The plan approval system now has enforcement at the tool execution level:
- After modifying more than 2 files, the agent MUST use plan_approval
- File-modifying tools (write, edit, delete, etc.) are blocked until
an approved plan exists
- Clear error message guides agent to create and submit a plan
This ensures agents can't bypass plan approval instructions.
- Added a system prompt builder to create dynamic prompts based on different modes (ask, code review, composer, debug, implement, plan, refactor).
- Introduced prompt tiers (balanced, fast, thorough) to tailor responses based on user needs
- Integrated multiple AI providers (Anthropic, Copilot, Google, Ollama, OpenAI) for flexible backend support.
- Updated agent and multi-agent services to utilize the new prompt system.
