Introduction
ARCUS (Any Repository Can Use Spec-driven development) is an agentic SDLC factory that transforms written user stories into reviewed, test-backed pull requests. Delivered as an installable agent-skills plugin for GitHub Copilot, Claude Code, VS Code, and OpenCode, ARCUS orchestrates a Spec → Code → Pull Request pipeline through one stateful orchestrator. Its default lean profile is orthogonal to the three modes (gated, intelligent, and afk); the full 5.x pipeline remains available as the opt-in thorough profile.
What is ARCUS?
ARCUS is your AI-powered software development lifecycle factory. By default it takes a markdown story through scaffold → plan → branch → task_1..N → code_review → context_sync → closure, producing tested, reviewed code and an opened pull request. Add --thorough to use the complete 5.x Brainstorm/Test Plan pipeline and its specialist review ceremony.
The system is built around a repository-agentifier and a single stateful orchestrator with two profiles (lean and thorough) and three modes:
repo-agentifier
Prepares your repository for agent-driven development by scanning its structure and building a shared context snapshot. It produces:
.context/directory — A token-efficient snapshot containing:repo_scope.md— Overview of repository structure and tech stackrepo_map.md— Navigation map of key directories and filesflows/*.md— Discovered business flows and patternstesting-patterns.md— Captured test conventions and frameworksdesign-and-coding-patterns.md— Captured design patterns, coding conventions, and a curated "Avoid" list
AGENTS.md— Navigation index for agent sessionsCLAUDE.md— Import directive to activate the context
This is a one-time setup per repository, re-run only after major restructuring or tech stack changes. See Context Engineering for how these five artifacts are built once, scoped per story, and synced on drift.
ARCUS is a three-tier capability library — atomic, plug-n-play capabilities; thin coordinators that sequence them; and one stateful orchestrator (arcus-controller) that owns the pipeline. See The Capability Library for how the pieces fit together. The orchestrator runs the same pipeline in three modes:
Gated (default) — arcus-controller
The default, user-driven mode. Start with arcus <STORY>. In lean, the default stop is after plan; in thorough, it remains after test_plan, implementation, and code_review. You can:
- Answer the open questions in your own words, all in one go
- Configure phase-boundary gates via
.arcus/config.json - Invoke stages individually —
generate test plan for <STORY>,code <STORY>,review <STORY>,close <STORY> - Pause and resume — your session checkpoint persists across agent sessions; on a cold resume, type
resume <STORY> - Answer the batched open questions — raised all at once, each presenting one Recommended option with a rationale, and you can always answer in your own words
Intelligent — arcus-controller
Surfaces only Brainstorm open questions; no phase-boundary gates ever fire. This is cloud's default behavior — every ARCUS cloud/CI run is scaffolded with --intelligent explicitly. Trigger locally with arcus <STORY> --intelligent.
AFK (autonomous) — arcus-controller
The hands-off mode of the same orchestrator. Open questions are recorded but never surfaced, so nothing ever stops. Trigger with forge <STORY>, afk <STORY>, run afk on <STORY>, or arcus <STORY> --afk.
Who is ARCUS for?
ARCUS is designed for development teams and individual engineers who want to:
- Maintain spec-driven development discipline without manual overhead
- Get AI assistance across the full SDLC, not just code generation
- Keep humans on the two decisions that matter: the open questions, and the pull request
- Produce consistently reviewed, tested, and documented changes
- Scale from small bug fixes to complex feature implementations
Next Steps
Ready to start? See the Quickstart guide to install ARCUS and run your first story.
Want to understand the pipeline stages in depth? Check out Pipeline Concepts for a detailed breakdown of how ARCUS works.