go mod download loads only the module graph a go.mod names; a vet, a test or a build then loads
packages shut off from the network and failed far from the fetch, in fewer words (the builder's
"pull_drop.go:11:2: module lookup disabled by GOPROXY=off"). After each directory's download the
fetch loads its packages with their tests shut off and fails naming each package whose load
would reach the network. The graph of a sibling laid beside is loaded once with the network
first, so a go.mod the composer's build reads and the graph's download skipped is fetched, then
proven the same way.
1966/AC178-2: a row of TestAVerifyThatFailsEndsTheStepNamingTheDirectoryAndTheLine; the beside
case in TestTheFetchDownloadsTheGraphOfASiblingModuleLaidWithTheTreesItRequires.
Claude-Session: https://claude.ai/code/session_01JG73KEq4zoDfM9WtYpcQgZ
ADR 0573 §2 builds every module against the latest published version of each mesh package it
uses, whatever its go.mod names. The fetch took versions only for packages the directory's own
go.mod names, so a package only a module it replaces by a directory requires (mesh-build-image
replacing buildstep, which imports the node-build-go seat package) was never taken: go mod
download passed and every shut-off step looked it up on the network. Now the requirements of
each module the directory replaces by a directory count too, taken the same way (the ask, its
repository's asked head, else the latest) and said as named by that directory.
1966/AC178-1: rows of TestARequirementAtTheZeroVersionIsTakenFromTheAskOrTheTrunk.
Claude-Session: https://claude.ai/code/session_01JG73KEq4zoDfM9WtYpcQgZ
Since buildstep's drop verb imports the node-build-go seat package, both modules' shut-off vet,
test and build looked it up on the network: their go.mod did not name it, so the fetch took no
version for it. go mod tidy adds it as an indirect requirement.
Claude-Session: https://claude.ai/code/session_01JG73KEq4zoDfM9WtYpcQgZ
mesh/check checking c39503e3227c on c4463ba47fcf | novox/mesh-catalog#650 in group fix/1966-seat-list-unread
claude_code_seat_list_again answered person.<name> and the healers, but every
call through the mesh MCP server is proved as its machine's tool runner
(issue 627); node-tools now stands for that session (issue 1966).
Claude-Session: https://claude.ai/code/session_01JG73KEq4zoDfM9WtYpcQgZ
ace's node-engine wrote seat-list-unread in its state findings, and no
condition rose: the keeper's observers did not name the node-engine. It now
reads mesh-host.findings, so ADR 0636 §1's condition is raised and clears at
the first full read (issue 1966).
Claude-Session: https://claude.ai/code/session_01JG73KEq4zoDfM9WtYpcQgZ
GOPROXY joined the Go seat, the public proxy and direct with commas, so go went on only after a 404 or 410: a contracts
package the seat does not hold, and a public proxy a node cannot reach, ended the fetch before direct. It joins them
with | now, so any failure goes on to the next. go words a proxy it cannot reach as "invalid version", which the pin
path read as an unknown revision and labelled "on no pull request head"; a network failure is never that now
(network is unreachable and no route to host are network failures).
Claude-Session: https://claude.ai/code/session_01JG73KEq4zoDfM9WtYpcQgZ
A lost event between a shared stream and a module's consumer could not be located: no tool read a consumer's delivered and acknowledged sequence or a stream's messages for one subject. nats_consumer reads from the monitoring document; nats_stream_messages asks STREAM.MSG.GET on the module's own credential, newest first and bounded.
Claude-Session: https://claude.ai/code/session_01JG73KEq4zoDfM9WtYpcQgZ
hand-acts read the whole log (84k entries, 47 MB) on every call and could never answer inside the call bound or
the bus payload. It now reads from the newest entry backwards, page by page, and stops at the window's start, at
900 KiB of acts or at 20 s, saying complete: false, cut and why. warrants is bounded the same way. mesh-healers'
start-up seed takes a partial answer and logs it, instead of retrying forever.
Quick fix: this (mesh-catalog only, about 1.5 h). Thorough fix: also the tool runner never freezing on an oversized
bundle line and logging unanswered calls (mesh-tools, a separate fix by another developer, lead's ruling Q3), and an
SDK-side bound. Chosen here: quick for this module, the thorough part running beside it; the byte cap is the
module's own (lead's ruling Q1, a `limit` argument is a later note).
Claude-Session: https://claude.ai/code/session_01NcbgfFfQ4Z9UE5jK32PmE5
The checker sends a large node's request packed, as it passes the bus's message limit; compose reads either form through
ComposeJob.RequestBytes. Only mesh-build-go serves compose. Delivered as a group with the mesh-delivery change of the same
branch name: an older mesh-build-go refuses a packed ask naming that it holds no request object.
Claude-Session: https://claude.ai/code/session_01NcbgfFfQ4Z9UE5jK32PmE5
Not read is not unneeded on the newest scope too (issue 940's rule): a module whose newest composition is
unpublished or refused is held at its newest build, said in provisioner_plans newest-held and bounded on the health.
A retiring key needed again cancels its ask on operator-channel (accept.cancel), retried until the router takes it.
Claude-Session: https://claude.ai/code/session_01NcbgfFSnn24M3GcqttoU7i
The composer finds a listen's machine port in a container mapping that names
it on either side (MachineSide). The program's listen 5000 matched the
registry container's inner 5000 in 127.0.0.1:5001:5000, so the store's address
became the container's loopback side and the program's own port was never
opened: the move to 83cd1a3d on 2026-11-08 23:25 UTC cut the artifact store off
(issue 1966 note 109). The program now binds 5100, where every machine fetches
(the artifact-store alias and the seats' reach agree), and the container serves
5001 inside and out. A row holds that no container names 5100.
Undo: revert this commit.
Claude-Session: https://claude.ai/code/session_01NcbgfFSnn24M3GcqttoU7i
jschoubben
mesh-tools