specrails-core

Specrails Core is a Node.js CLI and agent workflow framework for specification-driven software development. It installs baseline roles—architect, developer and reviewer—into a project and exposes provider-native commands for Claude Code, Codex CLI, Gemini CLI and Kimi Code. The intended flow is Specification → Architecture → Implementation → Review → Delivery; the architect produces an OpenSpec proposal, design and tasks, the developer implements agreed work, and the reviewer checks results against the spec and project verification commands. Installation from a project: npx specrails-core@latest init --provider claude (choose claude, codex, gemini or kimi). Non-interactive: npx specrails-core@latest init --yes --provider claude. Providers use native commands such as /specrails:implement, $implement, and /skill:specrails-implement. Kimi integration targets Kimi Code 0.27.0+ and uses .kimi-code/specrails/run-skill.mjs for headless activation. Core 5 installs roles and workflows directly without invoking a model for an enrichment phase. Installed surfaces include provider directories (.claude, .codex, .gemini, .kimi-code), openspec/ specs and change artifacts, .specrails/config.yaml and install-config.yaml, .specrails/rules/ and agent-memory/, .specrails/runtime/ and pipeline/, optional .specrails/local-tickets.json, and .specrails/profiles/. Standalone installs normally copy committable artifacts into the repo and use a shared framework store under ~/.specrails/; relocated or Desktop-managed workspaces can keep framework links and project artifacts outside the checkout. Managed files can be regenerated by updates; user extensions should use documented paths such as .specrails/profiles/** and custom-* roles. Updates are separate from refreshing the executable: npx specrails-core@latest update --dry-run, update, doctor. Existing provider selections are preserved by the current update implementation; the current source rejects an older CLI overwriting a newer installed framework. Migrating from Core 4 removes the enrich command, quick/full installation tiers and bundled non-core specialist agents; custom agents should be moved into protected custom-* roles and profile references updated. An opt-in programmatic agent runtime is included: a shared LangGraph workflow with per-role providers, durable recovery and local OpenAI-compatible models, exposed through a TypeScript API and CLI. Standalone automatic delivery can create branches, commits, pushes and pull requests when enabled; GitHub pull requests require an authenticated gh CLI. Hosts such as Specrails Desktop can set SPECRAILS_GIT_AUTO=false and own worktrees and delivery. Multi-repository contexts must explicitly identify selected repositories and the OpenSpec/backlog owner. Configuration, specs and run state are local files, but provider CLIs send supplied context to configured model services, and package installation, GitHub/Jira and MCP integrations may use the network. Model usage is billed by the provider. The project is MIT licensed and requires Node.js 20.19.0+ with Git and at least one authenticated AI CLI; development from source uses npm ci, npm test, npm run build, typecheck and coverage commands.