dsh-task-engine

Handing code changes to an AI is rarely about it failing to write something. It is about losing the thread: no aligned requirements, no design review, skipped verification, no one checking the commit before merge. dsh-task-engine pulls those scattered steps into a single engineering workflow workbench. The dev_task tool chains requirement review, design, development, delivery, and code audit into a verifiable, trackable task line, so every stage shows its prerequisites and deliverables on the surface instead of buried in chat history.

The workbench ships three preset flows: standard (six stages), agile light (four stages), and minimal (two stages). Pick one that matches the task size and go. You can also install your own skill folders and Markdown rules into project or personal directories and bind them to specific stages, so the AI works under your own conventions every time. The task ledger records implementation, verification, and review status in one place with keyword, stage, and risk filters. A task captures a flow snapshot at creation, so later config changes never mutate an in-flight task.

Built for engineers who develop personal projects locally with DeepSeek Harness: you want the AI to do the work, but every step should be auditable, anchored by your own rules, and left with a review trail. It does not replace your CI and does not make the final audit call for you. It simply keeps the process record and your methodology in one place so the next sprint starts without re-explaining from scratch.