Agent Arche

Agent Arche is a Codex skill harness that turns repeatable engineering workflows into portable, project-local AI instructions. It installs into a repo with a single npx command, either as a lightweight skills-only setup or as a fuller Codex workspace with skills, safety hooks, runtime config, and an Obsidian-style memory vault. The system replaces the older “multi-agent orchestration” model with progressive-disclosure skills: Codex sees small routing hints first, then loads the full workflow only when the task actually needs it. The harness covers implementation, bug diagnosis, TDD, code review, security review, UI/design work, SEO, PostgreSQL patterns, project startup, spec/ticket generation, git workflow, and handoff writing. Its default chain is intentionally conservative: implement the requested change, bring in tests/design/security only when relevant, review non-trivial edits, and avoid automatic commits, PRs, external issue creation, or background sub-agent spawning. Agent Arche also includes a low-overhead Codex runtime layer. A SessionStart hook loads compact project context, a PreToolUse hook blocks catastrophic shell operations, and optional repo-local memory preserves durable decisions, patterns, learnings, reviews, and session notes. The result is less of an autonomous agent swarm and more of a disciplined skill system: small enough to install anywhere, opinionated enough to improve real project work, and careful about context, safety, and token cost.