A replaced mesh-build-go left a check ask's step running and its start
sweep spared it (issue 1890, note 21). mesh-build-go, mesh-build-node and
mesh-build-image pin mesh-sdk 981a35b5, whose steps name their owner and
whose sweep ends another owner's at once; each runs the sweep through one
method that main calls and that its test runs through the conformance
kit's left-running case. The failurekit check now refuses a build module
whose tests do not run that case, test step or not.
47 packages carried the marker and were never kept: about 2000 tests per level ran on every check, whatever it
changed. The marker now lists, as paths or globs from the repository's root (? before a path whose absence is read),
what the package's tests read outside it; the digest hashes those files with the package's own non-Go files. A marker
that lists nothing, names an unclean path, or a line that matches nothing keeps today's never-kept digest.
The 46 markers list their reads; checks/deps and checks/trunk lose theirs (their tests read only temporary
directories). Digest form 6.
Quick fix of issue 1936, chosen by the operator 2026-10-26 21:56; the thorough fix (the Go test log's files kept with
each verdict) stays on the issue.
mesh/delivery-group group fix/capacity-follows-cpu-share rejected: composed together: the checker's verdict at the members' heads (63694fe9222d): unit: ..test
mesh/check checking 3f8b0b42480b on 7c5c29421f02 | novox/mesh-catalog#511 in group fix/capacity-follows-cpu-share
mesh-build-go, -node and -image declare build-hours (node=HH:MM-HH:MM in the node's local time,
empty = any hour) and hand it to the SDK puller (the operator, 2026-10-26: g14 builds only from
18:00 to 06:00); their SDK pin moves to mesh-sdk d00dac1, which has it.
Each of mesh-build-go, -node and -image declares cores-per-step (6, 2, 2) and hands it to the
SDK puller, which bounds the configured capacity by floor(slice quota cores / cores-per-step),
read live; thread limits and the status tools follow the effective capacity (issue 1806's
follow-up: three Go checks on shanks' 12-core slice starved past the 30-minute step bound).
mesh-build-go says its lint and test phases with the ask id (issue 1890).
mesh-build-node kept the first fifteen lines of node --test's output, so a passing line of the first file stood for the
failure. Both build modules now parse through the SDK (go test -json, node --test's TAP report written to a file, the
type-check and go vet as console failures), name themselves in the result, and run the SDK's captured cases; the
catalogue check failurekit refuses a node-build-* module with a test step whose tests do not run the kit.
mesh-catalog #495 gave read on every repository to the tools' account alone. The build modules
fetch over ssh as nox-builder (the forge-account provider's builders account, ADR 0466 §3), so
ace's and g14's fetch of jschoubben/car-hunter and jschoubben/snake still failed; the machines'
agents (nox-agent, ADR 0410 §1) could not read them either.
- gitea: readall.ts grants every account of one list, meshAccounts in tools-account.ts (mesh-tools,
nox-builder, nox-agent; a Go test holds the two Go names to internal/forgeaccount's constants),
on the same start pass, forge events and tools. gitea_unreadable_repos answers per account and
names the accounts the forge does not hold yet; gitea_grant_read takes accounts as well as
repositories. An absent account is asked nothing; the start pass runs again ten minutes apart
until every account was there once.
- mesh-build-go, mesh-build-node, mesh-build-image: pin mesh-sdk ff81bdf8 (fix/fetch-reads-all),
whose forgekey.Said keeps git's words past the bare "error:" line: the mirror's line said only
"git -C: error:". Their own git reasons (ls-files, ls-remote, rev-parse, repack) go through it too.
buildah writes STEP lines on stdout and its error on stderr, so the last line of both hid every image build's cause.
mesh-build-image's reason is now stderr's last lines (600 bytes) with the exit status, and the journal gets both tails;
mesh-build-node keeps stderr's end and stdout's start apart; mesh-build-go's cause skips go build's '# package' line.
Change A rebuilds the three build modules, so their runtimes restart once; it
changes nothing else about their builds until change B gives them the share.
Review round 1: a cap per module let three build modules take three shares
of the node. The share is now one node value held by the node-engine's
yielding slice; each step divides it by every build slot of the node, from the
composer's words, or assumes every build seat at its own capacity without them.
The operator's desktop slows while it builds (shanks at 92%, the Go build
runtime at 1650% of 24 processors). The build modules declare cpu-cap
(default shanks=50,g14=50,ace=75), which the composer turns into the runtime's
quota at idle priority; each step's go, node and buildah commands get
GOMAXPROCS and UV_THREADPOOL_SIZE for their part of it, and the status tools
show the limits and what the kernel holds the runtime to.
mesh-build-node's integration tests read the Node toolchain installed only where the node holds node-build-node, and
skipped on ace: its package now declares node-build-node beside its tests; the integration digest names it, a node
lacking it leaves the package unmet for the checker to place, and declines a placed ask. Re-pins mesh-sdk feat/890-placement.
On ace, which holds no Node build module, five integration tests each spent 70 s in npm's network retries and failed every mesh-catalog check run there (five verdicts since 00:07). The tests now copy the node's installed toolchain or are skipped as NOT RUN; the registry is never reached. Three tests sharing the fixture are one, so the toolchain is copied three times instead of five (10 integration tests in the module, now 8).
mesh/check checking 15a5fe09b62e on c51a46b802d3 | novox/mesh-catalog#488 in group fix/1921-every-failure-r
Pins mesh-sdk fix/1921-every-failure-r, whose check ask no longer stops a level at its first failing test step. A captured fixture with a failing test in each of two packages proves the build module's test step names both.
mesh/check failed: integration: meshbuildnode/cmd/mesh-build-node.TestAFixtureBundleBuiltTwiceHasOneDigest: no toolchain of this lock is installed und…
The review found a cut read took an arbitrary 1000 keys, so the newest steps could be missing. A step id is
base-36 milliseconds plus random digits, so its string order is its start order; the newest max(limit*4, 100)
keys are read now, 1000 at most. The toolchains read logs when it drops keys past its cap.
On a steps state of 1164 keys / 20 MB (mesh-build-node: 10034 keys) the read tools listed every key and
then read each value one round trip at a time, filtering by node only after the Get, so they timed out
at 30 s. Now the keys are filtered by the node's prefix before any Get, the values are read 16 at a
time, at most 1000 a call (the answer says it was cut), and one log line gives rows read and time.
build_*_toolchains reads its values the same parallel way. The kind: a read tool that reads a whole
state key by key; every read tool of the three modules is covered.
mesh/delivery-group group fix/steps-bucket-ages held: a member is held: novox/mesh-sdk@518dae5ec9c8 (merged without a passing check: only a person decides what…
mesh/check pass 4e82a824be7a on cdbe2945cf6e | unit 1994 run, 2322 kept, 5 skipped | integration 2016 run, 2341 kept, 9 skipped | over bound: unit, in…
Without it in Kept, stdio.KeptState refuses the state before the bus (mesh-sdk go/keeps.go
refusedUnannounced), so the pause could be neither read nor written.
mesh-build-go's steps bucket filled to the bus's 64 MiB cap (27,370 keys, none ever removed) and
refused the end of a passed check, so the check read failed; mesh-build-node's (17.8 MB) and
mesh-build-image's grow the same way. A step's key is per step: its reader reads its end within
the step's bound, and a check's verdict is kept by the checker, so two days is ample. The node's
pause key lived in the same bucket and would have ended by itself; it moves to a pauses state
without an age (mesh-sdk fix/steps-bucket-ages, Puller.PausePut). No node is paused today, so
nothing is carried over.
g14's Go builder sat idle at capacity 2 because nothing woke its parked
puller. mesh-build-go, mesh-build-node and mesh-build-image now pin mesh-sdk
b9911dc1, whose puller watches the config value capacity as mesh-config
resolved it and wakes on each change, and hand it that watch.
mesh/delivery covered after its failure: a later walk that contains it was delivered: plan-1792875397682979170 (novox/mesh-catalog at 6fb0ceee into main,…
mesh/check failed: unit: git.novox.be/novox/mesh-catalog/chec… | b60077850dfb on 3bec357968da | judged by the candidate's mesh-build-go at 47bcc1b5810e
mesh/check failed: novox: the check ran past its bound of 30m0s, stuck at: test the unit level of novox/mesh-catalog#458 on novox | 3ca48ff23005 on 89…
mesh-build-go, mesh-build-node and mesh-build-image refuse every fetch, test
and build, and say unhealthy, while their work or cache filesystem has less
than 20 GiB free (build-agent's floor), so the builder chooses another node
instead of filling this one. Each serves build_<x>_release, removing a
workspace once its caller's last step answered; a failed fetch removes its
own; a workspace never released goes two hours after its last use (ADR 0509)
in place of a day, never while a step holds it. Pins mesh-sdk's
fix/1839-build-floor (buildstep.Floor, buildstep.Workspaces).
mesh-build-go binds /var/lib/mesh-build-node-cache/toolchain read-only into the
integration level's one shut-off process when it is there; the unit level never.
mesh-build-node opens its installed toolchains to other accounts (folders 0755,
files 0644, writable by its owner alone; its cache 0711, so the npm cache beside
them stays its own), and realMachine copies the toolchain of its own lock from
there instead of npm ci from the registry, which the level cannot reach.
build-agent's check container has no bwrap; the build nodes, where the checker runs from
R4e, do. The sibling test and the real Node build skip naming it, and the pull-head pin
test still runs its fetch half and skips only the shut-off build.
jochen
jschoubben
mesh-tools