Skip to content

SKILL.md

Latest commit

 

History

History
88 lines (72 loc) · 5.37 KB

SKILL.md

File metadata and controls

88 lines (72 loc) · 5.37 KB
Raw
name coding-loop
description Enterprise-ready autonomous coding agent loop. Integrates isolated git worktrees, incremental build/test execution, and adversarial subagent checks to deliver verified code changes with zero human intervention.

🦾 Coding Loop: Autonomous Engineering Agent Skill

This skill allows an agent to run an autonomous execution loop for resolving software tickets, bugs, or feature additions. It implements a strict Maker-Checker pattern, running the target code changes in an isolated Git worktree and validating against linters, tests, and compliance checkers before proposing a pull request.

🔁 Execution Flow

When triggered, you must move through the following 6-step loop until the stopping condition is satisfied or the retry limit (default: 5) is reached.

                  ┌─────────────────────────────┐
                  │   1. Read Project Context   │
                  │   (VISION.md / ARCH.md)     │
                  └──────────────┬──────────────┘
                                 │
                  ┌──────────────▼──────────────┐
                  │ 2. Create Git Worktree &   │
                  │   Plan Implementation       │
                  └──────────────┬──────────────┘
                                 │
                  ┌──────────────▼──────────────┐
                  │    3. Execute Code Edits    │
                  └──────────────┬──────────────┘
                                 │
                                 ├────────────────────────┐
                  ┌──────────────▼──────────────┐         │ (tests fail)
                  │    4. Run Tests & Linters   │         │
                  └──────────────┬──────────────┘         │
                                 │ (tests pass)           │
                  ┌──────────────▼──────────────┐         │
                  │   5. Adversarial Verification│◄───────┘
                  │   (Subagent Review Rules)   │
                  └──────────────┬──────────────┘
                                 │ (verifier approves)
                  ┌──────────────▼──────────────┐
                  │  6. Propose PR & Walkthrough│
                  └─────────────────────────────┘

🛠️ Step-by-Step Instructions

Step 1: Read Project Context

Locate and parse the following project-level markdown context files if they exist in the root of the repository:

  • VISION.md: The overarching product goals and core rules.
  • ARCHITECTURE.md: Technical stack details, directory structure, and design conventions.
  • RULES.md (or coding-loop/templates/RULES.md): Hard restrictions (e.g., "no raw SQL", "no external auth libraries", "use specific error boundary patterns").

Step 2: Create Git Worktree & Plan

  1. Do not make edits directly in your master checkout. Spin up a new Git worktree on an isolated branch to isolate your dependencies and working files: 
    git worktree add -b feature/agent-loop ../worktrees/agent-loop main
  2. Write a clear, step-by-step implementation plan detailing the files you will modify and the tests you will run.

Step 3: Execute Code Edits

Perform code modifications inside the worktree directory. Use precise, targeted diff tools.

Step 4: Run Tests & Linters

  1. Run linting checks and the test suite: 
    npm run lint && npm run test
# Or the equivalent for Python, Go, Rust, etc.
  2. If failures occur: Extract the error logs, formulate a debugging plan, apply fixes, and re-run Step 4. Do not proceed to Step 5 until all tests and linters are green.

Step 5: Adversarial Verification (Subagent Check)

  1. Split the "Maker" (yourself) from the "Checker" (a subagent or separate prompt context).
  2. Instantiate a checker subagent with the verify_code.py parameters. Feed it the Git diff of your worktree and the test logs.
  3. The checker must check the diff against RULES.md and verify it contains no hidden security flaws, styling violations, or structural slop.
  4. If the checker rejects: Return to Step 3 with the checker's feedback.

Step 6: Propose PR & Walkthrough

  1. Once tests pass and the checker approves, push the branch and open a PR.
  2. Generate a structured walkthrough.md detailing the changes made, the files touched, and the test results.

🚫 Guardrails & Token Optimization

  • Max Iterations: Stop and alert the human if the loop exceeds 5 iterations without clearing the test suite. Do not burn tokens infinitely.
  • Fail Fast: If a basic typecheck or syntax compilation fails, do not write more code. Halt and address the syntax error first.