Facility 0.12 is now on main

[adrian-lorenzo]

We have merged PR #289, rebuilding Facility's execution model around persistent story workspaces.

We chose to merge this as a large 0.x change because we want to keep Facility nimble. At this stage, the fastest way to find the right product is to put our experiments in front of real engineering teams, observe how they behave in actual delivery work, and iterate from there.

Keeping compatibility with every experimental layer would make that feedback loop slower. We would spend more time preserving architecture that has not earned its place yet, and less time learning what Facility needs to become.

This release gives us a clearer foundation to test:

• one persistent workspace and conversation per story;
• resumable Claude Code and Codex sessions;
• complete development environments with Docker, Compose and browser testing;
• repository-owned agents, models and triggers under .agents/;
• MCP as the primary automation interface, with the UI using the same lifecycle;
• direct Git and GitHub maintainer workflows;
• turn-level Git evidence and a unified delivery timeline;
• costs, budgets, observability, analytics and the GitHub mirror in the control plane.

Some experimental implementations are no longer part of the current architecture, including receipts, internal write approvals, the template registry and several separate execution services. This does not change Facility's purpose. Humans, repository gates and reviewable evidence remain central to how it approaches AI-assisted delivery.

0.12 is intentionally a clean database boundary. There is no in-place upgrade from 0.11. Anyone already running Facility can stay on the previous version while evaluating the new model.

I am now validating the release against our deployment and importing data from the previous version. That validation covers the full workflow, including real Claude Code and Codex sessions, persistent environments, previews, GitHub delivery, costs and operational data. Any defects found during that process will be handled as focused hotfixes.

This rollout and its follow-up work are owned by me. Once the deployment validation is complete, I will share what worked, what required correction and what we learned from running the new model end to end.