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 failed: unit: ..test: ✔ a transient empty list for four passes retires nothing; five stable passes retire, data kept (418.253369ms) | d00da…
The operator, 2026-10-26 21:18: g14 builds only from 18:00 to 06:00. A build-hours setting names a
window per node in its local time (crossing midnight allowed, absent = any hour); outside it the
capacity judged by is 0, running steps finish, and the puller wakes on one timer at the next edge.
The change is said in the log and in capacity.<node> with the window and when it opens.
Since the build CPU cap (issue 1806) shanks' mesh-build-go kept capacity 3 inside a 12-core slice:
three full checks got about 4 cores each and ran past the 30-minute step bound (issue 1890).
The puller now judges by min(configured, floor(slice quota cores / cores-per-step)), the quota
read live from its cgroup, says it on change and writes capacity.<node> to the pauses state.
A run's phases (fetch and merge, digest, tests) are logged with the ask id and written to its
running step key; Asker.Wait hands every step value to a WithStepSeen hook.
mesh/check pass ff81bdf8cc8f on 960d6c34d25e | unit 5815 run, 345 kept, 18 skipped | integration 5603 run, 640 kept, 29 skipped | over bound: unit, in…
The mirror of jschoubben/car-hunter on ace said only "git -C: error:": git writes a bare
"error:" before the forge's message, and the reason kept the first non-empty line. forgekey.Said
keeps every line that says something, leaves bare prefixes out, joins them and bounds them at
600 bytes; every git and ssh-keygen reason of buildstep and forgekey goes through it, and the
mirror's commands are named by their verb, not by -C.
mesh/check checking fd7d93333e6b on 21252194d530 | novox/mesh-sdk#93 in group fix/daemon-dir-kept
Declining an ask whose probe failed would turn a truly broken toolchain into a
wait of up to the decline bound that then fails with an empty "needs a node
with" reason. With the node-engine no longer deleting a running build and
commands started in a directory that exists, the getwd cause is gone, so the
guard's refusal stays as it was. The decline refactor goes with it.
Since the grants part creates a manifest's buckets after a delivery (ADR 0547), a module restarted by that
delivery can open a new state before it exists; OptionalBucket answered nil and the part never looked again
(mesh-delivery's checker and builder, state hearing, 2026-10-26). AwaitBucket binds on the server's
STREAM.CREATED advisory, or at the next use, and writes the latest value kept per key. The copied composer's
agreement test leaves that one row out: the controller's list is no longer loaded since the take-over.
A build module runs in the directory its build was unpacked in. When the
node-engine replaced that build, every command the module started inherited a
deleted working directory, and the toolchain probe's failure became a final
refusal of the check that asked it (three checks on ace on 2026-10-26).
forgekey: ExecRunnerClean starts its command in the HOME it is given, or the
temporary directory; the other runners keep the process's directory while it
exists and use the temporary directory once it is gone.
buildstep: a guard whose probe failed answers a Refusal with Again set; the
puller declines such an ask (delivery.StepDeclined) so its asker puts it again
for another node or a later pull, instead of ending it refused. A version
mismatch stays a refusal.
The second half of ADR 0553, merged only once the node-engine that decodes
yields runs on every node (§5). The share is the node-engine's own config as
the node's layers give it; each step divides it by its own module's capacity.
A share out of range bounds builds at the default and never refuses the
node's composition. A change of the share recomposes the build runtimes, which
restart once.
Review round 2: change A must not slow builds before any slice exists, so a
module given no share sets no GOMAXPROCS. The slice's quota bounds the total;
threads only size one step, from the node's share over the module's own
capacity (shanks: 24 x 50% / 3 = 4). The share is the node-engine's own config.
Review round 1: the decoder type ships before any composer can emit it, so a
routine SDK bump never sends yields to an engine that refuses them; the
composer half moves to a later change. Every yielding process of a node goes
into one top-level slice, meshyield.slice, so the three build modules share one
quota instead of taking one each; the share is one node value, node-tools'
build-cpu-share, and each step's threads divide it by every build slot of the
node (words from the composer, a conservative fallback without them).
mesh/check being carried to 3203358c4cb5 on 2aae6f5580f2 | from mesh-delivery c8b26735c8cf | novox/mesh-sdk#92 in group fix/734-lost-ask
A deaf asker must be loud: SayHearing keeps each Hearing in the asker's state hearing, and HearingHealth answers the
module's health tool from it, unhealthy with why while DEAD_LETTERS is not heard.
An asker started before its grant was composed was refused once and stayed deaf until it started again, silently.
DeadLetterHearing tries in the background and again on an event (Again: a reconnect, an ask waited for while deaf), and
tells each change of Hearing so the module's health tool can say it. DeadLetterReads drops the flow control a pull
consumer does not use and deletes one consumer by name.
The checker's try and the builder's lanes waited for ever on an ask given up during a bus reload: no build node writes a
step for it. LostAskWaiter hands the asker a failed step on the node bus when DEAD_LETTERS keeps the ask's copy
(HearDeadLetters, granted by DeadLetterReads), and under NotTakenBound when no node took the ask within the bound.
A check ask given up during a bus reload left its try running for ever: the asker had no grant to learn that
DEAD_LETTERS holds it. DeadLetterReads grants an ordered consumer on DEAD_LETTERS filtered to the seat stream's copies,
and nothing else of the stream.
The operator's desktop slows while it builds: the Go build runtime took 1650% of
shanks' 24 processors with nothing holding it back. A build seat's holder now
runs its runtime at idle priority under its node's cpu-cap share (declaration
Yields, composed from Seat.Yields and the module's cpu-cap setting), and
buildcpu/buildstep give each step its part of that cap in threads.
A check ask's packages may need something only some build nodes have (mesh-build-node's integration tests read the
Node build module's installed toolchain). An ask now names its Needs; a holder lacking one ends the ask's step refused
as declined and drops the message for its asker to put again, skipping that worker until its next pull wait; a level's
packages whose declared needs the node lacks are left untested and named unmet, for the asker to place.
jschoubben
mesh-tools