ouroboros

Ouroboros provides a local runtime layer for AI coding agents. Instead of relying on a single free-form prompt, it guides work through an interview, crystallizes the answers into an immutable Seed specification, executes the task, evaluates the result, and feeds evaluation output into the next iteration. The main workflow is: - Interview: Socratic questions surface assumptions and produce an ambiguity score. - Seed: answers become a specification with acceptance criteria, ontology, and constraints. - Execute: tasks run through a Double Diamond-style decomposition process. - Evaluate: a three-stage gate runs mechanical checks, semantic evaluation, and multi-model consensus. - Evolve: reflection produces the next generation; the documented convergence target is ontology similarity of at least 0.95. The project exposes an MCP server and CLI commands using the ooo or ouroboros entry points. Setup can detect installed agent hosts and register the server where supported. Compatible runtimes listed in the README include Claude Code, Codex CLI, GitHub Copilot CLI, OpenCode, Hermes, Gemini, Kiro CLI, Pi CLI, OMP CLI, Zcode, Goose, GJC, Antigravity CLI, and Grok Build CLI. DeepSeek can be used through a dsh LLM backend or through a DeepSeek Harness plugin. Ouroboros describes itself in three layers. This repository is the OS/core layer containing the Seed, Ledger, Runtime, MCP interface, and safety boundaries. A separate plugin repository provides domain workflows, while Ourocode is a terminal UI for operating workflows across multiple agent CLIs. Installation scripts are provided for macOS, Linux, WSL 2, and Windows, with Python 3.12 or newer required for normal installs. The package can also be installed with pip, pipx, uv, or Homebrew, and host-specific plugins are documented for Claude Code and Codex. A persistent ralph mode can continue the evolutionary loop across sessions by reconstructing lineage from an event store. An uninstall command removes configuration, MCP registrations, and data.