Random Bucket System:
- Probabilistic events that trigger alongside user responses
- Random rolls before categorization to prevent AI bias
- Multiple random events can trigger simultaneously
- User bucket processed first, random events layer on top
- Metadata accumulates across all transitions
- Last transition's navigation wins
Implementation:
- activity.py: Core random bucket rolling logic
- activity_yaml_validator.py: Validation for random_buckets config
- research/guarded_ai.py: CLI simulator with random event display
- tests/unit/test_random_buckets.py: 22 comprehensive tests (all passing)
Fashion Empire Enhancement:
- activity40-fashion-empire-backrooms.yaml: Added random events to 4 zones
- fashion_emergency (5%): Urgent crises testing leadership
- creative_opportunity (10%): Breakthroughs rewarding innovation
- surprise_client (5%): VIP visitors recognizing reputation
- Random events enhance gameplay without hijacking user intent
Documentation:
- research/SPEC.yaml: Complete YAML specification with verbose comments
- All metadata operations (string concat, numeric ops, random)
- Random buckets with flow explanation
- Feedback prompts (multi-agent system)
- Processing scripts (pre_script, processing_script)
- Model overrides (classifier_model, feedback_model)
- Termination patterns and best practices
- Validation rules and examples
New Activities:
- activity-nuclear-power-plant-ai.yaml: Nuclear reactor control simulation
- activity-submarine-simulation.yaml: Deep sea exploration
- activity-unwaste-factory.yaml: Recycling facility management
Testing:
✅ All 22 random bucket tests passing
✅ YAML validation passing for all activities
✅ Deterministic triple-trigger test (100% probability)
Critical fix for guarded_ai.py:
- Add MODEL_NAME_{n} environment variable support
- Fixes hard-coded "model" string that breaks Azure OpenAI and other endpoints
- Falls back to "model" if MODEL_NAME_{n} not specified
- Some endpoints require actual deployment name in model parameter
Validator improvement:
- Allow feedback_prompts as alternative to feedback_tokens_for_ai
- Prevents false warning when using metadata_feedback_filter with new prompt system
Documentation:
- Added MODEL_NAME_{n} examples to CLAUDE.md
- Documented that Azure and similar endpoints need this variable
All 8 activities validated: 0 errors, 0 warnings
Allow activities to specify separate models for classification and feedback:
- classifier_model: Used for categorizing user responses into buckets
- feedback_model: Used for generating AI feedback and translations
Both fields can be set at activity level (defaults) and overridden at step level.
Updated activity37 to use:
- MODEL_1 (Hermes) for classification
- MODEL_3 (Qwen 3 Coder) for feedback
This allows using specialized models for different tasks, e.g., fast classification
with accurate feedback generation from domain-specific models.
Major improvements to battleship game feedback accuracy and user experience:
## New Multi-Prompt Feedback System
- Replaced single feedback with 3 specialized prompts: Shot Report, Ship Status, Game Over
- Each prompt has individual metadata filtering to see only relevant data
- Shot Report only sees hit/miss data, Ship Status only sees ship destruction data
- Added STFU token system to suppress empty messages (filtered out automatically)
## Technical Implementation
- Added per-prompt metadata_filter support in YAML structure
- Updated app.py and guarded_ai.py to handle prompt-specific filtering
- Legacy single-prompt system still works with transition-level filtering
- Added comprehensive test suite for feedback system validation
## User Experience Fixes
- Fixed TTS queue blocking JavaScript execution (async promises instead of await)
- Ship Status now correctly reports who destroyed which ship (role confusion fixed)
- Game Over only appears when game actually ends (no more random messages)
- Maintained dramatic storytelling while ensuring factual accuracy
## Battleship-Specific Improvements
- Ship destruction messages only appear when ships actually sink
- Clear separation of concerns: hits/misses vs ship destruction vs game over
- Eliminated false positive ship destruction reports
- Fixed role reversal where wrong player got credit for destruction
The battleship narrator now provides accurate, contextual feedback while preserving the dramatic naval warfare atmosphere.
- Updated validator terminal step detection to only flag truly terminal steps
- Fixed validator to accept integers and booleans in buckets (as supported by app.py)
- Fixed metadata_remove format in activity17 from dictionary to list of strings
- Added proper terminal section to activity3.yaml without questions/buckets
- Fixed missing restart transition and bucket in activity28
- Removed unused game_end transitions from battleship files
- Updated exit transitions to go directly to step_4 (goodbye step)
- Applied black formatting to validator code
All 30 activity YAML files now validate successfully with 0 errors and 0 warnings.
- Create universal activity_yaml_validator.py for validating activity configurations
- Add validation for metadata operations (metadata_add, metadata_remove, metadata_feedback_filter, etc.)
- Validate terminal steps cannot have questions or buckets
- Check Python syntax in processing_script and pre_script blocks
- Validate YAML structure, transitions, and logic flow
- Add 17 comprehensive unit tests with 100% pass rate
- Include test fixtures for validation testing
- Support both CLI and programmatic usage
