The dark software factory.
Software that writes itself — and gets smarter without you in between.
darq takes a GitHub issue and ships the merged PR — autonomously. Plan, implement, review, fix, merge, then Scenario-driven AI Testing (SAT) the result through three personas (junior / senior / maintainer). Every run extracts patterns into a learning store; the next run starts smarter.
cargo install darqRequires an ACP-compatible coding agent on your PATH (default: opencode).
Only tested on Linux (Ubuntu 25). macOS and Windows are not supported in v0.1.
# Start the daemon (one time)
darq daemon start
# Run a full pipeline against a real issue
darq run issue 42 --full
# Watch it live in the TUI
darq tuiissue → plan → implement → review ⇄ fix → merge → SAT → learn → next run
Seven workflow stations, one autonomous loop. The agent calls tools, opens a PR, and waits for itself to merge. After merge, three persona-driven judges score the result. Patterns flow into a vector store; future runs retrieve and inject them at plan time.
- SAT (Scenario-driven AI Testing) — quality gate is persona-based judgement, not just
cargo test. Three judges, blended score, threshold-gated verdict. - Learning loop —
ruvector+sona_engineextract reusable patterns from every successful run. Future plans cite them. - Alien-living TUI — 30 Hz oscilloscope shows the agent's heartbeat, breathing, and tool-use bursts as a live waveform. Three readable shapes — idle / thinking / producing. Built on ratatui.
- Daemon architecture — long-lived process owns DB, broadcaster, workflow engine. CLI commands are thin clients over a Unix socket.
Two-crate workspace:
darq-core— workflow engine, SAT, ruvector, sona, broadcasterdarq— daemon, CLI, TUI (binary crate)
Built on Rust 2024, ratatui, tokio, rusqlite (bundled), serde_yml.
For more depth, see docs/ and the inline AGENTS.md files per crate.
Licensed under MIT.
Built by @DogaOztuzun at dark-builders.