Atelier v0.9.7-beta

A polish release focused on observability, control, and pipeline correctness. Four merged PRs since v0.9.6-beta.

Already on v0.9.6-beta? Skip to the Upgrade section. The full list of behaviour changes is short — only one of them runs different work by default after upgrade.

Highlights

When an app is in an error state, the red banner at the top of the app page used to have a “View logs →” link that just navigated to the Logs tab. The original intent was to package the failure context and route it to Nova for analysis — that’s what it does now. Clicking sends the platform error message + the last 50 build-log entries to Nova in a kickoff message; Nova picks up from there.

The plain “View logs” link is still there if you just want to read the transcript yourself.

Downloadable logs with a time-period selector

The Logs tab gets a Download button. Behaviour depends on the active sub-tab:

Build sub-tab — downloads the persisted pipeline transcript as <app>-build.log.
Live sub-tab — pick a window from the dropdown (1 hour / 6 hours / 24 hours / 7 days / “all available”) then click Download. Snapshots recent pod stdout/stderr into <app>-<service>-last-<N>s-pod.log.

Useful for attaching to bug reports without copy-pasting from the browser. Capped at 10,000 lines per download to keep response sizes sane.

Drag-and-drop chat attachments

You can now drop .txt, .md, .json, .yaml, or .csv files onto the docs area in the build chat — same validation as the existing click-to-pick button. A violet outline + “Drop files to attach” overlay appears while dragging. Works on existing apps, the new-app create flow, both.

Pre-build review now runs by default on Standard profile

This is the one behaviour change worth knowing about before you upgrade. Previously the LLM-based pre-build review only ran on the Hardened profile. Now it runs on Standard too (and is genuinely off on Fast).

The pre-build review is now split into two layers:

Layer	What	When it runs	Cost
Layer 1 — programmatic	Static checks: Service-selector / Deployment-label match, nginx prefix-stripping vs backend routes, app-name-prefix on nginx hosts, secretKeyRef existence, frontend code using absolute `/api` paths	Always — every profile, no toggle	Free, <100ms
Layer 2 — LLM	Cross-resource reasoning a regex can’t do (port mismatches, logical contradictions, etc.)	Standard + Hardened by default; off on Fast	~1-3¢ per build, ~5-10s latency

After upgrade, builds on Standard apps will start including a 5-10s “Running infrastructure consistency review…” step they didn’t have before. If you’d rather opt out:

Switch to Fast in Settings → Build → Pipeline Profile (you’ll lose Layer 2 + post-build code review + CVE scanning), or
Toggle “LLM pre-build review” off explicitly in Settings → Build → Code Review (keeps the rest of Standard).

If you stay on Fast, the Pipeline Profile settings panel now shows a callout listing exactly what isn’t being checked, so you know what you’re opting out of.

New “Pre-build Review” LLM role

Settings → AI → LLM Profiles now has a Pre-build Review row alongside Build / Plan / Nova / Supervisor / Code Review. Pin a cheaper/faster model here (Haiku, gpt-4o-mini) for the focused pre-build pass without affecting post-build review. Falls back to whatever you’ve got assigned to Code Review if you don’t pin a separate model — so existing setups don’t need to change anything.

Two new programmatic checks for the recurring naming-drift bug class

Layer 1 grew two new checks (zero LLM cost), targeting the chinese-homework / messaging-app class of bug where the LLM drops the random hash suffix from the app name and writes references to a non-existent Service or Secret:

Check 8 — nginx proxy_pass host must include the app-name prefix.
Check 9 — secretKeyRef.name must reference a Secret in the manifests or the platform-injected <app_name>-secrets.

The agentic build prompt and the infra-review prompt also got a clearer reminder that the app name may include a hash suffix and must be used verbatim.

Tailscale guide

A new “Remote Access → Tailscale” section in the in-app documentation (open it from the Documentation link in the sidebar settings menu, or directly at http://atelier.home.arpa/docs.html). Covers:

Installing Tailscale on the cluster node.
Two DNS strategies for making atelier.home.arpa resolve over the tailnet (per-client hosts file vs. dnsmasq + Tailscale Split DNS).
Sharing the cluster with another person via Tailscale Node Sharing.

Subnet routing for reaching the wider home LAN over the tailnet isn’t covered yet — it’s a separate cluster-side configuration (tailscale up --advertise-routes=...) that will land in a follow-up doc update.

Upgrade

From any v0.9.x-beta:

curl -Lo upgrade.sh https://tryatelier.blob.core.windows.net/tryatelier/latest/upgrade.sh
chmod +x upgrade.sh
./upgrade.sh v0.9.7-beta

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

If you’re doing a fresh install rather than an upgrade, use:

curl -Lo atelier-install https://tryatelier.blob.core.windows.net/tryatelier/latest/atelier-install
curl -Lo atelier-core-image.tar.gz https://tryatelier.blob.core.windows.net/tryatelier/latest/atelier-core-image.tar.gz
curl -Lo atelier-ui-image.tar.gz https://tryatelier.blob.core.windows.net/tryatelier/latest/atelier-ui-image.tar.gz
curl -Lo mcp-fetch-image.tar.gz https://tryatelier.blob.core.windows.net/tryatelier/latest/mcp-fetch-image.tar.gz
chmod +x atelier-install
./atelier-install

mcp-fetch-image.tar.gz is required even though you may not use MCP yourself — atelier-install tags and pushes registry.atelier.local/mcp-fetch:latest unconditionally during finalisation, so the install aborts if it’s missing.

After upgrade, full operational docs are in the in-app Documentation panel — open it from the Documentation ↗ link in the sidebar settings menu (cog icon at the bottom of the sidebar), or go directly to http://atelier.home.arpa/docs.html.

Background work (no user-visible change)

The architectural epic #374 — Agent-Workflow Pipeline was filed. The eventual direction: every pipeline stage becomes an editable workflow step with declarative prompt + tools + permissions, instead of being compiled into the platform. No user-visible change in this release — the umbrella’s first wedge lands later, building on the review-and-revise flow (#353).

Feedback

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