CursorAI/Idea/Proj Rules GitHub CLI AI SLDC/MCP/Github Action Reasoning Model Planner

• Using #GitHub Action create plans and triage project status and backlog with a #LLM Reasoning Model using #MCP

  • Context

    • The parent idea, CursorAI/Idea/Proj Rules GitHub CLI AI SLDC/MCP describes how #GitHub/MCP could enable Cursor to utilize GitHub project planning functionalities. This would enable it to plan what to do and possibly keep an agent scratchpad in the issue comments whose history is more idiomatically recorded in the repository, the way that a real software engineer would.

  • Impetus

    • This idea extends the parent idea farther. Many Cursor power users swear that using a reasoning model to create the action plan is an essential first step to effectively using Cursor; see also CursorAI/Technique/Frontload Design. As of early 2025, it is a bit awkward to use reasoning models in Cursor.
      • It's not possible to use one's own API key to use OpenAI reasoning models:
        • CursorAI/Settings/API Keys/OpenAI API Keys

          • https://docs.cursor.com/settings/api-keys#openai-api-keys

          • OpenAI’s reasoning models (o1, o1-mini, o3-mini) require special configuration and are not currently supported with custom API keys. Note: this can cost more than pro and won't work for custom model features.
            

      • Also, using reasoning models requires usage-based pricing, which isn't feasible for every person or team; in particular, usage-based pricing may be disabled at the team-level:
        • CursorAI/Usage-Based Pricing/For Premium Models

          • As of early 2025, the premium models from CursorAI/Settings/Models only includes two reasoning models without CursorAI/Usage-Based Pricing:
            • • OpenAI/Model/o1/Mini
                • Provider: OpenAI
                • Premium: ✅
                • Agent: ❌
                • Pricing: $0.10
                • Note: 10 requests/day included on paid plan
            • • OpenAI/Model/o3/Mini
                • Provider: OpenAI
                • Premium: ✅
                • Agent: ✅
                • Pricing: Counts as 1/3 fast request
                • Note: This is the only reasoning model included in premium available to CursorAI/Agent
            • Note that out of these two of these, only o3-mini is able to be used in Agent mode, and according to LangChain/Blog/25/02/Benchmarking Single Agent Performance,
              • #o3-mini performs as well as #o1 and #Claude 3.5 Sonnet with smaller context, but sees a steeper performance drop as context grows
            • so we should expect tool calling accuracy, which, of course, includes #MCP tool calls, when the context grows, a situation which is typical in #CursorAI Agent Mode.

    • Idea - develop a GitHub Action which uses MCP to interact with GitHub

      • As a result of the above, we could develop a GitHub Action which is itself an MCP client that could utilize a #GitHub MCP server in the action itself to develop more advanced planning abilities. For #Example -
        • an execution plan for each issue
        • filing issues or sub-tasks
        • organizing the project backlog to sequence work
      • An API key with access to a reasoning model could be used to execute planning in the issues.
      • A Github Scheduled Workflow #cron could "wake up" and do Backlog Grooming on the project
      • When a git commit is pushed, it could trigger a workflow that would execute a "plan" with an LLM Reasoning Model against the Best Next Action, storing an Intermediate Work Product somewhere in the GitHub Project, such as in an Issue Comment
      • The local CursorAI model could use its own access to MCP to access the Intermediate Work Product left in the project to start from the instructions the LLM Reasoning Model left.

    • The difference between System 1 vs System 2 Thinking may be mirrored in fast, intuitive Local Cursor usage vs slow, CICD-initiated use of LLM Reasoning Models to conduct advanced Backlog Grooming and initiate CursorAI/Technique/Frontload Design

      • Inference Time Compute cost and latency grows with the length of the Chain of Thought. It may eventually be structurally important to bifurcate when and where we use reasoning models.
        • The cost structure of reasoning models encourage Cursor and likely also other AI Coding Tools to develop licensing models that restricts the use of reasoning models in their agent mode.
      • But there remains an important need to develop a well-reasoned plan on what to do. CICD may have a role in doing some System/2 Thinking pre-work and sequencing to help the System/1 Thinking CursorAI/Agent make the most of its resources.