rkstack

Engineering discipline system for AI coding agents.

One plugin. Install once, adapts to your project.

Scope	What it does
Workflow	Full cycle: idea, spec, plan, implement, verify, review, ship
Safety	PreToolUse hooks block destructive commands and scope-lock edits
Detection	rkstack detect scans your stack, caches results per session
Flow types	Web projects get visual QA, screenshots, and responsive checks
Platform-agnostic	Reads CLAUDE.md for commands, works with any stack

[!IMPORTANT] RKstack enforces discipline that prevents common AI agent failures: skipping tests, guessing at root causes, claiming things work without checking, and making destructive changes without warning.

---

Quick start

# Claude Code
/install-plugin rkstack@ccode-personal-plugins

Skills activate automatically based on what you're doing. Run /setup-project once to install project-level safety guards and working rules.

---

The flow

graph LR
    A["Idea"] --> B["brainstorming"]
    B --> DR1["dual-review"]
    DR1 --> C["writing-plans"]
    C --> DR2["dual-review"]
    DR2 --> D{"Execute"}
    D -->|inline| E["executing-plans"]
    D -->|subagents| F["subagent-driven"]
    E --> G["verification"]
    F --> G
    G --> H["code-review"]
    H --> I["finishing-branch"]
    I --> J["document-release"]

    style A fill:#f59e0b,color:#000
    style B fill:#8b5cf6,color:#fff
    style C fill:#6366f1,color:#fff
    style DR1 fill:#f97316,color:#fff
    style DR2 fill:#f97316,color:#fff
    style G fill:#10b981,color:#fff
    style H fill:#0ea5e9,color:#fff
    style I fill:#0f766e,color:#fff

Each step uses test-driven-development (RED, GREEN, REFACTOR). Bugs trigger systematic-debugging (5-phase investigation, 3-strike escalation). humanizer constraints activate during all prose writing.

---

Stack detection and flow types

At session start, rkstack detect runs scc on your project, identifies every language and tool in use, and caches the result to .rkstack/settings.json. The cache looks like this:

{
  "detection": {
    "flowType": "web",
    "stack": {
      "typescript": true,
      "css": true,
      "docker": true,
      "just": true
    },
    "stats": {
      "typescript": { "files": 46, "code": 9315, "complexity": 1150 }
    },
    "services": { "supabase": true },
    "repoMode": "solo",
    "totalCode": 44761
  }
}

The stack is a flat boolean map. Everything the project contains goes here: TypeScript, Python, Terraform, Docker, Ansible, Compose, just, mise. No hierarchy, no categories. /setup-project reads this to decide which guards and rules to offer.

The flowType controls how process skills behave. Today there are two:

Flow type	When	What changes
web	TS/JS + CSS or web framework config detected	Brainstorming suggests design systems. Plans include visual verification steps. Execution screenshots after UI tasks. Verification checks responsive layouts. Code review includes visual regression. Shipping gates on QA.
default	Everything else	Standard workflow. No browser-based checks.

More flow types will come as the system grows. If detection gets it wrong, override it: the overrides section in .rkstack/settings.json takes precedence over auto-detection.

Supabase detected (.mcp.json or supabase/ directory)? Skills also verify data via MCP after browser actions, checking that what the user sees matches what the database stored.

---

Dual-review: Claude writes, Codex reviews

Specs and plans go through a multi-round review loop before you approve them. Claude self-reviews first, then Codex reviews independently against the source code. Claude evaluates each finding, fixes valid ones, rejects false positives with evidence. Rounds repeat until Codex comes back clean or max rounds are reached.