The Agent Skills Directory

[SAFE]: Autonomous Execution Framework. The skill is designed for 'long-running' development spanning hours or days without human intervention. It explicitly instructs the agent to resolve ambiguities independently and minimizes user interaction after the initial planning phase. This level of autonomy is the primary intended purpose and is supported by structured state management in the .agent/ directory.
[SAFE]: Subagent Isolation. The orchestrator dispatches subagents using the Agent tool with isolation: "worktree". This is a robust security pattern that prevents subagents from having side effects on the primary environment and isolates task-specific changes.
[SAFE]: Self-Correcting Review Cycle. The workflow incorporates a mandatory 'Architectural Review' for every milestone. The reviewer subagent is specifically prompted to identify security issues, missing error handling, and architectural inconsistencies, providing a built-in automated security check.
[SAFE]: Prompt Ingestion Surface. While the agent and its subagents read configuration files from the project repository (.agent/standards.md, .agent/implement.md), the skill mandates verification steps, including testing, linting, and type checking, before any code is merged into the main branch.

long-running-agent