🦾 Coding Loop: Autonomous Engineering Agent

"Stop prompting your coding agents. Build loops that prompt your agents."

The standard approach to coding with AI is broken. You write a prompt, copy the code, run it, hit an error, copy-paste the error back to the model, and repeat. You are acting as the compiler, the QA engineer, and the git manager.

coding-loop is an enterprise-ready, installable agent skill that automates this entire feedback cycle. It wires your agent directly into your environment using isolated Git Worktrees, automated test runners, and an adversarial Maker-Checker subagent split.

---

⚡ Key Capabilities

• Isolated Execution (Worktrees): Spawns independent working directories for the agent to implement code. No file collisions with your working checkout.
• Fail-Fast Verification: Automatically runs your compilers, linters, and test suites. If a test fails, it self-corrects based on terminal output.
• Adversarial Security Checker: The agent writing the code is never the agent that approves it. A separate check subagent reviews the diff against your codebase security RULES.md before anything is committed.
• Token Budget Protection: Explicitly optimized for cost-effective frontier models (like DeepSeek V4) to prevent runaway infinite repair loops.

---

📂 Repository Structure

coding-loop/
├── SKILL.md                 # Procedural skill rules parsed by agents
├── README.md                # Human-facing instructions (this file)
├── scripts/
│   ├── run_loop.py          # Orchestrates worktrees, tests, and correction loop
│   └── verify_code.py       # Adversarial linter, regex, and compliance checker
└── templates/
    └── RULES.md             # Custom project rule constraints template

---

🚀 Quick Start & Installation

To install this skill into your local agent environment (such as Hermes or Claude Code):

# Register the skill with your agent CLI
agents skill install file:///Users/matthewbishop/BishopTech.dev/bishoptech-skills-for-agents/coding-loop

1. Configure Project Rules

Copy the templates/RULES.md to the root of your project:

cp coding-loop/templates/RULES.md /your/project/root/RULES.md

Edit this file to define absolute constraints (e.g., "All React components must use functional exports", "No raw SQL statements allowed").

2. Run the Loop

Run the orchestrator script, pointing it to your project root, a description of the task, and the command to run your test suite:

python3 coding-loop/scripts/run_loop.py \
  --project-path /your/project/root \
  --prompt "Fix the auth timeout token expiration bug" \
  --test-command "npm run test" \
  --max-retries 5

---

⚙️ How it Works under the Hood

[Issue Triage] ──► [Worktree Setup] ──► [Agent Edits Code]
                                                 │
                                                 ▼
[Push branch / PR] ◄── [Adversarial Check] ◄── [Test Runner]
                         (verify_code.py)

1. Isolation: The script creates a temporary branch and checkout using git worktree.
2. Execution: The agent performs code modifications based on the issue description.
3. Validation: The orchestrator executes the provided --test-command. If it returns a non-zero exit code, the logs are fed back into the agent to patch the code.
4. Checker Gate: verify_code.py performs static analysis and rule checks on the git diff.
5. Human Gate: Once verified, the worktree is cleaned up and a Git branch/PR is pushed.

---

🧠 Best Practices for Token Efficiency

Tip

Optimizing Costs

• Budget limit: Always set the --max-retries flag. The default is 5.
• Keep test suites narrow: Run specific test files (e.g., npm run test -- auth.test.js) rather than your entire integration suite to speed up the loop.
• Use cheap models for checking: Configure the orchestrator to run code generation with a frontier reasoning model, but run the checker script with a fast, high-concurrency model like DeepSeek Flash.