MVP smoke test: the real client against the real server, on the build host #115
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
3 participants
Notifications
Due date
No due date set.
Dependencies
No dependencies set
Reference
Grimoirix/grimoire-of-digital-signage#115
Loading…
Reference in a new issue
No description provided.
Delete branch "%!s()"
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?
Before the product owner's test week on a real Ubuntu mini PC (#29), the real client must have run against the real server at least once. Today the tests check the server only. Nobody has run the whole system together.
What to do
On the build host, in a disposable setup (container or VM, no host install):
devthe waydeployment.mdsays (container image)./: first-run setup, log in.Write each step as a script or a checklist in
docs/so #29 can follow the same steps on the mini PC.Done when
dev, and the commit is named here.Depends on
#113, #114
Order: write the steps and the script now; run them once #113 and #114 are merged in dev. Do not wait idle for them.
Per the comment on this issue: wrote the checklist and the script now, without waiting idle for #113/#114.
docs/mvp-smoke-test.md— the 5-step checklist, each step with the exact commands/API calls and what to confirm, cross-checked line-by-line against the current handler code (crates/server/src/{auth,devices,media,playlists}/handlers.rs) rather than guessed. Notes which parts are inherently manual (the browser) and which the script below covers.scripts/mvp-smoke-test.sh— automates everything that doesn't need a human eye on a screen: builds and starts the server the packaged way (packaging/server/README.md's own install sequence), creates the first user and logs in by API, builds the realgrimoirix-signage-client(thegtk-rendererfeature, not the headless dev binary) and runs it underXvfb, waits for it to go pending and then accepted after approval, generates one small real.mp4on the spot withgst-launch-1.0(videotestsrc ! x264enc ! mp4mux, no network fetch, nothing checked into the repo) plus a test image with ImageMagick, uploads both, builds and assigns a playlist, and confirms two separate things from the client's ownplayback:log lines: the two items actually play, and a playlist change reaches the client with no restart.Two things found while writing this, beyond the known #113/#114 gap:
dockergroup — onlyvivyandfjrunnerare (getent group docker).scripts/mvp-smoke-test.shchecks for this at preflight and fails with that exact explanation rather than a confusing Docker error three steps in. Whoever ends up running this script once #113/#114 land will need eitherdockeraccess on an agent account, or to run it asvivy.WebSocketsync) are fully implemented ondevtoday — I read throughcrates/server/src/{auth,devices,media,playlists,ws}andcrates/client/src/{sync,player,bin/grimoirix_signage_client}.rsto write the script's API calls and log patterns. The only two gaps really are #113 (panel stillUSE_MOCK = true) and #114 (no static route, Docker image doesn't buildfrontend-web). So steps 1, 3, 4 and 5 of the checklist do not need either dependency and are already provable; only step 2's browser half does.I did not run the script end to end — that needs both the dependencies above and the Docker access above. I did validate every
jq/curlpayload against the actual Rust request structs, and the image/video generation commands, by hand on this host.Branch:
test/GRI-1885-mvp-smoke-test. Pushing once the local gates pass.Pushed and opened PR #117: #117 — local gates (
scripts/local-gates.sh) passed before push. CI is queued on the forge (ETA ~41m); will report the verdict here. Next action after that: merge if green, then this task's remaining work is the actual run — blocked on #113/#114 (panel) and on docker group access for an agent account (see my previous comment).Status update.
PR #117 (this checklist + script) had CI red on
cargo deny(bans/licenses/sources — the required gate, not the non-blockingadvisoriesjob tracked separately in GRI-1872). My branch only addsdocs/mvp-smoke-test.mdandscripts/mvp-smoke-test.sh, no dependency changes, anddev's own tip is green, so this reads as the same runner-side transientcargo denyfailure already retriggered today on another branch in this repo (commits42115fe/d0fec86). Pushed an empty retrigger commit (7685dfa); CI is running again now, queue position 0.Dependency status:
cargo deny/cargo deny advisoriesshape, but its requiredcargo denyjob passed — only the non-blockingadvisoriesjob is red, same as GRI-1872.cargo denyjob is also red with the same transient shape as mine.Still blocked on: #113/#114 landing in
dev, and ondockergroup access for agent accounts (noted in PR #117's description) to actually executescripts/mvp-smoke-test.shend to end.Next action: wait for #117's CI retrigger to resolve, then for #116/#118 to land, then run the smoke test for real.
Status update.
PR #117 (this checklist + script) merged into
devat0d7a7ea— the retrigger resolved the transientcargo denyfailure, no real issue.Dependency status:
dev.tower-http,mime_guess,http-range-header,tokio-util) to serve the panel, so its requiredcargo denyfailure may not be the same transient shape as the ones retriggered today — nyssa's thread; not digging into it further here. It's also mid-review on the GRI-1884 no-hardcoding scope (database URL as a Compose profile/env default).New blocker found while trying to actually run the script: no agent account on this build host is in the
dockergroup (onlyvivyandfjrunnerare, andfjrunner's membership was itself flagged as a security problem in GRI-619).scripts/mvp-smoke-test.shneeds a reachable Docker daemon to build and start the server the packaged way. Filed as GRI-1902 (board), assigned to the System Administrator, child of GRI-1888 — this is an access/privilege decision, not mine to make.Next action: wait for #116 to land in
dev, and for GRI-1902 to resolve, then runscripts/mvp-smoke-test.shfor real and report the result (pass/fail per step) here with the commit it ran on.Status update.
Confirmed PR #117 (checklist + script) landed: merge commit
0d7a7eais indev's history (verified withfj landed GRI-1885, exit 0). Working tree for that branch removed now that it's merged.Still blocked on the actual run, same two blockers as before:
#114-> PR#116(nyssa, GRI-1886) is still open, not merged.pc assigncall had never been made, so the task sat assigned to me with nobody woken. Fixed: it is now actually assigned to aetherius (System Administrator).Next action: wait for #116 to land and for GRI-1902 to resolve, then run
scripts/mvp-smoke-test.shfor real and report pass/fail per step here with the commit it ran on. Nothing else for me to do on this task until one of those lands.Status update.
#116(nyssa, GRI-1886): CI failed on requiredcargo denywith new real dependencies added (tower-http,mime_guess,http-range-header,tokio-util), so unlike the transient retriggers on#117/#118this one may be a genuine deny finding. nyssa just pushed a retrigger commit ("cargo deny passes locally") at 19:14 and CI is running again now — her PR, her call, not touching it.blockedfor intervention. Escalated to the CTO on that issue since it's a stuck agent, not a decision I can make.Still can't run the actual end-to-end smoke test until both clear. Re-checking in ~20 min.
Status update, no change in blockers.
#114-> PR#116(nyssa, GRI-1886): retrigger CI (pushed 19:15) is still failing on the requiredcargo denyjob (not just the knownadvisoriesone), same as the first attempt. Her branch, her call — not touching it.Still can't run the actual end-to-end smoke test until both clear. Re-checking in ~20 min.
Container runtime for this issue (GRI-1904): rootless Docker for
astraea, not ready yet.astraeawill get its own Docker daemon, running asastraea(rootless). Thedockergroup is root-equivalent, so it stays out of this. The script and document are in grimoire-infra PR #309. Applying it needs root and an Aegis review, so it is waiting on those. I will post here when--verifypasses asastraea.Once it is applied:
unix:///run/user/1007/docker.sock. A Paperclip run's$HOMEis not your account home, so a CLI context will not be found. Set it in the script:export DOCKER_HOST="${DOCKER_HOST:-unix:///run/user/$(id -u)/docker.sock}"scripts/mvp-smoke-test.shline 63 anddocs/mvp-smoke-test.mdsay "an account in the docker group". After this change that wording is wrong. The requirement is a reachable daemon./workspace/home/astraea/.local/share/dockerand count against your own memory/swap caps.Status: PR #116 (nyssa) was blocked on a shared infra issue — Rust 1.99.0 toolchain download timing out from static.rust-lang.org, confirmed by the CTO as affecting dev's own CI too, not the branch. dev's tip (
cd02e96) now passes cargo deny cleanly, so the download appears reliable again; nyssa's PR should re-check. GRI-1902 (docker group access, the other blocker on actually running the smoke test) is rerouted: aetherius couldn't start a session in my closed workspace, so the chief of staff moved that work to GRI-1904 in a workspace aetherius can open; it's running there now. GRI-1888 stays blocked on both landing. Re-arming monitor.No change on either blocker (#114/PR #116, GRI-1904). Independently re-verified: PR #116's cargo deny red is the same infra cause already diagnosed (Rust 1.99.0 toolchain download timing out from static.rust-lang.org, also red on dev itself) — not a dependency issue;
cargo deny check bans licenses sourcespasses cleanly against origin/dev tip for me too. Nothing to retrigger until the download is reliable. GRI-1904 has concrete progress: aetherius's rootless-docker script is in grimoire-infra PR #309 (CI green), waiting on an Aegis review and Vivy to apply it as root. Re-arming to re-check in ~20min.Status: dev's own tip (
0d7a7ea) is red on cargo deny again — this is the same infra toolchain-download flakiness, now with a real fix in flight: vivy's PR #119 (GRI-1761, run cargo deny without downloading the pinned toolchain) has cargo deny itself passing, CI still running the rest. Once #119 merges, dev and nyssa's #116 should both go green and #116 can retrigger.Docker-access blocker (GRI-1902/1904): aetherius has written and tested a rootless-Docker provisioning script (grimoire-infra PR #309), check/apply/verify/rollback modes, idempotent, refuses docker-group/root/uid0 accounts. Needs Aegis security review (two host-wide widenings: uidmap setuid binaries, an AppArmor profile for rootlesskit) then Vivy to merge and run --apply as root for astraea. Nothing actionable for me on either blocker right now.
Next action: wait for #119 CI + merge, then retrigger #116; wait for Aegis review + Vivy's --apply on GRI-1904, then run scripts/mvp-smoke-test.sh for real.
Update: PR #119 (vivy's cargo-deny toolchain fix, GRI-1761) is now CI-green —
CI #119 successat 20:08, cargo deny passing (only the pre-existing non-blocking advisories job is red, per GRI-1872). It is mergeable and just needs to land.Once #119 is in dev, nyssa's PR #116 (still failing
cargo denyon its stale base) needs a merge-from-dev + retrigger to go green.Docker access (GRI-1902→1904): no change since last check — aetherius's rootless-docker script (grimoire-infra PR #309) still has zero reviews and is unmerged, still waiting on Aegis security review and Vivy's root-level apply.
Nothing I can act on directly (merging and security review are outside my role). Flagging to the CTO that #119 is ready to merge.
Update: PR #119 (GRI-1761, cargo-deny fix) merged into dev at
d323b98. #116 (nyssa, GRI-1886) merged dev into itself and now picks up the fix — cargo deny passes there now; cargo test and frontend-web still running, only the pre-existing non-blocking cargo-deny-advisories job (GRI-1872) is red. Should be mergeable once those finish.Docker access (GRI-1902/1904, grimoire-infra PR #309): unchanged — Aegis has a review request open but has not submitted it yet, and Vivy's root-level apply is still pending. Still the other blocker before I can actually run scripts/mvp-smoke-test.sh end to end.
Nothing actionable for me directly this check; re-arming the wait.
Both dependencies landed: PR #118 (#113) and PR #116 (#114, nyssa) are merged into dev — dev tip is
4734441.Docker-access blocker (GRI-1904/PR #309, grimoire-infra): Aegis approved it at 20:53 (review #428, pinned to 40db0db9). It now needs only Vivy's merge +
sudo .../provision-rootless-docker.sh --apply --yes --user astraea, then--verify. Aetherius is coordinating that directly on GRI-1904 with a wake armed for 21:34Z.Once astraea has a docker socket, I'll run scripts/mvp-smoke-test.sh for real against dev tip
4734441(or later) and report here. Nothing actionable for me until the apply lands.Container runtime is ready for astraea (GRI-1904). grimoire-infra #309 merged (258e2f5);
provision-rootless-docker.sh --apply --user astraearan as root and--verifypassed: rootless daemon, DockerRootDir under /workspace/home/astraea, astraea cannot write the root socket,docker compose upworks.In each run:
export DOCKER_HOST=unix:///run/user/1007/docker.sock($HOME in a run is not the account home, so the CLI context is not found). Then runscripts/mvp-smoke-test.shagainst dev tip.Rootless Docker is confirmed working for astraea (GRI-1904 applied). Running
scripts/mvp-smoke-test.sh --keepfor real now, on dev tip (4734441) -- the full stack: packaged server container, real gtk-renderer client under Xvfb, playlist assignment, playback confirmation. Will report the result and open a separate issue per bug found, per the Done-when criteria. (Separately: PR #120 against this branch is CI-green and still open, docs-only wording fix for the now-resolved #113/#114/docker-group references -- no behavior change, not a blocker for this run.)@astraea now has a Docker daemon of its own: rootless, and not via the
dockergroup. (GRI-1904, grimoire-infra #309)What was set up on the build host (applied by vivy as root at 21:36Z,
--applyexit 0 with every verify probe ok):uidmapandslirp4netns. An AppArmor profile so/usr/bin/rootlesskitmay create a user namespace, which Ubuntu 24.04 needs.loginctl enable-linger astraea, so astraea's user manager, and the daemon with it, runs without a login session.docker.servicerunning as uid 1007. Socket/run/user/1007/docker.sock, data root under/workspace/home/astraea/.local/share/docker, on the work disk.astraea, inuser@1007.service's own cgroup, so the GRI-489 memory caps cover the containers too. The root daemon's socket staysroot:docker 0660and astraea is not indocker, so the two daemons share no socket and no images.To run the smoke test, as astraea:
Set
DOCKER_HOSTexplicitly in every run. A Paperclip run has no login session, soXDG_RUNTIME_DIRmay be unset, and withoutDOCKER_HOSTthe CLI falls back to the root socket. That fails with permission denied, which is correct, but misleading.Things to know:
up --buildpulls every image into astraea's own store. Nothing is shared with the root daemon's cache, so the first run is slower.http://localhost:8000works as the script expects. It is one port for the whole host, though, so two stacks at once will collide.scripts/mvp-smoke-test.shline 16 ("an account in the docker group") and thedocker pserror on line 63 are now wrong advice, and the docker group is root-equivalent (GRI-30, GRI-619). Please reword them to say "a reachable daemon, e.g.DOCKER_HOSTset to your rootless socket", and the same indocs/mvp-smoke-test.md, "Before you start".build-host/rootless-docker.mdin grimoire-infra.My local uncommitted workaround (running
cargo clean -p shared -p clientafter the real source was already COPY'd in) did not fix the Dockerfile cache bug — "Removed 0 files", then the same unresolved-import errors. Superseded: nyssa found and fixed the real bug properly on GRI-1906 — the clean has to run right after the dummy build and before the dummysrc/is deleted, sincecargo cleanstill needs the manifest present. PR #121 is CI-green.PR #121 is now stuck on an org question, not a technical one: the CTO declined to merge it (
packaging/server/Dockerfileis product-side, out of their remit under the current GRI-72/GRI-1099 scope block) and opened a question card on GRI-1906 asking the product owner to merge, reassign, or hold. Still unanswered as of this update.Nothing for me to do on GRI-1885 until that lands in dev. No containers were ever started by my failed workaround attempt (the build failed before compose could run), so there was nothing to tear down; cleaned up the empty scratch dirs. Waiting on GRI-1906/PR #121, then resuming the real smoke test on dev tip.
PR #121 (GRI-1906 Dockerfile fix) merged into dev (
bb58677). Synced to that tip and confirmed the real fix is in place. Relaunched scripts/mvp-smoke-test.sh --keep against dev tipbb58677with DOCKER_HOST=unix:///run/user/1007/docker.sock, detached so it survives this turn. Currently building the server image. Will report pass/fail and file any new bugs separately once it finishes.Ran the real end-to-end smoke test for real this time (dev tip
bb58677, rootless Docker as astraea). Server image builds cleanly, but container creation fails the moment Docker populates the new named volume from image content:Error response from daemon: failed to copy xattrs: failed to set xattr "system.posix_acl_default" on .../volumes/server_server-state/_data: invalid argumentRoot cause is environmental, not a code bug in this repo: the volume's data directory carries a default POSIX ACL naming the host account vivy (uid 1001), which is outside astraea's rootless-Docker subuid range (558752:65536) — a uid outside a user namespace's mapping can't be represented in an ACL written from inside that namespace, which matches the EINVAL. I could not find what sets that ACL (ruled out both GRI-1904's provisioning script, which explicitly never touches ACLs, and the GRI-604 cross-agent reclamation grant, which is scoped to gravewarden and to depth 0-1 only).
Filed as its own issue, assigned to the System Administrator (aetherius) who owns the rootless-Docker provisioning: Grimoirix/grimoire-infra#321 — blocks this one.
The half-up compose stack and generated config from this run are left in place on the build host (/tmp/mvp-smoke-test.xOHTSF, --keep) for whoever investigates #321. Nothing to fix in this repo yet; resuming once #321 clears.
Unblocked: grimoire-infra#321 cleared (stray default ACL stripped from your Docker store; a named volume populated from the server image now works). Run
scripts/mvp-smoke-test.shagain on dev tip with DOCKER_HOST=unix:///run/user/1007/docker.sock. File each new bug as its own issue.Unblocked again, 00:30 UTC. The
config.toml: Permission deniedwas not the bind mount. Three findings:1. The image was broken, same root cause as grimoire-infra#321.
server-server:latestwas built at 23:21, before the ACL was stripped from~/.local/share/docker. BuildKit's overlay layers sit in that store, so every file aRUN/COPY --fromstep created inheriteddefault:other::---:/etc/grimoire-signage,/var/lib/grimoire-signage, the panel dir and the server binary were allroot 0750. uid 999 could not enter/etc/grimoire-signage(lsof it: Permission denied), so the config file mode never mattered. Stripping the ACL later does not give theotherbits back, and the build cache carried the damaged layers.2. Builds could not resolve DNS for the same reason.
buildkit/executor/resolv.confwas created 21:38 under the ACL, mode0640, so apt's_aptuser could not read it:Temporary failure resolving 'deb.debian.org'on every--no-cachebuild. Set to0644.Done as astraea:
docker builder prune -af, resolv.conf fixed,docker compose build --no-cache server. New image: all those paths0755.3. A real bug in
scripts/mvp-smoke-test.sh— yours to fix. Step 1 writes.envwith onlyPOSTGRES_PASSWORD, soCOMPOSE_PROFILES=bundled-db(from.env.example) is missing and thedbservice never starts. The server then loops onfailed to lookup address informationfordb. Fix: write both lines, e.g.Checked with that profile set:
dbhealthy,/health→{"status":"ok"},/api/v1/setup/status→{"needs_setup":true}. Stack is down again (down -v), nothing left running.Next: fix the
.envline on your branch, rerunscripts/mvp-smoke-test.sh --keep. Note the script'spackage-server-image.shstep uses the cache, which is now clean.Applied the
.envfix:scripts/mvp-smoke-test.shstep 1 now writes bothCOMPOSE_PROFILES=bundled-dbandPOSTGRES_PASSWORD=....Moved off the merged
docs/GRI-1885-unblock-deps-notebranch onto a freshfix/GRI-1885-env-compose-profile(origin/dev tip2a8a3de), committed as6637187.Relaunched
scripts/mvp-smoke-test.sh --keepdetached (setsid, survives the turn) against the rebuilt clean image and this fix branch. Will report pass/fail here once it finishes, push and open a PR, and file any new bugs as separate issues in the MVP milestone.Run4 (dev tip
2a8a3de+ fix/GRI-1885-env-compose-profile's one-line.envfix, detached PID 2039650, log/tmp/mvp-smoke-test-GRI-1885-run4.log) got further than any prior run: build, volume-populate and the db container all passed with Vivy's rebuilt image. The server container still restart-loops on the sameconfig.toml: Permission deniedas the 23:52 report -- a freshly-generated working-tree file created after the 00:31 image fix still carries the inherited home-dir default ACL, so the root cause (thedefault:ACL on/home/astraeaitself) is confirmed still live, not just a stale-image artifact. Evidence posted on grimoire-infra#321: Grimoirix/grimoire-infra#321 (comment)Nothing left for a Test Writer to do until that lands -- it needs a root-level ACL fix on the home directory, same ask as before. Blocking again on grimoire-infra#321 (aetherius).
01:15 (Vivy): the run4 restart loop is fixed. The server is up,
/healthanswers{"status":"ok"}, and the panel is served on :8000.Cause:
packaging/server/config.tomlwas created at 00:32 under the inherited home ACL (grimoire-infra#321), so it was0660withother::---. It is bind-mounted read-only, and the server in the container runs as a non-root uid. Under rootless Docker that uid counts as "other" on the host file, so it could not read it. The image was fine.What I did: as astraea,
chmod 0644andsetfacl -bon that one file, thendocker restart server-server-1. Nothing else changed. The containers are still running (--keep).For your branch
fix/GRI-1885-env-compose-profile: addchmod 0644 config.tomlinmvp-smoke-test.shright after the script writes the file. In the compose path this file holds no secret (the DB url comes fromSIGNAGE_DATABASE_URL), and the script then does not depend on the umask or ACL of whoever runs it. Then carry on from the step after server start and report pass/fail here.Vivy's config.toml fix (0644 + ACL strip on the bind-mounted file) confirmed: server came up healthy on :8000.
Added the same fix to scripts/mvp-smoke-test.sh (commit
8af8ddcon fix/GRI-1885-env-compose-profile):chmod 0644+setfacl -bon config.toml right after it's copied from the example, so future runs don't trip the home-directory default ACL from grimoire-infra#321.Tore down the manually-patched stack and relaunched the full automated script clean (run5) to get a real end-to-end pass on the fix rather than relying on the hand-patched state. Currently building the server image; will report the result here.
Run 5 (dev tip
8af8ddc, with the config.toml ACL fix) got further than every prior run: build, volume populate, db, server all up, client built and connected, device pending, approved through the admin API — and then rejected:sync: server says Rejected: already enrolled on another connection, backing off for minutes. This is no longer an infra/ACL blocker; it looks like a real race/bug in the enrolment handshake (handle_temporary_reconnect / issue_token in crates/server/src/ws.rs + devices.rs, likely compounded by the socket being dropped without a WebSocket close frame after approve()/pend()). Full evidence, code pointers and a repro note are on #123, labelled Area/Backend + Kind/Bug + Lang/Rust + Priority/High so it routes to a backend engineer.\n\nStack left up for inspection: server-server-1 / server-db-1 (compose project server) in /home/astraea/repos/wt-GRI-1885, client log at /tmp/mvp-smoke-test.iTBIty/client.log. Not touching crates/server or crates/client myself — that is #123's owner's call, not a Test Writer's. Blocking #115 on #123.Vivy 02:00: good find on #123, Nyssa has it (GRI-1912, running). One open item for you while you wait: your script fix (
8af8ddcon fix/GRI-1885-env-compose-profile: COMPOSE_PROFILES=bundled-db + config.toml chmod) is not on the forge. Push it and open the PR to dev; I review and merge it. When #123 lands, rerun the smoke test on dev tip.Pushed the script fix Vivy flagged: PR #124 (fix/GRI-1885-env-compose-profile, head
8af8ddc) — COMPOSE_PROFILES=bundled-db + config.toml chmod/ACL fix, both already verified end-to-end on this host. Local gates pass; CI now running on #124.Separately, #123 (the reconnect-rejected bug that stopped run5) is being worked as GRI-1912 by nyssa.
Next: once #124 is merged and #123/GRI-1912 lands, rerun the smoke test on dev tip for the real end-to-end pass.
No change needing action. PR #124 (script fixes) is CI-green on the required gate, ready to merge but not yet merged. PR #125 (nyssa, GRI-1912 backend reconnect fix) is still mid-CI. Will rerun the full smoke test on dev tip once both land.
Run6 against dev tip
54ea3ea(after both PR #124 and GRI-1912/PR #125 merged) got further than ever: build, volume-populate, server up, db healthy, client connects, pending, approved. But the reconnect-rejection bug reproduces again at the exact same point — client goes straight to 'Rejected: already enrolled on another connection' on its first post-approval reconnect, never 'Accepted'. GRI-1912's graceful-close fix did not resolve the client-observed symptom end-to-end with the real client; reopened #123 with the new evidence (same milestone, assignee nyssa unchanged). Stack left up for inspection in /home/astraea/repos/wt-GRI-1885. Blocking GRI-1885 on #123 again.Run7 (dev tip
c789d2b, with #126's approve-token fix) is the furthest the smoke test has ever gotten: build, server up, db healthy, client enrolls, gets approved, reconnects as Accepted with no rejection loop — #126 confirmed fixed. Playlist arrives at the client (sync: server says Accepted: welcome back), upload and assignment succeed server-side.It then fails step 5 for a different, more fundamental reason: the client never receives any file bytes.
Player::nextreportsFallbackReason::CannotLoadforever (playback: nothing in the playlist can be shown yet) — the playlist is known and an item is due, but the cache never has it.Root cause, filed as #127 (MVP milestone, assigned nyssa): the server never sends a
TransferOffer.grep -rn TransferOffer crates/server/srcis empty.push_to_device(playlists/handlers.rs) pushesMediaManifestandPlaylistPushonly;ws.rs:5says outright "Transfers are later work." The client's half of the protocol (session.rs::on_offer/on_binary, chunked resume, hash verify) is fully implemented and unit-tested but has never been exercised against a real server, because the server has no code path that constructs aTransferOfferfor any device, ever. This is specced inplan/shared/protocol.md("Media transfer") but was never built server-side.This is the core "device downloads and plays content" feature, and it's missing, not flaky — not something I should patch as Test Writer. Stack left up for inspection:
server-server-1/server-db-1in/home/astraea/repos/wt-GRI-1885, client log at/tmp/mvp-smoke-test.Yrmzrd/client.log.Blocking GRI-1885 on #127. Will rerun the full smoke test once media transfer is implemented server-side.
PR #128 confirmed merged into dev (
91e72c4): server-side media transfer (TransferOffer, chunk streaming, resend of assignment on Accepted reconnect) is implemented. Closing #127 as fixed. Relaunched the smoke test (run8) against dev tip91e72c4— will report pass/fail here once it finishes.Run8 (dev tip
91e72c4, with #127/#128 server-side media transfer landed) is the furthest yet: enroll, approve, upload, assign, and initial playback (image + video both played) all pass for the first time — the core download-and-play feature works end to end.It then fails on the last step: changing the already-assigned playlist's items. The client never receives the update and loops the original playlist forever (confirmed via 30s+ of client log with no new frame).
Root cause:
replace_items(PUT /playlists/{id}/items,crates/server/src/playlists/handlers.rs:236) never pushes to the live connection — onlyassign_playlistand theAccepted-reconnect resend do.plan/shared/protocol.md:104requiresPlaylistPush"once on assignment and again on every change to the device's assigned playlist." Filed as #129, assigned nyssa.Stack left up for inspection in
/home/astraea/repos/wt-GRI-1885(compose projectserver); client log at/tmp/mvp-smoke-test.oCVzze/client.log; full run log/tmp/mvp-smoke-test-GRI-1885-run8.log.Blocking on #129; will rerun the smoke test once the playlist-edit push lands.
(Side note, not blocking:
fj issue-open --label/fj label --addboth 403 for me with "token does not have at least one of required scope(s): [read:organization]" — #129 was opened with no labels and milestone unset as a result. Not spending more time on it; flagging in case it's a wider token-scope gap.)Run 9 pass on
eb0de5bis good news, thank you. But please do not close #115 yet.The smoke test covers image + video only (step 4 of
scripts/mvp-smoke-test.sh). The product owner's must-haves for 11-20 are PDF and web page URL (2 campuses each); image and video are the first things we cut. So the end-to-end path that matters has not run yet.Please extend the script, in the same PR as the docs result or a follow-up before closing:
convert xc:white xc:white lobby.pdf, orgs), upload viaPOST /api/v1/media, and check the server returns page images./healthor a local static page so no internet is needed).wait_for_logfor PDF page playback (each page as an Image item) and for the web page item.Any failure: file a forge issue as before; I triage and route it.
PR #132 (PDF render-dir fix) merged into dev at
b71fd85. Synced wt-GRI-1885 (merge commit0698f94), no leaked processes/stacks from run10. Relaunched scripts/mvp-smoke-test.sh --keep detached under setsid (run11, PID 323803, own session, survives this turn) with DOCKER_HOST=unix:///run/user/1007/docker.sock. Log: /tmp/mvp-smoke-test-GRI-1885-run11.log. This exercises PDF (2-page) + web-page items for the first time, with every previously found bug fixed on dev. Will report pass/fail once it finishes.Run11 is a full PASS on dev tip
b71fd85(scripts/mvp-smoke-test.sh --keep): enroll, approve, upload of an image, a video, a 2-page PDF, and a web page URL, playlist assignment, and playback of all four item kinds — including the two must-haves for the test week (#29), the PDF pages and the web page — plus a live playlist-edit follow-through, all end to end in one run.This also confirms #131/#132 (PDF render-dir deleted before pages were read) has no recurrence.
Writing up docs/mvp-smoke-test.md's Result section now (describing run11 instead of run9, which only covered image+video), running local gates, then opening the PR to dev. Full log held locally at /tmp/mvp-smoke-test-GRI-1885-run11.log; posting the tail here for the record:
Pushed and opened PR #133 (dev <- docs/GRI-1885-smoke-test-pass, commit
db2c029), carrying the run11 Result write-up and the script's PDF+web-page extension. Local gates passed before push; forge CI is running now (queue position 0).Done. Smoke run 11 on dev
b71fd85passed in full: enrol, approve, upload, assign, playback of PDF pages, web page, image and video, live edit. Script and doc merged in #133 (dev3ef69ad).