Repository navigation
Releases
v0.22.0
Minor Changes
- 7dc71f9: The fleet facts job (A11-283):
fleet facts submit --boxes <n|all|names> --label <key>is a new kind of job in the fleet job service. It books its boxes like a capture's, wakes each one, runs the read-only playansible/facts.ymlon it (the first play the service runs, ADR 0005) and records the box's make, model and machine type, serial, BIOS version and date, CPU, RAM, each disk, free space on the system drive, Windows build, TPM and Secure Boot state. The play changes nothing, and that is held three ways: a static test of its modules, the runner's check that the callback'schangedis 0, and the script's own$Ansible.Changed = $false. The facts are kept in/etc/a11ign/fleet-facts.jsonbeside the inventory, one entry per box withreadAt; a box that did not answer keeps its last facts, marked stale.fleet facts showprints them as a table. A facts job is written at store version 3, so an older service fails it by name. Nothing on any box changes; the fleet stays frozen.
v0.21.0
Minor Changes
- e1d40ca: The lab reaches the fleet job service, and the lab catalogue hands its capture jobs a label and boxes (A11-263, step 2 of 4).
ansible/fleet-jobs-lab-access.ymlmakes/root/.ssh/a11y-fleet-jobson the lab (never copied off it), puts its public half in the control plane'sauthorized_keysasrestrict,command="… src/fleet-jobs/gate.ts", checks control out on the lab at/opt/a11y-fleet-cli, writes/etc/a11ign/fleet.envand/usr/local/bin/fleet, and reads back from the lab thatfleet lease list --jsonanswers and thatidover the key is refused.src/fleet-jobs/gate.tsrunsremote.tsonly in the formcontrolPlaneTransportsends (single-quoted arguments, no shell) and refuses every other command with exit 2; ADR 0004 gives the reasons. Inlab-job.yml,labelis a new published parameter, REQUIRED by every job that captures (capture,capture-only, bothcapture-acceptance,capture-real-pages,retrain,everything,evidence-check,stability,gate-stability,gate-probe-order,capture-check): a launch without it (or with one that is not a ticket key) is refused naming the parameter, bylab:jobbefore any box is woken and by the playbook. Those jobs also setA11Y_FLEET_BOXES(theworkersorworkername or count, elseready),A11Y_FLEET_LABELandA11Y_FLEET_WAIT(the job's own timeout); a lab older than step 3 never reads them.lab:pipelinetakes--label=<ticket key>and passes it to the stages that take one (a pipeline that captures is refused without it, before its unit starts), and a--qualify-sharun that names no label is booked asqualify-<row>. Revert: remove the three variables and thelabelparameter from the catalogue, the--labelflag fromlab-pipeline.ts, the default label fromqualification-run.ts, andgate.tswith its play; remove thea11y-lab-fleet-jobsline from root'sauthorized_keyson the control plane.
v0.20.0
Minor Changes
- 60f729c: The fleet job service takes jobs of cases, for lab's captures (A11-263, step 1 of 4).
fleet capture submit --cases <file>submits a job whose every capture has its own id, URL and/capturebody, sent exactly (the page job's defaults are not merged in), and an optionalgroup: every case of a group is captured on one box, one after another and in file order, and a group is never split, even after a restart (a group moves only whole and only while untouched). A group may carryskipWhen, which skips it, with nothing sent and no record written, when the box's/healthenvironment matches the named fields.fleet job recapture <id> --cases a,b --label <key>prints a new job ID for named cases of adonecase job; each case is pinned to the box its group was captured on, the job asks for those boxes by name so it queues and leases like any named job (A11-244), and its records are<case>.a<attempt>.json. A job of cases or a re-capture is store version 2 (the service reads 1 and 2; a page job is still version 1, and an older service fails a case job by name). The manifest of a case job gainscases,recaptureOf, each capture'scase,groupandattempt, and each box's wholeenvironment. A job of pages behaves as before. Revert: remove--casesandrecapturefromcli.tsandremote.ts; a page job reads and runs as it did.
v0.19.0
Minor Changes
- fb99b46: Fleet booking, step 2 of 2 (A11-244): the scheduled fleet work takes leases from the lease ledger too, so it never lands in the middle of a capture. The hourly drift check (
fleet-drift.ts --record) takes a sharedreadlease on every box nomaintainlease holds; it reads beside a capture, because it sends onlyGET /healthandGET /progress, the reads that already run during every capture. A box undermaintainis not asked that hour: it is recorded inskippedas unread with its holder, and its last differences are carried. The daily freeze (a11y-fleet-freeze-daily.service) now runssrc/fleet-leases/freeze-daily.ts, which takes an exclusivemaintainlease per box, runsfreeze.yml --limit <the free boxes>as batches, waits for the rest, and stops after 6 hours, naming any box it never got inleasedinruns/fleet-freeze-last-run.json(carried into the hourly record'sfreeze) and failing the unit. The idle power-off produces itslease-pending(per box: a livecaptureormaintainlease, or a queued job that named it) andbatch-queued(a job waiting for "any N" orreadyboxes) signals from the ledger, and takes a non-queueingmaintainlease across eachsleep.ymldispatch. Re-runpnpm run fleet:freeze-scheduleafter merge to install the freeze unit's newExecStart.
v0.18.0
Minor Changes
- c2a9683: Fleet booking, step 1 of 2 (A11-244): every fleet job leases its boxes through one lease ledger on the control plane, so jobs queue and never collide. The ledger is
/var/lib/a11y-fleet-leases/ledger.json(src/fleet-leases/: a versioned JSON file written under a pid-checked mutex, with the booking rules, the estimated-start walk and the file kept apart).captureandmaintainleases are exclusive;readis shared withcapture; a ledger that cannot be read, or is of a version this code does not know, reads as every box held. The job service's in-memory claim is gone: a job joins the queue at its submission time (so a restart keeps every job's place), starts only when it holds the turn for boxes it can have (jobs are served in order; a job that names boxes reserves only those; a job for any N boxes orreadyholds the head of the line so it is never starved), releases each box as soon as its own share of the pages is done, releases the lease when it ends (done, failed or cancelled), renews it every 30 s, and loses it 5 minutes after the last renewal if its service dies.fleet job statusshows a queued job'splace: n of m, itsestimated start(orunknown (why)) and who it waits on, in text and in--json(queue);fleet lease list [--json]shows the leases and the queue. A box that leaves a running job, by three failed captures in a row or by a current difference in the hourly drift record (runs/fleet-drift.json), fails only its own unsent pages (<box> left the fleet: <why>, no longer moved to the other boxes), is named indropped, and has its lease freed once its capture in flight ends; a box the drift record lists is not offered to a new job, and a job that names one fails at start.a11y-fleet-jobs.servicegains thea11y-fleet-leasesstate directory. Step 2 adds the scheduled work (the drift check, the daily freeze, auto-off).
v0.17.0
Minor Changes
- a2c01f8: The fleet job service (A11-243): fleet work is a job you submit and follow by job ID, not a script in your own shell.
a11y-fleet-jobs.serviceis a long-lived process on the control plane that owns a job store (/var/lib/a11y-fleet-jobs, files with one writer each) and runs each job's captures against the fleet over the workers' own/capture, so a job outlives the agent that submitted it.fleet capture submittakes the pages, the boxes (a number,readyor names), the capture options of the current scripts, a freeze reference and a label, and returns the job ID at once.fleet job status [--json]gives the state, progress per box, an estimated finish and the results' location;fleet job wait --timeout <duration>always ends (0 done, 1 failed with the reason, 3 cancelled, 124 timeout with the progress and what is left, 2 usage or unreachable);fleet job cancel,list --labelandfetchcomplete it. Results are one record per page and pass plus amanifest.json(pages, boxes with their builds and versions, the pinned versions, start and finish) in a fixed place per job, copied to the agents host byfetch. Two jobs never share a box, a box that fails three captures in a row is dropped from its job, and a restart resumes every running job and takes again only the captures the dead process had sent. The CLI reaches the service over the control plane's ssh (a newcontrolPlaneInvocationbesidesshToControlPlane), so there is no new port. It is put on the plane bypnpm run fleet:jobs-install, which runs the service from its own checkout,/root/fleet-jobs, and reads the heartbeat back. Nothing on any box changes; the fleet stays frozen.
v0.16.0
Minor Changes
- 5a7a48f: The fleet upgrades only in one deliberate whole-fleet run, and is watched between runs (A11-205, step 3 of 5; the decision is
docs/adr/0001).fleet:patchis now a chain:pnpm run fleet:patchprints the plan (which box has nothing to upgrade and what each other box is missing), andfleet:patch -- --applyrunsdeploy.yml, thenpatch.ymlone box at a time, then a read of EVERY box againstdeploy-pins.json, stopping at the first failure and naming the box.patch.ymltakes its target only from the manifest (Windows release, build and exactly the manifest's KB throughwin_updates'accept_list; the newest build any box reports is no longer the target), refuses a box above the manifest by name and never rolls it back, installs the manifest's Node, NVDA and Edge and the display driver pin through the role's own task files, freezes again inalways:, and ends by reading every box against the whole manifest. A newfleet:drift(src/fleet-drift.ts) reads every box's Windows build, Edge, NVDA and worker code against the manifest and names every difference (exit 3; a box that does not answer is unread, exit 4 when none does);fleet:freeze-scheduleinstallsa11y-fleet-drift.timer, hourly and read-only, which writesruns/fleet-drift.json, andfleet:watchmirrors it toruns/fleet-drift-mirror.jsonfor the chairman's announcement channel. A box that differs is out of capture:lab:jobreads the workers it needs just before it dispatches and leaves a drifting one out of a whole-fleet pool asdrift_excluded(a job that names one is refused, and a read that cannot be made refuses). The patch window's unit names are removed from the unit-drift derivation. Adds control's first ADR and the README sections on the deliberate upgrade.
v0.15.0
Minor Changes
- 5cc99d6: Nothing on a fleet box upgrades by itself (A11-205, step 2 of 5).
ansible/roles/worker/tasks/freeze.ymlholds every lever that stops a self-updater, values included and none of them indefaults/main.yml: Windows Update (NoAutoUpdate,UseWUServerwith an unreachableWUServer/WUStatusServer,DoNotConnectToWindowsUpdateInternetLocations, the target release held at the manifest's, the driver exclusion,wuauserv,UsoSvcandWaaSMedicSvcatStart=4, the Update Orchestrator, Medic and Store scan tasks disabled, Medic's TrustedInstaller-owned keys handled through a one-shot TrustedInstaller task), the Edge updater (services off,MicrosoftEdgeUpdate*tasks unregistered, every updater binary renamed to*.a11y-frozen, one outbound block rule per binary), the Store (AutoDownload=2,InstallServiceoff) and Defender's signature updates (reported, not fought). The Windows Update deferral (30 days quality, 365 days feature) is deleted from every box. The role runs the freeze on every provision;ansible/freeze.ymlre-applies only the freeze to every reachable box and names the ones it did not reach;fleet:freeze-scheduleinstallsa11y-fleet-freeze-daily.timerand removes the patch window.fleet:freeze-proof -- --limit=<box>shows each lever holding from the box. The Edge install is a block whosealways:freezes, the 15-minute Edge guard now only reports, NVDA's update check and the add-on store's are turned off by a newa11y_nvda_updatesmodule and an NVDA that is not the manifest's is refused, andpatch.ymllifts only the Windows levers and freezes again inalways:. The 28-day patch window is removed:patch-schedule.yml,fleet:patch-schedule,--scheduledand its window flags, the patch run record andfleet-watch'spatch-window-missedline.
v0.14.0
Minor Changes
- 425aeb1:
deploy-pins.jsonis the fleet's one pins manifest (A11-205, step 1 of 5). It gainswindows(release, build and the cumulative-update KB),edge(version, MSI URL and hash),nvdaandnode(version and a checksum per architecture), beside the worker release andcore.commitit already held. The worker role derivesworker_edge_*andworker_node_*from it and keeps no Edge, Node, NVDA or Windows literal; the Intel display driver pin stays indefaults/main.ymlas the one declared exception.src/fleet-pins.tsis the one reader.fleet:statuscompares every reachable box with the manifest, field by field (Windows build, Edge, NVDA, worker code), and names the box, the field, what it has and what the manifest wants (fleet OFF THE MANIFEST); any difference, or a manifest that cannot be read, exits 3, and--jsoncarriesdriftbesidebuildSplit. Its advice no longer says to converge forward to the newest build. Renovate's two managers andfleet-watch's Edge-pin reading read the manifest. Nothing moves a box: this step changes what is compared, not what is installed.
v0.13.48
Patch Changes
-
b785ff8:
fleet:deployregisters thea11ysrvtask before it restarts it, reads the registered action back and refuses a box whose task does not start the published launcher (A11-204). The deploy used to restart a task it never registered, so a box whose task still started the laidpackages\nvda-worker\src\run-server.cmd(an orphan since a11ign/a11ign#5546, v0.9.0 on a11y-worker-11) kept serving that tree through the restart and through the rebootverify-code.ymlescalates to: the logon trigger starts the registered action, not the code the deploy laid. The two registration steps move unchanged fromroles/worker/tasks/tasks.ymlintoroles/worker/tasks/worker-task.yml, which the role imports anddeploy.ymlruns, so the registration is defined once.fleet:statusreads each box's worker code (/health.code) as a fourth field of the build split: a code split, or a reachable box that reports no code, printsfleet BUILD SPLIT — Worker code: <hash> on <boxes>; …and exits 3, and an agreeing fleet names its one code in thefleet builds AGREEline. The oldfleet SPLIT — N different code hashesline, which named no box, said "Nothing is broken" and exited 0, is gone;codesstays in--json.