Skip to content

spec.md

Latest commit

 

History

History
66 lines (53 loc) · 3.42 KB

spec.md

File metadata and controls

66 lines (53 loc) · 3.42 KB
Raw
  • no actions visible under the model suggests the following actions
  • don't think we need an action xml tag, all of those are kind of actions

Project Initialization

  • Write a spec.md for each project to define requirements.
  • Use aider to generate structured output based on the spec.

Task Management

  • Split tasks into small, manageable pieces to increase success rates.
  • Use pytest for testing and to identify failures.
  • Include feedback from test results in prompts for better coordination.

Monitoring and Coordination

  • Use tmux for session management, starting new sessions with custom endings to avoid conflicts.
  • Display and refresh stats constantly for easy monitoring.
  • Coordinate with agent-aider-worktree without modifying it to preserve its integrity.

Planning and Execution

  • Create a plan tree using XML to structure tasks.
  • Execute tasks based on the generated plan.
  • Track task dependencies and show progress.
  • Allow the agent to modify its own plan as needed.
  • Enable the agent to indicate task completion or request further input from the user.

User Interaction

  • Implement an interactive mode with chat functionality for direct user interaction.
  • Require user confirmation for critical actions to ensure control.
  • Handle multi-line inputs properly to process pasted content effectively.

Technical Implementation

  • Stream and display reasoning with proper formatting for transparency.
  • Manage terminal output to preserve history and avoid overwriting existing text.
  • Integrate with different models and APIs for flexibility.
  • Use deepseekr1 for reasoning and evaluation of results.
  • Handle API overload issues, possibly by using native DeepSeek APIs when necessary.

Code Maintenance

  • Keep complexity low to ensure maintainability.
  • Refactor code into smaller, independent modules for easier testing and development.
  • Fix bugs related to API overload and module imports.
  • Avoid modifying existing code like agent-aider-worktree; duplicate functionality if needed.

Enhancements

  • Include system information (e.g., date, time, timezone) in the context for better awareness.
  • Implement multi-step execution to handle complex tasks over multiple interactions.
  • Maintain conversation history to provide context for the agent.
  • Add a vim-like interface for command navigation (e.g., using j, k to move through commands).
  • Implement persistent memory for the agent to retain and modify information over time.
  • Allow the agent to edit files using XML tags (e.g., search and replace with filename specifications).
  • Enable the agent to update its plan based on new information learned during execution.

Context Management

  • Format input messages to the model in XML to set a consistent example.
  • Provide an XML schema for the agent’s responses to clarify expected output structure.
  • Exclude reasoning tokens from the model’s context to avoid confusion, as they weren’t part of training.
  • Truncate shell command outputs to the last 5000 characters per command to manage context size.
  • Log command execution details (e.g., auto-run, confirmed, rejected) for valuable context.

Debugging

  • Print the complete message sent to the model for debugging purposes.
  • Remove unnecessary status messages (e.g., "generating plan", "end of reasoning") and rely on color-coding.

Alternative Approaches

  • Explore using litellm with deepseekr1 as an alternative to aider for building the agent.