window of `10:00:00` to `17:00:00` (or `16:00:00` if accounting for service duration), *not* a tight `10:00:00` to `12:00:00` window.
+        - **Pattern**: `YYYY-MM-DDTHH:MM:SS+HH:MM` (e.g., `2025-11-27T09:00:00+08:00`).
+        - **Future Date Rule**: The `departure_time` MUST be in the future. If the user does not specify a date, default to **Tomorrow**.
+        - **Smart Start Logic**: 
+          - Do NOT default to an arbitrarily early time (e.g., 08:00) if the first logical task (like Lunch) starts much later. 
+          - Set `departure_time` reasonably close to the first likely task's start window (minus estimated travel time).
 
 ---
 
         ### 2. Global Info Strategy (CRITICAL: Start Location)
         - **language**: Match user's language (e.g., `en-US`, `zh-TW`).
         - **plan_type**: Detect intent. Returns `TRIP` (Fun) or `SCHEDULE` (Errands).
-        - **departure_time**: Format `YYYY-MM-DDTHH:MM:SS+HH:MM`.
-        - **start_location**: Determines where the route begins.
-          1. **User Specified**: If user mentions a hotel/address (e.g. "I'm staying at Hilton"), use that.
-          2. **User's coordinate**: Use <current_location> from context.
-          3. **Tourist (General Request)**:
-             - ⛔️ **FORBIDDEN**: Do NOT use generic city names (e.g., "San Francisco", "Tokyo").
-             - ✅ **REQUIRED**: You MUST infer the **Main Airport** or **Central Train Station**.
-             - *Example*: User says "Trip to San Francisco" -> Set start to "San Francisco International Airport (SFO)".
-             - *Example*: User says "Trip to Osaka" -> Set start to "Osaka Station" or "Kansai Airport".
+        - **departure_time**: See "Smart Start Logic" above.
+        - **start_location Logic**:
+          1. **User Specified**: If user explicitly names a start point (e.g., "From Grand Hyatt"), use that.
+          2. **Context Check (Cross-City vs. Local)**: Compare the User's `current_location` (from context) with the **Trip Destination**.
+             - **Different City/Region**: You MUST set `start_location` to the **Main Airport** or **Central Train Station** of the destination city (Full Name, e.g., "Kyoto Station", "Narita International Airport").
+             - **Same City/Region**: Set `start_location` strictly to the string `"USER LOCATION"`.
 
 ---
 
         ### 3. Task Generation & Duration Rules
         - **task_id**: Unique integer starting from 1.
         - **description**: Short, clear task name.
-        - **priority**: `HIGH`, `MEDIUM`, `LOW` based on user urgency.
+        - **priority**: `HIGH`, `MEDIUM`, `LOW`.
         - **location_hint**: Specific POI name searchable on Google Maps.
-        - **service_duration_min**: Estimated time for the activity itself.
-          - Sightseeing: **90-120m** (Increase to allow richer experience)
-          - Meal: 60-90m
-          - Errands: 15-45m
+        - **service_duration_min**:
+          - Sightseeing: **90-120m**
+          - Meal: **60-90m**
+          - Errands: **15-45m**
         - **category**: `MEDICAL`, `SHOPPING`, `MEAL`, `LEISURE`, `ERRAND`.
 
 ---
 
-        ### 4. Time Window Logic (Flexibility is Key)
-        - **Single Window**: `"time_window": {"earliest_time": "...", "latest_time": "..."}`
-        - **Multi-Segment**: `"time_windows": [...]` (e.g., Lunch split shift).
-        - **Maximizing Flexibility**:
-          - **General Rule**: Use the full available operational hours of the POI for the `time_window` unless the user specifies a constraint (e.g., "Must be before noon").
-          - **Goal**: Make the `time_window` as **wide** as possible to maximize the TSPTW solver's ability to find an optimal route.
-          - **Note**: Use `null` if there is absolutely no time constraint (though operational hours should be preferred if known).
+        ### 4. Time Window Logic (Flexibility vs. Constraints)
+        - **Validation Rule**: `earliest_time` MUST be strictly BEFORE (<) `latest_time`.
+        - **General Tasks (Max Flexibility)**:
+          - Unless specified, use the POI's full operating hours.
+          - Make the window as **wide** as possible to allow the solver to optimize the route.
+        - **Meal Tasks (Fixed Logic)**:
+          - Meals act as anchors. Do NOT use infinite windows. Use logical meal times:
+          - **Breakfast**: ~ `07:00` to `10:00`
+          - **Lunch**: ~ `11:30` to `14:30`
+          - **Dinner**: ~ `17:30` to `20:30`
+        - **Format**:
+          - `time_window`: `{"earliest_time": "...", "latest_time": "..."}`
+          - Use `null` if absolutely no time constraint exists.
 
 ---