為可測性 export 三個內部函式（isSafeRepoPath、extractBalancedJSON/extractJSONText、normalizeText）。
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- Implemented role parsing and loading from markdown files, including attributes like name, side, focus, badge, color, and personality.
- Created functions to build prompts for analysis, line location, and verdicts based on roles.
- Added tests for role management functionalities to ensure correct parsing and loading of roles.
- Developed usage tracking for AI assistant interactions, including token usage and rate limits.
- Implemented functions to extract and record usage data from various LLM providers.
- Added tests for usage tracking functionalities to validate correct accumulation and reporting of usage statistics.
jiantw83