cc-sdd is a spec-driven development harness that turns approved specifications into long-running autonomous implementation for AI coding agents.
The tool addresses the challenge of maintaining coherent, large-scale AI-assisted development by treating specifications as explicit contracts between system components rather than command documents. It works by installing an agentic software development lifecycle workflow as Agent Skills, covering discovery, requirements, design, task decomposition, and autonomous implementation. Each task runs independently with test-driven development, feature flags, dedicated review, and auto-debug passes that investigate root causes when implementation stalls. Implementation learnings propagate forward through notes in task files, allowing safe resumption after interruption.
The project suits teams building substantial features where humans need to approve architectural boundaries before agents implement. It works across eight AI coding agents—Claude Code, Codex, Cursor, Copilot, Windsurf, OpenCode, Gemini CLI, and Antigravity—with the same 17-skill set on each platform. The tool emphasizes boundary-first spec discipline, where design documents include explicit file structure plans and tasks carry boundary and dependency annotations. For multi-spec initiatives, the batch workflow catches cross-spec contradictions and interface mismatches before implementation begins. The approach assumes that explicit contracts at the right granularity enable parallel human and agent work without constant synchronization, with code remaining the source of truth.
Development activity shows sustained focus on the core workflow: the v3.0 rework centered on Agent Skills and long-running autonomous implementation, introducing the discovery entry point, the autonomous implementation command, and boundary-first discipline. The project maintains compatibility with earlier spec formats while introducing new structural patterns. Support spans multiple natural languages in documentation, and migration guidance exists for users upgrading from earlier versions. The tool carries no external dependencies and spawns subagents through each platform's native primitives.
