executor_skill

Overview

This skill executes GSD plans atomically with per-task commits, deviation handling, checkpoint protocols, and state management. It runs plans, enforces one-task-one-commit discipline, pauses at checkpoints or auth gates, and produces SUMMARY.md and updated STATE.md artifacts. It is built to continue safely from checkpoints and to document any automatic deviations applied during execution.

How this skill works

The executor loads the project state (.gsd/STATE.md) and the provided PLAN.md, parses frontmatter, tasks, verification and success criteria, then chooses an execution pattern: fully autonomous, checkpointed, or continuation. It runs each task according to its type: auto tasks are executed and committed, checkpoint tasks stop and return a structured checkpoint payload. Deviations (bugs, missing critical features, blockers) are handled automatically per defined rules or escalated as architectural checkpoints. After completion it writes SUMMARY.md, updates STATE.md, and records commit hashes for each task.

When to use it

• Automate and execute a PLAN.md with clear tasks and verifications
• Run multi-step workflows that require atomic per-task commits and traceability
• Enforce safety by stopping at visual/decision/auth checkpoints
• Resume work after a previously reached checkpoint using continuation agents
• Document and track automatic fixes and deviations for audits

Best practices

• Provide a complete PLAN.md with frontmatter and explicit checkpoint tasks when human input is required
• Keep task scope narrow so each task maps to a single commit
• Include verification commands and success criteria for every task
• Add only necessary context files in plan to avoid loading unrelated history
• When an architectural change is required, expect a decision checkpoint rather than an automatic change

Example use cases

• Deploy a multi-step feature rollout where each step must be committed and verified
• Fix and harden code paths discovered during execution (auto-apply bugs and critical fixes) and record them in SUMMARY
• Run QA checkpoints that require human visual verification before continuing
• Resume an interrupted plan after authentication or decision checkpoints by verifying prior commits and continuing from the resume point
• Execute CI-like maintenance tasks across phases with clear progress and state tracking