Atelier v0.9.9-beta

A small, focused release: the Upload Local, Clone Repo, and From Gitea (AI Build) flows now have a review-and-revise step before the build runs, matching the chat workspace’s planning loop. One PR since v0.9.8-beta.

Already on v0.9.8-beta? Skip to the Upgrade section. There are no behaviour changes that affect existing apps — the new step only applies when you create new apps via Upload / Clone / From Gitea.

Highlights

Review-and-revise for Upload, Clone, and From Gitea

Until now, New App had a planning loop where the LLM sketched what it intended to build and you could refine it via chat (“actually use Postgres”, “the backend listens on 3000 not 8080”) before clicking Build. The other source-driven flows — Upload Local, Clone Repo, and From Gitea (AI Build) — jumped straight to the build with no review step. Anything the planner got wrong meant aborting and starting over, or letting the build fail and using the auto-fix loop.

This release closes that asymmetry. When you upload, clone, or pick a Gitea repo in AI Build mode, Atelier now:

Stages the source in memory (no Gitea write yet).
Runs a planner pass over the file tree + key config files (Dockerfile, package.json, pyproject.toml, go.mod, Cargo.toml, README, etc.).
Lands you in the chat workspace with the planner’s analysis as the first assistant turn:
- What I’ll build — the services it intends to create
- Detected services — language/framework/port for each
- Build steps — the concrete commands the Dockerfile will run
- Manifests I’ll generate — the K8s resources
- Concerns and assumptions — anything it had to guess that you should confirm

You then refine the plan via chat the same way New App works. Click Build when the plan looks right and the existing build pipeline runs against the source you reviewed. The chat history is persisted on the new app so post-deploy refinements continue naturally.

”Skip review” for users who already trust their setup

Each modal has a “Review the plan before building” checkbox, default on. Uncheck it for the legacy single-step behaviour — the original endpoint still works unchanged. Direct Deploy from Gitea is always single-step regardless (it’s the explicit “no LLM involvement” path).

Cost

Review adds one planner-tier LLM round-trip per stage + one per chat refinement. With the Plan role configurable separately under Settings → AI → LLM Profiles (since #373), you can pin a cheaper/faster model here — Haiku, gpt-4o-mini, or a local LLM — without affecting the rest of the pipeline.

Behaviour change worth knowing

None for existing apps. The new flow only runs when you create new apps via Upload / Clone / From Gitea. Anything already built or scheduled stays exactly as it was.

If you’re a “build it now, fix it later” user, untick Review the plan before building in each modal once and you’re back to the v0.9.8-beta behaviour.

Upgrade

curl -Lo upgrade.sh https://tryatelier.blob.core.windows.net/tryatelier/latest/upgrade.sh
chmod +x upgrade.sh
./upgrade.sh v0.9.9-beta

~30 seconds. Running app pods aren’t restarted; only the Atelier control plane (atelier-core + atelier-ui) rolls.

What’s next

#386 — first-class offline / air-gapped install + upgrade mode: in-progress epic so the offline tarball path becomes properly versioned end-to-end. Tracked separately.
#374 — agent-workflow pipeline: the umbrella epic this PR (#353/#387) was the first wedge of. Pipeline stages will become editable workflow steps over the next few releases.

Bug reports and feedback in the testers Discord channel. Thanks for trying it.