github artifact-keeper/artifact-keeper-web v1.12.0
Artifact Keeper Web 1.12.0

one hour ago

The web UI for backend 1.12.0. Highlights: MCP Registry and Agent Plugins repositories, a vulnerability drill-down behind every dashboard count, durable acknowledgements and the proxy enforcement preview, instance quotas, a scan queue view, live migration progress, and the TOTP step after LDAP and SSO sign-in.

Sponsors

Thank you to our sponsors for supporting ongoing development of Artifact Keeper.

Backers

Become a sponsor to support the project and get your name listed here.

Thank You

  • @nicola-preda for the vulnerability drill-down and clickable dashboard counts (#1143, #1144), the scan queue page (#1137), live migration progress and Start All (#1102), the read-only upload size limit (#1140), the Nexus Basic Auth source form (#1139), hiding not-applicable scans (#1100), Maven versions for proxy-cached files (#1099), the RT-15 E2E seed fix (#1146), the Storage card fallback before storage stats run (#1169) and paging the Permissions list past 100 (#1182).
  • @coddan for keeping a record_only proxy scan action (#1122), keeping the age gate's instance lock on save (#1126), the session refresh on a fresh load after the access cookie expires (#1125), showing hidden and bidi characters in stored text (#1127), the Maven POM detail dialog (#1123) and showing the Environments tab only where it can be used (#1124, merged as #1211).
  • @imanlotfimahyari for counting proxy storage once in the storage totals (#1093).
  • @loyalpartner for buf CLI steps that buf can run in the protobuf setup guide (#1065).
  • @der-eismann for reporting that a Maven POM row opened no detail dialog and could not be deleted (#1035), and that Maven repositories showed an Environments tab that does nothing (#1037).
  • @Kimahriman for reporting that the PyPI Tracks tab was only on virtual repositories (#1066).
  • @cervotoc for reporting that the UI looked signed out on a fresh load after the access cookie expired (#1084).
  • @knowinglyAnonymous for reporting that a stored record_only proxy scan action showed as fail-open and was rewritten on save (#1069).
  • @allanmc for reporting that a Nexus migration source connection tested green with an API token and then failed (artifact-keeper#5479, #1139).

Compatibility

  • Backend 1.12.0 needs web 1.12.0, or at least web 1.11.3 (artifact-keeper#5663) - Backend 1.12.0 reports a package's download_count as null when it cannot tie the package to its downloads, which is the case for a Remote or Virtual Maven package cached through a streamed pull-through (artifact-keeper#5419). Web 1.11.2 and earlier format that count without a null check, so against backend 1.12.0 /repositories, /packages and every Remote or Virtual Maven repository page show "Something went wrong". Web 1.12.0 shows a dash for a count that is null, missing or zero (#1085), and so does web 1.11.3, which has the same fix (#1178). Web 1.12.0 also carries the UI side of other backend 1.12.0 changes: the TOTP step after LDAP, OIDC and SAML sign-in (artifact-keeper#5484), the "finished in another browser" SSO message (artifact-keeper#5486), the read-only upload size limit (artifact-keeper#5751), the repository-scoped acknowledgements (#1072) and the removal of the pull replication mode (#1082). Upgrade the web UI together with the backend.
  • Repository delete answers 202 on backend 1.12.0 (#1048, artifact-keeper#4311) - Backend 1.12.0 accepts a repository delete with 202 Accepted and removes it in a background job. Web 1.12.0 says the deletion has started, then reports when the job finishes or fails, or that it is still running in the background if it has not finished after 30 minutes or its status cannot be read. Against backend 1.11.x a delete still answers 200 or 204, and web 1.12.0 shows "Repository deleted" as before.
  • Web 1.12.0 also runs against backend 1.11.x - No screen that works on backend 1.11.x stops working. Against backend 1.11.13, the E2E interaction and role suites pass, except the cases that test backend 1.12.0 behaviour: the RT-15 MCP rendering suite, which needs the mcp format, and the upload limit refusal. The screens for 1.12.0-only features:
    • hide themselves: MCP Registry is not offered as a format, Vulnerabilities is not in the sidebar and the dashboard counts stay plain numbers. Against 1.11.13 and earlier the quota section and banner stay hidden too; backend 1.11.14 has the instance quotas (artifact-keeper#5912), and they work there.
    • say the backend is too old: Scan Queue.
    • show a load error: Acknowledgements and Proxy Blocked, which stay in the Security sidebar.
    • Agent Plugins is still offered in the create dialog, and the backend refuses to create it.
    • Settings > Storage shows the stored upload size value read-only, so it cannot be edited from the UI. On 1.11.x that value was never enforced; the limit in force there is MAX_UPLOAD_SIZE.

Added

  • A vulnerability drill-down behind the dashboard severity counts (#1141, artifact-keeper#5865) - Every security dashboard counted CVEs and none of them could be opened, so "7 high" was a number with no way to see which seven. /security/findings lists exactly the rows a tile counted: filterable by severity, state and repository, with a group-by-artifact view, a row click through to the CVE's blast radius and a cell click through to the artifact. Filters live in the URL so a tile can deep-link into a pre-filtered list, and each one is validated before it is sent on. A row reads Acknowledged only while a live decision covers it (artifact-keeper#5153), and its tooltip says who decided, why and until when. The advisory link is derived from the identifier through the shared VulnIdLink, so no link target comes from API data. An older backend without the endpoint gets an upgrade notice instead of a raw 404. Vulnerabilities is in the Security sidebar group, not gated on a scanner, since findings outlive the scanner that found them, but shown only when a one-row request to the endpoint succeeds: it answers only an instance administrator with an unscoped credential, so a repository-scoped admin token (403) and an older backend (404) do not see the entry, and opening the page directly explains the 403.
  • The dashboard vulnerability counts are clickable (#1142) - The Overview severity bars, its Open and Acknowledged counts, and the security dashboard's Open Critical / Open High / Open Findings tiles and per-repository severity and Ack'd cells all led nowhere, so a number was the end of the trail. Each one now links to the matching /security/findings list, scoped to the same filters that produced the number: the severity counts exclude acknowledged findings, so those links carry state=open and return exactly what was clicked. The tiles use StatCard's href (#822) rather than an onClick, so they are focusable and open with the keyboard. A zero count stays an inert span: there is no empty list to send anyone to, and Fixed is derived by diffing scans so it has no row list at all. The counts link only for a user the list endpoint (artifact-keeper#5865) answers, which is an instance administrator with an unscoped credential: for a repository-scoped admin token, or on a backend without the endpoint, they stay plain numbers. Trade-off: the per-repository cells come from the stored security score, recalculated after scans and decisions, while the list is live, so the two can differ briefly.
  • Instance storage and transfer quotas on the dashboard, in a banner and on refused uploads (#1149) - The backend can now enforce an instance-wide storage quota and a monthly download transfer quota set by the administrator, but the UI showed neither, and an upload refused over a quota failed with a generic message. The dashboard now has a Quotas section: storage used against the quota, with a percentage and a bar, and "Transfer this month" with the date the period resets (in UTC). Administrators' figures come from /api/v1/admin/stats and everyone else's from /api/v1/system/quota. A banner at the top of every page warns when either quota reaches its warning threshold, when the storage quota is full (new uploads are refused, downloads still work), when the transfer quota is past 100 % and in its grace allowance, and when downloads are refused until the period ends. It links the help page the administrator configured, if any. Each notice can be dismissed for the session and comes back if its state gets worse. An upload refused with 403 STORAGE_QUOTA_EXCEEDED / QUOTA_EXCEEDED, 413 PAYLOAD_TOO_LARGE or 429 TRANSFER_QUOTA_EXCEEDED now shows the server's message and help link, for single and chunked uploads, and a refused chunk is no longer retried. The upload picker uses max_upload_bytes from the quota response for its client-side size check, as well as the system config. Against a backend without these fields, or one that answers 404 on /system/quota, nothing changes. Trade-off: usage is refreshed every five minutes and after a refused upload, so the banner can trail the server's state by that much; a bare 413 without the structured code (a proxy in front of the backend) still gets the fixed "exceeds the maximum upload size" text, since its body is only a reason phrase.
  • Migration progress keeps moving after a reload or from another tab (#1101) - Progress only advanced while the event stream of the tab that clicked Start stayed open, so reloading the page or opening it elsewhere showed a frozen job until the next manual refresh. The job list and the open job's items now poll every 3 seconds while a job is running or assessing. A Start All button starts every pending or ready job, and a sticky alert names each job you started when it finishes. Trade-offs: Start All starts every job at once and the backend has no limit on how many migrations run in parallel, so a long list of large jobs competes for the same workers and upstreams; the finish alert covers only jobs started from the open page and is lost when you navigate away.
  • MCP Registry repositories (#1076, artifact-keeper#5253) - The backend adds an mcp format that serves the MCP registry API v0.1 at /mcp/<repo>/v0.1/..., so VS Code's MCP gallery, Copilot's registry policy and mcp-publisher can use an Artifact Keeper repository instead of the public registry, and by default refuses any server whose packages are not hosted or proxied by the same instance. It can now be created as "MCP Registry" (offered only when GET /api/v1/formats lists mcp; Local, Staging, Remote and Virtual), with its namespace allowlist, package policy (own-repos, or open, labelled as a catalog whose packages are not resolved or gated) and backing repository per package type in the create dialog and the Settings tab. The repository opens on a Servers gallery of cards (transport, source, secrets and risk badges) with search, filters and an Install sheet; there is also a cross-repository MCP Servers page in the sidebar. A server page has Overview, Install, Packages & Remotes, Configuration, Versions, Security, Lineage and Tools tabs. The Install panel writes configurations for Claude Code, VS Code (plus the registry's gallery serviceUrl), Cursor, Claude Desktop, Windsurf and plain JSON from the coordinates the registry serves; a value typed into a secret field is never put in a snippet or stored, the snippet carries a placeholder the client expands. A version clients cannot see shows a "Why can't my client see this?" panel: a repository holder gets the reason the backend gives (pending review, refused by policy, a package it runs is held or gone, age gate), anyone else is told only that it is not served, because the registry deliberately answers them with the same 404 as an unknown name. The Setup Guide covers the VS Code gallery, Copilot's policy URL, publishing with curl or mcp-publisher, and why Claude Code has no registry subscription. Settings also carries the later-phase options the backend validates and echoes: deny globs, and for Remote repositories the posture, auto-approve namespaces, floating OCI tag policy and sync scope (allowlist or mirror), plus the remote probe, sandboxed introspection, OCI name-label and signed-bundle rules; an instance administrator can lock the posture and probe settings. For Remote repositories the curation page has an approval queue over the backend's catalog API: Each row shows risk, transports, secrets and lint from the catalog's summary. You can approve or block a name, pin one version, re-decide a version whose upstream content changed, decide a whole selection at once, and approve or block by full-name glob after the backend previews the matching names. The queue also has a catalog of every upstream name with "Request approval". A dashboard tile counts what awaits a decision. On a server page in a Local or Staging repository any signed-in user gets Deprecate, Mark active and Delete version, which need only repository write (Delete version uses the registry's soft delete, DELETE .../versions/{version}), and a refusal says whether write or a namespace grant is missing; a Remote or Virtual repository, which refuses every version write with 405 mcp_read_only, offers none of them, there or on its Artifacts tab (#1088). The curation link and probes stay with repository administrators. The source badge says whether a version is hosted, proxied or mirrored, and whether the backend verified it against the official registry. Elsewhere: Tools snapshots with their provenance (declared, CI-reported, observed), the health roll-up and lint findings on the Security tab and in the lineage graph, signed MCPB bundles, and namespace grants as permission rules with mcp_namespaces conditions. Each of these surfaces hides itself when the backend answers 404, so an older backend shows only the hosted registry screens. The namespace grant requirement is offered in Settings only, because the repository API ignores fields it does not know and the create dialog could not tell whether it was applied. Package filters in the gallery apply to the loaded pages, since the registry API has no parameters for them. An MCP repository's Artifacts tab and artifact dialog offer no generic Delete, which the backend refuses with 400 on MCP repositories; they open the server in the Servers tab instead, where a version is deleted through the registry's soft delete.
  • Agent Plugins repositories (#1074, artifact-keeper#5251) - The backend adds a hosted agentplugins format for AI coding-agent plugin bundles. It can now be picked as "Agent Plugins" when creating a repository, its /agentplugins/... routes (including the token-in-path /agentplugins/t/<token>/... form) are proxied to the backend, and the repository opens on its Packages tab, where the install snippet is ak agent plugins install <name> --repo <repo> --harness claude. The Setup Guide has an "AI agents" category with three tabs: Claude Code (claude plugin marketplace add with the repository's claude-marketplace.json, and, for a private repository, either an extraKnownMarketplaces entry in Claude Code settings that sends the read token as a header, which is the recommended route, or a read token in the URL path, with a note that the token is then kept in Claude Code's marketplace record and shows in server access logs), Cursor / Codex / VS Code (through ak agent plugins install --harness), and Publish (ak agent plugins publish or curl -T). Trade-off: only Local repositories exist for this format, but the create dialog does not limit the repository type per format, so it still offers Remote and Virtual, which the format does not support yet.
  • Proxy Blocked page, with a preview of what enforcing would refuse (#1072, artifact-keeper#5232) - record_only exists to try enforcement out, but nothing in the UI showed what it would refuse, and the proxied files a gate refuses today had no page either. Security now has a Proxy Blocked page, and a Preview enforcement toggle on it asks the backend to grade each repository's stored verdicts as if it enforced. The summary keeps the two kinds of number apart: files whose stored verdict would be refused, and files with no usable verdict (no verdict, stale, or clean but graded nothing), which fail_closed scans on the next pull and may serve, refuse or withhold. Those three are shown as "up to", since they come from the catalog and include files the format serves without gating. The dashboard gains Proxy Blocked and Proxy Would Block tiles. The page needs an instance administrator with an unrestricted credential, like the API, and the preview runs only when asked for because it is slow on a large catalog. Trade-off: the preview reads stored verdicts only, so a pull can still differ after a CVE database update, for a format that reuses a verdict per coordinate, and on a Virtual route; the page says so.
  • Acknowledgements page, expiry and scope on the acknowledge form (#1072) - An acknowledgement is now a decision that can outlive any finding row, and there was nowhere to see one or undo it without such a row. Security now lists decisions and revokes them by id. The acknowledge form takes an optional expiry, and offers "Every repository" to an administrator who may act everywhere, which it learns by asking the decisions list, the same check the backend applies.
  • Promotion result says whether the scan verdict travelled (#1072) - A promotion can succeed and leave the target seeing an unscanned artifact, or leave an acknowledgement behind, which a target policy with block_unscanned or the target's own thresholds then act on with no explanation. The promotion dialog now shows the backend's reason and stays open when that happened. Approving a promotion request does not report this in its response, so the approvals page is unchanged.
  • Curation tab on a Remote MCP repository for its administrators (#1087) - The backend lets a Remote MCP repository's administrators approve, block, pin and bulk-decide its catalog rows, but the approval queue was only on /curation, which needs an instance administrator, so they had to use the API. A Remote MCP repository's page now has a Curation tab with that repository's queue (Awaiting decision, Approved, Blocked, Catalog, glob approve/block) for anyone its repository response reports mcp_caller_is_repo_admin for, and the server page's "Open in curation queue" opens it. /curation, with every repository, and the lock switches in Settings stay with instance administrators.
  • A live view of the durable scan queue (#1135, artifact-keeper#5775) - With the backend's new opt-in scan queue (SCAN_QUEUE_ENABLED) the scans run on a worker, and the only way to tell whether they were draining, stuck or failing was its metrics. With SCAN_QUEUE_ENABLED set on the backend (artifact-keeper 1.12.0 or later), /security/queue (admin only) shows the depth per state, the age of the oldest pending job, and the rows themselves with their kind, target, attempt count, claiming worker and last error. It polls every 10s and says so on the Refresh button. A finished job is deleted, so the page only ever shows pending, running and failed work. When the backend runs with the queue off the page says so, since scans then run in-process and never show up here, and a backend without the endpoint gets a notice instead of an error and is not polled again. Scan Queue is in the Security sidebar group, not gated on a scanner, since the page itself says when the queue is off. The page is read-only, as the backend endpoint is: a failed job cannot be retried or removed from here, and a hosted-artifact job shows its artifact id as text, since there is no artifact page to link it to.

Changed

  • The C1 shield logo and favicons (#1070) - The sidebar, login page, error page, favicon and app icons now use the C1 shield mark, with a 96 px variant for high-density screens and icon.png / apple-icon.png for browsers and home-screen shortcuts. The unused Next.js boilerplate SVGs are gone from public/.
  • Pull replication is no longer offered (#1082, artifact-keeper#5293) - The backend never implemented pull replication: a pull subscription created no sync tasks and never replicated, and the backend now refuses pull with 400. Sync Policies offers Push and Mirror, and shows pull as a disabled "not implemented" item only on a policy saved with it, so editing it does not render a blank mode. The per-repository replication mode offers Push, Mirror and None, with Pull listed as disabled "not implemented". The replication tooltip no longer says Pull fetches from the peer.

Fixed

  • Package versions and downloads read as the backend reports them, and a rejected file is badged Rejected (#1085) - A rejected file in a repository's file list was badged "Quarantined"; it is now badged "Rejected", a conda withdrawal "Withdrawn", and a released or expired hold shows no badge. Versions of revision- or tag-style formats (Hugging Face, ML models, Git LFS, generic, the OCI family and container images) show as stored, so a Hugging Face revision reads main instead of vmain, and a Go version that already starts with v is not doubled. A package or version download count that is null, missing or zero shows as a dash, or is left out, instead of "0 downloads" or an error; this is the fix that keeps the repository pages working against backend 1.12.0 (see Compatibility). Trade-off: since backend 1.12.0 a 0 is a real count, and it is still shown as a dash here; showing it as 0 is a follow-up.
  • The Hugging Face install hint downloads through this repository (#1080, artifact-keeper#5290) - The hint was huggingface-cli download <id> with no HF_ENDPOINT, so the copied command downloaded from huggingface.co, bypassing the registry, and used the deprecated CLI name. It is now HF_ENDPOINT=https://<host>/huggingface/<repo> hf download <id> --revision <version>, with a comment line about HF_HUB_DISABLE_XET=1 for the Xet 404. Without a known host or repository it uses placeholders, never huggingface.co.
  • LDAP, OIDC and SAML sign-in asks for the TOTP code on accounts that enrolled one (#1147, artifact-keeper#5484) - Backend 1.11.14 and 1.12.0 stop treating a directory password or an IdP assertion as a complete sign-in for an account with TOTP enrolled: LDAP login and the OIDC/SAML code exchange now answer with the same TOTP challenge as the local login, and set no session. This UI did not know that answer, so such a user who signed in through LDAP or SSO was sent back to the login page with no session and no message. The LDAP form and the /callback page now show the existing two-factor form and finish the sign-in at POST /auth/totp/verify; accounts without TOTP sign in as before. A background silent-SSO check (prompt=none) that meets a TOTP account stays signed out instead of prompting; the user signs in with the SSO button and enters the code then. Run this UI with backend 1.11.14 / 1.12.0 or later for TOTP accounts that use LDAP or SSO; against an older backend nothing changes.
  • Sizes are labelled in binary units (#1149) - Every size in the UI is computed in powers of 1024 but was labelled KB, MB, GB or TB, so it disagreed with tools that use decimal units by up to 10 %: 11,274,290,864 bytes showed as "10.5 GB" where it is 11.27 GB, or 10.5 GiB. Sizes now read KiB, MiB, GiB, TiB and PiB; the numbers are unchanged. Sizes above 1024 TiB now show in PiB instead of a four-digit TiB figure.
  • Saving the upload size limit on Settings, Storage always failed (#1132, artifact-keeper#5751) - Backend 1.12.0 reports the upload limit actually in force (MAX_UPLOAD_SIZE, 10 GiB by default) and answers 400 to any change of max_upload_size_bytes through the settings API, so every save from the MB/GB editor failed. Before that, the page showed and edited a stored 100 MB that nothing enforced. The row is now read-only like its neighbours, shows the limit from the backend ("No limit" for 0) and says it is set with the MAX_UPLOAD_SIZE environment variable and applied on restart. The Storage tab's read-only rows are now also labelled for screen readers. Trade-off: against an older backend the row still shows the stored value it reports, which that backend did not enforce.
  • A Nexus source connection offers Basic Auth only (#1138, artifact-keeper#5479) - The Add Source Connection form defaulted to API Token, which the backend ignores for Nexus and replaces with empty Basic credentials, so the connection tested green and later requests failed. Picking Nexus now switches to Basic Auth and hides API Token, with a note that a Nexus user token goes in as its name code and pass code.
  • Scans that did not apply to an artifact's format no longer crowd out the real results (#1098) - Security > Scans hides not_applicable rows unless picked in the status filter, and says how many it hid. The status filter gains its accessible name. Trade-off: the backend has no parameter to exclude a status, so the page filters the rows it was sent: a page can list fewer rows than the page size, or none (it then says why instead of "No scan results found"), the total still counts the hidden rows, and real results can still sit on later pages behind not-applicable ones.
  • A proxy-cached Maven file shows its version instead of a dash (#1097) - In a Maven or Gradle repository, every proxy-cached file showed a dash in the Artifacts tab's Version column, because a cached record carries no version, so two releases of the same jar looked identical. The column now falls back to the file's Maven coordinates: the backend-parsed metadata first, the group/artifact/version/file path second. Trade-off: a file outside the usual layout still shows a dash.
  • A clear message when an SSO sign-in is finished in another browser (#1148, artifact-keeper#5486) - Backend 1.12.0 completes an OIDC or SAML sign-in, and redeems its exchange code, only in the browser that started it, using an HttpOnly cookie scoped to /api/v1/auth/sso. A /callback?code=… link opened in any other browser is refused with 401, and the callback page showed only "Authentication failed. Please try again." It now says the sign-in was finished in a different browser than the one that started it, or expired, and to start again from this browser. The exchange request already sends cookies (credentials: "include"); a test now guards that, since the sign-in fails without it. Trade-off: the backend has no dedicated error code for this refusal, so the page matches the start of its message, and a reworded backend message falls back to the generic text.
  • Acknowledging or revoking a finding on a scan copied to other repositories (#1072) - Backend 1.12.0 makes an acknowledgement a decision scoped to a repository, and answers 409 when a scan has copies elsewhere and the request names no repository. That includes the source of every promotion, so those findings could not be acknowledged from the UI. Acknowledge and revoke now name the repository being viewed, and any remaining 409 gets a message saying what to do. Opening a reused scan from a list now carries which copy it is: such a row reports its source's id, so the page used to show the source's rows and would have acted on the source's repository. It now lists the copy's own rows and says it is a copy. Trade-off: the scan details at the top of that page are still the original scan's, since the backend reports no others.
  • A revoke no longer reads as complete when it is not (#1072) - A revoke undoes what the caller has authority over and leaves copies elsewhere in force. The UI said "Acknowledgment revoked." regardless, so an administrator could believe a risk had ended while it stayed accepted in another repository. The answer's count is now shown in a warning that stays until dismissed, on the finding, by-CVE (Reopen) and by-id paths.
  • Storage totals no longer count proxy cache bytes twice (#1039) - Analytics and Dashboard added proxy bytes to backend totals that already include them. Use the reported total and subtract proxy bytes for the Local tooltip and Storage Trend split; repository breakdowns and artifact/download counts retain their separate-bucket accounting.
  • Cargo setup guide gave an index URL cargo cannot use (#1129, artifact-keeper-site#179) - The Setup Guide told users to set index = "<origin>/cargo/<repo>/index". Without the sparse+ prefix cargo treats that URL as a git index, and with the /index suffix cargo asks for /cargo/<repo>/index/config.json, which the backend does not serve, so setup failed with config.json not found in registry. The guide now gives index = "sparse+<origin>/cargo/<repo>/", the registry root where the backend serves config.json and the index files. The cargo install note on an MCP server's Install panel had the same /index/ suffix and is fixed the same way.
  • Deleting a repository says the deletion has started and then reports how it went (#1048, artifact-keeper#4311) - From backend 1.12.0 a repository delete answers 202 with a background job. The toast now says the deletion has started, follows the job and reports completion, or a failure with the backend's reason and a Retry that deletes again. A backend that deletes at once (1.11.x) still gets "Repository deleted".
  • A Maven .pom row opens its detail dialog, so a POM can be inspected and deleted (#1035) - In a Maven repository's Grouped artifacts view, every file of a component opened the artifact detail dialog except the .pom, whose row was only a download link. That made the POM the one file you could not delete from the UI, so removing a bad release left the POM behind and the GAV still existed for Maven clients, and you could not see its size, SHA-256, download stats or SBOM and Security tabs. The POM row is now the same "Open details" button as the others and keeps its POM badge; the one-click download from #442 stays as a separate icon link beside it. Trade-off: downloading the POM is now an icon rather than the filename itself, so clicking the filename opens the dialog (which also has a Download button).
  • A repository's Storage card shows its size before the storage stats have run (#1131) - until the backend's stats refresher has visited a repository, GET /repositories/{key}/storage answers zeros with no computed_at, so a fresh remote that had already cached files read "Logical size 0 B" under a header saying "564 KiB used". The card now shows the repository's own storage_used_bytes in that case, with a line saying the breakdown is not computed yet. A virtual repository, which stores nothing itself and never gets a breakdown, shows the size held by its members instead.
  • The PyPI Tracks tab is on local and staging repositories (#1066) - The Tracks tab, where an administrator declares a PEP 708 tracks upstream for a project, was shown only on virtual PyPI repositories, but the backend accepts a tracks declaration only on the local or staging repository that owns the project, so every declaration made from the tab failed and the repository it belongs on had no tab. The tab now shows on local and staging PyPI repositories and no longer on virtual or remote ones; a virtual repository still applies the tracks declared on its local members.
  • Opening the UI after the access cookie expired no longer looks signed out (#1084) - The access cookie lasts as long as the access token (30 minutes by default) while the refresh cookie lasts longer. Opening a fresh tab, a bookmark or a typed URL after the access cookie had expired showed the signed-out shell (no user menu, no admin sections, a Sign In button) even though the refresh token was still valid: the page's data requests recovered on their own, but the first /auth/me call was never retried, so the shell stayed anonymous until you pressed F5. The initial load now refreshes the session once on a 401 from /auth/me and asks again before deciding the visitor is anonymous. That refresh is shared with the one a data request triggers (they join a single in-flight request), so the two cannot race and invalidate each other if the backend rotates refresh tokens. Trade-off: an anonymous visitor now makes one extra request, a refresh that returns 401, on every fresh page load; there is no way to tell from script whether the httpOnly refresh cookie exists.
  • The protobuf setup guide gave buf commands buf rejects (#1064) - It named the module <host>/proto/<repo>/myorg/mymodule, which buf refuses because a module name is exactly registry/owner/name, used buf push --registry, a flag buf does not have, and authenticated with buf registry login, which cannot validate a token against Artifact Keeper. A local or staging protobuf repository now gets steps that work with backend 1.12.0: give the repository a hostname whose root proxies to /proto/<repo>/, name the module after that hostname, set BUF_TOKEN=<token>@<host>, then buf push --create and buf dep update. A remote or virtual repository is told that the buf CLI works with hosted repositories only, since the backend answers the buf CLI there with unimplemented. Thanks to @loyalpartner for the fix.
  • The Permissions page pages its list instead of stopping at 100 (#1181) - the page asked for per_page=1000 in one request, the backend caps per_page at 100, so only the first 100 rules were listed and the rest could not be reached. The table now pages on the server, 100 rules a page at most, with the usual page size picker
  • Hidden and bidi characters in stored MCP and repository text are shown visibly, and over-long text is cut (#1107) - Repository descriptions, MCP tool, resource and prompt names and descriptions, a CI capability report's provenance label, and a Remote catalog's upstream title, description, request note and evaluation reason were displayed exactly as stored. They were inert (React escapes them) but misleading: a name such as \u202egnp.exe reads as exe.png, and zero-width or control characters make two different strings look identical. One shared formatter now shows control characters, the bidi embeddings, overrides and isolates, line and paragraph separators, the zero-width space, word joiner, BOM and Unicode tag characters (which can hide an instruction in plain sight) as a visible [U+XXXX], with a tooltip saying so, instead of dropping them. Text longer than 1000 characters (200 for names and titles) is cut with an ellipsis and the whole text is in the tooltip. It is applied to the MCP gallery cards, the server and package detail, the capability lists, the install and environment panels, the Remote approval queue and upstream catalog, the MCP server name in the gallery card and detail header, and the repository and staging repository headers. Trade-offs: the zero-width joiner and non-joiner are not escaped, because emoji sequences and Persian text need them, so a hidden character built from those two is not flagged; the bidi marks (LRM, RLM, ALM) and subdivision flags such as England's are not escaped either, because Hebrew and Arabic text and those emoji need them; homoglyphs (a Cyrillic \u0435 for e) are not detected, as there is no way to tell them from real text; the raw server.json viewer still shows the stored JSON as it is. The RT-15 e2e spec now expects the escaped rendering (it uses the same formatter).
  • A record_only proxy scan action is no longer rewritten to fail-open (#1069) - A repository set to record_only showed as fail-open, and saving any scanning field, for example the severity threshold, stored fail_open over it, so a repository that never blocked started blocking with nothing on screen saying so. The switch is now a Proxy scan action select with Fail open, Fail closed and Record only. A mode the UI does not recognise is shown as unknown and kept, and a save sends only the fields you changed. A stored severity threshold the list does not offer, such as info, is now shown and kept instead of leaving the select empty.
  • Saving a repository's age gate no longer re-locks a policy that was handed back (#1038) - Backend 1.12.0 lets an instance administrator lock a repository's age gate policy or hand it back to the repository's administrators, and a save that leaves out instance_locked locks it. The Age Gate settings never sent the field, so an instance administrator who changed only the minimum age re-locked a handed-back policy. The settings now show a Lock this policy switch with the current state, and every save sends the lock back as it is unless you change it.
  • Saving the scan settings no longer re-locks a proxy scan policy that was handed back (#1185) - Backend 1.12.0 lets an instance administrator lock a repository's proxy scan policy or hand it back to the repository's administrators, and a save that leaves out instance_locked locks it. Neither the Scanning & enforcement settings nor the repository's Security tab sent the field, so an instance administrator who changed, for example, the severity threshold re-locked a handed-back policy. Both now send the lock back as it is, and the settings show a Lock the proxy scan policy switch with the current state. A repository with no scan settings yet is still locked by its first save, as the backend decides.
  • The Environments tab is only shown where it can be used (#1037) - Since 1.11.0 every repository showed an Environments tab to any signed-in user, including Maven, where it did nothing useful: the backend stores pixi.lock, conda-lock.yml, package-lock.json, Cargo.lock, poetry.lock and uv.lock and has no Maven, Gradle, container, Helm, Debian, RPM, NuGet, Go or generic lockfile, the advisory feed is empty without registered environments, and the cross-repository "which environments contain this component" lookup is not about the repository you are viewing (it is also on Security > Blast radius). The tab now shows for conda, npm, cargo and PyPI-family repositories (including conda_native, poetry and jupyter), and for any repository that already holds a registered environment, so data registered under an unrelated repository stays reachable. Trade-offs: for the other formats the web makes one extra request per repository page to see whether anything is registered (the same cached request the tab uses); yarn and pnpm repositories do not get the tab, since only package-lock.json is a supported npm lockfile.
  • The Permissions page says a column sort orders the current page only (#1187) - Since the list pages on the server, sorting by Principal, Target or Created only reordered the rules on screen, which read as a sort of the whole list. The backend lists rules newest first and takes no sort parameter, so when the rules span more than one page the table now says the sort applies to this page.
  • The Permissions table shows when the next page is loading (#1188) - After a page change the old rules stayed up, looking current, until the new page arrived. They now stay up dimmed, with the table marked busy for screen readers, until the next page loads.
  • Hidden and bidi characters are shown visibly in package descriptions and the remaining MCP names (#1202) - Package descriptions on the packages list, the package detail page and a repository's Packages tab were rendered exactly as stored, so a bidi override or zero-width character in a description was invisible; they now use the same formatter as the MCP text (#1107). The MCP server name's namespace and short name in the gallery card and detail header, and the server name in the lineage view, now get the [U+XXXX] tooltip and marker instead of a plain escaped string, and the accessible labels that carry an MCP name (the input schema of a tool, the Install button, the lineage list, and the catalog's Request approval and Select controls) are escaped too. Three gaps in the formatter are closed: when text is both cut and escaped, the tooltip now says hidden characters were found as well as showing the whole text; text is cut only between graphemes, so a joined emoji or a flag at the length limit is dropped whole instead of split; and only the three subdivision flags Unicode recommends (England, Scotland, Wales) keep their tag characters, so a flag-shaped sequence with any other tag letters is escaped. Trade-off: a regional flag such as \u{1F3F4} + usca that a platform might draw is now shown with escaped tags.
  • A remote instance's 401 on /auth/me no longer refreshes the local session on every load (#1204) - With a remote instance active, /auth/me goes through the proxy, so a 401 there can come from the remote refusing its stored API key rather than from the local session. The on-load refresh added for #1084 treated both the same and refreshed the local session on every page load. A 401 on /auth/me while a remote instance is active now first asks this backend's own remote-instance list, which needs the local session: only a 401 there leads to the refresh. Trade-off: with a remote instance active, a load that gets a 401 from /auth/me makes one extra local request.
  • The repository delete job is no longer polled after you leave the page or sign out (#1190) - After a delete that answers 202, the Repositories page polls GET /api/v1/repository-deletions/{id} and reports the outcome in a toast. The poller kept running after the page was left or the user signed out, and after a sign-out every read failed as unauthenticated. It now stops when the page unmounts, replacing the toast with a note that the deletion keeps running in the background, and on sign-out, dropping the toast. No read is made after either, and a pending wait between reads is cut short.
  • Deleting a repository another user is already deleting says so (#1191, artifact-keeper#4311) - A repeated delete of a repository whose deletion job is running returns that job, but backend 1.12.0 shows a job's status only to the user who started it and to instance administrators, and answers 404 to anyone else. A non-admin who deleted a repository someone else was already deleting got "Its progress could not be read". The toast now says the repository is already being deleted, that someone else started the deletion so its progress is not shown here, and that it keeps running in the background. The poller stops on that 404 instead of retrying it.
  • A stored proxy scan action or severity threshold the UI does not list can be chosen again (#1194) - The Scanning & enforcement settings show a stored value they do not offer, such as an unknown proxy scan action or an info threshold, as an extra option, but it disappeared as soon as you picked something else, so the only way back was to reload. The option now stays, and choosing it again counts as no change. The Proxy scan action select no longer offers Record only against a backend older than 1.11.0, which saves it as fail-open.
  • The scan settings no longer send a stale instance lock, and the Security tab shows the lock (#1213) - Both the Scanning & enforcement settings and the repository's Security tab send the stored instance_locked back on every save, so saving while the other form's change was still being refetched could send the old lock. Save now waits for the refetch to finish. The Security tab shows whether the proxy scan policy is locked and points to Settings to change it. On backend 1.12.0 or later, the settings now say that the first save on a repository with no scan settings yet locks the policy, which the backend decides.
  • Storage cards no longer show a negative Local share, and the Storage card fallback keeps what still works (#1214) - The Local half of the storage totals on the dashboard and the analytics page (total minus proxy bytes) is clamped at 0, so an old snapshot with a hosted-only total and backfilled proxy bytes reads 0 B instead of "--". A virtual repository on a backend that does not send member_storage_used_bytes gets the virtual-repository note with the members' size shown as unavailable, instead of 0 B and the generic note. Before a repository's storage breakdown is computed, administrators still see the "Instance unique" figure and the "Estimate" button for reclaimable space (not on a virtual repository, which has nothing to reclaim).
  • The protobuf setup guide's nginx snippet serves the buf hostname over TLS and accepts larger pushes (#1193) - The buf CLI only connects over HTTPS, so the snippet is now a listen 443 ssl server block with certificate lines, and the step says the hostname needs a TLS certificate. It also sets client_max_body_size 256m, the limit of the backend's protobuf routes, since nginx's 1 MiB default answered a larger buf push with 413.
  • Every Unicode default-ignorable character is shown visibly in untrusted text (#1226) - The formatter for package and MCP names and descriptions escaped control and bidi characters but let other invisible characters through, so pay\u2062pal (an invisible operator) looked exactly like paypal. It now escapes every Default_Ignorable_Code_Point in Unicode 18.0.0, including the invisible operators, the soft hyphen, the combining grapheme joiner, the Hangul fillers, the Mongolian and Khmer invisibles and every variation selector. A zero-width joiner and an emoji variation selector are kept only inside a valid emoji sequence (a joined emoji, an emoji with its presentation selector or skin tone, a keycap), and the zero-width joiner and non-joiner only between two letters of a script that needs them, such as Persian or Devanagari. The Hebrew and Arabic bidi marks stay as before. The MCP Overview's Metadata > Name field now uses the same formatter. Trade-off: an ideographic variation selector after a CJK character is now shown escaped.
  • A repository delete-job poller that starts after you leave the page or sign out is stopped too (#1228) - The Repositories page stopped its delete-job pollers when it unmounted or the user signed out (#1190), but a poller that started afterwards, from a DELETE still in flight or from Retry on the failed-job toast, ran on. It now stops at once, with the same note (after leaving the page) or no toast (after sign-out). When the stop lands just after a read returned a finished job, the outcome toast is no longer shown either. The failed-job toast, which stays up until dismissed, is now removed on sign-out, so its Retry cannot be used by the next session.
  • Saving the repository Security tab no longer overwrites scan settings saved from Settings (#1229) - The Security tab sent its whole form on every save, so after scan settings were saved in the Settings tab, a save from a Security tab with older edits wrote back its stale Enable Scanning and severity threshold. It now sends only the fields you changed, as the Settings tab does since #1122, and Save stays disabled until something changed; a repository with no scan settings yet still gets the whole form on its first save. On backend 1.12.0 or later the Security tab now says, as Settings does, that this first save locks the proxy scan policy. In Settings, a backend version that does not parse (such as dev) no longer hides the Record only proxy scan action, and Record only is no longer offered for a moment before the backend version has loaded.
  • Package names, the MCP Publisher namespace, and a delete Retry after sign-out are handled like the rest - The package page's breadcrumb, browser tab title and heading, the MCP server breadcrumb, and the MCP Publisher > Namespace field now render the name through the untrusted-text formatter, so hidden and bidi characters show as [U+XXXX]. A repository DELETE that returns after you sign out no longer shows its outcome, so an already-failed job's Retry toast cannot reach the next user. The Security tab's note that the first save locks the proxy scan policy now also shows on a backend dev build. (#1231)

Notes

  • The MCP Remote sync trigger is typed for its 202 answer (#1091, artifact-keeper#5404) - Backend 1.12.0 runs a manual MCP Remote sync in the background and answers POST .../mcp/sync with 202 {repository, status, statusUrl}; the API client now types that answer and the running flag of the sync state. No component triggers a sync yet, so nothing user-visible changes.
  • E2E fixes for main's stack (#1145, #1151, #1162) - The RT-15 seed no longer puts javascript: in a URL template's scheme position, which the backend refuses since artifact-keeper#5546, and the sidebar visual baseline shows the MCP Servers entry. The lifecycle empty-state locator is scoped to the main landmark, since its regex also matched the sidebar. The monitoring health log check no longer matches "Nov" in row timestamps as an empty state, which failed it every November. Test-only.
  • Stored MCP content is now tested inert in the real UI, as a release gate (#1105, artifact-keeper#5472) - The red-team assessment could only reach the backend's JSON and Swagger, so case RT-15 (hostile MCP server, capability, status and audit text rendered to users and administrators) was untested in the HTML application. A new Playwright suite stores markup, javascript:/data:/vbscript: URLs, bidi, zero-width, control and homoglyph text and prompt-injection phrases through the live backend (publish, CI capability report, status change, a Remote synced from a hostile upstream registry, an approval request and an approval reason), then opens every MCP view that renders them in Chromium as an administrator and as an unprivileged reader, with and without the Content-Security-Policy, and fails on script execution, an active node, an active-scheme URL attribute or any request to the attacker origin. No defect was found; nothing user-visible changes. The hostile upstream runs from a compose overlay (docker-compose.e2e-mcp-upstream.yml) that the interaction job and scripts/run-e2e.sh now apply: it puts the upstream and the backend on a small network of their own and lets the backend's SSRF guard admit that one address, leaving ak-e2e-network as it was. Both set RT15_REQUIRE_UPSTREAM=1, so a Remote case that cannot be seeded fails instead of skipping; the E2E job is advisory, so its result still has to be read. Bidi and control characters still display as stored (inert, but able to reorder what a reader sees); that is tracked separately.
  • The release gate is called with one named secret and a read-only token (#1154) - docker-publish.yml called artifact-keeper-test's release-gate.yml@main with secrets: inherit, which hands that workflow every secret this repository holds, and with the repository's default token permissions. It now passes only GHCR_DOCKER_CONFIG, the one secret the gate declares, and grants contents: read and packages: read. The release, tutorial, visual-baseline and publish workflows also pass inputs and step outputs to their shell steps through env: instead of inline ${{ }}, which made a value part of the script; behaviour is unchanged. Trade-off: the gate call still follows artifact-keeper-test's main rather than a pinned commit, until that repository has branch protection and the 1.12.0 gate changes have landed.
  • List-page E2E search locators are scoped to the main landmark (#1062) - On a first load, React's streamed server copy of a page can sit for a moment in a hidden <div id="S:n"> under <body> next to the client-rendered page, so getByPlaceholder(/search/i) on Packages, Staging and Repositories sometimes matched two inputs and failed in strict mode. The search locators in the list-page specs and page objects now look inside <main> only. Test-only.
  • The E2E check for paging the Permissions list now runs on every stack (#1189) - It only clicked Next page when the backend held more than 100 rules, and the E2E seed creates fewer, so CI never exercised paging. The spec now serves 150 rules from a mocked list and checks that Next page loads rules 101 to 150 and ends on Page 2 of 2. Test-only.
  • The lifecycle E2E empty-state check targets the empty state's heading (#1219) - "policies table renders or shows empty state" looked for /no.*polic|no.*data|no.*result|empty/i inside <main>, which also matches the page description ("not just one repository. Unassigned policies"), so it resolved to two elements and failed strict mode when another spec had already created a policy. It now looks for the "No lifecycle policies" heading. Test-only.
  • E2E navigation waits for React's streamed segments (#1192) - On a first load, React's streamed server copy of a page can sit for a moment in a hidden <div id="S:n"> under <body>, where an unscoped locator also matches it and fails strict mode (#1062 scoped the list-page search locators to <main> for this). The shared navigateTo helper now also waits until no [id^="S:"] segment is left, and the Search and OpenSearch search specs navigate with it instead of a bare page.goto. Test-only.
  • The npm aliases yarn, pnpm and bower keep the Environments tab hidden until an environment is registered (#1215) - Their clients write yarn.lock and pnpm-lock.yaml, which the backend does not parse and rejects at registration, and bower has no lockfile, so the tab would offer nothing there. A repository of any format that already holds an environment still shows it. The one /environments lookup that decides the tab for such formats stays, since the repository record carries no environment count. Code comment and tests only.
  • E2E specs check the 1.12.0 fix wave against the real backend (#1069, #1185, #1038, #1084, #1048, #1035, #1037, #1066) - Those fixes shipped with unit tests against a mocked backend. New Playwright specs now run them against the dev backend image in CI. A stored record_only proxy scan action, and a scan policy or age gate an instance administrator handed back (instance_locked: false), survive saving an unrelated field from the Settings tab and the Security tab, checked by re-reading the stored row through the API. A page loaded with only the refresh cookie, after the access cookie has expired, shows the signed-in shell. Deleting a repository shows the started toast and then the completed one. A Maven POM row opens its detail dialog. The Environments tab is hidden on Maven and shown on npm, and the PyPI Tracks tab shows on a local repository and not on a virtual one. Each spec creates and removes its own repositories or user. Test-only.

What's Changed

  • brand: adopt the C1 shield logo and favicons by @brandonrc in #1071
  • feat(security): durable acknowledgements and the proxy enforcement preview by @brandonrc in #1073
  • feat(agentplugins): register the Agent Plugins format and its setup guide by @brandonrc in #1075
  • fix(packages): point the Hugging Face install hint at this repository with HF_ENDPOINT (#1080) by @brandonrc in #1081
  • fix(replication): stop offering the pull replication mode, which the backend refuses (#1082) by @brandonrc in #1083
  • fix(packages): badge rejected files as rejected and show Hugging Face revisions and downloads as the backend reports them (#1085) by @brandonrc in #1086
  • feat(mcp): MCP Registry UI (artifact-keeper#5253) by @brandonrc in #1077
  • fix(mcp): hide version writes on Remote and Virtual MCP repositories by @brandonrc in #1089
  • feat(mcp): Curation tab on a Remote MCP repository for its administrators by @brandonrc in #1090
  • fix(mcp): type the 202 answer of a Remote sync trigger by @brandonrc in #1092
  • test(e2e): render stored MCP content inert in the HTML UI (RT-15) by @brandonrc in #1106
  • test(e2e): fix the RT-15 seed and refresh the sidebar baseline by @nicola-preda in #1146
  • fix(settings): show the upload size limit read-only by @nicola-preda in #1140
  • feat(quota): show the instance storage and transfer quotas and quota refusals by @brandonrc in #1150
  • feat(migration): live progress, start all and finish alert by @nicola-preda in #1102
  • feat(scan): add a scan queue page by @nicola-preda in #1137
  • fix(migration): offer Basic Auth only for Nexus sources by @nicola-preda in #1139
  • fix(security): hide not_applicable scans unless selected by @nicola-preda in #1100
  • fix(repositories): show Maven version for cached files by @nicola-preda in #1099
  • fix(auth): say when an SSO sign-in is finished in another browser (artifact-keeper#5486) by @brandonrc in #1148
  • feat(security): add a vulnerability drill-down page by @nicola-preda in #1143
  • feat(security): link the dashboard counts to the list by @nicola-preda in #1144
  • test(e2e): scope the lifecycle empty-state locator to the main landmark by @brandonrc in #1152
  • fix(auth): ask for the TOTP code after LDAP, OIDC and SAML sign-in (artifact-keeper#5484) by @brandonrc in #1147
  • ci: pass the release gate only its one secret with a read-only token, and move run: interpolations into env by @brandonrc in #1154
  • test(e2e): stop the monitoring empty-state regex matching month names by @brandonrc in #1161
  • chore(release): web 1.12.0 prep: CHANGELOG section with Sponsors, Thank You and Compatibility, version bump by @brandonrc in #1176
  • chore(release): web 1.12.0 notes: web 1.11.3 also has the null download_count fix by @brandonrc in #1180
  • fix(setup): give cargo a sparse+ index URL at the registry root by @brandonrc in #1130
  • fix(repositories): open the detail dialog from a Maven POM row by @coddan in #1123
  • fix(analytics): count proxy storage once in totals by @imanlotfimahyari in #1093
  • fix(repositories): show the used size until storage stats are computed by @nicola-preda in #1169
  • fix(repositories): show the PyPI Tracks tab on local and staging repositories by @brandonrc in #1183
  • fix(auth): refresh the session on load when the access cookie has expired by @coddan in #1125
  • fix(setup): emit a buf module name buf can parse by @loyalpartner in #1065
  • fix(permissions): page the list instead of stopping at 100 by @nicola-preda in #1182
  • fix(lane-e): repository delete job toast and scoped E2E search locators by @brandonrc in #1184
  • fix(ui): show hidden and bidi characters in stored MCP and repository text by @coddan in #1127
  • fix(scan): keep a record_only proxy scan action instead of rewriting it to fail_open by @coddan in #1122
  • fix(age-gate): keep the instance lock when saving, and show it by @coddan in #1126
  • fix(scan): keep the instance lock when saving scan settings, and show it by @brandonrc in #1186
  • fix(repositories): only show the Environments tab where it can be used by @brandonrc in #1211
  • chore(release): web 1.12.0: fold the fix wave into the 1.12.0 section by @brandonrc in #1217
  • fix(permissions): paging follow-ups: sort note, page loading state, E2E paging check by @brandonrc in #1220
  • fix: show hidden characters in package descriptions and MCP names; remote 401 on /auth/me no longer refreshes the local session by @brandonrc in #1223
  • fix: repository delete job follow-ups and E2E strict-mode fixes by @brandonrc in #1221
  • fix(scan): 1.12.1 follow-ups for the scan selects and the scan lock by @brandonrc in #1224
  • fix(repositories): 1.12.1 storage card, Environments tab and protobuf guide follow-ups by @brandonrc in #1222
  • test(e2e): check the 1.12.0 fix wave against the real backend by @brandonrc in #1225
  • fix: untrusted-text default-ignorables, delete-poller stops, Security tab patch saves (1.12.1) by @brandonrc in #1230
  • fix: untrusted package names, sign-out delete toast, dev-build first-save note (#1231) by @brandonrc in #1232
  • chore(release): prepare 1.12.0 by @brandonrc in #1233

New Contributors

Full Changelog: v1.11.2...v1.12.0

Don't miss a new artifact-keeper-web release

NewReleases is sending notifications on new releases.