Atelier v0.22.0-beta

This release tightens the loop between Atelier and the agents that drive it. Recurring alerts you can quiet, platform events you can push to your own monitor, build feedback (lint and vulnerabilities) that actually reaches the agent building the app, and a tidier token store. Along the way the platform’s own use of a language model became purely optional: every build path now runs with no LLM at all.

Coming from v0.21.x? Straight upgrade — one command. Nothing you already run changes. This is a beta: the outbound event-webhook system is new and getting a soak before the next stable tag.

Quiet a recurring alert

A persistent condition — a node sitting at 85% pod capacity, say — used to re-alert on every supervisor sweep with no off switch short of disabling the whole check. Now you can snooze any alert from the Approvals → Alerts view: for an hour, a day, a week, or until it clears. A snoozed alert stops notifying but is still recorded, and an escalation to a higher severity breaks through so you never miss something getting worse. Un-quiet it any time.

Push platform events to your own monitor

New outbound event webhooks let an external monitor — your own agent, a dashboard, anything that speaks HTTP — watch Atelier and react, with no human relaying. Register a URL, a shared secret, and the event types you care about, and Atelier POSTs a signed JSON event (HMAC-SHA256 over the body, the same scheme as its inbound webhooks) when something happens: an alert fires, an app crashes, an image is held for a critical CVE. Delivery is best-effort with retries, and the secret is never returned once set. Configure it through the admin API today; a Settings panel and the full set of build/deploy lifecycle events follow. This release covers the alert-sourced events.

Build feedback that reaches your agent

When an agent builds an app by pushing code, it needs to see what the build found. Lint findings already landed in the build log; now vulnerability scan results do too — a Scan: N critical, M high line with a pointer to the full per-CVE report, so the agent (or you) can bump a base image without digging through the UI. The scan runs after deploy, so the line arrives a few seconds behind the build completing. The lint gate is also honest about skipping an unsupported stack now (Lint: no lintable languages detected) instead of going silent, so you can tell “passed” from “not run”.

The platform builds with no LLM — for real

Since Atelier stopped authoring code, a language model has been optional — and this release makes that genuinely true and clearly documented. Every build path — bring a container image, clone a repo, or push your own code — is a deterministic container build with no LLM involved. A model only unlocks the assistive features (the Nova ops agent, crash diagnosis, the supervisor’s analysis, post-build code review), and only the optional Build with AI agent authors an app from a description. We removed a leftover post-build summary that quietly spent an LLM call on every build to produce output that went nowhere, and corrected the install docs that wrongly implied you needed a model to build anything.

Token hygiene

Expired and revoked API tokens are now pruned automatically instead of accumulating in the token list — handy if you mint short-lived or per-session tokens.

Skills

The bundled agent skills (atelier-build, atelier-operate) gained guidance on reading lint and CVE feedback from the build log, and a small published manifest (skills-version.json, next to the skills bundle) lets an agent check whether its local skill copy is stale — and pull a fresh bundle only when it is — without downloading the whole archive to find out.

Upgrade

For everyone — one command:

curl -Lo upgrade.sh https://tryatelier.blob.core.windows.net/tryatelier/latest/upgrade.sh
chmod +x upgrade.sh
./upgrade.sh v0.22.0-beta

It rolls atelier-core, atelier-ui, and the MCP server pods. Your running app pods aren’t touched; folders, pins, alert channels, public-access configuration, and MCP server deployments carry over unchanged.

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