Test 01 / Software · In development
dev-orchestrator
Open development. Bounded autonomy. An open-source workflow kit you copy into a repository, carrying product intent through to a merge-ready pull request while the owner stays on product decisions and the merge.
About the project
Coding agents can increasingly write the software. The harder problem is trusting them to keep working without becoming their full-time operator.
The project now has two layers. A harness-agnostic kit you drop into any repository: an AGENTS.md template, a workflow contract with scripts and git hooks, role procedures written as skills, and thin command wrappers. And Dev Orchestrator itself, which moves that contract out of prompts and into code.
The flow is a short list of commands. Shape the idea into an MVP document, derive the supporting docs, plan a milestone, open the issues, then implement one issue at a time. The owner approves the documents, the issue list and the final merge.
What makes it trustworthy is that the rules are enforced rather than requested. Issue gates are machine-readable and parsed by scripts. Test-first boundaries are held by agent edit permissions and by check scripts, not by instruction. The review loop is capped at three rounds and reviewers re-run the checks themselves. Implementation ends in exactly one explicit state, including blocked on the owner, so "finish no matter what" is no longer an option.
Autonomy is deliberately bounded. Agents can make routine engineering decisions inside an approved scope, but material product, architecture and security decisions stay with the owner. The final merge stays human too. The goal is to move the human up a level, from operating coding agents to working primarily on product, design and consequential decisions.
The project is being built as open-source software. The first release is local and CLI-first, using GitHub with OpenCode and Codex, and additional coding harnesses and models are planned as the core workflow proves itself.