atomic-execution

The atomic-execution skill provides a structured methodology for managing long-running software development tasks by preventing the common issue of context degradation. Designed for engineers and developers working on complex features or roadmap implementation, this agentic skill converts high-level objectives into granular, atomic XML execution plans. By isolating individual tasks, the engine ensures that the model maintains peak reasoning capabilities throughout the development lifecycle without becoming overwhelmed by the state of previous, unrelated sub-tasks.

• XML-based task decomposition: Automatically extracts atomic task blocks from ROADMAP.md or design.md specifications, ensuring clear boundaries for each unit of work.

• Verification-gated workflow: Implements a mandatory verify step within each XML task, where agents must run tests or validation scripts before marking a task as complete.

• Isolated State Management: Integrates with STATE.md to persist progress and metadata across sessions, preventing context bloating while maintaining a clear audit trail of decisions, blockers, and project status.

• Git-integrated safety: Works seamlessly with micro-commit and git-worktree-swarm, ensuring that every small success is persisted immediately into version control.

• Context optimization: Operates on a 'divide and conquer' principle, allowing the model to focus entirely on one atomic action at a time, drastically improving the quality of generated code and reducing hallucinations.

• Requires an established project structure including a ROADMAP.md or similar spec file.

• Best utilized during the implementation phase of software development to transition from design to functional code.

• Inputs include roadmap files, design specifications, and terminal commands for verification.

• Expected output is a series of validated code changes, updated project state files, and commit logs that reflect incremental progress.

• The skill is specifically designed to handle long-running phases where monolithic feature generation would otherwise exhaust the context window.