Skip to main content

Changelog

Notable customer-visible changes to the Moss SDK (@viamoss/moss-sdk) and the Moss dashboard. From v0.13.0, the SDK version tracks the Moss platform version.

0.16.26 — 2026-11-18

SDK only: @viamoss/moss-sdk 0.16.26 is on npm. There are no new server images or Helm chart. SDK 0.16.26 works with a 0.16.25 server, and hosted Moss is unchanged.

Fixed

  • The Next button beside the chat input on guide steps is now as tall as the input and lines up with it. Before, it was slightly shorter and sat a few pixels lower than the input box.

0.16.24–0.16.25 — 2026-11-16

0.16.24 and 0.16.25 are described together. 0.16.24 wasn’t published to npm or rolled out to hosted Moss, so the SDK and hosted Moss go from 0.16.23 straight to 0.16.25. 0.16.24’s self-hosted images and Helm chart were published, but install 0.16.25 instead.

Added

  • New appearance settings. All of them are optional, and set under appearance.chrome in your integration’s code; none is in the dashboard. Unset, nothing looks different.
    • relatedArticles.variant: 'list' shows the related articles under an answer as plain rows, with titles in the primary text color, a small sentence-case “Related articles” label and a document icon in the primary color. 'cards', the default, keeps the bordered cards.
    • guideOffer.label and guideOffer.icon: the text and icon of the button that offers a guide. The icon is 'emoji' (🚀, the default), 'outline' (an outline rocket in the button’s text color) or 'none'.
    • message.paragraphGap: the space between paragraphs, lists and other blocks in an answer, as a CSS length. Unset, they sit directly on one another.
    • startScreen.queryIcons: your own icon for each built-in suggested-question icon name. Every question uses document unless zeroState.suggestedQueryIcons is 'per-query'. A name without an entry keeps the built-in icon.
    • startScreen.queryFontWeight, queryGap, articlesPaddingX and articlesRowGap: the weight of the suggested questions and the gap between them, and the side padding of the recommended articles and the gap between their rows.
    • inputBar.paddingY and inputBar.lineHeight: the input text’s top and bottom padding, and its line height as a number ('1.5'), a length ('20px') or a percentage ('150%').
    • inputBar.focusBorder: the input’s border color while it has focus.
    • icon on the footer links: 'external-link', 'ticket', 'none', or your own element, such as an inline SVG. The Help center icon follows its label, and the Open a ticket icon precedes it. Unset, Help center keeps ”↗” and Open a ticket has no icon.
    • guide.pageTint: dims the page around a guide’s highlight. true lays a 36% dark tint over it. An object sets color, opacity (0 to 1), or both. The chat never dims, clicks still reach the page, and an open modal dialog stays clear. The tint holds through step changes and lifts on a step with nothing highlighted. Off by default.
    • guide.targetBadges: 'step' numbers the highlight’s badge with the guide’s step number, as in the guide bar. Every target of a step shows that same number, so several targets don’t read as separate steps, and targets in one group, such as a row of tabs, share one badge. 'within-step', the default, keeps today’s numbering.
    • guide.nextInBar: false moves Next from the guide bar to beside the input box. Next moves beside the input on every step, click steps included, and hides while Moss works and once the guide is finished. The guide bar then shows only the plan and its progress, and the chat’s step loader shows that Moss is working. Unset or true, the default, keeps the guide bar’s Next. With both nextInBar and nextBesideInput set to false there’s no Next at all, and development builds warn in the console.
    • loader: 'steps' shows a short list of what Moss is doing while it works, such as “Reading the page” and “Planning the next step”. Screen readers hear the row in progress, except while the guide bar is already announcing “Working…”. 'line', the default, keeps the one-line indicator.
  • Next beside the input box. On a guide step with nothing to click, and on file-upload and confirmation steps, Next also shows beside the input box. Other click steps look the same as before, and it doesn’t show once the guide is finished. The guide bar’s Next stays. To keep Next only in the guide bar, set chrome.guide.nextBesideInput: false.
  • Self-hosted: Zendesk calls through a proxy. A new backend setting, ZENDESK_USE_ENV_PROXY, is off by default. When it’s on, Moss’s Zendesk calls (connecting, syncing the ticket form and filing tickets) go through the http:// proxy named in HTTPS_PROXY. A value without a scheme, such as proxy.example.com:3128, is read as http://, and NO_PROXY is honoured. An https:// or malformed proxy value isn’t used: the request isn’t sent, and the sync banner or the connection check says this server couldn’t reach Zendesk through the proxy. With the setting off, nothing changes.

Changed

  • You can type while Moss answers. The input stays editable during a reply. Send and Enter wait until the reply has arrived, and the text stays in the box. The input is disabled only while the connection is down.
  • Next sends what you typed. With text in the input, a guide’s Next sends it as your message. With an empty input, Next goes to the next step, as before.
  • No related articles while a guide runs. Replies during a running guide don’t show the related-articles list. The answer that offers the guide keeps it. Sources that your application’s instructions ask for in the answer text are unchanged.
  • The guide bar shows when Moss is working. While a request runs, a spinner and “Working…” take Next’s place, and screen readers announce it. When the request ends, Next comes back, with the focus if it had it.
  • The guide offer’s default label is “Start guide”, and it’s translated. It was “Guide me”, in English for every user. Korean users now see 가이드 시작.
  • One line of input text is centred on the send button. When chrome.inputBar is set, in code or in the dashboard, the send button sits a few pixels higher. With several lines, it stays level with the last one.
  • Zendesk connection errors name the cause. With host-brokered filing, when Test & connect or Test connection can’t read the end-user ticket form, the message now says why: the HTTP status, a DNS failure, a timeout, a TLS failure, a failed connection, or a reply that isn’t a ticket form. A 404 also asks you to check the subdomain. Zendesk’s reply itself is never shown.
  • Zendesk sync errors name the cause too. When a ticket-form sync fails, the “Last sync had a problem” banner, in Configure > Integrations > Zendesk and on the Ticket intake page, now says why: the HTTP status Zendesk answered, a DNS failure, a timeout, a TLS failure, a failed connection, or a reply that couldn’t be read. It no longer shows text from Zendesk’s reply.
  • Guide steps and guide offers.
    • When a step only asks the user to check something, such as a setting or a status, and the page shows the check passing, Moss says so and gives the next step in the same reply. If the page shows something else, Moss says what it shows and stays on the step.
    • Choices on the page, such as radio buttons, selects and links, are highlighted, and the user’s pick moves the guide on.
    • A step with nothing to click ends by pointing to Next, without saying where Next is on the screen.
    • Moss treats a “where is…” question as a how-to. When the answer lists several places in your application, Moss names the most common one and offers a guide there. Questions about places outside your application get no offer. The offer is the answer’s last sentence, before any Sources list your instructions ask for.
  • Browser extensions send the SDK version they bundle, so they get the current guide behavior. Extension builds sent no SDK version, so their sessions got the older guide behavior whatever the application’s stateless_turn_policy setting said. Every extension built from this release sends it. Once its new build is installed:
    • Ending a guide in the chat ends it for Moss too, so the next question isn’t answered inside the guide the user left. That covers closing the chat when closingChatEndsGuide is on, and filing a support ticket during a guide.
    • Asking Moss to stop (“stop guiding me”, “exit the guide”) ends the guide.
    • A guide with no progress ends on its own: after 2 hours if it was only offered, after 24 hours if it was running.
    • Moss decides on each question whether to answer it or guide, instead of keeping the mode an earlier question set.
    An extension switches only when its new build is installed. Pages that load a released SDK from npm or the CDN already had this behavior. An application with stateless_turn_policy: false keeps the older behavior.
  • SDK pre-releases get the current guide behavior. A pre-release of any version after 0.10.0, such as 0.17.0-beta.3, is now treated like a release. A page that loads one switches when the backend is upgraded, with no new install. Before, every pre-release got the older behavior. A pre-release of 0.10.0 itself still does.
  • Several highlighted places are named, not numbered. With guide.targetBadges: 'step', every highlight of a step shows the step’s number. When a guide step highlights more than one place, Moss’s reply now names each one (“Click Search or Share”) instead of numbering them 1, 2, 3. This needs the SDK from this release. It also covers the reply that starts a guide and highlights its first step.
  • No database migrations in this release.

Fixed

  • A guide’s Exit button is back. Open the guide bar’s step list: “Exit guide” is at the end of it, and asks before it ends the guide. Exiting removes the guide’s highlight, and Moss ends the guide too.
  • After a guide’s Exit, clicking the element its step highlighted no longer sends the step.
  • After pressing Next, clicking the highlighted element no longer sends the step a second time. If Next’s request fails, the step comes back, so Next can be pressed again.
  • After exiting a guide with the keyboard, the focus stays in the message input.
  • The chat panel fits short browser windows. In a window shorter than the panel, the panel opened with its header, and the New chat and Close buttons, off screen. It now fits inside the window, shrinking if it has to, and gets its full size back in a taller window. The same goes after a resize, after Reset Chat Position, and for a position saved on an earlier visit.
  • Screen readers name the chat header’s buttons. New chat and Close, and on the classic header Reset Chat Position and Change Language, now have names.
  • Screen readers hear once that Moss is working. With the default loader, a running guide was announced twice, and nothing was announced while a request ran after a guide had finished. Each is now announced once, and the default loader no longer re-announces as its text animates.
  • Coming back to a session too old to resume. The “Starting fresh” card no longer has a working indicator behind it, and nothing can be sent to the old session from behind the card.
  • One session per first open. Loading a page or opening the chat for the first time starts at most one session. Before, it could start a second, empty one. The app’s settings are also read once per page load, not twice.
  • Dashboard: long application names. The desktop sidebar keeps its width, and at phone width the pages no longer scroll sideways.
  • openChat and toggleChat open the chat at the position you pass. The position applies to that open only and isn’t saved. If the user moves or resizes the chat during that open, their placement is kept.
  • Next stays available when a click step’s setup gives up. When a step’s target couldn’t be found, or Next or Exit was pressed while it was being set up, Next beside the input could stay hidden for the rest of the guide.
  • Guides work on more pages whose forms have fields named after built-in page properties or methods, such as submit, closest or matches.
  • A request the model provider’s safety filter blocks gets its own answer. When a filter such as Model Armor or Gemini’s safety settings blocks a request or its reply, Moss answers “A safety filter blocked this request, so I can’t answer it. If it keeps happening on this page, please contact your administrator.” It used to ask the user to rephrase. A blocked request counts as blocked, not as an error, and the session shows a blocked step.
  • SDK rate limits: an unreachable limit store is logged as an error. With SDK_RATE_LIMITS_ENABLED on, a run of skipped checks is logged as an error whatever the cause. Before, only timeouts counted, so a store that refused connections was logged as a warning only.
  • Each question reaches Moss once. With the conversation log that’s the default since 0.16.1, every typed question went twice into the conversation Moss reads. So Moss took a conversation’s first question for a later one and didn’t work out the user’s goal for the first answer, and a very short question searched the help articles with its words repeated. The first answer stays as fast as before, and Moss now works out the goal alongside it.
  • Clicking a guide’s first step right after Moss offers it no longer ends the guide with “I couldn’t recover this guide.”
  • Pressing Next while a click step is still setting up now moves the guide on. Before, the press could be lost, and the guide bar kept showing “Working…”.
  • With closingChatEndsGuide on, closing the chat right after pressing Next or sending a message no longer leaves the chat busy when it’s reopened.
  • Guide highlights no longer fail on pages whose forms have fields named after built-in page properties.
  • Self-hosted: deleting a user’s data and missing buckets. A deletion is no longer certified complete when the user’s files were stored in a bucket that the dashboard backend can’t find. The certificate notes the missing bucket, and the request can be retried once the bucket is configured.
    • Helm: global.storageBuckets.recordings and .screenshots now set the bucket names for both backends. Rendering fails when the two backends name different buckets, so helm template, install and upgrade stop with a message that names the setting. Set each name to the bucket the install already uses; nothing moves.
    • Upgrading with --reuse-values: add --set global.storageBuckets.recordings=rrweb-recordings, or upgrade without --reuse-values. An install that never changed these names and upgrades without it needs no action.
    • Known limit: for an application that recorded sessions before 0.16.1 and has since set session recording to anything but on, recordings from before 0.16.1 may not be checked.

Security

  • Python, CVE-2026-119170 (High): fixed. The backend and worker images require the Python 3.14.10 build that fixes it.
  • soupsieve 4.0 in the backend and worker, which fixes CVE-2026-126190 and CVE-2026-126191 (High).
  • opentelemetry-api 1.45.2 in the backend and worker, which fixes CVE-2026-126360 (Medium).
  • URL rules: page addresses are matched against URL rules with a length limit and a time limit, so an unusual address can’t hold up the service.

Known issues

  • The 0.16.23 known issues still apply, except the one about a guide’s Exit control and the note about a missing recordings or snapshots bucket, both fixed above.

0.16.23 — 2026-11-13

Added

  • Closing the chat can end a running guide. With closingChatEndsGuide: true in the SDK config, closing the chat while a guide runs ends the guide, as its Exit control does. The highlight comes off, and a page refresh brings nothing back. It applies to the close button, a data-moss-trigger toggle, AssistantButton, closeChat() and toggleChat(). A guide only offered, a guide already finished, a guide running in another tab and Moss’s own closes are left alone. It is off by default, and nothing changes until you turn it on.
  • New optional backend settings, all off or unchanged by default:
    • SDK_REQUIRE_ORIGIN: when on, SDK requests that carry no Origin or Referer, and every request for an application with no allowed domains, get 403 ORIGIN_REQUIRED. Leave it off unless every application has its list filled in.
    • SDK_RATE_LIMITS_ENABLED: per-minute limits on SDK session starts, queries, streams and recording uploads, kept in Redis. A slow or unreachable Redis lets requests through. Current SDKs don’t wait out a 429, so a limited request fails its turn.
    • LLM_REQUEST_TIMEOUT_SECONDS (default 45): the first attempt’s timeout for each model request in a chat turn.

Changed

  • Allowed domains are checked by exact host, on every SDK request. Check each application’s list before you upgrade:
    • An entry matches only that exact host: example.com no longer matches app.example.com. For subdomains, add a *. entry such as *.example.com. A *. entry matches subdomains at any depth but not the bare domain, so list both if Moss runs on both.
    • The check covers every SDK request from the page, not only the first configuration load. A page the list doesn’t cover gets 403 DOMAIN_NOT_ALLOWED, and Moss doesn’t load there.
    • An entry with a port matches only that port. Browsers leave out the default port, so an entry ending in :443 never matches an https page.
    • Origin: null, which browsers send from sandboxed iframes and local files, is refused when the list isn’t empty.
    • An empty list is unchanged.
  • Screenshot uploads need the user’s token. SDK 0.16.23 sends it. Older SDKs fall back to sending the screenshot inside the chat request, a few MB each, so check that your ingress accepts bodies that size on /api/v1/query until the SDK is upgraded. Upgrade the backend and dashboard first, then the SDK.
  • Saving a guide from the SDK needs the authoring grant. Save as Guide in debug mode and in Moss Studio now needs a token with authoring: true, and saves only the token user’s own session. An ordinary end-user token gets 403, and the button says the user may not save guides. The dashboard’s own save is unchanged.
  • A signed token with an empty subject is refused. A JWT whose sub is empty, blank or not a string gets 401 INVALID_TOKEN, instead of running as the user named in the request. Tokens with a real subject are unaffected.
  • A model request that hangs ends sooner. Each request a chat turn makes has a timeout, then one retry with what is left of the turn’s 120 seconds. A hung request now ends at about 118 seconds with “I’m sorry, that took too long to answer. Please try again.” instead of the high-traffic message. A slow but working request completes on its retry.
  • Deletion requests that fail are marked failed, with details, instead of staying in progress, and can be retried.
  • No database migrations in this release, and the Helm chart’s values are unchanged.

Fixed

  • A native modal dialog opened with showModal() no longer turns the page grey under Moss when the page styles ::backdrop, and the chat no longer inherits the dialog’s text layout, so its messages wrap.
  • A page whose form has a control named after a form property, such as parentElement or id, no longer hangs the tab or breaks page capture.
  • Deleting a user’s data now completes when their chats have analytics insights, evaluation scores, monitor flags, action logs, page snapshots or guides. It also removes their model-call logs and proactive-guide events.
  • A rate-limited model request that the client gave up on is retried.

Security

  • Redaction. Redacted content no longer reaches Moss through a parent’s text, accessible names and labels, interaction HTML, element fingerprints, the validation summary or a clicked element’s text. This holds across open shadow roots and same-origin frames. Redacted alerts and fields are listed by tag only, so the summary still shows that they’re there.
  • glibc, CVE-2026-84243 (High): fixed. All four images ship glibc 2.44-r17, which includes the 2.44-r16 fix.
  • Python, CVE-2026-119170 (High): not affected. It applies only when OpenSSL runs in FIPS mode, and no image installs a FIPS provider. Scanners may still report it.
  • Chrome for Testing 157.0.8103.78 in the worker, which fixes CVE-2026-125350 (High) and the advisories published on 11 November.
  • Python dependency updates, and addresses containing a NUL character are refused before any lookup.
  • The SDK package is published with an exact, integrity-checked npm.

Known issues

  • After a guide’s own Exit control, a click on the step’s target still sends the step. Closing the chat with closingChatEndsGuide on doesn’t have this gap.
  • Self-hosted: if RECORDINGS_BUCKET or DOM_SNAPSHOTS_BUCKET names a bucket that doesn’t exist, the deletion certificate reports the erasure complete without the request’s note about the missing bucket. Check that both name your real buckets. A fix is planned for 0.16.24.
  • The 0.16.22 known issues still apply.

0.16.22 — 2026-11-10

Added

  • Returning after inactivity. Configure > Behavior > Session Behavior has a new field, “When a user returns after inactivity”: “Ask to continue or start new” (the default, as before) or “Start a new chat”. With “Start a new chat”, a user who comes back to a session past its inactivity timeout (30 minutes by default) or its max age (2 hours by default) gets a new chat, with no card and no rating. This applies when Moss loads into a timed-out session, and when the user reopens the chat on a page left open, with the launcher, openChat() or a data-moss-trigger element. The previous chat can’t be reopened.
    • In the SDK, sessionConfig.onReturnAfterInactivity takes 'new_chat' or 'ask'. A value in your config wins. Left unset, the application’s dashboard setting applies, delivered at session start by backend 0.16.22 and later; without one, 'ask'. For npm embeds, only this field comes from the dashboard: the inactivity timeout, max age and showNotification stay the ones in your config. A config built with convertBackendConfig() takes all four from the dashboard.
    • Update the backend and dashboard before the SDK. An older backend drops the setting on save without an error, and an older SDK ignores it. SDK 0.16.22 on an older backend asks, unless your config sets the field.
    • Moss’s own opens, such as askMoss() and proactive or onboarding flows, don’t check. openChat() followed by sendMessage() sends the message in the new chat once it connects. Moss holds one such message, and logs a console warning when it can’t send it.
    • Guide steps taken with the chat closed, and activity on the same session in another tab, count as activity. A guide idle past the timeout ends with its session; if it ran in this tab, Moss Plan Finished reports it with outcome: 'abandoned' under the old session’s sessionId.
  • Moving the chat window. The chat window also moves from the empty space in the conversation and around the message box, not only from its header and edges, so a window whose header is off screen can be dragged back. Messages, links, buttons, cards, the guide bar and the scrollbar don’t drag, and a drag selects no text. On touch, the conversation scrolls, and the window moves from the header, the edges and the area around the message box.

Changed

  • The worker refuses private addresses on its own fetches. Image downloads, sitemaps, discovery’s page sampling, and Document360 discovery and scraping now reach public addresses only. Check these before you upgrade:
    • A forward proxy is no longer used for Document360, image and sitemap fetches. These fetches connect directly and ignore HTTP_PROXY and HTTPS_PROXY, and there is no setting yet to send them through a proxy. If your worker’s egress requires one, contact us before upgrading. Firecrawl, the LLM and embedding providers, Pinecone, storage uploads, Slack and the browser’s page loads during discovery are unchanged.
    • Images and sitemaps on an intranet stop working until their host or range is listed in the new worker setting WORKER_FETCH_ALLOWLIST (comma-separated hosts and CIDRs, such as wiki.corp.example.com,10.20.0.0/16). It is empty by default.
    • localhost needs its own entry. Cloud metadata addresses can never be allowed. An invalid entry stops the worker at boot. Your object storage (S3_ENDPOINT) needs no entry.
    • A refused image is recorded as failed and keeps its original address. A discovery start URL on a refused host fails at once. A downloaded body that isn’t an image is never stored.
  • Document360 API host. It must be a Document360 host (*.document360.io or *.document360.com, such as apihub.document360.io). Any other host is refused when it’s saved, and one saved earlier is refused at sync, with a message to enter it again. A help center on its own domain is fine, because that isn’t the API host.
  • Zendesk subdomain. It must be the account name only, such as acme, not a full address. Upper case is accepted and stored in lower case.
  • CSV exports (conversations, traffic logs and the audit log). A cell that starts with =, +, -, @, a tab or a carriage return gets a leading apostrophe, so spreadsheets open it as text, never as a formula. That includes negative numbers: -1 exports as '-1.
  • Slack alerts. Alert text is escaped and length-limited. Worker job alerts carry the outcome, an error code and counts, but not the raw error text; the details stay in the job’s record in Moss.
  • New optional setting: WORKER_FETCH_ALLOWLIST, on the worker. In the Helm chart’s values.yaml, only commented examples for it are added.
  • Support tickets during a guide. While a ticket card is open, the guide bar is hidden. Cancelling the card brings the guide back on the same step. When the ticket is filed, or its outcome is unconfirmed, the guide ends and stays ended: onEvent reports Moss Ticket Submitted, then Moss Plan Finished with outcome: 'abandoned'. A ticket reported as failed leaves the guide running. A guide started after the ticket shows its bar as usual.
  • The ticket card opens on Moss’s reply. When Moss drafts a support ticket, the chat opens on Moss’s message above the ticket form, not on the bottom of the form.
  • Rating after a filed ticket. Once a support ticket is filed, “Was this conversation helpful?” appears at once, under the ticket confirmation, so the ticket number and its link stay in view. Answering it leaves the conversation as it is, and New chat doesn’t ask again for that conversation. The next message or guide step removes it. It doesn’t appear for a ticket that failed, came back unconfirmed or was cancelled. A rating that can’t be sent brings the buttons back.
  • A guide step after a long pause. When a user idles past the inactivity timeout mid-guide, then clicks a step’s target that loads a new page, the new page keeps the session and continues the guide, with no card to continue or start new.
  • Saving session settings. A configuration save merges sessionConfig with the stored settings: a save that sends only some session settings keeps the others, and a setting sent as null is cleared. A save with sessionConfig: null no longer clears them all.
  • Default loading-indicator selectors. With the loading-indicators stability layer on and its default selectors, skeleton, spinner and loader classes match in any case.
  • Every SDK change in this release works on any backend, except the dashboard setting for returning after inactivity, which needs backend 0.16.22.
  • No database migrations in this release. The one new environment setting is the worker’s optional WORKER_FETCH_ALLOWLIST, above.

Fixed

  • Moss no longer reads a page that is still drawing its loading placeholders. Before it sends a turn’s page context, the SDK waits up to 4 seconds for visible placeholders, such as skeletons, to clear, so answers describe the loaded page instead of saying it is still loading. With none on screen there is no wait. Placeholders that never clear cost one wait, and are then skipped until they’re gone or the page’s address changes. The settings are under stability.placeholderGuard (enabled, timeoutMs, excludeSelectors, additionalSelectors and selectors). When you set stability.loadingIndicators.selectors, the wait uses that list.
  • A turn that starts while an earlier one is still waiting for the page to settle now waits too. It could read the page with no wait.
  • A guide step whose click opens a new tab, through a target=_blank link, window.open() or a link that redirects, brings the chat up in the new tab at the next step, and the step is sent once. A same-tab redirect, or a page that rewrites its address as it loads, no longer stalls the guide. Switching tabs right after a step’s click no longer ends the guide. This fixes the 0.16.21 known issue for a page whose address changes as it loads.

Security

  • Integration hosts for Document360 and Zendesk are now validated, and CSV exports open safely in spreadsheets.
  • Alerts and audit records are hardened against untrusted input.
  • Worker fetches of private addresses (High). A crawled page could point an image or sitemap link at an internal address, and the worker would fetch it. The worker now refuses private, loopback, link-local and metadata addresses on its own fetches, and stores only bodies that are images.
  • The images are rebuilt on current base-image packages, with no dependency changes.

Known issues

  • Some valid images are refused and keep their original address: a UTF-16 SVG, an SVG that ends in a comment containing </svg>, an AVIF or HEIC whose last box has size 0, and an SVG with a prefixed DOCTYPE name.
  • The 0.16.21 known issues for a guide whose starting answer is lost, in one tab or with a second tab, still apply, as do the other two 0.16.20 known issues for guides in more than one tab.
  • “Retry failed pages” counts and retries only pages that couldn’t be fetched. To run pages that failed indexing again, refresh the source or edit the page.
  • With “Start a new chat”, a chat left open while the user is away keeps its conversation: Moss checks only as it loads and as the chat opens. Calling openChat() and then askMoss() in one handler, past the timeout, starts one extra session that is never used.
  • If the page reloads mid-guide while a ticket card is open, the step’s highlight comes back with the guide bar hidden, and a click on it completes the step. When a ticket is filed in a tab that only shows the guide, the tab running it reports Moss Plan Finished but keeps the guide bar until a later answer clears it.
  • Placeholders inside shadow roots aren’t seen. A placeholder the page re-creates on every render is waited on at each turn; exclude it with stability.placeholderGuard.excludeSelectors.
  • The guide stays in the first tab when a step’s new tab can’t be tied to the click: the tab opened in the background on a different address, the page opened it only after the step was sent, a page-wide no-referrer policy or window.open(url, '_blank', 'noreferrer') hid where it came from, or it was reached through more than one intermediate page. A target=_blank link to a page without Moss waits out the step’s window, 30 seconds by default, as since 0.16.19.
  • Within 5 seconds of a step’s click, a tab the user opens by hand after a link that sends no referrer, or another link the user opens from the same page, can take the step.

0.16.21 — 2026-11-06

Changed

  • A source whose pages can’t be indexed ends in Error. It used to end Ready with no content or with part of it. Each such page shows “Re-indexing failed” in its page view. A refresh that fails this way keeps the current version serving, with all of its content. The processing count on a source advances while pages are indexed.
    • A Ready source that still has a page from an earlier run that can never be indexed stays Ready. Until that page is refreshed or edited, each later processing run records the error on it and sends a failed notification.
    • The failed notification (CHUNKING_FAILED) and the worker’s log say how many pages failed, as N/M pages failed chunking (K indexed)., then give the latest error as one of a fixed set of descriptions, never the raw error text.
    • This also covers the 0.16.19 known issue on Vertex AI eu and us: a source on gemini-embedding-001 now ends in Error instead of Ready with no content.
  • Ingestion errors keep no raw error text. A page’s or a job’s error in the ingestion API, and a source’s last error, are now a fixed description: an error code, an HTTP status, an exception type or a provider name, such as OpenAI. They no longer include an exception’s message or a provider’s response. Errors stored before the upgrade are returned the same way. What the dashboard shows is unchanged: it shows these errors in dev mode only. The worker’s log keeps the full message.
  • Document360 API token on save. Spaces and line breaks at the ends of a token are removed when it’s saved. A token with a line break, or another character that isn’t printable ASCII, inside it is refused, with a message to paste it again. A token saved earlier with a trailing line break is now sent without it.
  • No database migrations or new settings in this release.

Fixed

  • Direct Zendesk filing sends nested branch fields. Fields that a conditional field opens at any depth, such as a recording type under a product area, are now sent. Tickets on those branches used to be held back as missing a required field.
  • Host-brokered ticket card. When a branch of the ticket form requires a text field that Moss fills with the user’s role, such as “Your Role”, the card now shows the role the user gave. It used to show an empty row.
  • Plan events with Moss open in more than one tab (SDK only).
    • A guide step the user completes with a click, in a tab that was only showing the guide, now reports Moss Plan Step Completed, and Moss Plan Finished when it was the last step. The click takes the guide over in that tab, as in 0.16.20. In 0.16.20 the tab could report neither event for a step that doesn’t open another page, which covers most “Click …” steps.
    • Pressing Start Guide on an offer after another tab has already started that guide no longer reports Moss Plan Started, or the other tab’s completed steps, a second time. The tab only shows the guide, and takes it over when the user continues it there, with Next or a step.

Security

  • Browser in moss-worker. Chrome for Testing moves from 156.0.8078.80 to 157.0.8103.38, which fixes 26 Critical and 54 High advisories in 156. The other three images don’t include the browser.
  • glibc, CVE-2026-84243 (High): no fix yet, assessed. Your scanner may report it in glibc 2.44 in all four images, and no fixed glibc is published yet. Exploiting it needs an attacker-controlled LANGUAGE, an active message locale other than C or POSIX, and a crafted message catalog the process can read. The 0.16.21 images run with the C locale, set no locale variable and install no other locale definition. The fixed glibc ships in the first release after one is published. If your deployment sets LANG, LANGUAGE, LC_ALL, LC_MESSAGES or LOCPATH, or mounts locale files, contact us and we’ll reassess.
  • sqlite-libs, CVE-2026-119586 (High): not affected. The advisory is in the SQLite command-line shell’s .recover command. No image includes the sqlite CLI, and the library in the backend and worker images has no recover functions. Your scanner may still report it, with no fixed version.
  • Error text in the ingestion API. Through 0.16.20, the ingestion API returned a page’s or a job’s stored error text to any member of the source’s organization. In rare cases that text repeated a credential: a Document360 token saved with a line break in it, or a token in a source’s URL. 0.16.21 stores no such text, and returns only the fixed description for errors stored before the upgrade. If a Document360 token was saved with a line break, rotate it.
  • Backend dependency security update (idna), to its first patched version, 3.15, for a Medium advisory.
  • The images are rebuilt on current base-image packages. The moss-backend and moss-worker builds also place a copy of each installed Python package’s own SBOM where scanners look, so image scans see those packages’ components. Nothing changes at run time.

Known issues

  • This replaces the first 0.16.20 known issue. If the backend starts a guide but the answer doesn’t reach the tab, that tab only shows the guide. This happens when the answer to a typed yes is lost, or when the SDK’s automatic retry of a 503 answers instead of the first attempt. The tab doesn’t continue the guide after a reload until the user continues it there, with Next or a step. It never reports that guide’s Moss Plan Started, even then: Next or a step there reports only Moss Plan Step Completed. When the answer to Start Guide is lost to a network or server error, pressing Start Guide again runs the guide in that tab and reports Moss Plan Started once.
  • If a tab’s Start Guide gets no answer because of a network or server error, and another tab starts the same guide before this tab’s next answer, both tabs report Moss Plan Started. After a guide overview, the same applies to any guide another tab starts before the user presses Start Guide again.
  • The other two 0.16.20 known issues for guides in more than one tab still apply.
  • A guide step into a page that redirects or changes its address as it loads, such as a changed query value, may not bring up the chat in the new tab; the guide then carries on in the first tab after 30 seconds. In the same tab, the chat comes back but the guide can stall. Fixed in 0.16.22.
  • “Retry failed pages” counts and retries only pages that couldn’t be fetched. To run pages that failed indexing again, refresh the source or edit the page.

0.16.20 — 2026-11-04

Added

  • Pinecone private endpoint (AWS PrivateLink). With PINECONE_USE_PRIVATE_ENDPOINT=true (off by default), sdk-backend, dashboard-backend and the worker still look the index up on api.pinecone.io, then send all data-plane traffic (search, fetch, upsert and delete) to the index’s private host, <index>.svc.private.<...>.pinecone.io, which Moss reads from the index description; there is no host to configure. api.pinecone.io must stay reachable from all three workloads.
    • There is no fallback to the public host. If the index’s Pinecone project has no private endpoint, Moss doesn’t connect to the index: chats answer without the application’s knowledge base, deleting a source fails, and chunking jobs fail before they touch a source’s pages or vectors. Each logs an error that names the index and the setting.
    • Set it on all three workloads, with the upper-case name and the same value. true, t, yes, y, on and 1 turn it on, in any case; false, f, no, n, off and 0 leave it off. Any other value, including an empty one or one with surrounding spaces or quotes, stops the backends from starting and fails every chunking job.
    • Each process logs Pinecone data-plane host: private (or public) the first time it opens the index. The line shows the host Moss picked, not that it is reachable.
    • Before turning on Pinecone’s private-endpoint-only restriction for the project, check that Pinecone answers through the private host, since a wrong DNS record or security group still logs private. On sdk-backend, one chat turn that uses the knowledge base logs the private line, then Pinecone multi-namespace search completed with results=N/M and N above 0, and no search failed. On the worker, one knowledge-source refresh logs the private line and the source ends Ready with more than 0 chunks. On dashboard-backend, which may not log the line for a while, confirm the variable is set on its deployment. With the restriction on, Pinecone rejects data-plane requests that don’t come through the endpoint, so a workload still on the public host loses retrieval or indexing.
    • The setting was tested against a mocked Pinecone API, not a live PrivateLink endpoint, so the check above is the test on your install.

Changed

  • New optional setting: PINECONE_USE_PRIVATE_ENDPOINT. In the Helm chart’s values.yaml, only commented examples for it are added, on sdk-backend, dashboard-backend and the worker; the chart README documents it. With the setting off, Moss sends the same Pinecone requests to the same hosts as 0.16.19.
  • No database migrations in this release.

Fixed

  • With Moss open in more than one tab, a guide runs only in the tab where the user started it. Another tab can show it, and takes it over only when the user continues it there. Plan events (Moss Plan Started, Moss Plan Step Completed, Moss Plan Finished) come only from the tab the guide runs in, so each guide counts once; in 0.16.19 a second tab could report them again. A guide exited from another tab is reported as Moss Plan Finished with outcome: 'abandoned' by the tab that ran it. The fix needs both the 0.16.20 backend and SDK 0.16.20.

Security

  • Backend dependency security updates (h2, hpack, langgraph, langgraph-sdk, soupsieve) and worker dependency security updates (multidict, soupsieve), each to its first patched version, for Medium advisories.
  • The moss-dashboard and moss-dashboard-oidc builds install one exact, hash-checked pnpm, 10.34.7, instead of a floating pnpm at build time.
  • There is no other Dockerfile change. The images are rebuilt on current base-image packages.

Known issues

  • If the answer that starts a guide doesn’t reach the tab, for example after a retried 503, that tab only shows the guide. It reports no Moss Plan Started, and doesn’t continue the guide after a reload until the user continues it there, with Next or a step.
  • When another tab takes a guide over, with Next or a step there, the original tab still reports Moss Plan Step Completed for the other tab’s steps if its user asks something later. A New chat in the original tab still ends the guide on the device and reports it finished there.
  • If the tab running a guide is mid page load when a tab that only shows the guide exits it, no tab reports that Moss Plan Finished.

0.16.19 — 2026-11-03

Added

  • SDK analytics events. Pass onEvent in the provider config to forward Moss usage to your analytics, for example onEvent: (e) => window.analytics?.track(e.name, e.properties). Each event is { name, properties }, named in Segment’s “Object Action” form. Twelve events cover the chat, questions and answers, errors, guided plans, feedback and support tickets. MOSS_EVENT_NAMES lists them, and the SDK README documents each event’s properties. Moss Question Asked carries source: 'user' | 'host', where 'host' marks a question sent with askMoss(), counted once as its request leaves. Question text is left out unless you set eventQueryText to 'include' or a function. A handler that throws is logged and ignored, and leaving onEvent unset changes nothing.
  • Starting Moss on demand. For hosts that mount Moss only when the user asks for it, shouldStartMoss() from @viamoss/moss-sdk/start (no React) tells the page, as it loads, whether to mount Moss. It returns true when the chat was open in this tab before the page changed, while a guide is running in this tab, while an askMoss() question waits in this tab, or when a guide step the user took is waiting to continue on this page (30 seconds by default). Mount Moss when it returns true, and don’t open or toggle the chat yourself. It needs stability.afterReload.enabled: true; if you set stability.afterReload.pendingTTL, pass the same value. It only reads browser storage and never throws.
  • Opening Moss with a question. askMoss(question), also exported from @viamoss/moss-sdk/start, hands a question from your own assistant or help widget to Moss: Moss starts a new chat, opens it and sends the question as the user’s first message. It returns { accepted: true, id, queued } or { accepted: false, reason: 'empty' | 'unavailable' }, and never throws. With queued: true, mount Moss without opening the chat within 30 seconds (ttlMs changes that). Call it in the same window and on the same origin as Moss, and open the chat with openChat(), not a toggle or a [data-moss-trigger] click: a toggle while a question waits closes the chat and drops the question. A chat or guide in progress ends as New chat ends it.
  • Vertex AI eu and us multi-regions. With GOOGLE_AUTH_MODE=vertex and GOOGLE_CLOUD_LOCATION=eu, chat, gemini-embedding-2 embeddings and image annotation all go to aiplatform.eu.rep.googleapis.com; us works the same way on aiplatform.us.rep.googleapis.com. Set the location on both backends and the worker, and allow egress to that host from all three. Vertex on global or a single region sends the same requests to the same hosts as before.
  • OIDC session renewal. With OIDC_REFRESH_TOKENS=true (off by default) on the moss-dashboard-oidc container, the dashboard renews the user’s ID token with a refresh token, so sessions outlast the IdP’s ID-token lifetime (60 minutes on Okta). Allow the Refresh Token grant for the Moss application in the IdP first and, for an Okta custom authorization server, offline_access in its access policy. Without the access-policy change, every sign-in fails once the setting is on.
  • security.txt. The backend serves an RFC 9116 file at /.well-known/security.txt, without sign-in and in every service mode, and /security.txt redirects there. The contact defaults to mailto:security@viamoss.ai; set SECURITY_TXT_CONTACT to a mailto: or https: URI to list your own.

Changed

  • Image annotator default model. With ANNOTATOR_LLM_MODEL unset, the worker uses gemini-3.8-flash at low thinking, replacing gemini-2.5-flash, which Google retired on 2026-10-20. An explicit gemini-3.8-flash also moves to low thinking, with a 4,096-token output limit. Other models are unchanged.
  • OIDC sign-in. Users in neither OIDC_ADMIN_GROUPS nor OIDC_MEMBER_GROUPS see an Access Denied page at sign-in. If the dashboard-serving backend’s OIDC_AUDIENCE lacks the OIDC client ID, users return to sign-in after each sign-in; add the client ID. Users signed in before the upgrade sign in once more.
  • Dashboard toasts appear at the bottom right, in both dashboard images. They move up when they would cover the Copilot or Playground message box, or the Save and Discard bar on Suggested Content.
  • In the SDK, each tab keeps its own chat and guide. A tab opened with window.open() starts with the chat closed and no guide, and a guide continues after a reload only in the tab it runs in. When a guide step opens a new tab, the chat opens there if it was open in the first tab.
  • New optional settings: OIDC_REFRESH_TOKENS on the dashboard and SECURITY_TXT_CONTACT on the backend. GOOGLE_CLOUD_LOCATION also accepts eu and us. In the Helm chart’s values.yaml, only the commented ANNOTATOR_LLM_MODEL example changes.
  • No database migrations in this release.

Fixed

  • OIDC dashboard sessions no longer outlive their ID token. Before, about an hour after sign-in on Okta, every API call failed with 401 while the dashboard still showed the user as signed in. Now the next page load goes to sign-in (“Your session has expired”) and then back to the same page; with an active IdP session, usually without asking for credentials.
  • The dashboard’s retry after a 401 asks for a fresh token, and a sign-in the IdP refuses says so on the sign-in page.
  • LLM_THINKING=minimal on gemini-3.8-flash is sent as low. Every chat request with it failed with a 400 before.
  • In the SDK, a chat start that fails after New chat has replaced it no longer leaves the new chat’s message box disabled with no Retry.

Security

  • API action identity. Configured API actions take the user’s company and user ID from the signed SDK token, never from the conversation. The company comes from company_id (or companyId) in the JWT metadata claim. Company-type fields (company_id, tenant_id, organization_id, org_id and their X- header forms) are filled by the server, and an action with one is refused when the token carries no company. user_id is filled with the signed-in user, and a call that names another user is refused. Applications without API actions see no change.
  • Role-limited procedures. A procedure with allowed_roles runs only for users whose token signs one of those roles in metadata.role. Until the host signs role, such a procedure is hidden from every user. A malformed allowed_roles value admits nobody.
  • A missing or blank APP_ENV now means production. Workers that don’t set it report production in telemetry, so dashboards that filter on local may shift.
  • The images are rebuilt on current base-image packages, with no dependency changes.

Known issues

  • With Moss open in two tabs during one guide, the tab that isn’t running the guide also emits Moss Plan Started, Moss Plan Step Completed and Moss Plan Finished for it, so those can be counted twice. With one tab the counts are right. Fixed in 0.16.20.
  • gemini-embedding-001 isn’t supported on eu or us: ingestion with it fails, and chats answer without the application’s knowledge base. An application’s embedding model can’t be changed, so moving to gemini-embedding-2 means a new application whose sources are ingested again.
  • The chart’s chat default, gemini-3-flash-preview, is served on global only. On eu or us, set LLM_MODEL_NAME on both backends to a model Google serves there, such as gemini-3.8-flash.
  • Analytics chat, the daily insight sweep, monitor scoring, procedure authoring, self-healing and session evaluations call the Gemini Developer API with LLM_API_KEY, whatever GOOGLE_AUTH_MODE says. On eu or us, leave LLM_API_KEY unset on both backends so these calls fail instead of leaving the region, and set INSIGHT_SWEEP_ENABLED=false on the dashboard backend.
  • Okta has been reported to leave the groups claim out of refreshed ID tokens. The dashboard then doesn’t use them, and sessions still end every hour. Try OIDC_REFRESH_TOKENS on a staging install first.

0.16.18 — 2026-11-03

Security

  • Backend and worker image security update: libexpat 2.10.0 (CVE-2026-113351, CVE-2026-115506), replacing the patched library shipped in 0.16.16. The worker keeps refusing sitemaps that declare a DTD.
  • Dashboard dependency security update (prismjs 1.30.0, fflate 0.4.9).

Changed

  • The Guidance Flows empty state now says only that there are no flows.
  • No database migrations in this release.

0.16.17 — 2026-11-02

Changed

  • Build fix only. The images run the same software as 0.16.16, and deployments already on 0.16.16 don’t need to upgrade.
  • No database migrations in this release.

0.16.16 — 2026-11-02

Changed

  • Support tickets: the card comes first. With host-brokered filing, once a ticket is called for, Moss shows the editable ticket card as soon as it has a subject and a description. It no longer writes the ticket out in chat, asks for a confirmation there, or asks for an email address or other contact details. Nothing is filed until the user submits the card, and Moss says a ticket was filed only once the filing is confirmed. Remove procedure text that asks for a confirmation in chat.
  • Ticket fields are pre-filled only from what the user said, in both filing modes, and values for anything outside the ticket’s fields are dropped. The user can still change every field the card shows. The impact level follows the user’s own words in the subject, the description and their messages, so the request type no longer raises it. Critical is pre-filled when the user reports an outage, or says work has stopped or everyone is affected. Without such a signal, a host-brokered card leaves the impact empty and a direct ticket defaults to Medium.
  • The Filing switch is the single on/off for tickets. Off, in the Filing row of Configure > Integrations > Zendesk, applies in every procedure, in the default reply, to a ticket request under way and at Submit, and direct connections no longer dry-run. Procedures also offer no tickets while a connection is Paused or otherwise can’t file. Only an Admin can see or change the switch.
  • Before upgrading, decide each Zendesk connection’s Filing setting. A connection that shows Off while its procedures still offer tickets stops offering them at the upgrade. Switch host-brokered connections to Live only after the upgrade, because on earlier releases Live also turns on a default-reply ticket offer that asks for the user’s email address. To keep the default reply from offering tickets, set procedures.default_reply_policy to false in the application’s configuration; this also drops its suggestion to contact support. Where a direct connection’s Zendesk form requires the requester’s role, pin a value for that field on the Ticket intake page, or have the ticket procedures ask for the role.
  • SDK 0.16.16 types the outcome’s ticketId as string | number and sends it as text.
  • LLM_CLIENT_TYPE no longer accepts baseten, and the provider-failover settings LLM_AUTOROUTER_ENABLED and LLM_FALLBACK_PROVIDERS are ignored. The backend uses the one provider that LLM_CLIENT_TYPE names.
  • Traffic-log audit events from Bearer sign-ins record auth_method jwt instead of clerk_jwt. Older events keep clerk_jwt, so update any SIEM filter on the value.
  • The worker refuses sitemaps that declare a DTD.
  • Upgrading from 0.16.7 also brings in 0.16.8 through 0.16.15, none of which has a database migration or a Helm values change.
  • No database migrations in this release. The Helm chart’s values.yaml changes only in a comment.

Fixed

  • A repeated Submit of the same card before your backend’s answer is recorded, for example after a lost response, is accepted and handed to your backend under the same handoffId. Deduplicating on handoffId gives the user one ticket. A repeat with changed values is still refused.
  • Your backend can return ticketId as a number or as text, and a whole-number errorCode or errorDetail is kept as text. These were refused before, and the outcome went unrecorded.
  • The dashboard’s Broker contract page describes what your backend receives: a timeout, a callback that throws or an answer without a boolean ok leaves the outcome unknown, not failed; custom-field values are strings; and the ticket body is the description followed by the last 12 chat messages, each cut at 400 characters.

Security

  • Admin-only access. These need an Admin: changing, testing or importing an integration’s configuration, including Zendesk and custom API connections and actions, which Members can still read; the traffic logs, which now also work with OIDC sign-in; accepting AI-instruction proposals; and reading deletion requests. Members see Accept disabled on instruction proposals and Save disabled on Suggested Content. A Member refused on the audit log or the traffic logs is recorded as an auth.denied audit event.
  • Integration credentials are kept out of application reads, the SDK session response, logs and audit events, and the session response carries only the settings the SDK uses. Audit lists and exports mask credential fields in earlier events as [REDACTED]. An Admin API client that sends back a configuration it read keeps the stored credentials, unless a block’s address, such as the Zendesk subdomain, changed.
  • Image security updates:
    • Brotli (CVE-2026-111497), all four images. No path lets untrusted input reach the decoder.
    • Chrome for Testing 156.0.8078.80, worker. Affected in principle, since the worker browser loads the help sites you configure; fixed by the update.
    • PostgreSQL client libpq-18 in place of libpq-16, backend and worker. Not affected.
    • Expat, patched library (CVE-2026-113351), backend and worker. Refusing sitemaps that declare a DTD mitigates CVE-2026-115506, since the worker never parses a DTD.
    • Python 3.14.10 (CVE-2026-115557), backend and worker. The image build refuses an older Python.
    • CA certificates (CVE-2026-111943), backend and worker, with the distrusted emSign roots removed from the system and Python stores.
    • pyasn1 0.6.5, Pillow 13 and certifi 2026.10.23, backend and worker.
    • Next.js 15.5.28 (GHSA-2464-fhr4-w9xw, GHSA-p3m5-59mp-jv5r) and Node.js 24 (CVE-2026-105482, CVE-2026-106069), dashboards.
    • Playwright 1.63.0, worker.
    • An unused LLM client library, which carried its own OpenSSL, is removed from the backend.

Known issues

  • Outage detection depends on wording: some outage reports aren’t read as Critical, and some other reports are. The user can change the impact on the card wherever the ticket intake shows it (by default, under Review details).
  • With Filing Off, a card drafted before the switch shows “Couldn’t submit right now. Nothing was sent — please try again.” at Submit; switching back to Live lets it through. The procedure Playground and Copilot’s simulations ignore the switch.

0.16.15 — 2026-10-19

Fixed

  • Required fields that a conditional branch of the Zendesk form opens, but that no intake role maps, appear on the ticket card as empty required rows, so the user can fill them and a validation error points at the row. Fields the Ticket intake page fills another way (a pinned value, the assistant, or your backend) keep their handling. A stored intake picks these fields up at its next form sync.

Security

  • Worker browser update to Chrome for Testing 155.0.8059.84.
  • Backend and worker dependency security updates: psycopg 3.3.6 and psycopg2-binary 2.9.13, which bundle OpenSSL 3.5.8, and cryptography 50.0.2, which bundles OpenSSL 4.0.3.

Changed

  • No database migrations in this release.

0.16.14 — 2026-10-15

Security

  • Base image security refresh for all four images: OpenSSL 4.0.3 (CVE-2026-54873, CVE-2026-72897, CVE-2026-84782, CVE-2026-84784) and zlib (CVE-2026-85091).

Changed

  • No database migrations in this release.

0.16.13 — 2026-10-15

Fixed

  • Changing a dropdown on the ticket card refreshes the fields its value requires, nested conditional fields included, so the fields a branch opens appear on the card. Values from branches that are no longer selected are dropped at submit.
  • When validation rejects a field, the card marks that field’s row with the reason and opens it, instead of showing only a generic note.

Changed

  • The refresh after a dropdown change, the per-field reasons and opening rejected rows need SDK 0.16.13. Older SDKs keep working: nested fields are offered when the draft already carries their parent values, and a rejected row is marked “This value wasn’t accepted.”
  • No database migrations in this release.

0.16.12 — 2026-10-15

Fixed

  • Refreshing or reprocessing a page that now produces fewer chunks no longer leaves its old trailing chunks in knowledge search. The worker’s cleanup before re-chunking failed against Pinecone’s preview fetch API, logged Batch pinecone cleanup failed, falling back per-page, and deleted nothing. First ingests and full source refreshes were unaffected apart from the warning. After upgrading, refresh an affected page, or run a full source refresh, to remove its stale chunks.

Changed

  • No database migrations in this release.

0.16.11 — 2026-10-14

Security

  • Worker browser update to Chrome for Testing 155.0.8059.39.

Changed

  • No database migrations in this release.

0.16.10 — 2026-10-13

Security

  • Worker dependency security updates (anyio, PyJWT, CairoSVG, fsspec, urllib3).

Changed

  • First complete on-prem image set since 0.16.7. It carries the 0.16.8 and 0.16.9 fixes.
  • No database migrations in this release.

0.16.9 — 2026-10-13

The worker image for 0.16.9 was not published because it was held by the vulnerability gate. On-prem deployments should use 0.16.10 or later, which carries every 0.16.9 fix.

Security

  • Backend dependency security updates (anyio, PyJWT, CairoSVG, langgraph-sdk, urllib3).
  • Dashboard dependency security updates: sharp 0.35.5 (GHSA-wq5f-xc86-pv6w) and source-map-js 1.2.2 (CVE-2026-93749).

Changed

  • No database migrations in this release.

0.16.8 — 2026-10-13

The images for 0.16.8 were not published because they were held by the vulnerability gate. On-prem deployments should use 0.16.10 or later, which carries every 0.16.8 fix.

Fixed

  • Knowledge search returns documents again. Pinecone’s preview search API began rejecting a reserved field that Moss requested, so every search returned no documents and the assistant answered without the knowledge base.

Changed

  • No database migrations in this release.

0.16.7 — 2026-09-17

Added

  • Multi-architecture on-prem images. Every on-prem image tag (moss-backend, moss-worker, moss-dashboard, moss-dashboard-oidc) resolves to an OCI image index with linux/amd64 and linux/arm64 images. Cosign signatures and CycloneDX SBOM attestations cover the index digest and each architecture digest. Mirror the full index together with its signature and attestation tags; docker pull and docker push flatten it to one architecture.

Changed

  • Each image tag resolves to an index digest instead of a single-image digest. Verification by tag is unaffected. Deployments that pin a digest re-pin once, to the index digest or to the architecture digest they want.
  • The SDK package loads in CommonJS projects: main and the require condition point at a CommonJS build with its own type declarations.
  • getAppearance and the appearance preset objects are no longer declared by the SDK; select a preset with the appearance config option.
  • SDK tokens whose metadata claim is present but not an object are rejected with 401 INVALID_TOKEN. A non-empty list, string or number caused a server error before, and an empty one ([], "", 0, false) was accepted.
  • The worker drops unused dependencies, nltk among them.
  • Base image refresh.
  • No database migrations in this release.

Fixed

  • FloatingActionBar, fetchAppConfig, convertBackendConfig, mergeAppearance, MOSS_TRIGGER_ATTR and installModalDialogEscape can be imported from the SDK package. Its type declarations listed them, but the bundle did not carry them.

Security

  • Dashboard dependency security update: Next.js 15.5.24 and sharp 0.35.4 (CVE-2026-75604, GHSA-2xp9-vwfh-vxw4, GHSA-rgj7-g3m4-5g8c).

0.16.6 — 2026-09-05

Added

  • Application export and import. Admins can export a versioned bundle from Configure > General and import one from the application list through an Import dialog. The bundle covers the authored application alongside its configuration: procedures, API connections and actions, integration connections, monitors, intent tags, and the self-healing schedule. Credentials are never included and are re-entered on the target. Integration connections in direct-filing mode are reported for completion rather than created. Import runs in one transaction, and the dry-run report lists what would be created, remapped identifiers, every value written differently from the bundle, and a re-entry checklist. The dashboard export produces the full bundle by default, and the configuration-only bundle remains available. Guidance flows are not part of the bundle. The endpoints are GET /api/v1/applications/{id}/export and POST /api/v1/applications/import?dry_run=true.
  • Dashboard-managed appearance controls (first pass). A new theme slice and a footer background tint can be edited from the dashboard. The SDK applies dashboard-saved values over the host prop per leaf, so precedence is preset, then host prop, then dashboard value. Hosts must upgrade the SDK to 0.16.6 to see dashboard theme values.
  • The SDK chat window can now be dragged from the full header and from the window edges.
  • Dashboard toasts have a dismiss button, and error toasts stay until dismissed.

Changed

  • Clearing the last leaf of a configuration slice now removes the slice instead of storing an empty object.
  • The SDK package no longer ships sourcemaps in the tarball and now carries a public README and package metadata.
  • Upgrade order for split deployments: backend, then dashboard, then SDK.
  • The application prune script now covers every application-owned table, removes stored DOM snapshot objects, resolves tables through the configured database schema, and refuses to run when it detects cross-application references or cannot count a table.
  • No database migrations in this release.

Known issues

  • For a stored footer that contains only links, the public configuration endpoint reports enabled: false.
  • Importing into an organization whose install already holds an application with the same name is refused. Application names are currently unique per install.

0.16.5 — 2026-09-02

Security

  • Worker dependency security update (nltk).

Changed

  • First complete on-prem image set that carries the 0.16.4 fixes. On-prem deployments that skipped 0.16.4 should upgrade directly to 0.16.5.
  • For split deployments, upgrade the backend before the dashboard.

0.16.4 — 2026-09-02

The worker image for 0.16.4 was not published because it was held by the vulnerability gate. On-prem deployments should use 0.16.5, which carries every 0.16.4 fix.

Fixed

  • The dashboard pages for Procedures, Custom API Actions, Integrations, Monitors, and Proposals now work on deployments that sign in with OIDC.
  • Per-prompt zero-state icons configured in the dashboard now render in the SDK exactly as configured, and turning them on takes a single save.
  • The public SDK configuration endpoint now tolerates unknown zero-state keys instead of failing.

Changed

  • Base image refresh.

0.16.3 — 2026-08-27

Fixed

  • SDK tokens that carry an audience (aud) claim are no longer rejected when the deployment does not require one.
  • Reasoning levels that the configured model does not support are now coerced to the nearest supported level instead of failing the request.
  • Saving the dashboard Appearance or Security form no longer overwrites SDK appearance presets, and each form saves only its own fields.

0.16.2 — 2026-08-23

Security

  • Base image security refresh. This is the first published on-prem image set of the 0.16 line.

0.16.1 — 2026-08-23

The images for 0.16.1 were not published because they were held by the vulnerability gate. On-prem deployments should use 0.16.2 or later, which carries every 0.16.1 fix.

Fixed

  • The public SDK configuration endpoint is now served in every backend service mode, so split deployments that expose only the SDK backend can bootstrap the SDK.
  • The public key field on the dashboard Security page no longer overflows its container.

Security

  • Dashboard dependency security update (nanoid).

0.16.0 — 2026-08-17

Added

  • Support Ticket Handoff — the assistant can collect a support request, show the user a review card, and hand the confirmed ticket to your application to file through your own backend. See Support Ticket Handoff. Ticket filing is offered only on applications where it is configured and only to sessions whose host registered the onSupportTicket callback.
  • Session detail — the dashboard session page groups the transcript more clearly and closes several display gaps.

Fixed

  • Multiline input (Shift+Enter) now also works in the compact guide bubble.
  • Dashboard deep links that cross organizations now recover to the application’s owning organization instead of failing.

0.15.4 — 2026-08-10

Security

  • Dependency security updates.

0.15.3 — 2026-08-10

Added

  • Knowledge base progress — documentation sources show a segmented progress bar with separate scrape and process phases.

Fixed

  • Continuing a guide no longer records a duplicate user turn in the session transcript.
  • On-prem OIDC sign-in hardening, including support for multiple accepted audiences.

0.15.0–0.15.2 — 2026-08-06

Added

  • Session transcripts redesigned — the dashboard renders sessions as a chat stream with paired stimulus/response cards, user role labels, and date dividers.

Fixed

  • Soft line breaks in assistant responses now render correctly in chat, streaming, and guide instructions.

0.14.0 — 2026-08-06

Added

  • Multiline chat input — Enter sends, Shift+Enter adds a line.
  • Assistant responses in dashboard session transcripts render as markdown.
  • Citations from Document360 sources now link to the public portal URL.

0.13.0 — 2026-08-05

Changed

  • The SDK version now tracks the Moss platform version.
  • Session resume card redesigned: centered rating prompt and a compact choice between continuing and starting a new session.

Fixed

  • Prevented stale responses from appearing after starting a new chat.
  • Prevented resuming onto a stale pending interaction.