---

name: agents description: Reference WorldArchitect agent routing priority and modal lock patterns.

Agent Routing & Modal Lock Patterns

COMPACTNESS RULE: Keep this file under 200 lines. Link to code instead of duplicating implementation details.

Error Handling: Warnings only - no assertions, retries, or default content. Log issues explicitly and let validation catch failures rather than silently backfilling missing data.

Agent Routing Priority ($PROJECT_ROOT/agents.py)

Routing order in get_agent_for_input():

1. GodModeAgent - Manual override (GOD MODE: prefix)
2. Character Creation Completion Override - Forced exit from modal
3. MODAL LOCKS - CharacterCreationAgent, LevelUpAgent (highest priority active state)
4. CampaignUpgradeAgent - State-based (ascension ceremonies)
5. CharacterCreationAgent - State-based (new character or level-up)
6. Active Combat - State-based (in_combat=True)
7. Semantic Intent Classifier - Primary routing brain (FastEmbed)
8. Explicit Mode Override - API parameter
9. StoryModeAgent - Default fallback

Modal Lock Enforcement

Critical Pattern: Stale Flag Guards

Problem: Leftover flags (e.g., level_up_pending=True) from previous modal sessions can incorrectly reactivate modals.

Solution: Check for explicit False flags that indicate intentional deactivation:

# In agents.py get_agent_for_input():
level_up_in_progress = custom_state.get("level_up_in_progress")
if level_up_in_progress is False:  # Explicit False = stale guard
    level_up_modal_active = False

level_up_pending_flag = custom_state.get("level_up_pending")
if level_up_pending_flag is False and not bool(level_up_in_progress):
    level_up_modal_active = False

Why: None (unset) vs False (explicit guard) distinction prevents modal reactivation from stale data.

Modal Scoping: Cross-Modal Flag Isolation

Problem: Character creation lock was protecting ALL custom_state flags, including level-up flags.

Solution: Scoped protection per modal type (world_logic.py _should_protect_field()):

• Character creation modal: Only protects character creation flags
• Level-up modal: Only protects level-up flags
• Prevents cross-modal interference (char creation forcing level_up_in_progress=True)

Level-Up Activation Triggers

Modal activates when ANY of these are true:

1. level_up_in_progress=True (explicit in-progress flag)
2. level_up_pending=True (pending flag from rewards)
3. rewards_pending.level_up_available=True (rewards system signal)

Prevented by stale guards: If level_up_in_progress=False or level_up_pending=False explicitly set, modal won't activate.

Finish Choice Injection (world_logic.py)

Modal exit enforcement via server-side choice injection:

# In _inject_modal_finish_choice_if_needed():
# 1. Uses pre-update state (not post-update) to prevent race conditions
# 2. Checks modal is active (no stale guards blocking)
# 3. Adds canonical finish choice as LAST item in planning_block.choices
# 4. Uses structured_fields["planning_block"] not structured_response (preserves injection)

Key fix: Use current_game_state_dict (pre-update) instead of updated_game_state_dict to avoid race condition where LLM sets level_up_in_progress=False in same turn.

Time Freezing During Modals

Level-up and character creation modals freeze game time:

# In world_logic.py should_freeze_time():
is_char_creation = not custom_state.get("character_creation_completed", False)
is_level_up_mode = custom_state.get("level_up_in_progress", False) or \
                   custom_state.get("level_up_pending", False)
return is_char_creation or is_level_up_mode

Common Issues & Fixes

| Issue | Root Cause | Fix Location | Fix | |-------|-----------|--------------|-----| | Modal bypass | level_up_pending=True didn't activate lock | agents.py:2871-2876 | Added pending flag to activation logic | | Routing inconsistency | Injection used different detection logic than routing | agents.py + world_logic.py | Unified stale flag guards | | Cross-modal interference | Char creation protected level-up flags | world_logic.py _should_protect_field() | Scoped protection per modal | | Finish choice position | LLM response overwrote server injection | world_logic.py:5414 | Use structured_fields not structured_response | | Stale reactivation | Old flags reactivated completed modals | agents.py:2864-2876 | Check explicit False flags |

Testing Patterns

Unit Tests ($PROJECT_ROOT/tests/):

• test_rev_439p_modal_bypass.py - Activation with level_up_pending=True
• test_rev_0g1y_inconsistent_detection.py - Stale flag guard consistency
• test_world_logic.py::TestModalLockFlagScoping - Cross-modal isolation

E2E Tests (testing_mcp/creation/):

• test_level_up_modal_flow_real.py - Full modal lifecycle with real LLM
• test_character_creation_agent_real_e2e.py - Character creation flows

Quick Reference

Agent matching check order:

1. State-based: matches_game_state(game_state)
2. Intent-based: Semantic classifier or matches_input(user_input)
3. Explicit: Mode parameter override

Modal debugging:

• Check level_up_in_progress, level_up_pending, rewards_pending.level_up_available
• Look for stale flags: explicit False values blocking activation
• Verify character_creation_completed not interfering with level-up

See also:

• .claude/skills/character-creation-modal-exit.md - Character creation specific patterns
• $PROJECT_ROOT/agents.py - Full implementation (lines 2739-3200)
• $PROJECT_ROOT/world_logic.py - Modal injection and protection (lines 1800-1900, 5400-5500)