The bus package's doc names the proof, so node-tools rebuilds against the SDK version the same merge call publishes; go.mod keeps mesh-sdk at the zero version.
Builds: design 47 H §5 step 7, G §370
Review round 1 (P7): the runner's bind copied the SDK's claim wrapper and its
caller lookup. It now calls process.ConsumeOnce, which claims, pulls and
releases (at a stop, or once a drain or the bus ended the loop, before its
Closed is closed), naming the claim for the module whose hearing it is.
Builds: design 47 H H-A-197, H-A-196
A tool runner built before H-A-132/133 ran two pull loops on the indexer
bundle's one durable, each event reached one loop only, and nothing said so;
the SDK's guard (H-A-196) could not see it, since the runner bound through the
bus client directly. hearing.bind now claims each durable with
process.PullOnce, naming the runner's hearing and the part that asked it, and
releases the claim as the loop is stopped or drained, or as the bus ends it,
so a re-serve (which stops its old loop first) binds again without a refusal.
The heartbeat test's second consume pulled the same durable: it gets its own.
Builds: design 47 H H-A-197, H-A-196
Review round 1. A ledger binding heard while a module was loaded again
made refreshRepairs serve the repairs on the reload's stand-in, whose
subscriptions nobody stopped: the repairs were served twice, and still
answered after the runtime stopped. The stand-in is now marked as
reloading and the refresh leaves it alone; the reload serves the module's
repairs as its binding is when the load ends. The reload test holds a
bundle's start, issues the binding inside the reload, and asks a repair
after the runtime stopped: without the change it still answered (6 of 6
runs, GOMAXPROCS=1 and 24).
No test failed without keep's stop waiting for the heartbeat's goroutine.
The heartbeat test now holds one Say and stops meanwhile: stop must not
return before that Say has; without the wait it failed 50 of 50 runs at
GOMAXPROCS=1 and at 24.
Builds: design 47 H §774, H-A-160; principle 11
When a module's ledger binding came or went, refreshRepairs stopped every
subscription of its tools and served them all again, so a call in that
gap was dropped the same way as on a membership (H-A-158). serve is now
the one routine that changes a module's subscriptions, by tool address:
it keeps those still named, subscribes the new, then stops those no
longer named; refreshRepairs and the first serve both go through it, so
no tool is served twice or not at all.
The ledger test's later binding now comes and goes ten times, and the
module's own tool must answer the first call after each, within the
tests' 2 s bound; it failed without the change at GOMAXPROCS=1 and 24.
Builds: design 47 H §774, H-A-158, H-A-160; principle 11
TestTheRuntimeSaysItIsThereEveryInterval slept 70 ms and wanted three
beats at 20 ms: a loaded machine that did not schedule the heartbeat in
time failed it (main.20's integration step on the build node); and its
stop returned before the heartbeat's goroutine ended, so a beat could be
said after it ("said after it was stopped"). The ticks are now the
caller's (keep, beside Keep's own ticker), the test drives them and waits
on each beat, and stop returns once nothing more is said. Pinned to one
processor at idle priority beside six busy loops, the old test failed 13
of 100 runs; the new passes 100 of 100.
TestTwoWritersIncrementingOneCounterOnTheBusLoseNoUpdate bounded a
writer's tries at 1000, a guess: on one processor the scheduler lets one
writer win every round and the other lost its 1000 tries every time. A
write is lost only to the other writer's write, so the bound is the
other's count plus one.
Builds: design 47 H §774; principle 11
On every membership of a module, HandleCalled stopped each subscription
of a tool and subscribed again on the same subjects; a call the bus
routed in that window was dropped by the client in silence, and its
caller waited the runner's whole bound (30 s). On the build node this
was the 30 s of TestAReplacedBundlesOldToolStaysGoneOnTheNextMembership
and of TestTheLedgerRepairsAreToolsOfEveryLedgerOwner, and most of the
integration level's 72 s. Now a membership keeps every subscription whose
subject and queue it still names, subscribes the new ones, and only then
stops those it no longer names (make before break).
The test asks the served tool once after each of five memberships, with
the tests' own call bound of 2 s (principle 11); it fails without the
change. The runtime tests' call helper and the ledger listing wait that
bound, not the runner's 30 s.
Builds: design 47 H §774, H-A-158; principle 11
In a module's own runtime the branch follower (first views) and the
bundles each bound the module's durable <node>_<module>_tools with a pull
loop of their own, so each event reached one of the two: the catalog
revision of main.1 reached the follower and never the indexer's bundle,
whose revision verb answered none-yet. The runtime now binds the consumer
once, for the bundles, and hands each event to the follower after the
bundles took it (runtime.WithHearer); the machine's runtime keeps the
follower's own binding.
Builds: design 47 H §752, H-A-113, H-A-132
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
The fixture state mesh-delivery-checker.owed, which the retiring checker keeps, is the assigner's `pins`: the three tests of the
file assert the same of a state watched by a module that is granted it.
Builds: design 47 H §764 (The checker), H-A-61
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
The row of an ask the router holds back as the fourth open one goes, and the router's refusal row names another refusal: the
operator channel keeps no limit on open asks, and its record no state `held-back` (mesh-sdk, mesh-contracts).
Builds: design 47 H §764 (The operator channel's limit), H-A-57
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
The `admitted` hook, its parameter of serveSeats and its comments go; the test rows that built definitions with Callers and Class
go with the field mesh-sdk removes. A seat's verb is answered for every caller.
Builds: design 47 H §752, §764 (Every seat contract), H-A-41
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
The console's surface holds the runner's seat definitions; a seat verb's call waits its `timeout`, and a timeout's text names the
bound waited, not a fixed 30s.
Builds: design 47 H §760, H-A-43
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
Tool gains RunWithin and Within; a bundle's tool call waits `within` where the seat verb it serves names a timeout, else
CallTimeout. serveSeats takes the bound of a verb's key beside `chunked`.
Builds: design 47 H §760, H-A-43
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
The runner reads the `timeout` (whole seconds) of a seat's verb from the seat list's definitions beside `chunked` and hands it to
the bus connection's ask (AskWithin, AskAsWithin, AskAsChunkedWithin); a verb that names none, a seat the definitions do not hold
and a module's tool wait the runner's own bound, RequestTimeout.
Builds: design 47 H §760, H-A-43
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
A bundle's ask of seat:mesh-wants.input reached the holder with no `call` and was refused ("call is one token of 16 to 64 ..."), so no
composition step passed: the runner chooses the call only for a verb its seat list says sends chunks, and a verb the list did not name
(a list read once, unread, or stale) was taken for one that does not. chunkedAsked reads the list again once and, where the verb is still
not named, refuses the ask saying why. A first read of the list that ended without an answer is now said in the log.
Builds: design 47 G A-249 (the runner receives a chunked seat answer for the bundle), ADR 0636 §1 (the seat list is read, not guessed)
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
The operator, 2026-11-15 23:25 CET: every module, agent and operator may call anything; gates and guards can be added
later. The runner refused every call of the knowledge base and the issue tracker while it had not read their seat
definitions. The approval mark stays, held by the gate.
Live on novox every gitea tool was refused: each of gitea's eight bundles overwrote the gate's registered set, so the last one, which does not carry gitea_branch_protection_loosen, made the mark unserved.
A verb asked with no machine (the indexer's seed asks node-build-go's steps) is
answered by any one holder, on the tool and the caller-named subject.
Builds: lead answer A-393 (design 47 G)
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
The mesh-catalog seat definition publishes catalog-revision on
mesh.seat.mesh-catalog.event.catalog-revision.<branch>, so the bus key is
mesh-catalog.catalog-revision.<branch>; the exact-key lookup missed and the event
was dropped silently. Look up the key and its shorter prefixes (as mesh-sdk's
views.adoptedHears does) and say an event no view hears once per key.
Builds: design 47 G (catalog half's revision events)
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
mesh/check failed (2): mesh-delivery-checker: no published of the tagged version of go-meshconditions of me…; and 1 more | 6be5679a0d58 on a96df478d498
The capability stream-publish is stated under mesh-sdk's word (buscalls.StreamPublishWord) and in the runtime artifact's states, so G's test that the stated words are the manifest's states holds; the publish uses G's exported JetStream.
Builds: design 47 G §545, §622; design 42 zfs-owns-zfs A-42, A-44
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
mesh/check being carried to df2162e381f5 on af8385495688 | from 447a8d035935, mesh-catalog 117f933dacf0, mesh-delivery 7221a1e5cd7f, mesh-host 0193949…
Builds: design 42 zfs-owns-zfs §25, lead answer A-44
mesh/check checking 481a83e11c4f on af8385495688 | novox/mesh-tools#69 in group feat/47-G-join11
The tests G added fold into existing ones as subtests (design 47 G §626, A-306: 171 tests down to 154, integration 9
down to 7). TestMeshWaitRefusesLoudlyAndNeverLoops runs its groups of one-second rows as parallel subtests, each
on a runtime of its own, 15 s down to 5 s (A-311: the one-second floor of a wait's timeout stands, no clock is
injected). The go.mod no longer requires meshconditionsclient, which no package loads (T4). Node names in the
lines G adds are role names (T5).
Builds: design 47 G §626, design 34 every-event §19, A-306, A-311; review round 2 findings T2, T3, T4, T5
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
The tests of A-249 and A-257 served the verb on one connection and asked it from another at once; the
subscription was still in the holder's buffer when the ask reached the bus, so the ask was answered 503
no responders (13 of 300 runs with -cpu 1 before, 0 of 2000 after). The runner flushes after serving
(runtime.go:352, :807); the tests now do as it does.
Builds: design 47 G A-249, A-257
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
The join's importers name the delivery repository's moved contracts and sdk
(MV1), node-tools' tests use mesh-sdk's gate, the duplicate node-engine claim
rule of the reader release gives way to G's, and the assigner names its first
watched branch, never a constant. The seat holders' AnswerSeat callers are
area 6's.
Builds: design 47 G join 11 (merge of fix round 11)
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
The seat verb approvals is the gate's own, served wherever a marked verb is: the runner answers it by caller from the
calls its gate holds, with no bundle naming it; a marked verb already reaches the gate by caller. Tested with zsh's
node-login-shell execute as the case.
Builds: design 47 G §677, §697; lead answers A-264, A-265, A-266, A-271
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
A seat verb whose definition says chunks, served by a bundle that answers the whole file as the member `file`, is sent by the
runner with buscalls.SeatChunks.Send and answered with the chunk count (Conn.AnswerChunked, called from serveSeats). A chunked verb
asked of one machine (`@<node>`) is received in chunks as the verb is.
Builds: design 47 G A-257, A-249, A-254, §681
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
The approval unit table's in-memory state held `<node>.act.sha256:<hex>`, a key the live bus
refuses, so the table passed while every marked call failed on the bus. meshtest.State now
refuses a key outside letters, digits and -/_= in dot-separated names, as the bus does, checked
with mesh-sdk's keys.CheckName per name; a row of the refusals table holds it.
Builds: design 46 7-approval-mark §13, §22; lead answer A-209
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
A capability reading must be tied to the runner that made it, so the grants
part can tell an older runner's reading from the running one's.
Builds: design 46 7 §6 (the announcement carries the runner's fingerprint), lead answer A-203 (key `fingerprint`)
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
Design 46 7-approval-mark §22's integration test: the runner's gate wired as main wires it, on a
real bus with stand-ins of the operator channel's router and the hand-act seat: a marked call
asked, approved, run once, recorded, listed; a restart between the run and warranted's answer
recorded once; a Decline runs nothing; no state bound or no holder of operator-channel runs
nothing. It failed first on the act's key: `<node>.act.sha256:<hex>` is a key the bus refuses,
so every marked call was refused on the live bus. The digest is now escaped with
keys.EscapeName (lead answer A-209). The runner's two start tests are one, so the module gains
no test (design 47 G §419, §626).
Builds: design 46 7-approval-mark §13, §22; design 47 G §419, §626; lead answer A-209
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
The runtime's load test gains rows (run with the Python fixture, before the bundles' skip): a bundle whose
marks name a tool it does not register serves no tool of its module, and one that registers it serves.
mesh_call's failure words move into Surface.callFailed, whose rows hold that only a marked tool reaching no
responder names withheld and mesh-grants.grants_user_list.
Builds: design 46 7-approval-mark §4, §6, §22
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
Where serveSeats answers a seat's verb it reads the verb from the runner's seat list (seatdef's
EffectiveCallers): an unproven caller is refused, and every caller while the seat's definition is unread,
names no such verb or opens it to none; a proven caller is among them, since the grants part composes the
call subject only for the effective callers. A module's tool on the subject naming no caller is unchanged.
Builds: design 47 G §610; lead answers A-165, A-179
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
At each start (and when the hand-act seat's holder is heard again) the gate pends each acted, unrecorded
ask's call-unanswered key in its held writer, bound five minutes, and ends the pending once its warranted
is made again, so a restart neither clears nor raises the condition of a run still unrecorded.
Builds: design 46 7-approval-mark §16; design 47 G §384; lead answer A-154
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h