You are an expert in using the pi coding harness -- github and source code and agentic systems in general.
I would like your help in designing a "task" system to be used from within the "pi" coding harness that allows PI and AI agents to follow a "spec-driven" development style for a given task.
Unless you recommend a different approach, I imagine this working as an agentskill which works through concrete high-level phases of definition, understanding, planning and implementation for any task based on a repeatable conversation between the agent and the user. I would like the output of this (skill definition, prompts) to be portable so that I can use it in other codebases and with other agents.
The high-level phases of working through a task:
- Task Definition - Prompt the user to describe the task itself. What is the problem? What are the expectations for when the solution has been provided? Is this task complex enough to require research, or is it small enough to get the necessary context during the planning phase?
- Research - Only needed for complex large tasks, or when working in a large messy codebase. In this phase, the agent learns as much as it needs to learn in order to plan the implementation of the task.
- Plan - Identify a concrete and detailed step-by-step implementation plan which will be used to implement the task to the user's specifications.
- Review/Revise with User - In this interactive phase, work with the user to review, revise and reword the plan to ensure that the plan is sounds, covers all requirements, and accounts for any and all concerns related to the task. This is expected to be a loop of review/revision/repeat until the plan is finalized.
- Implement - Once the plan is sound, this should require no more user attention. The agent implements the plan.
- Code Review - Once the task has been implemented and the code appears to work. Change roles to become a thorough, experienced code reviewer and offer a deep code review of the generated output.
- Respond to Code Review Suggestions - Take the output from the code review and address critical and important issues. For suggestions or lower-priority suggestions, interact with the user to decide which code review suggestions to implement. Once any changes have been made, we will loop back to step 6 and have the code reviewer review again until the user is satisfied that no more changes are needed.
- Final Polish - To a last pass to ensure that all formatting and linting are clean. Search specifically for any remaining indicators of AI-generated code and fix last issues.
Given the 8 phases above, I would like your help in creating a set of prompts for each phase. We will have a "tasks" directory, and each new task should have its own subdirectory. Each phase above should write markdown files into this directory with the output of that phase.
DO NOT IMPLEMENT YET. With the understanding above, the goal is to generate a "task-system-design-plan.md" file with a plan proposal for how to implement the above. I expect that for each of the 8 phases, there will be interaction between you and the user to put more definition around the prompts and specifics of each phase. Please interactively work with the user to define each phase above correctly.