What problem does it solve? Unattended AI agents invoked by CI schedulers or cron jobs fail in ways interactive sessions never do: they stall on approval prompts, improvise deterministic steps, re-trigger on their own output, and report green runs while silently skipping work. This Skill codifies the architecture and operational rules that keep automation-invoked agents and recurring routines correct, bounded, and auditable. ## Core Features & Use Cases - Agent architecture rules: Hard-code deterministic steps in the workflow around the agent, bound the agent's write surface with post-run diff verification, converge all failure modes to one triage state, and isolate committing subagents in worktrees. - Routine folder structure: Standardizes routines as a folder with routine.md (or task.md/task.json) bracketed by preconditions.sh and postconditions.sh scripts, with rules for what belongs in prose versus scripts. - Coded invariant checks: Ships checks.mjs and declared-checks.json that enforce the prose rules, including a blocking check that flags GitHub REST token usage in MCP-only sessions and a routine-structure validator. - Use Case: When building a nightly GitHub Actions routine that dispatches an agent to triage issues, apply this Skill to structure the routine folder, wire eligibility gates, avoid self-triggering loops, and ensure failures converge to human triage. ## Quick Start Use the unattended-agents skill to review my scheduled routine folder and verify its structure, failure handling, and GitHub access patterns.
