modal-agent-pattern

Modal Agent Pattern - Strict Character Creation & Level-Up

Purpose: Design specification for "modal" agents that cannot be exited until the user explicitly chooses to exit. This prevents accidental escapes from character creation and level-up flows.

Problem Statement

Current Behavior (Problematic)

1. Character Creation can be bypassed: Users can start a campaign and immediately begin the story without creating a character
2. Classifier interference: The intent classifier can route away from CharacterCreationAgent to other agents (dialog, combat, etc.) based on user input
3. No explicit exit mechanism: No forced "Exit Character Creation" choice that must be selected
4. Level-up escapes: Users can exit level-up flows without completing ability score improvements, feat selection, or spell choices

Root Cause

• get_agent_for_input() routing logic allows semantic classification to override character creation/level-up state
• Priority 5 (semantic classifier) can route to Dialog, Combat, Rewards agents EVEN when Priority 4 (CharacterCreationAgent state check) should be active
• No enforcement of completion choices in the LLM prompt

Solution: Modal Agent Pattern

Core Principles

1. MODAL AGENTS Agents that implement the modal pattern:

• CharacterCreationAgent (for new character creation AND level-ups)
• LevelUpAgent (new dedicated agent for level-up flows)

2. MODAL CONSTRAINTS When a modal agent is active:

• User CANNOT leave until they explicitly select an "Exit" choice
• Semantic classifier is DISABLED (bypassed entirely)
• Only two agents can be accessed:
  • The modal agent itself (CharacterCreationAgent or LevelUpAgent)
  • GodModeAgent (always accessible via "GOD MODE:" prefix)

3. EXIT MECHANISM Two ways to exit modal agents:

• LLM-generated choice: Modal agent MUST include "Exit Character Creation" / "Exit Level-Up" as a choice in planning_block.choices
• Python-side injection: If LLM forgets to include exit choice, Python backend injects it automatically

4. ROUTING PRIORITY OVERRIDE When modal agent is active (based on game state):

• Priority 1: GodModeAgent (unchanged)
• Priority 2: MODAL AGENT CHECK (NEW - BLOCKS ALL OTHER ROUTING)
  • If character_creation_in_progress == true, ONLY route to CharacterCreationAgent or GodModeAgent
  • If level_up_in_progress == true, ONLY route to LevelUpAgent or GodModeAgent
• Priority 3+: All other routing (SKIPPED during modal flows)

Implementation Design

1. State Flags

Character Creation Modal State:

"custom_campaign_state": {
  "character_creation_in_progress": true,  // MODAL LOCK - prevents exit
  "character_creation_stage": "mechanics", // Current stage
  "character_creation_completed": false    // Final completion flag
}

Level-Up Modal State:

"custom_campaign_state": {
  "level_up_in_progress": true,           // MODAL LOCK - prevents exit
  "level_up_stage": "asi_selection",      // Current stage
  "level_up_from_level": 3,               // Level before level-up
  "level_up_to_level": 4                  // Target level
}

2. Agent Routing Logic (get_agent_for_input)

NEW Priority 2: Modal Agent Lock

def get_agent_for_input(
    user_input: str,
    game_state: GameState | None = None,
    mode: str | None = None,
    last_ai_response: str | None = None,
) -> tuple[BaseAgent, dict[str, Any]]:
    """
    Priority:
    1. GodModeAgent (Manual override)
    2. MODAL AGENT LOCK (NEW - CharacterCreation/LevelUp active)
    3. Character Creation Completion Override
    4. CampaignUpgradeAgent (State-based)
    5. Semantic Intent Classification
    6. API Explicit Mode
    7. StoryModeAgent (Default)
    """

    # 1. God Mode (always accessible)
    if GodModeAgent.matches_input(user_input, mode):
        return GodModeAgent(game_state), {...}

    # 2. MODAL AGENT LOCK (NEW)
    # This check MUST come before ALL other routing (except God Mode)
    if game_state is not None:
        custom_state = get_custom_state(game_state)

        # Check for modal agent lock flags
        if custom_state.get("character_creation_in_progress", False):
            logging_util.info("? MODAL_LOCK: CharacterCreationAgent locked - ignoring classifier")
            return CharacterCreationAgent(game_state), {
                "intent": constants.MODE_CHARACTER_CREATION,
                "classifier_source": "modal_lock",
                "confidence": 1.0,
                "routing_priority": "2_modal_char_creation"
            }

        if custom_state.get("level_up_in_progress", False):
            logging_util.info("? MODAL_LOCK: LevelUpAgent locked - ignoring classifier")
            return LevelUpAgent(game_state), {
                "intent": constants.MODE_LEVEL_UP,
                "classifier_source": "modal_lock",
                "confidence": 1.0,
                "routing_priority": "2_modal_level_up"
            }

    # 3. Character Creation Completion (check if user explicitly exited)
    # ... existing completion logic ...

    # 4-7. All other routing continues as normal
    # ... existing routing logic ...

3. LLM Prompt Instructions

Updated character_creation_instruction.md:

Add new section after "Core Principle":

## CRITICAL: Modal Agent Constraints

**YOU ARE IN A MODAL DIALOG - USER CANNOT LEAVE WITHOUT EXPLICIT EXIT**

While character creation is in progress:
1. **CLASSIFIER IS DISABLED**: User input will NOT be routed to other agents
2. **ONLY TWO AGENTS ACCESSIBLE**:
   - CharacterCreationAgent (you)
   - GodModeAgent (via "GOD MODE:" prefix only)
3. **REQUIRED EXIT CHOICE**: You MUST provide an explicit exit choice in EVERY response:
   - "Exit Character Creation" (complete and return to story)
   - OR "Continue Creating Character" (stay in this mode)

**MANDATORY CHOICE FORMAT:**
```json
"planning_block": {
  "choices": {
    "option_1": {
      "text": "Finish Character Creation",
      "description": "Complete character and start the adventure"
    },
    "option_2": {
      "text": "Continue Editing",
      "description": "Make more changes to the character"
    }
  }
}

If you forget to include the exit choice, the Python backend will inject it automatically.

Why This Matters

• Users cannot accidentally escape character creation by saying things like "let's start the adventure" before character is ready
• Semantic classifier cannot route dialog-like queries ("tell me about my character") to DialogAgent
• Combat-like queries ("what weapons do I have?") stay in character creation context
• Level-up flows cannot be interrupted mid-ASI selection

**New dedicated prompt: `level_up_instruction.md`:**

```markdown
# Level-Up Mode - Strict Modal Flow

**Purpose:** Process level-ups with full D&D 5e rule compliance. User CANNOT leave until level-up is complete.

## CRITICAL: Modal Agent Constraints

**YOU ARE IN A MODAL DIALOG - USER CANNOT LEAVE WITHOUT EXPLICIT EXIT**

[Same modal constraints as character creation]

## Level-Up Flow

### Step 1: Announce Level-Up
Present level increase, XP totals, and what will be gained.

### Step 2: Hit Points
- Roll hit die OR take average
- Add Constitution modifier
- Update hp_max

### Step 3: Class Features
Present new features gained at this level.

### Step 4: Ability Score Improvement (ASI) or Feat
**MODAL CHECKPOINT**: User MUST make a choice:
- Increase ability scores (+2 to one, or +1 to two)
- Select a feat

**REQUIRED CHOICES:**
```json
"planning_block": {
  "choices": {
    "option_1": {"text": "Increase Ability Scores", "description": "..."},
    "option_2": {"text": "Choose a Feat", "description": "..."},
    "option_3": {"text": "Exit Level-Up", "description": "Complete level-up and return to story"}
  }
}

Step 5: Spellcasting (if applicable)

Process new spell slots, spells known, cantrips.

Step 6: Completion

Once all choices made, present summary and REQUIRE exit choice:

"planning_block": {
  "choices": {
    "option_1": {"text": "Complete Level-Up", "description": "Finish and return to adventure"},
    "option_2": {"text": "Review Changes", "description": "Double-check selections"}
  }
}

### 4. Python-Side Exit Choice Injection

**New utility: `$PROJECT_ROOT/modal_agent_utils.py`**

```python
"""
Modal Agent Utilities - Exit Choice Injection

This module ensures modal agents (CharacterCreation, LevelUp) always
provide explicit exit choices, even if the LLM forgets to include them.
"""

from mvp_site import logging_util

EXIT_CHOICE_CHAR_CREATION = {
    "text": "Exit Character Creation",
    "description": "Complete character creation and start the adventure"
}

EXIT_CHOICE_LEVEL_UP = {
    "text": "Exit Level-Up",
    "description": "Complete level-up and return to the story"
}

def inject_exit_choice_if_missing(
    response_dict: dict,
    agent_type: str  # "character_creation" or "level_up"
) -> dict:
    """
    Inject exit choice into LLM response if missing.

    Args:
        response_dict: LLM response dictionary with planning_block
        agent_type: Type of modal agent ("character_creation" or "level_up")

    Returns:
        Modified response_dict with exit choice guaranteed
    """
    # Get planning_block
    planning_block = response_dict.get("planning_block", {})
    choices = planning_block.get("choices", {})

    # Determine which exit choice to use
    exit_choice = (
        EXIT_CHOICE_CHAR_CREATION
        if agent_type == "character_creation"
        else EXIT_CHOICE_LEVEL_UP
    )

    # Check if any choice matches exit intent
    has_exit_choice = False
    exit_keywords = ["exit", "finish", "complete", "done", "return"]

    for choice_key, choice_data in choices.items():
        if isinstance(choice_data, dict):
            text = choice_data.get("text", "").lower()
            if any(keyword in text for keyword in exit_keywords):
                has_exit_choice = True
                break

    # Inject exit choice if missing
    if not has_exit_choice:
        logging_util.warning(
            f"? MODAL_AGENT_INJECTION: LLM forgot exit choice for {agent_type}, "
            f"injecting automatically"
        )

        # Find next available option number
        existing_keys = [k for k in choices.keys() if k.startswith("option_")]
        if existing_keys:
            max_num = max(int(k.split("_")[1]) for k in existing_keys)
            new_key = f"option_{max_num + 1}"
        else:
            new_key = "option_1"

        # Inject exit choice
        choices[new_key] = exit_choice
        planning_block["choices"] = choices
        response_dict["planning_block"] = planning_block

    return response_dict

Integration point in world_logic.py or response handler:

# After LLM response parsing, before returning to user
if current_agent_type == "character_creation":
    response_dict = inject_exit_choice_if_missing(
        response_dict,
        "character_creation"
    )
elif current_agent_type == "level_up":
    response_dict = inject_exit_choice_if_missing(
        response_dict,
        "level_up"
    )

5. Frontend Changes (Optional Enhancement)

UI-level enforcement:

• When character_creation_in_progress == true, disable "Send" button unless:
  • Input matches exit choice text, OR
  • Input is "GOD MODE:" prefix
• Show modal dialog backdrop to indicate locked state
• Display "You are in Character Creation Mode" banner

Edge Cases & Handling

Case 1: User tries to escape with dialog-like input

Input: "I want to talk to the bartender" Current behavior: Classifier routes to DialogAgent New behavior: Modal lock returns CharacterCreationAgent, responds with:

[CHARACTER CREATION MODE - Active]

You're currently creating your character. Before you can interact with the world,
we need to finish building your character.

Would you like to:
1. Continue Character Creation
2. Exit Character Creation (if character is ready)

Case 2: User tries combat-like action

Input: "I attack the goblin" Current behavior: Classifier routes to CombatAgent New behavior: Modal lock returns CharacterCreationAgent, responds with:

[CHARACTER CREATION MODE - Active]

You're eager to jump into action! But first, let's finish creating your character
so you have the stats and abilities you'll need in combat.

Current Progress: [stage summary]

Ready to exit character creation?
1. Finish Character Creation (if ready)
2. Continue Building Character

Case 3: LLM forgets exit choice

LLM response: Only provides character customization choices, no exit option Python injection: Automatically adds "Exit Character Creation" as final choice User sees: All LLM choices + injected exit choice

Case 4: Level-up with multiple stages

Scenario: Level 4 → Level 5 requires ASI + new spell selection Behavior:

• Turn 1: ASI selection (choices: +2 to one, +1 to two, choose feat, exit)
• Turn 2: Spell selection (choices: spell A, spell B, spell C, exit)
• Turn 3: Review summary (choices: confirm, review again, exit)
• Modal lock persists until user selects exit choice

Case 5: God Mode Override

Input: "GOD MODE: skip character creation and start at level 5" Behavior: GodModeAgent takes over (Priority 1), can bypass modal lock Justification: God Mode is admin/debugging tool and should have full override

Benefits

1. Prevents accidental bypasses: Campaign ARrfJ39LhNEi5rcGq1c7 issue would be impossible
2. Clear UX: Users know they're in a modal flow and how to exit
3. Classifier isolation: Semantic classifier doesn't interfere with creation flows
4. LLM safety net: Python injection ensures exit choice always exists
5. Consistent pattern: Same approach for character creation AND level-ups

Implementation Phases

Phase 1: Core Modal Lock (Minimum Viable)

• [ ] Add character_creation_in_progress check at Priority 2 in get_agent_for_input
• [ ] Update character_creation_instruction.md with modal constraints
• [ ] Test that classifier is bypassed during character creation

Phase 2: Exit Choice Injection

• [ ] Create modal_agent_utils.py with inject_exit_choice_if_missing
• [ ] Integrate injection into response handler
• [ ] Test LLM responses with/without exit choices

Phase 3: Level-Up Agent

• [ ] Create dedicated LevelUpAgent class
• [ ] Add level_up_instruction.md prompt
• [ ] Implement level_up_in_progress modal lock
• [ ] Update routing logic for level-up flows

Phase 4: Enhanced UX (Optional)

• [ ] Frontend modal dialog backdrop
• [ ] "Character Creation Mode" banner
• [ ] Disable irrelevant UI actions during modal flow

Testing Strategy

Unit Tests

• test_modal_agent_routing.py: Verify Priority 2 modal lock
• test_exit_choice_injection.py: Verify Python injection logic
• test_character_creation_modal.py: End-to-end character creation with modal lock

Integration Tests

• Create campaign → Verify cannot escape char creation with dialog input
• Level-up flow → Verify cannot escape mid-ASI selection
• God Mode override → Verify can bypass modal lock

Edge Case Tests

• LLM forgets exit choice → Verify injection works
• Multiple level-up stages → Verify modal lock persists
• Campaign with pre-populated character → Verify review mode works

Related Files

Agent routing:

• /Users/$USER/projects/worktree_worker8/$PROJECT_ROOT/agents.py (lines 2552-3164)

Character creation:

• /Users/$USER/projects/worktree_worker8/$PROJECT_ROOT/agents.py (lines 830-1151, CharacterCreationAgent)
• /Users/$USER/projects/worktree_worker8/$PROJECT_ROOT/prompts/character_creation_instruction.md

Intent classifier:

• /Users/$USER/projects/worktree_worker8/$PROJECT_ROOT/intent_classifier.py

State management:

• /Users/$USER/projects/worktree_worker8/$PROJECT_ROOT/game_state.py

References

• CLAUDE.md: LLM decides, server executes (lines 10-30)
• Intent classifier architecture (intent_classifier.py lines 1-176)
• Agent routing priority (agents.py lines 2552-2585)