zypherhq/agent-orchestration-skill

This repository packages a Skill for Codex, OpenAI's agent runtime, that aims to keep a multi-agent coding session organized and bounded. The author calls it a 'root-only' skill: it is meant to be loaded by the parent session that hands work to smaller worker sessions, and the worker sessions are forbidden from loading it or spawning further children of their own. The point is to prevent runaway fan-out where a parent asks a worker for help, the worker asks two more workers, and the token bill plus the confusion grow out of control. The skill ships as a single Markdown contract file with a set of rules the root session has to follow. Examples include batching work by user flow or area of the codebase instead of one worker per file, matching reasoning effort to the ambiguity and risk of the task, classifying a failure before retrying it, and sending each worker a short 'Dispatch Packet' with only what that worker needs instead of the full plan. The recommended workflow is short. The root session classifies the task by file count, surface area, risk, ambiguity, and verification needs, then picks one of five sizes from XS to XL. Non-trivial work gets a 'run ledger' folder that records the run, and medium or larger work gets a small dependency graph that has to pass a gate before any worker spawns. Around the Markdown rules the repo ships sixteen Python helper scripts. None of them call an AI model. They include a task classifier, a batcher, a budget checker, a DAG planner, a plan gate, a dispatch compiler, a run ledger writer, a handoff validator, a handoff router that flags overlapping file ownership, a failure classifier, a test command discoverer, a quality gate runner, a worktree creator for isolated runs, a small notepad for durable learnings, and a linter for installed skills. Installation is one 'npx skills add' command pointed at the GitHub URL, and the OpenAI manifest is set so that the skill only runs when explicitly requested, not automatically.