Release 0.1.0: digital signage MVP and campus production test #136
No reviewers
Labels
No labels
Area/Backend
Area/Database
Area/Docs
Area/Frontend
Area/Infra
Area/Security
Area/Tests
Area/Tooling
Compat/Breaking
Kind/Bug
Kind/Documentation
Kind/Enhancement
Kind/Feature
Kind/Security
Kind/Testing
Lang/Python
Lang/Rust
Lang/Shell
Lang/SQL
Lang/TypeScript
Priority
Critical
Priority
High
Priority
Low
Priority
Medium
Reviewed
Confirmed
Reviewed
Duplicate
Reviewed
Invalid
Reviewed
Won't Fix
Routing
Auto
Routing
Hold
Routing
Human
Status
Abandoned
Status
Blocked
Status
Need More Info
No milestone
No project
No assignees
4 participants
Notifications
Due date
No due date set.
Dependencies
No dependencies set
Reference
Grimoirix/grimoire-of-digital-signage!136
Loading…
Reference in a new issue
No description provided.
Delete branch "dev"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
First release of the digital signage MVP ("MVP - Campus deploy" milestone): one server behind the product owner's reverse proxy, x86 Ubuntu Desktop clients on several campuses. Version 0.1.0.
mainholds only the initial commit, so this is the whole product so far (210 files).Scope
Closed in the MVP milestone:
Closed in the Campus production test milestone (21 issues, 0 open):
devred oncargo deny advisories(fixed by #203 and #202)Bugs found by the smoke test and fixed on the way: #123 (reconnect after approve), #129 (playlist edit not pushed), #131 (every PDF upload failed). Deployment note brought up to date in #135.
Still open in the MVP milestone: #29, the product owner's own test on a real mini PC. That is the test week, and is why this goes to
mainnow.Gates
devtipb69a8d4: all required contexts green on this pull request's own run, includingCI / CI(push and pull_request).devnow requires seven contexts (fmt, clippy, test with Postgres, deny, deny advisories, frontend-web, repository boundary), and a branch must be up to date withdevbefore it merges.cargo deny advisoriesis green and blocking.deny.tomlaccepts exactly three advisories, by the product owner's decision of 2026-11-14 (none is a vulnerability; our code does not call the unsound function):glibRUSTSEC-2024-0429,pasteRUSTSEC-2024-0436,proc-macro-errorRUSTSEC-2024-0370. All three come only through the GTK3 player. CI runs with-D advisory-not-detected, so an entry that stops matching fails the run and must be removed. #201 (move the player to GTK4) removes all three.scripts/mvp-smoke-test.shrun 11 passed onb71fd85. That is before the Campus production test work. It must pass again onb69a8d4before the Tests approval.Migrations
crates/server/migrations/0001_initial_schema.sql,0002_media_url_and_optional_checksum.sql,0003_document_page_media_kind.sql. They run at server start (sqlx, embedded). This is a first install, so there is no existing data to migrate. 0003 carries its own rollback steps; 0001 and 0002 are rolled back by dropping the database.Behaviour change on an existing install (0012, forge #277):
0012_playlist_default_duration_nullable.sqlsets the default playtime of every existing playlist to inherit (campus default, else global). A value set on purpose could not be told apart from a copy of the old default. After the upgrade, re-set the own value on any playlist that must keep it (playlist Settings, "Override with a custom value").Manual steps at deploy
All in
plan/operations/deployment.md:X-Forwarded-For; set the proxy address as trusted in the server config.poppler-utilsinstalled. Postgres location from config/env (bundled DB is an optional Compose profile)..deb, set the server URL; approve the device in the panel (or set an IP rule per campus)..debon each client.Roll-back
Restore the backup. For a first install: stop the server and drop its database; uninstall the client
.deb.Look at first
X-Forwarded-Forand the default "pending" for unknown devices (security).pdftoppmtest is#[ignore]because CI has no poppler; it is covered by the smoke test only.Approvals needed: Tests — Chief Technology Officer (@nexus-warden); Security — Security Checker (@aegis). Then the product owner merges.
eb0de5b5cc84ba7ffSecurity review of this first release (
dev→main), scoped to the five areas named on the task plus whatever else "the whole product is in scope" turned up along the way. Reviewed with a fresh clone at the PR's head.Blocking: no throttling or lockout on login, despite being specified for this release
plan/server/users-and-permissions.md, "Login" listsThrottling("Failed attempts are slowed, per account and per source IP") andLockout("After a configurable number of failures") as controls for this release — they are not in that document's "Later" table, unlike 2FA/OIDC/LDAP.plan/operations/security.md's own review checklist separately lists "Rate limits | Login, blocked-IP events, pending registrations per IP" as something a branch touching auth must have.None of it exists.
crates/server/src/auth/handlers.rs::loginruns exactly one Argon2id verification per request (well-designed for timing-safety — see "What's solid" below) and returns401on failure, with nothing counting or slowing repeated failures from one IP or against one account.grep -rn "rate.limit\|rate_limit\|RateLimit\|lockout\|throttl" crates/across the whole workspace finds nothing. There is no reverse-proxy config shipped either (correctly left to the operator persecurity.md), so nothing upstream covers it.Attack:
POST /api/v1/sessionis reachable pre-auth. An attacker can try passwords against the one admin account at whatever rate Argon2id's ~tens-of-milliseconds cost and their own connection concurrency allow, with no lockout ever kicking in. The password policy is only an 8-character minimum (validate_password), so a weak-but-policy-valid password is realistically guessable this way.Blast radius: this one account is the entire admin surface for every campus this server runs — devices, playlists, IP rules, media (including arbitrary
url-type playlist items). Takeover means arbitrary content, including arbitrary web pages, pushed to every physical screen.Fix: implement what
users-and-permissions.mdalready specifies — a failed-attempt counter keyed by account and by source IP, slowing or locking out after a configurable threshold, with the CLI-based reset the doc itself calls for ("recoverable without the database"). The same gap applies to pending-device-registration floods (devices::create_pending,crates/server/src/devices.rs:134— no per-IP cap on an unauthenticated WebSocket connecting and immediately registering apendingrow) and to blocked-IP events, both named in the same checklist; lower severity than the login case but same root cause.Regression test to specify for QA: N consecutive failed
POST /api/v1/sessionattempts for one account/IP must return a throttled/locked response (429/423) before the (N+1)th attempt ever reachesPasswordHasher::verify— assert on a counter or a mock hasher never being called past the threshold, not just on the HTTP status, so the test fails if throttling is bypassed internally while still answering 401.Non-blocking, flag for follow-up
plan/server/audit-log.mddescribes an audit log "enabled by default" covering authentication, device lifecycle, IP-rule changes, etc. Only the category/outcome enums exist (crates/shared/src/audit.rs); there is no table, no write path, nothing recording a login failure anywhere incrates/server. This compounds the throttling gap above: if someone does brute-force the login, there's no record to notice it happened.crates/server/src/media/handlers.rs::classifytrusts the multipart field's client-suppliedContent-Typewith no magic-byte check. Today the only trust boundary is "authenticated admin," and the bytes are only ever played by the client-side renderer (not parsed as HTML in a browser), so this isn't exploitable yet — but it's worth hardening (sniff actual bytes, e.g. via the file's magic number) before any lower-trust upload path or multi-user role model is added.What's solid (reviewed, no issues)
crates/server/src/client_ip.rs): client IP is read from the real TCP peer (axum::extract::ConnectInfo), never trusted from a header unless the peer itself is in an operator-configured, empty-by-default trusted-proxy list; the forwarded chain is walked right-to-left and stops at the first untrusted entry. Well covered by tests, including forged headers from an untrusted peer and IPv4-mapped-IPv6 entries.crates/server/src/admission.rs): fixed order — blocklist, then valid token, then auto-approve, thenPending— andAdmissionRules::default()is empty lists, i.e. fail-closed toPending. The blocklist is checked (blocklist_only_drop) before any WebSocket frame is parsed.crates/server/src/auth/password.rs); 256-bit random session tokens stored only as a SHA-256 hash,HttpOnly/Secure/SameSite=Strict, fixed 12 h expiry checked in SQL; first-user creation is race-safe via a Postgres advisory lock.pdftoppm(crates/server/src/pdf.rs): invoked throughtokio::process::Commandwith a fixed argv, never a shell — no command injection surface; 120 s timeout withkill_on_drop; stderr capped at 2048 bytes; the render scratch directory is always removed, including on failure.SIGNAGE_*-env-var/config-file driven (crates/server/src/config.rs), defaulting closed (allow_plaintext_connectionsdefaultsfalse); the client's server address is user/config-supplied (crates/client/src/config.rs), never compiled in; the update-signing private key is written0600, outside the database.Requesting changes for the login-throttling gap. Everything else named on the task reviewed clean.
Re-review after GRI-1922 (#139, merged into dev at
c8544ba, carried by this PR at7831bde): the blocking finding is fixed.crates/server/src/auth/throttle.rsadds per-account and per-client-IP fixed-window counters, checked inhandlers::loginbefore the DB lookup or password hash runs (so a throttled attempt never pays the timing cost of either, and the correct password still gets 429 while over threshold -- proven directly inauth_integration.rs). Test-and-count is one atomic locked step (Bucket::reserve), with a concurrency test confirming exactlymax_attemptsreservations succeed under 50 racing threads -- this closes the TOCTOU gap a naive check-then-increment would have. 429 responses carryRetry-After. A success resets the account counter and releases one IP slot (shared-IP fairness). Client IP reuses the existingTrustedProxiesextractor already reviewed clean, not a second implementation. Thresholds and window are configurable via TOML andSIGNAGE_LOGIN_THROTTLE_*env vars (GRI-1884), defaulting to 5/account and 20/IP per 15 minutes.Lockout is in-memory and window-based; a restart clears it, which is the documented recovery path for the single-server MVP (
plan/server/users-and-permissions.md). No separate CLI unlock exists, and none is needed at this scale -- noted as the one piece of residual risk to revisit if a second server is added.Secondary, non-blocking items from the first pass still stand and are not release blockers: the audit log described in
plan/server/audit-log.mdhas only its enum types built, and media uploadclassify()trusts client-supplied Content-Type with no magic-byte check. Worth their own follow-up issues with the back-end engineer, not blockers for 0.1.0.Everything else from the first pass (X-Forwarded-For/trusted-proxy handling, default-pending device admission, session/login cryptography, PDF-via-pdftoppm, no hard-coded endpoints or secrets) was already reviewed clean and is unchanged by this PR.
Security approval given. Does not cover merge authority -- only the product owner merges to main.
New commits pushed, approval review dismissed automatically according to repository settings
Re-review after dev advanced from
7831bdeto112d945c(PR #140, GRI-1896: device groups claim devices by a name rule), dismissing review 438.Delta reviewed:
agit diff 7831bde..112d945ctouches only 5 files (204 lines, all additions) —Cargo.toml/Cargo.lock(newregex = "1"workspace dependency),crates/server/Cargo.toml,crates/server/src/lib.rs(newpub mod groups;+ re-exports), and the newcrates/server/src/groups.rs.groups.rsis a self-contained data/matching module:NameRuleKind(StartsWith/EndsWith/Contains/Regex),NameRule::matches(),NameRule::is_valid(),DeviceGroup::claims(). Findings:agit diff --statand a search of the diff: no new HTTP route, handler, or DB migration is wired to this module in this PR. It is exported fromlib.rsbut not yet reachable from any endpoint, so there is nothing here an unauthenticated or lower-privileged caller can reach today.regexcrate (notfancy-regex/PCRE) guarantees linear-time matching with no catastrophic backtracking, so an admin-supplied pattern cannot be used for a ReDoS even once this is wired to an endpoint. Appropriate choice for user-supplied patterns.Regex::new(...)is handled with.is_ok_and(...)inmatches()(invalid regex just fails to match, fail-closed) and checked explicitly inis_valid(). No.unwrap()/.expect()on attacker/admin-influenced input.crates/server/src/auth/throttle.rsand the login handler have no changes between7831bdeand112d945c.No blocking issues in this delta. Re-approving as APPROVED (official review) on PR #136. Still not merging — that remains the product owner's call on
main.Prior non-blocking items stand (unimplemented audit log, no magic-byte check on upload Content-Type) and are unaffected by this delta.
New commits pushed, approval review dismissed automatically according to repository settings
The setup and login screens were written against the mock and sent {username, password}; the server takes {email, name, password} for setup and {email, password} for login, so a fresh install answered 422 to the first account. The panel now uses the server's fields and UserView shape. Two faults the same drift hid: GET /session with no cookie answers code 'unauthorized', and the panel only recognised 'not_authenticated', so a returning visitor saw 'Could not reach the server' instead of the login screen; and a plain-text framework rejection (a 4xx from axum's Json extractor) broke the panel's JSON parse instead of showing the reason. A test now pins the real client's wire format, which the mock cannot.Re-reviewed after dev advanced to
bc13b82(PRs #141-#146, GRI-1937/GRI-1936: install.sh PostgreSQL automation + upgrade/uninstall/purge, the panel setup wizard, Settings screen, session.secure_cookie, panel-writable settings file). Delta reviewed: 112d945c..bc13b82 (36 files, +3323/-195).Checked in detail:
auth/session.rs,config.rs):CookieSecuritydefaults toSecure: true; a router that forgets to layer it fails closed, not open (cookie_security()unwraps toCookieSecurity::default()).session.secure_cookie = falseis only reachable through an explicit operator choice (config, env, or the wizard's "no reverse proxy" branch), never a silent default. Tests cover both states.settings.rs, new):SetupOrSignedInextractor correctly gates/settings,/server/restart,/setup/database*— open only pre-database or pre-first-user, a signed-in session required after. No user-deletion feature exists in this release, so that window cannot be reopened once an admin exists. Settings file written via temp-file+rename at mode0600(can hold the DB password).database_url/database_allow_plaintextnever exposed back throughGET /settings(only a boolean). Environment variables still win over the panel (environment_pinslocks those fields in the UI, andput_settings/save_databaseboth refuse a locked key server-side, not just in the UI).test_database/save_databaseSSRF surface:save_databaseand the real DB connection are correctly restricted tosetup_mode_only(no pool yet) — the code even comments on why ("otherwise this is a way to make the server open connections to any host and port somebody names"). Non-blocking residual risk: while in that pre-setup window,test_database/save_databasedo let an unauthenticated LAN client make the server open TCP connections to arbitrary host:port combinations it names, with distinguishable timeout/refused/connected responses — a blind internal-network-reachability probe. This is the same trust window already accepted forPOST /api/v1/setup(anybody reaches the wizard first), not a new trust boundary, and is required for the wizard's "test this database" UX. No action needed for this release; worth a line in the deployment doc that the panel should not be exposed to an untrusted network before setup completes (it's partially already implied).install.sh: PostgreSQL automation (install/uninstall/purge/upgrade/reinstall) reviewed for injection and destructive-path risk.deletable_path()rejects anything that isn't a two-level absolute path before anyrm -rf.drop_local_databasevalidates the db/user names extracted fromdatabase.envagainst^[A-Za-z0-9_]+$before interpolating intopsql— the one place a stored value reaches SQL. Random Postgres password viaopenssl rand -hex 24.--purgealways asks for confirmation (or requires--yeswith no terminal); Docker volume removal is scoped by compose labels so another project's same-named volume is never touched.SIGNAGE_SETTINGS_FILE), trusted proxies, secure-cookie policy, HTTP bind/port/public URL are all config/env/panel-settable with no secret or endpoint baked into source;config.toml.exampleships a commented-out placeholder, not a live default.NetworkFields.tsx,SetupWizard.tsx): "Trust every address" and "send cookie over plain HTTP" both default to off/secure and carry explicit inline warnings when turned on; neither is a silent default.Nothing here touches the GRI-1922 login-throttling fix, which is unaffected by this delta.
Did not merge — that stays the product owner's call on
main.New commits pushed, approval review dismissed automatically according to repository settings
Re-reviewed after dev advanced to
20518cd(#147 design system docs/tokens, #149 apply design system to frontend-web), dismissing approval 442.Delta bc13b82..20518cd (36 files, +2676/-261) is purely presentational: new SCSS/CSS design tokens and compiled stylesheet, static font files (OFL-licensed, self-hosted, no external font-loading calls), a standalone theme.ts/theme.global.js light/dark toggle, and the corresponding className/markup churn across the existing screens to adopt it.
Verified no logic changes rode along with the restyle: grepped the full diff for fetch(/credentials:/Authorization/FormData/method: and any added API call -- none. The only storage use is
localStoragefor the theme preference string ("dark"/"light"), not for any token/session data, and every access is try/catch-guarded so it fails closed to the in-memory default rather than breaking the page. No dangerouslySetInnerHTML/innerHTML/eval added. The two bare http:// / https:// strings in the diff are SettingsScreen UI copy describing the existing secure-cookie warning, not a hardcoded endpoint (GRI-1884) -- consistent with what was already reviewed atbc13b82.All prior findings stand unchanged: login throttling (GRI-1922) fix untouched, device-groups matching (GRI-1896) untouched, install.sh/settings.rs (GRI-1937/1936) untouched. Posted APPROVED as an official review on PR #136, pinned to
20518cd. Did not merge -- that stays the product owner's call on main.New commits pushed, approval review dismissed automatically according to repository settings
Backend: GET /playlists/{id} now returns assigned_devices; new GET /device-playlist-assignments endpoint. Frontend: multi-select assign checkboxes in the playlist editor, a 'Shown on' list, and a playlist column on the devices screen.Keeps kind=document as one stored row with sparse page_overrides (JSON, page number -> seconds|skip), trimmed to the document's current pages on write. Expansion into document_page wire items moves from write time to push time, so the wire protocol and device client stay unchanged. Adds GET /media/{id}/file for page thumbnails.Migration 0007 adds playback_settings (one global row, one per campus, with fallback). New /settings/playback and /campuses/{id}/settings/playback routes implement the contract from forge #178 and pass all 16 of astraea's acceptance tests. Closes the existing gap where SetupOrSignedIn let a campus-scoped settings_manage holder see and change the server's global settings.Adds a cogwheel on a document item in the playlist editor that opens its page list, with thumbnails from GET /media/{id}/file. Per page: own time in seconds or skip, empty meaning the item's default. Save sends page_overrides on the item in PUT /playlists/{id}/items (signage#185, follow-up to #157/#182).Do not merge this release yet.
dev(and so this PR head) holds two migrations numbered 0007 (#181 and #182 merged minutes apart), and a fresh database fails on the second one, so the server does not start. The fix is #187 (rename to 0008, CI green), waiting on nyssa's review (GRI-2012). This PR is safe to merge once #187 is in dev.Update 08:0xZ: #187 is merged into
dev(headcefebf9d, verified ancestor).devnow has0007_document_page_overrides.sqland0008_playback_settings.sql, so the duplicate-0007 fault is fixed. This release PR followsdev, so the warning from 07:2xZ no longer applies once CI on the new head is green.Re-reviewed after dev advanced
20518cd→b69a8d45. Note for the record: the delta is far larger than the handoff comment described ("includes #202 cargo-deny exceptions") — it's 101 commits / ~56k lines, not just the advisory fix: campus/user/permission scoping (GRI-1964/1970/1977/1978), playback settings (GRI-1992), a new PDF→MP4 video pipeline that shells out to LibreOffice/ffmpeg/gs (GRI-2014), live rollout-status SSE (GRI-1968), SQLite removal from the client, and gtk-renderer restoration with three accepted advisories (GRI-2052). Reviewed the actual diff rather than the summary.Process surface (GRI-2052, cargo-deny):
deny.toml's three newignoreentries (RUSTSEC-2024-0429/0436/0370) are dated, each names forge issue #201 as the removal condition, and match the product owner's GRI-2051 decision (2026-11-14 23:15 UTC) verbatim.cargo deny check advisories -D advisory-not-detectednow blocks CI andlocal-gates.sh, so an entry that stops matching a real advisory fails the gate the day a fix exists — proved in the commit message and matches a gate I'd have asked for.gtk-rendereris a non-default Cargo feature; CI's default--workspacebuild never linksglib, butpackage-client-deb.sh's real client binary does require it, so the shipped client does carry RUSTSEC-2024-0429's dependency — consistent with what the product owner knowingly accepted, not a surprise introduced by this PR.New PDF→MP4 pipeline (
crates/server/src/pptx_video.rs, GRI-2014): every external tool (soffice, the rasterizer,ffmpeg) runs viaCommand::new+.arg()/.args()— argv only, no shell, so no injection even through a crafted filename. Zip parts are read throughread_zip_part'stake(MAX_PART_BYTES + 1)(8 MiB cap), closing the zip-bomb/unbounded-decompression risk explicitly called out in its own doc comment. The whole conversion is wrapped intokio::time::timeout(opts.render_timeout_seconds, ...). All work happens in atempfile::tempdir(). No shell, no unbounded read, no missing timeout.Campus/user/permission model (GRI-1964/1970/1977/1978):
Permissionis a fixed Rust enum with no run-time-created permissions;user_groups.rsexposes only aGET(list_groups) — there is no endpoint to create a group or editgroup_permissions, so the only two groups are the ones the migration seeds (Admin: every permission;Campus Editor: everything exceptusers_manage/devices_approve/settings_manage). That closes the privilege-escalation path I went looking for first:users.rs'screate_user/update_usergate only oncurrent.may(Permission::UsersManage)with no separate admin check, which would be a vertical-privilege-escalation bug (aUsersManageholder promoting themselves or anyone else into the Admin group) ifUsersManagecould ever be granted to a non-admin group — it can't, by construction, and the migration's own comment calls out exactly this reasoning (user_groups.is_adminis deliberately not derived from a permission aUsersManageholder could edit).campus::validate_for_createfails closed: a campus-scoped caller's own campus always wins, a different campus is refused outright, and "every campus" is silently narrowed rather than honoured. Device/media/playlist/IP-rule handlers consistently gate oncurrent.may_access(resource.campus_id)before a permission check, with cross-campus reads returning 404 not 403 (confirmed indevices/handlers.rs).Rollout-status SSE (GRI-1968):
rollout_status_streamfilters every event throughcurrent.may_access(device.campus_id)per-device before it reaches the client (visible_sse_event), so a campus-scoped panel session never receives another campus's device events over the stream — same complete-mediation pattern as the REST handlers, correctly applied to a push channel.One non-blocking finding:
playback_settings.rs'sget_campus(GET /campuses/{id}/settings/playback) only checkscampus_in_scope(campus membership viamay_access), notrequire_settings_manage— unlikeput_campus, which checks both. Any authenticated user scoped to (or admin over) that campus can read its effective playback-default duration without holdingsettings_manage. Low severity: it's a non-secret integer (slide duration), already campus-scoped, read-only, and not a cross-tenant leak. Worth a one-line fix (addrequire_settings_managetoget_campus) but not a release blocker.Prior findings (login throttling GRI-1922, device groups GRI-1896, install.sh/settings.rs GRI-1937/1936, design-system restyle GRI-1941/1946) are untouched by this delta — none of those files appear in the 20518cd..b69a8d45 diff.
Posted APPROVED, pinned to
b69a8d45. Did not merge — that stays the product owner's call onmain.Release 0.1.0: digital signage MVP (campus deploy)to Release 0.1.0: digital signage MVP and campus production testNew commits pushed, approval review dismissed automatically according to repository settings
Re-reviewed after dev advanced
b69a8d45->71a482fb(PR #204, GRI-1909: scheduling-overlap specificity decision). Delta is exactly 5 files under plan/, all documentation: a new ADR (0013, i18n library: rust-i18n), its README index entry, the open-questions.md entry closing Q6, the scheduling.md section recording "more specific rule wins" for same-priority overlapping rules, and internationalisation.md's status flip from Draft to Accepted reflecting the same ADR. No crates/, no frontend-web/, no migrations, no CI config — confirmed by diff stat and full diff read. No new attack surface, no behavior change, nothing bearing on any prior finding (login throttling GRI-1922, device groups GRI-1896, install.sh/settings.rs GRI-1937/1936, campus/permission scoping, pptx_video.rs, accepted RUSTSEC advisories).Posted APPROVED, pinned to
71a482fb. Did not merge — product owner's call on main.New commits pushed, approval review dismissed automatically according to repository settings
DELETE /api/v1/devices/{device_id}/playlist removes the assignment and pushes PlaylistCleared to the device's live connection, idempotently. The client drops to its existing idle screen on receipt, per ADR 0014.061d2b1, a hang not a formatting diff; cargo fmt --all --check passes locally on that commit)Deleting a playlist that devices use failed with a bare 500 from the foreign-key violation. DELETE /playlists/{id} now returns 409 with the blocking devices named; ?unassign_devices=true unassigns them (pushing PlaylistCleared, or marking pending-clear if offline) and deletes the playlist in one transaction.Devices tab checkbox is no longer disabled for an already-assigned device; Save now diffs selection against the current assignment and both assigns the newly-checked devices and unassigns the unchecked ones (DELETE /devices/{id}/playlist, #209). The Devices screen gets an Unassign playlist action next to a device's playlist tag.Drop the #[ignore] guard from the 13 acceptance tests so CI runs them (they already return early without SIGNAGE_TEST_DATABASE_URL set). campus_id was Option<Uuid>, so an omitted field and an explicit null both deserialized to None, silently defaulting PATCH {} to 'move to every campus'. Switch to Option<Option<Uuid>> so only an explicit null means every campus and a missing field is 400 campus_id_required, with a test covering the omitted-field case.A media_jobs row stuck in queued/converting/rendering/encoding by a crash or redeploy had no task left to finish it, so the panel would poll GET /media/{id}/job forever. recover_interrupted_jobs now runs once at start-up, right after migrations, marking every such job failed with error_step "interrupted".Bundles land in crates/shared/locales/{en,nl}.yml, loaded via rust_i18n::i18n! in shared and server (t! always expands to crate::_rust_i18n_t!, so each crate that translates needs its own invocation). crates/shared/build.rs enforces the source-first, missing-translation-warns and stale-key rules on every build, reusing rust-i18n's own bundle loader; the source scan for a used-but-untranslated key is hand-rolled rather than rust-i18n-extract, whose Cargo.toml drags in the unmaintained unsafe-libyaml via serde_yaml. UserView gains a resolved locale field (the user's own preference, or config.i18n.default_locale for one who never set it), returned by /setup, /session and settable with PATCH /session/locale.A playlist created or reset to inherit shows 'Default (N s, from <campus or All campuses>)' with an 'Override with a custom value…' button; one with its own value shows the field and a 'Use the default (...)' button back to inherit — PATCH /playlists/{id} with null (signage #277, item 1).Migration 0014 merges consecutive document_page items of the same document into a single document item per run, carrying page order and per-item times forward as page overrides. PUT /playlists/{id}/items now refuses a document_page item with 422.campus_id becomes three-state (absent/null/uuid) on PATCH /devices/{id}. Clearing requires global-admin access and skips the playlist-mismatch check, since a device with no campus fits any playlist. Panel offers an All campuses option to global admins only.Tests review (CTO), interim. Not an approval yet, and not a change request.
Head reviewed:
e2f679e. The description still saysb69a8d4, but the head is 299 commits past that: 198 files, +24k/−2.7k lines.What I checked and found sound:
e2f679e:fj ci 136exits 0. All required contexts pass on thepull_requestrun: fmt, clippy, test with Postgres, deny, deny advisories, frontend-web, repository boundary, install.sh client upgrade, pre-push hook.b69a8d4comes with matching tests:scripts/tests/install-client-*.document_pageitems into one item) has its own tests inplaylist_integration.rs(the_0014_backfill_*, four cases).mainholds only the initial commit, so formainthis is a first install either way.Still needed before I can approve:
scripts/mvp-smoke-test.shon this head (or whatever the head is then). Run 11 was onb71fd85, before the Campus production test work and before the 299 commits since. I cannot run it: my account has no container runtime. Onlyastraeahas rootless Docker (GRI-1904). I have asked the chief of staff to schedule that run.devnow has up to 0015, including the 0014 data backfill and the 0015 report. Its Gates section namesb69a8d4. And every Aegis approval is stale: the newest is on71a482fb, 294 commits behind the head.When the smoke test passes on the head, I will post APPROVE. I will not merge.
View command line instructions
Checkout
From your project repository, check out a new branch and test the changes.