repo-agent-orchestration

Repo Agent Orchestration is an unofficial community Codex Skill (MIT licensed) that keeps multi-agent repository delivery conventions inside the repository itself instead of a central service. Its stated goal is finishing repository work across appropriate contexts with clear ownership, proportional verification and recoverable handoff. The project explicitly notes it grants no permissions, unlocks no product capabilities and does not reproduce any proprietary orchestration system. The documentation frames three ownership modes — direct, delivery and architected — as descriptions of who owns decisions rather than fixed numbers of conversations. Guidance covers delegating separable work, continuing dependency-ready work instead of polling unchanged status, keeping implementation and review judgments independent, and verifying actual Git identity and exclusive paths. Internal writers are described as sharing their owner's tree with disjoint scopes, while independently running App writers use separate registered repository-local worktrees. Independent design review is required by explicit user or repository requirements, or for material cross-cutting or irreversible risk. Installation is performed from the source checkout with scripts/install_repository.py pointed at a repository path. It copies the Skill into .agents/skills/repo-agent-orchestration and idempotently updates a marked block in the repository's root AGENTS.md, preserving rules outside that block. A --dry-run flag previews writes, and --check detects installed-file or profile drift without modifying the target. The installer accepts options covering main branch, worktree root, branch prefix, root worktree policy, task host policy, controller model policy, per-role model settings, shared integration paths, continuity policy and external gates. Model defaults are app_default, which omits model and thinking overrides; upgrades preserve existing explicit bindings except a retired installer default that migrates back to app_default. An optional structured adapter supports explicitly requested App tasks where the host supports a saved-project, local route. Its constructor and validator handle eight packet kinds — binding, write, review, update, design_handoff, delivery_update, design_reopen and design_decision — sharing a schema. The documentation states the adapter checks path containment, saved-project identity, current Git branch and commit, read-only review and preserved report settings, but does not prove task authorization, actual delivery, review quality or acceptance, and that a report's DELIVERY field is an intended route rather than a receipt. Validation relies on Python's standard-library unittest discovery, an example validator and a local demo, with CI targeting Python 3.11 and 3.13; Git is required for installer and worktree tests. The demo creates a temporary Git repository and isolated writer trees and proves local Git and contract checks only, not App creation, message delivery, effective models or real agent behavior. Behavioral cases are supplied for exercises that string assertions cannot prove. Cleanup guidance requires an explicit decision for every owned worktree and branch, removing only validated clean, integrated or explicitly abandoned residue through Git's worktree flow, and retaining unsafe or useful residue with exact coordinates; no age-based cleanup, forced deletion of recoverable work, remote-branch deletion, background daemon or automatic permission escalation is included. The layout separates the installable Skill, installer, example validator, local demo, behavioral inputs, an optional desktop runbook and deterministic tests.