Pipeline Kokoro
Date Published
Overview
Pipeline Kokoro is an automated development pipeline engine that orchestrates Claude Code agents through a structured, multi-phase workflow. It breaks complex software projects into stages — from requirements gathering through production deployment — with each stage executed by specialized AI agents that follow skill-based instructions. The engine manages the full lifecycle: planning, parallel execution with git worktree isolation, automated code audits, assembly, and reporting.
Architecture
The engine is built in Rust across four crates: kokoro-core (types and state), kokoro-agent (Claude Code spawning and git operations), kokoro-engine (state machine and 6-phase orchestration), and kokoro-cli (binary entry point). All persistent state lives in flat JSON/YAML/Markdown files under .pipeline/ with no database, keeping everything inspectable and git-friendly.
The console is a TypeScript stack with an Express API server that reads pipeline state from the filesystem, and a React/Vite/Tailwind frontend organized into six sections: pipeline design, progress monitoring, review gates, knowledge base, automation, and settings. An MCP (Model Context Protocol) server enables Claude Code to interact with the pipeline directly from the coding assistant.
Key Features
Visual Pipeline Designer — Node-based editor built with React Flow for composing pipelines with drag-and-drop modules, typed edges (execution and dataflow), and configurable parameters per stage.
18 Skill Modules — Self-contained instruction sets covering scaffolding, implementation, testing, QA, security auditing, financial review, environment setup, and production deployment. Each skill is a standalone SKILL.md that agents execute.
Plan Review & Discovery Gates — AI-generated plans and discoveries are reviewed and approved before execution proceeds. Annotation support enables inline feedback that feeds back into the agent loop.
Autonomous Development Mode — Task queue with an automated runner that picks up tasks, creates branches, executes the full pipeline, and opens pull requests. Supports daemon, scheduled (cron), and one-shot modes.
Audit & Quality Assurance — Per-unit code audits with verdict tracking (approved, needs revision, blocked), automatic fix iterations, and manual override controls.
MCP Integration — 20 tools exposed via Model Context Protocol for AI-assisted pipeline management directly from Claude Code, covering state queries, command dispatch, review gates, task management, and documentation access.
Technical Decisions
Rust was chosen for the engine’s state management and orchestration logic — no async runtime needed since agents are spawned as subprocesses via the Dispatcher trait. The filesystem-as-database approach (JSON/YAML/Markdown) keeps all state inspectable, diffable, and git-friendly. The console server is stateless, reading from .pipeline/ on each request with no caching layer.
Git worktree isolation ensures parallel sub-agents never conflict — each agent works in its own branch and worktree, with the engine handling merge and assembly. All command writes use atomic file operations (write to .tmp, then rename) to prevent corruption under concurrent access.