agent-loop

You can automate GitHub issue implementation by using an autonomous loop that claims issues, executes work in isolated worktrees, and runs multi-stage adversarial code reviews. This handles the end-to-end lifecycle from implementation to validation without manual intervention.

Multi-stage code review works by orchestrating a convergent, multi-pass review process using Codex and Claude to validate and refactor code against repository standards. It ensures strict dependency gating and inline finding traceability before any code is published.

Yes, you need git, gh, jq, python3, and timeout installed. These dependencies are required to manage worktree isolation, execute local workers, handle issue tracking, and enforce bounded convergence during the review process.

Yes, you can process multiple GitHub issues in a single run by specifying the issue numbers and a maximum iteration count. The autonomous worker sequentially claims, implements, and validates each issue within isolated worktrees.

Worktree isolation is the practice of executing local worker implementations in separate git worktrees. It prevents branch conflicts and ensures that automated code review and validation processes remain non-mutating and independently verifiable.

You should avoid using an autonomous implementation loop for issues requiring subjective architectural decisions or manual design input. The loop is designed for high-velocity, high-volume backlogs requiring strict dependency gating and bounded convergence, not open-ended design tasks.