RooPrompts/currentContext.md

Roo Modes Enhancement Project: Current Context

This document serves as a scratchpad and progress tracker for defining and refining Roo Code modes.

Mode Development To-Do List

Default Modes to Enhance:

• Code Mode
• Ask Mode
• Debug Mode
• Orchestrator Mode

New Custom Modes to Create:

• QA Tester Mode
• Enhanced Planning Mode
• Deep Research Mode
• Deep Thinker Mode
• Code Reviewer Mode
• ReScript Master Mode
• Haskell God Mode

Initial Analysis & Findings from roo.md

• Core Customization Mechanisms:
  • Custom Modes: Defined via slug, name, roleDefinition, groups (tool access + fileRegex), whenToUse, customInstructions. Can be global (custom_modes.json) or project-specific (.roomodes).
  • Custom Instructions: Applied globally, per-workspace, or per-mode. File-based instructions (.roo/rules/, .roo/rules-{mode-slug}/, .roorules, .roorules-{mode-slug}) offer structured ways to manage these. Precedence matters.
  • System Prompt Override (.roo/system-prompt-{mode-slug}): Advanced "footgun" feature for complete control over a mode's system prompt.
• Key Features to Leverage for Modes:
  • Tools: Access controlled by groups in mode definition. Understanding available tools (read_file, search_files, list_files, list_code_definition_names, apply_diff, insert_content, search_and_replace, write_to_file, execute_command, browser_action, MCP tools like use_mcp_tool and access_mcp_resource, workflow tools like ask_followup_question, attempt_completion, switch_mode, new_task) is crucial.
  • MCP (Model Context Protocol): Essential for modes requiring external data or actions (Deep Research, ReScript Master, Haskell God). Servers like Context7 and Playwright are mentioned. Roo can also create MCP servers (roo.md:409).
  • Context Mentions (@): Useful for providing specific context to modes (e.g., @problems, @/path/to/file.ts).
  • Shell Integration: For modes that need to run commands.
  • Sticky Models: Each mode remembers its last used AI model.
  • Boomerang Tasks (new_task tool): Orchestrator mode uses this to delegate. Could be relevant for complex modes breaking down their own work.
• User's Specific Mode Requirements (Summary):
  • QA Tester: Memory bank, code/test case analysis, test execution/writing.
  • Enhanced Planning: Triggered by repeated failures; CoT, ToT, research, step-by-step planning.
  • Deep Research: In-depth research using MCPs (Context7, Playwright).
  • Deep Thinker: Thought-oriented analysis, precursor to Enhanced Planning.
  • Code Reviewer: Understand project (memory bank), iterative file review, review.md for notes.
  • ReScript Master: Assist with large ReScript monorepo, grep/code search, use rescript-llm-full.txt.
  • Haskell God: Assist with large Haskell repo, robust file reading, grep search.

General Notes & Research Topics

• Prompt engineering best practices for AI assistants/agents.
• Examples of system prompts for similar specialized AI roles (e.g., from GitHub, articles).
• How to effectively use "HIGHEST PRIORITY" or similar emphasis in prompts.
• Optimal fileRegex patterns for different modes.
• Interaction patterns between modes (e.g., Deep Thinker -> Enhanced Planning).
• Sequential thinking MCP tool usage for planning/analysis modes.