State on Disk
Project vision, requirements, roadmaps, and progress live in .planning/ as structured files. Context reconstructs from disk, not from the conversation window.
型を練習する
/ˈkɑːtɑː/
Spec-driven development
for any runtime
Structured phases for Claude Code, OpenCode, Gemini CLI, and Codex.
Install from terminal
npx @kata-sh/orc@latest Scroll
Structured phases,
fresh contexts
Slash commands that structure development into discuss, plan, execute, verify.
Fresh Subagent Contexts
Each task executes in a dedicated subagent with a fresh 200k token window. Your main session stays lean while heavy work happens in focused agents.
Atomic Git Commits
Each completed task gets its own commit immediately. Clean history, easy bisect, independently revertable. Every change traces back to a plan.
Parallel Wave Execution
Plans group into waves by dependency. Independent plans in the same wave run in parallel; dependent waves run sequentially. Maximizes throughput without conflicts.
Multi-Runtime
Installs into Claude Code, OpenCode, Gemini CLI, and Codex. Same slash commands, same planning system, same output across all supported runtimes.
Plan Validation
Plans are validated against requirements before execution. A plan checker loops until the plan passes. Prevents drift before any code is written.
Discuss, plan, execute,
verify
Each step produces structured files in .planning/. Context stays fresh. Quality stays high.
01
/kata:new-project Initialize Project
Runs an interview to capture goals, constraints, and tech preferences. Optionally spawns parallel research agents. Produces scoped requirements and a phase roadmap.
For existing codebases, run /kata:map-codebase first.
Output:
PROJECT.mdREQUIREMENTS.mdROADMAP.mdSTATE.md 02
/kata:discuss-phase 1 optional Discuss Phase
Captures implementation decisions before planning. Identifies gray areas based on what's being built and asks targeted questions. Output feeds into research and planning.
Output:
{phase}-CONTEXT.md 03
/kata:plan-phase 1 Plan Phase
Researches the implementation domain, produces 2-3 atomic task plans, then validates plans against requirements. Plans are sized to fit a single context window.
Output:
{phase}-RESEARCH.md{phase}-{N}-PLAN.md 04
/kata:execute-phase 1 Execute Phase
Groups plans into waves by dependency. Independent plans run in parallel; dependent waves run sequentially. Each plan executes in a fresh subagent context. Each task gets an atomic git commit.
Output:
{phase}-{N}-SUMMARY.md{phase}-VERIFICATION.md 05
/kata:verify-work 1 optional Verify Work
Extracts testable deliverables from the phase and walks through them one at a time. Failures trigger automated debug agents that diagnose root causes and produce fix plans.
Output:
{phase}-UAT.md 06
/kata:complete-milestone Complete Milestone
Archives the milestone, updates PROJECT.md, tags the release. Start the next iteration with /kata:new-milestone.
Repeat steps 2-5 per phase →
/kata:complete-milestone → /kata:new-milestone Built for every development scenario
From new projects to legacy codebases, simple features to complex systems—Kata adapts to your workflow.
Greenfield Projects
Start from scratch with AI-powered development. Describe your project vision in natural language, and Kata creates a complete roadmap with phases, milestones, and atomic commits. Build entire applications from idea to deployment.
Brownfield Projects
Map existing codebases before adding features. Kata analyzes your architecture, patterns, and conventions, then plans new work that fits naturally. No need to start from scratch—Kata adapts to your project.
Complex Multi-Phase Features
Tackle features that span multiple sessions and files. Kata breaks complex work into manageable phases with checkpoints, maintains context across sessions, and delivers clean atomic commits for each completed task.
Systematic Debugging
Debug issues with scientific method and persistent state. Kata creates hypotheses, runs experiments, and tracks findings across debug sessions. No more losing context when debugging complex issues.
Slash commands for
every step
Installed into your runtime. Run /kata:help to see them all.
/kata: for Claude Code & Gemini · /kata- for OpenCode · $kata- for Codex
Core Workflow
/kata:new-project Interview, research, requirements, roadmap /kata:discuss-phase [N] Capture decisions before planning /kata:plan-phase [N] Research, plan, validate /kata:execute-phase <N> Execute in parallel waves, verify /kata:verify-work [N] User acceptance walkthrough /kata:audit-milestone Verify milestone definition of done /kata:complete-milestone Archive and tag /kata:new-milestone [name] Start next iteration Navigation
/kata:progress Current position and next action /kata:help All commands /kata:update Update with changelog preview Phase Management
/kata:add-phase Append phase to roadmap /kata:insert-phase [N] Insert between phases /kata:remove-phase [N] Remove and renumber /kata:list-phase-assumptions [N] Preview approach before planning /kata:plan-milestone-gaps Create phases to close audit gaps /kata:map-codebase Analyze existing codebase before planning Utilities
/kata:settings Configure model profile and workflow agents /kata:set-profile <profile> Switch quality / balanced / budget /kata:quick Ad-hoc task with Kata guarantees /kata:debug [desc] Systematic debugging with persistent state /kata:add-todo [desc] Capture idea for later /kata:check-todos List pending todos /kata:pause-work Save handoff state /kata:resume-work Restore from last session /kata:health Validate .planning/ integrity Common Questions About Kata Orchestrator
How do I install Kata Orchestrator?
Install from npm:
The installer prompts for your runtime (Claude Code, OpenCode, Gemini CLI, Codex, or all) and scope (global or local). Verify with
npx @kata-sh/orc@latestThe installer prompts for your runtime (Claude Code, OpenCode, Gemini CLI, Codex, or all) and scope (global or local). Verify with
/kata:help.What is multi-agent orchestration in Kata?
Kata uses thin orchestrators that spawn specialized agents for different tasks. When you start a project, Kata creates dedicated agents with fresh 200k context windows for research, planning, execution, and verification. Your main context stays lean while heavy work happens in focused agents. This prevents context degradation and maintains quality throughout long sessions.
How does Kata create atomic git commits?
Kata commits each completed task immediately and separately. When you complete a feature, Kata breaks it into logical units—each gets its own commit with a descriptive message. This creates clean git history that's easy to review, bisect, and revert. No more massive commits with unrelated changes.
Can I use Kata with an existing codebase?
Yes! Kata includes a codebase analysis workflow that maps your existing patterns, architecture, and conventions before planning new work. It adapts to your project structure, coding style, and tech stack. You can start using Kata on any project—greenfield or brownfield.
What does 'spec-driven development' mean?
Spec-driven development means planning before executing. Kata creates atomic task plans with dependencies and verification steps before writing any code. Plans are validated against requirements, then executed in parallel waves with fresh subagent contexts. This prevents drift during implementation.
How is Kata different from using Claude Code directly?
Without Kata, long sessions lose context and produce inconsistent results. Kata adds structure: project state on disk instead of in the context window, fresh subagent contexts per task, automatic atomic commits, and a systematic discuss-plan-execute-verify workflow. Works across Claude Code, OpenCode, Gemini CLI, and Codex.
Which coding runtimes does Kata Orchestrator support?
Kata Orchestrator works with Claude Code (
/kata: commands), OpenCode (/kata- commands), Gemini CLI (/kata: commands), and Codex ($kata- commands and skills). The same workflows and planning system works across all supported runtimes. Check out Kata Desktop for a standalone workspace, or Kata CLI for a terminal coding agent.