provisioner_compose took 2.6-13.6 s per node with nothing changed, CPU
bound. The node is now composed at its wanted and its newest builds over
one read (frombus.PlansHeldTwice, whose composer resolves each machine
once), its own plans, provisions and refusals and the providers' entries
are read from the watched copy, its own writes taken in at once, and a
provider's manifest once by digest: no read by key per composition is
left. The result says where the time went, phase by phase, and a slow
composition that changed nothing says it in the log.
BenchmarkComposeNode times it over a mesh captured with the read tools
(MESH_COMPOSE_CAPTURE): novox 11.7 s to 0.49 s, ace 0.49 s, g14 and
shanks 0.39 s; plans and provisions identical to before over nine
compositions of the four nodes.
Needs mesh-sdk feat/compose-once (5d26c65); go.mod still pins the old
mesh-sdk until the group pins its head.
Claude-Session: https://claude.ai/code/session_01NcbgfJ9Gum47zkFKWRaovL
Each node's reflector ranked its own list on its own timer, and on 2026-10-31 one node's
lagging first mirror left it a Node.js version behind the rest. The list is now the
module's state mirrors, key list: the node that opens a week's round ranks it with the
module's reflector options, keeps only mirrors whose last sync is no older than the
newest found, and the round's key names the list's revision. Every keeper writes
/etc/pacman.d/mirrorlist whole from the key through the root helper's new verb
mirrorlist, when the key changes and when the file differs, and never writes an empty
or unreadable list. reflector.timer is declared stopped and disabled, and reflector.conf
saves nothing. pacman_mirrors answers the shared list, its ranker and age and whether
this node's file matches; with refresh it asks this node's keeper to rank again,
refused while a round runs, naming it. The round's integration test proves the opening
write and both nodes' files.
Claude-Session: https://claude.ai/code/session_01NcbgfJ9Gum47zkFKWRaovL
The read node-engine.packages named a bucket node-engine_packages that nothing writes:
the engine's manifest is mesh-host's, so its state is mesh-host_packages, read as
disk-load reads mesh-host.data. The view still names which source answered, the
engine's key or the controller's mirror.
Claude-Session: https://claude.ai/code/session_01NcbgfJ9Gum47zkFKWRaovL
mesh/delivery-group group feat/l5-l7-main rejected: composed together: the checker's verdict at the members' heads (c598e176f3b2): unit: modules/gitea.the chec…
mesh/check failed: unit: modules/gitea.the checker's owed is what this module reads, and it consumes no check.status: "Cannot read properties of undef…
ADR 0580 clause 2 as revised: the holder hears haves by watch as it hears resolved values, so a have's change recomposes
its node from the copy it already holds.
Claude-Session: https://claude.ai/code/session_01NcbgfJ9Gum47zkFKWRaovL
Composing one node read every placement, have, module and resolved key one at a time and took past 30 s with the bus
local (ADR 0580 clause 2): the holder now reads them from mesh-sdk's watched copy, its watch handlers queue the nodes a
change touches after the copy took it, and a resolved value's change recomposes its node. mesh-sdk pinned at ba8b8f2.
Claude-Session: https://claude.ai/code/session_01NcbgfJ9Gum47zkFKWRaovL
mesh/check checking 432326264f6c on 4b16c193f0fb | novox/mesh-catalog#585 in group feat/judge-once
The test copies aMesh's log, dated from 2026-10-01 for a fixed clock, but runs the holder on time.Now; since
2026-10-31 every entry fell outside the log's 30 days, so the sweep rightly kept every copy as unrecorded and the
mirror act was missing. The sweep is right; the fixture is shifted to now (issue 1955, L7a).
Claude-Session: https://claude.ai/code/session_01NcbgfJ9Gum47zkFKWRaovL
mesh/delivery-group group feat/judge-once rejected: composed together: the checker's verdict at the members' heads (5c42ba1fb5f3): integration: artifactstoreto…
mesh/check failed: integration: artifactstoretools/cmd/artifa… | 3e08301b27ea on 4b16c193f0fb | judged by the candidate's mesh-build-go at 8dd51540f73a
ADR 0573 §2 and §4 have a build take the latest published version of
each mesh package, or a group's own head, whatever go.mod names (issue
1585, S15). The fetch writes each asked version into the directory's
temporary modfile before downloading: a trunk commit is required at its
Go version with its checksums for go mod verify, a pull request's head
is placed as a pin on a head is, a newer go.mod version is kept and
said, and the answer and the log name each.
Claude-Session: https://claude.ai/code/session_01NcbgfJ9Gum47zkFKWRaovL
mesh/check pass 0c46f47ae57e on 444c2903af7b | unit 4510 run, 4259 kept, 49 skipped | integration 4589 run, 4361 kept, 51 skipped | over bound: unit, …
A key deleted while disk-load was not watching (a restart, a reboot, a machine asleep) never
arrives as a delete. Once the watch has handed over every current key, the readings of modules
with no key are retired once; an empty set is skipped and said, since the engine has not written.
Claude-Session: https://claude.ai/code/session_01NcbgfJ9Gum47zkFKWRaovL
A round no longer retires readings missing from the watched set, which may be empty or partial;
a module's readings are retired when its key of mesh-host_data is deleted or names fewer items.
A round refused before the watch caught up runs once it has, not an hour later. disk_round reads
this machine's key alone, as its grant allows.
Claude-Session: https://claude.ai/code/session_01NcbgfJ9Gum47zkFKWRaovL
It watches its node's keys of mesh-host_data instead of composing every module's items from
about 370 resolved keys per round, so a round no longer takes 30 to 45 minutes across the tunnel
and a manifest its mesh-sdk cannot judge no longer silences it. A changed key measures its module
at once; each round writes how far it got to the per-node state rounds, said by disk_round.
Pins mesh-sdk feat/data-items.
Claude-Session: https://claude.ai/code/session_01NcbgfJ9Gum47zkFKWRaovL
jschoubben
mesh-tools