Bug fixing and debugging for ANY error, crash, loss divergence, gradient explosion, distributed hang, NaN, or unexpected behavior. Covers quick fixes and full protocol with 5-phase investigation. Trigger: 'fix bug', 'fix error', 'broken', 'crash', 'doesn't work', 'fails with', 'loss NaN', 'training hangs', 'OOM'.
Feature development with cross-module impact analysis. Covers trainer hierarchy, model adapters, reward pipeline, config system, sample dataclasses, and distributed training paths. Trigger: 'add feature', 'implement', 'refactor', 'reorganize', 'new capability'.
Complete workflow for adding a new RL training algorithm. Covers paradigm selection, TrainingArguments subclass, trainer implementation, registry, example config, and verification. Trigger: 'add algorithm', 'new trainer', 'new training method', 'implement algorithm'.
Mandatory pre-commit code review gate. Checks constraint violations, cross-module consistency, and implementation quality. Trigger proactively when changes span multiple files or touch shared infrastructure. Trigger: 'review', 'check before commit'.
Complete workflow for adding a new model adapter. Covers analysis, sample dataclass, adapter implementation (4 abstract methods + per-modality encoder overrides), registry, example YAML, and verification. Trigger: 'add model', 'support new model', 'integrate model', 'new adapter'.
Complete workflow for adding a new reward model. Covers pointwise vs groupwise design, __call__ contract, registration, YAML config, multi-reward setup, and verification. Trigger: 'add reward', 'new reward model', 'custom reward', 'scoring function'.