autonomous-skill

You can run long-running tasks across Claude Code sessions by using headless mode to spawn child sessions in a bash loop, or hook mode to intercept session exits and feed prompts back. This coordinates sustained, unattended work.

The best way to decompose complex projects is using structured task decomposition, which creates a task_list.md with phased subtasks. This allows the deterministic task runner to coordinate sustained work across multiple sessions.

Autonomous task execution supports TDD and refactoring through lightweight Ralph-style iteration, which uses a fixed prompt to iterate tasks instead of structured subtask lists. Progress is tracked via progress.md.

Hook mode intercepts session exits by using a Stop hook to feed prompts back inside the session, creating an in-session loop. This enables continuous workflow automation without spawning external child sessions.

Yes, you need jq and perl installed as dependencies to manage multi-session task progress. These tools support the deterministic task runner and local state management required for completion detection.

Frontmatter-driven discovery is used for multi-session task coordination because it enables deterministic task execution and local state management. It ensures proper task progress tracking and completion detection across tools.