v0.5.0

• Fixed: a provider that cannot answer now says so in one sentence, instead of turning the room
  red. An expired key, a rate limit and a model that is no longer there all arrived the same way:
  a failed step in the work trail wearing the provider's own error text, with the engine's "Node
  execution failed for agent_reason:" in front of it. Nothing in that said whose problem it was or
  what to do — and a key you could replace in thirty seconds should never look like a crash. The
  chat now shows one calm line and at most one action, chosen by what actually went wrong: a
  rejected key offers Reconnect provider, a missing model offers Change model. A rate limit
  or an outage says the limit is theirs, with nothing to click, because there is nothing on your
  side to fix. Failures that are not the provider's — a broken flow, a tool that threw — read
  exactly as before, so the two are never confused for each other.
• New: a failure you can simply send again. Where the provider was rate-limiting, unreachable,
  or failed for a reason we could not name, the card now offers Send again, which re-sends your
  last message as a fresh turn. Never automatic, and it disarms the moment you press it. It is
  deliberately withheld where repeating cannot help, and withheld again whenever part of the turn
  had already taken effect — a page created, a brand staged, a picture made. Sending those words
  twice would do that work twice, so the card explains the absence rather than risking it.
• Fixed: an answer the model was cut off in the middle of is reported as a failure, not served as
  the answer. When a turn hit the model's output limit, the provider handed back what it managed
  to write — a sentence stopping mid-thought, or a tool call it never finished emitting, which then
  simply never ran. Nothing asked whether the turn had finished, so you could get half an answer, or
  a confident reply about work that was never done, with every step marked successful. Atelier now
  reads the provider's own finish reason and names it: "this turn was too long — start a new
  conversation, or ask for it in smaller pieces." Nothing is inferred from an answer's length, so a
  provider that reports nothing behaves exactly as before.
• Fixed: a conversation no longer gets stuck saying the operator is still working. If the
  process running a turn died part-way through — the model had answered, but the work after it never
  happened — the thread went on believing that turn was in progress and refused every later message,
  permanently, while telling you to give it a moment. The refusal now checks whether any work is
  genuinely in flight rather than trusting a flag: if nothing is running, the thread is released and
  your message goes through. A turn that really is still working is untouched, however long it takes.
  Recovery previously waited on a maintenance task that runs every few hours.
• New: every room says what it can do — and the Media room is no longer taken away. A small row
  above the composer names things in plain words: Write (titles, descriptions, copy), Describe
  (reads a picture you give it), Draw (makes one that didn't exist). Each room shows only the
  verbs its own assistant can spend, so the General studio no longer advertises pictures it has no
  tool to make. A chip you cannot use dims and says why, and the two reasons read differently: this
  chat has no tool for it (nothing to fix), or the site has no provider behind it — which, for an
  administrator, links to the setup wizard, where a key and a model are actually captured. The Media
  room used to vanish entirely when no image provider was connected, which removed a working
  conversation to protect one tool inside it; naming an image or writing its alt text from your own
  words needs no image provider at all. It now always opens, and the assistant is told the same
  facts the chips show, from the same source, so it cannot offer you a picture this install can't make.
• Fixed: a Google key now draws on its own, and an install that can't draw is told why.
  Connecting a Google AI Studio key as Gemini gave you chat and vision and an empty image picker —
  picture models were reachable only under a second provider entry, backed by the very same key, with
  nothing on the page saying so. Gemini now offers its own picture models, so one key turns on the
  Media studio's AI rail without connecting anything else. Separately, when the only thing connected
  is a shared gateway forwarding to other services, the image role no longer sits empty and silent:
  it says pictures need a provider connected directly, and names one. (Google's picture models need
  billing enabled on the key; a free key lists them and refuses at use time — which now arrives as
  one calm line rather than a red room.)
• Fixed: asking the General room to build a landing page opens the right room, carrying your
  sentence with it. Its most prominent starter ask answered with four cards and "pick a room to
  open it" — and picking Pages landed you on an empty composer, where you retyped what you had just
  written. A specific ask now gets a single card naming the room it belongs to, and opening it hands
  your own sentence back to you in that room's composer, ready to edit and send. Never sent for you,
  used once and gone. "Show me around" still shows the whole map.
• Fixed: the first rebrand of a session actually repaints the preview. Asking the design
  assistant for a new look from outside the Design System studio produced a run of "Applied to
  preview" cards over a preview still showing the old brand; reloading and asking again worked. The
  new tokens were staged before the preview existed to receive them — it now adopts whatever is
  already staged the moment it opens. A replayed card from an earlier session reads as a record
  ("staged earlier — no longer active") rather than as live state.
• Fixed: accepting "Start a new thread" after Publish no longer leaves the page unsaveable.
  Publishing and taking the offer of a fresh conversation froze the studio — Discard, Save draft and
  Publish all dead, only Archive alive — while it still claimed the conversation was wrapped up,
  right after you had started a new one. The verdict was cached and refreshed only when a
  conversation got wrapped up, never when you moved to a different one. It is now read from the
  conversation you are actually in, which fixes the same staleness on every other way of leaving a
  wrapped-up thread: the sidebar, a link, archiving then switching.
• New: a provider's credential can come from the environment. Set ATELIER_<PROVIDER>_API_KEY
  (and ATELIER_<PROVIDER>_ENDPOINT where a base URL is needed) and Atelier uses it — nothing is
  written to the database, so the secret appears in no dump and no backup, and rotates by changing
  the variable and restarting. Such a provider shows in the wizard as Set by this server, with
  no key field and no Disconnect, since this process cannot unset a variable it does not own; both
  actions previously reported success and changed nothing. A second form,
  ATELIER_DEFAULT_<PROVIDER>_API_KEY, is a default rather than a rule: it is copied into the
  credential store once, on the first boot after it is set, and is an ordinary key from then on —
  the owner can replace it, and a Disconnect stays disconnected across restarts. The variable's name
  is the whole configuration surface; there is nothing else to set.
• Fixed: your email API key survives an update. The appliance re-imports configuration on every
  boot, and that import deleted any stored key not present in the shipped configuration — including
  the mail key, leaving outgoing email with an empty credential after every update. Stored
  credentials are now treated as the runtime state they are, and are neither exported nor deleted.
• Changed: three unused modules are gone. Layout Builder, the Views UI and Views data export are
  uninstalled. None was used by anything Atelier ships, and Layout Builder in particular is a second,
  contradictory page-building model sitting in the admin UI next to the one the product is built on.
• Changed: a room's starter suggestions come from its own configuration and nowhere else. The
  console also carried a hardcoded fallback list, so which suggestions a room offered depended on
  which of the two happened to resolve — and they said different things. A room's chips are its own
  promise, so a room with none configured now shows none rather than borrowing asks it cannot
  fulfil. The composer still takes any request, and the room's welcome still says what it does.

---

Container image: ghcr.io/aincient-labs/atelier-cms:v0.5.0 · Install · Full changelog