cc-sdd: Long-running spec-driven implementation for AI coding agents

Package README: English | 日本語 | 繁體中文

Turn approved specs into long-running autonomous implementation

One command installs an agentic SDLC workflow as Agent Skills: discovery, requirements, design, tasks, and autonomous implementation with per-task independent review on hosts with native subagents. Works across 8 AI coding agents, with the same 17-skill set on each.

👻 Kiro-inspired. Similar spec-driven, agentic SDLC style as Kiro IDE. Existing Kiro specs remain compatible and portable.

What's new in v3.0

cc-sdd v3.0 is a rework around Agent Skills and long-running autonomous implementation.

• /kiro-discovery as the new entry point. Discovery routes new work into one of: extend an existing spec, implement directly with no spec, create one new spec, decompose into multiple specs, or mixed decomposition. It writes brief.md and, when needed, roadmap.md, so you can resume a workstream without re-explaining scope.
• /kiro-impl for long-running autonomous implementation. When native subagents are available, each task gets a fresh implementer running TDD (RED → GREEN) behind a feature flag, an independent reviewer, and an auto-debug pass when blocked or after repeated review rejection. Otherwise, implementation and review run inline. Learnings propagate via ## Implementation Notes in tasks.md. Each iteration handles one task; recorded task state supports resumption after checking unfinished work and active workers.
• Boundary-first spec discipline. design.md now includes a File Structure Plan that drives task boundaries. Tasks carry _Boundary:_ and _Depends:_ annotations. Review and validation look for boundary violations, not just style issues.
• /kiro-spec-batch for multi-spec initiatives. Turn a roadmap into multiple specs by dependency wave, in parallel when native subagents are available, with cross-spec review to catch contradictions, duplicated responsibilities, and interface mismatches.
• Agent Skills across 8 coding agents. 17 skills per install, loaded on demand (progressive disclosure). Claude Code and Codex are stable; Cursor, Copilot, Devin Local / CLI, OpenCode, Gemini CLI, and Antigravity are in beta. Native subagents are used when available; hosts without them execute sequentially with inline review.

Full skills-mode workflow and /kiro-impl internals: Skill Reference.

Upgrading from v1.x or v2.x? See the Migration Guide.

Why cc-sdd?

cc-sdd treats the spec as a contract between parts of the system, not a master command document handed to the agent. Code remains the source of truth. Specs make the boundaries between parts of the code explicit so humans and agents can work in parallel without constant synchronization.

The bet: explicit contracts at the right granularity let AI-driven development at team scale move faster, not slower. Agents write the spec, humans approve the contract at phase gates, code is what ships.

Boundaries are not overhead. They are what lets you move freely inside while protecting the outside.

Full rationale, trade-offs, and when-to-use / when-not-to-use: Why cc-sdd? A philosophy note.

Quick Start

cd your-project
npx cc-sdd@latest

The default installs Claude Code Skills with English docs. To pick another agent or language:

npx cc-sdd@latest --codex-skills --lang ja      # Codex, Japanese
npx cc-sdd@latest --cursor-skills --lang zh-TW  # Cursor IDE, Traditional Chinese

Supports 8 AI coding agents (Claude Code and Codex stable; Cursor, Copilot, Devin Local / CLI, OpenCode, Gemini CLI, and Antigravity in beta) and 14 languages. See Supported Agents for the full list.

Then, in your agent:

/kiro-discovery <idea>

Not sure where to start? Start with kiro-discovery. It routes your request and tells you what command to run next.

Common workflows

You want to...	Skills mode
Start a new feature or product-sized idea	kiro-discovery → kiro-spec-init → kiro-spec-requirements → kiro-spec-design → kiro-spec-tasks → kiro-impl
Extend an existing system	kiro-steering → kiro-discovery or kiro-spec-init → optional kiro-validate-gap → kiro-spec-design → kiro-spec-tasks → kiro-impl
Break down a large initiative	kiro-discovery → kiro-spec-batch
Implement a small change with no spec	kiro-discovery → direct implementation

Legacy /kiro:* command modes are still available (--claude, --cursor, etc.) but are deprecated. See the Migration Guide for the upgrade path.

For larger approved task sets, run kiro-impl. Native subagents provide per-task implementation, independent review, and auto-debug when available; otherwise the workflow runs in the main context with inline review.

See It In Action

Example: build a new Photo Albums feature.

/kiro-discovery Photo albums with upload, tagging, and sharing
# discovery writes brief.md (and roadmap.md when multi-spec) and suggests the next command
/kiro-spec-init photo-albums
/kiro-spec-requirements photo-albums
/kiro-spec-design photo-albums
/kiro-spec-tasks photo-albums
/kiro-impl photo-albums
# autonomous: fresh implementer/reviewer when subagents are available; inline otherwise

Typical spec outputs (under 10 minutes):

• requirements.md: EARS-format requirements with acceptance criteria.
• design.md: architecture with Mermaid diagrams and a File Structure Plan.
• tasks.md: implementation tasks with boundaries and dependency annotations.

On a host with native subagents, /kiro-impl runs tasks with TDD (RED → GREEN) behind feature flags, independent review, and auto-debug on failure. Inline execution follows the same review protocol in the main context.

Supported Agents

Eight cur