The pin judge removed main.23 and its entries when main.24 became current, while the assigner had
just written main.23's builder as the want: the node-engine asked its entry and was told none is
kept, holding the want for good. A passed pipeline-ended now pins the pipeline's kept revision
under a record release.<branch>.<module>, one per mover, beside the pipeline's own pin; the seat
mesh-wants' pipeline-wanted of a newer pipeline on that branch (its wants written to the end,
tier-late included) frees that record from every revision below the newer one, for each module the
newer revision's movers name. The indexer consumes pipeline-wanted (its manifest says so). A
revision a release pin holds keeps its entries, as removeUnlisted already does for any kept revision.
Tests: rows in the pins table (a passed pipeline's revision kept once no longer current; a failed
one's goes), the freeing in the pins test (a release past a failed pipeline, freed by the next
passed one's pipeline-wanted, twice changing nothing), and a pipeline-wanted row in the events
test. TestIndexerWhole now reads main.11's release beside its validate base, and waits for the held
source-moved's acknowledgement before it reads the consumer (the write comes first). No test added.
Builds: design 47 H-A-168, H-A-170 (hq design/47-H1-answers-59 87b43ec8), H-A-147, H §749
A run on request copied its pipeline's source onto its build step ask,
so the builder waited for the run's candidate fingerprints in its
picture of the branch, where they never come, and failed every run past
the bound. The seat mesh-builds' contract leaves a run's source absent:
the runner now does. A revision the builder's view takes in whole that
changes no entry now also wakes a waiting ask, told by the catalog
client's new Held.
Builds: design 47 H H-A-162, H-A-155
With nothing numbered of the builder's own, the tool answered that a floor
would be asked at the next claim and looked again, doing nothing.
Builds: design 47 H H-A-159, G §369
The process gives the builder its picture of each watched branch (the
source revision the branch's view holds) and the bound of a step ask's
wait for its revision, read from the setting revision-wait-minutes
(default 2) carried as a file under the state dir like
toolchain-wait-minutes.
Builds: design 47 H H-A-155, H-A-156, H-A-141
A step ask reached the builder before its picture of main's catalog
revisions held the revision the pipeline builds; the fingerprint was
built as no first branch's build, reserved no version, and the step
failed on builder_reserve_done's words. The ask now waits, woken by each
entry the picture takes in, until the picture holds its source revision
and the asked entry from its branch; past the bound it fails naming the
revision and builds nothing.
Builds: design 47 H H-A-155, H-A-141
Every look numbered one more patch in the state versions (claims carry no
source revision since H-A-48): nodeengine 0.1.5 to 0.1.36, beside the
reservation's numbers.
Builds: design 47 H H-A-159, G §348, §349, §369
The reservation's published of nodeengine 0.1.4 (9724da7a) was taken into the
state versions' entry 0.1.4, numbered for an older fingerprint, which turned
published and kept the old digest c0ef9a3d.
Builds: design 47 H H-A-159, G §369, A-415; ADR 0605 §4
Every revision the validate test judged with a base had it at its number minus one, so the test
passed with the case before read as the number before (review round 1). Revision 12 now names base
7 with 8 to 11 never kept: it ends manifest-refused against base 7's manifests, and could not run
when the case before is read as the number minus one. No test added.
Builds: design 47 H-A-154 (hq design/47-H1-answers-59 01842bbe), H-A-151
A take-in that moves nothing keeps no revision for its source number, so validate's case before,
read as the step's revision minus one, named a revision never kept and could not run. The take-in
now writes, on the catalog revision it keeps, the field base: the newest kept revision below it,
the one it pins for its pipeline, in the same write; absent on the branch's first take-in. The
assigner's validate reads the base from the revision it already asks for; where the revision names
none the case before is empty and its end says so, never a could-not-run.
- mesh-delivery-contracts: the seat mesh-catalog's CatalogRevision gains base (schema, fixture,
generated Go and TypeScript by seatgen).
- mesh-delivery-indexer: keepBranchEntries names the base it pins; the pin test's rows check it.
- mesh-delivery-assigner: before reads the base; the validate test asks a revision with none, and
both stand-ins name the base as the take-in does.
Builds: design 47 H-A-151, H-A-154 (hq design/47-H1-answers-59 01842bbe), H-A-147, H-A-108 (3), design 47 H §743
TestIndexerCountsOnlyThePlacedBuildNodesToolchains failed on the build node under load
(pipeline main.18) in two ways, both of the test, not of the indexer's count of the
placed build nodes' toolchains:
- it restarted the indexer as soon as the module was written, before its take-in
ended: the stop cancelled the take-in, its source-moved stayed unacknowledged, and
the branch's durable (one message at a time) held the next one behind it for its
acknowledgement wait, past the test's 20 s bound. Each restart now waits for the
source-moved's acknowledgement; TestIndexerWhole's restart the same way.
- a source-moved said right after a restart could reach the stopped program's pull,
still counted until the bus closed its connection. restart now returns only once
the bus has closed every connection of the stopped program, heard on the system
account's disconnect events, with no sleep.
Under load (8 runs side by side, GOMAXPROCS=1) the test failed 5 of 80 runs before,
2 of 240 with the first change alone, and 0 of 600 with both; TestIndexerWhole 0 of 600.
No test added.
Builds: design 47 H §743
The leaving-out took every assigned module the case's manifests lacked,
while the judgement said only those the watched branches' newest
revisions list nowhere, read again at judgement time. Now the set is read
once from the newest revisions when the run starts, written on the run,
and used for the leaving-out, the finding and every step's judgement and
the resume. The candidate never enters it: a module the change drops
while still assigned stays in and fails the composition as its own
fault. While no newest revision is followed nothing is left out.
Builds: design 47 H H-A-152, H-A-153
The judgement of every step reading a run, and the resume after a
restart, read only the stored run; the set of assignments left out is
written there when the run starts, so it is not kept in a step's memory.
Builds: design 47 H H-A-153
A stale assignment left out of a machine's composition is now a finding the
condition keeper carries: kind assignment-unlisted, scope delivery, id
<node>.<module>, a warning, raised by the composition step that leaves it
out. The pass over the assignments clears it when the assignment is gone or a
watched branch of the module lists it again, so the clear never waits for the
next step on that machine.
Builds: design 47 H H-A-149, H-A-148
The retired checker's assignment on the control-node made every composition
there refuse "nothing provides", so every composition step on that machine
ended could not run. A module assigned to a machine that the case's manifests
do not hold (no watched branch lists it) is now left out of the case's
assignments as well, said in the journal and at the judgement as unlisted, and
the machine is judged on what remains: a stale assignment is a fault of the
assignments, not of the change judged.
Builds: design 47 H H-A-148, H-A-143; ADR 0578 §2
The pin judge removed the branch's previous revision in the very write that took the next one in
(pipelines main.13 and main.15 on 2026-11-20), so the validate step's case before could not be read
and the composition was never asked. The take-in now pins the newest revision it kept before its
write under the same pipeline record as its own, and the pipeline's end frees the record on every
key it pins; a take-in with no revision before it pins only its own and says so.
Builds: H-A-147 (hq design/47-H1-answers-59 310f8955ba16a1ca8ae9661d8196cab16e3d9597), design 47 H §743, §749, H-A-108 (3)
Under load (8 copies side by side) TestTheGrantsPartWhole panicked at
integration_test.go:151 at abb7cfcf too: st.Consumer's error was dropped
and a consumer deleted after ConsumerNames listed it gave a nil one.
Builds: design 47 H §746 (the pipeline's integration step)
The subtest "a runtime denied the marks is delivered no call on a marked
subject" published the caller-named call on mesh.mod.shell.call.execute.asker~a,
a token the composer never grants (buscompose.CallerToken gives a~asker
for the user a.asker), so the server refused it every time; the test
passed only because it read the refusals with a non-blocking select
right after Flush, before the client's error callback had run. Under
load the callback won the race and the pipeline's integration step
failed (main.12, 15:50). The call now uses CallerToken, and a probe on
another caller's token is published last: the server refuses in order,
so waiting for the probe's refusal waits on every refusal asserted.
Builds: design 47 H §746 (the pipeline's integration step), G A-203 (AM-D4)
The kept stream of the runner's pipeline-ended was made at G's
definition, sourced by a three-part subject; H1 moved the event to
<branch>.<source>, so the stream took nothing in since 08:13 and the
assigner, which hears a kept event from its kept stream, heard no end.
createKept left every held kept stream as it was. It now brings a held
one's source filter from the events stream to the definition's in place
(a stream update, its messages and other settings kept, never made
again) and says the held and the composed filter in the journal and the
pass; the same holds for pipeline-started and wait-held.
Builds: design 47 H H-A-145 (hq design/47-H1-answers-59 0787704b), §746
Pipeline main.8 failed TestBridgeSaysOneHeadMovedPerMove and TestBridgePerformsWhatIsOwedOnce on
shanks; both fail the same way at 403fb306, before the catalog client change, under load.
- The bridge read every head again, start true, whenever the registrations' first view said its
values after the start had begun, so a head moved just after the start was said as a start.
Only a registration adding a watched branch is adopted now (Watched.Set says so), decided under
one lock with the start's read.
- The head-moved test moved the head before the restart's read was answered, and read the state
heads before the bridge wrote it; it now waits for the stand-in's answer and watches heads.
- The owed test stopped the bridge while an outcome's publish was on its way (heard again after
the start), and an act said while the bridge stops can be handed to the stopped bridge's pull,
which the bus hands again only after the consumer's acknowledgement wait (30 s) past the
test's 20 s bound. It now hears the bridge's acknowledgements before the stop and waits the
consumer's acknowledgement wait for the act said meanwhile.
Builds: design 47 H §739; design 47 G §532; design 47 H-A-48
Review round 1: an unlisted module was lost where the composer blamed a module.
It is now said in the failed or could-not-run words where one is set, else as a
warning. Once registrations are held, a module none names has no watched
branches and is unlisted, never composed from the branches the client follows;
before the first registrations the gate's source's branches stand in, in its
order. A module is judged unlisted by its own watched branches.
Builds: design 47 H H-A-143, H-A-144, G §532
The composition step judged a machine at the step's one branch revision, so an
app registered on another watched branch was missing and the machine failed
"nothing provides", ending every mover there "could not run". Each module's
entry now comes from the newest revision of the first of its watched branches,
in its registration's order, that lists it; the candidate's revision replaces
its branch's newest for the modules it moves. A module assigned to the machine
that no watched branch lists is said in the machine's reason.
Builds: design 47 H H-A-143, G §532
A reader whose picture held the earlier model's revision 528 skipped
revisions 1 to 7 of the new numbering in silence: the builder built
every pipeline step as a candidate's, unversioned, so nothing was
released and no want moved. An event below the held revision is now
the state made again: said with both revisions, the view's made time
moved on (so the follower forgets the old numbering's revisions heard
and is not left behind for good), and the first view asked again. An
event equal to the held revision is still dropped.
Builds: design 47 H H-A-141 (hq design/47-H1-answers-59 3d722934), §750, §754
As the grants part gives a module's consumer on a seat event's stream kept by age:
without a start from now a made durable would replay every past source-moved of the
branch past the start's guard; the current revision is read at the start already.
Builds: design 47 H §743, H-A-134 (hq 8d30cc42), H-A-129, H-A-130
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
catalog-revisions kept trunk.<n> and branch.<branch>.<n> objects beside the
branch-keyed revisions, the newest trunk one marked current, so a reader heard
revision 528 while its picture held 1. A branch's take-in now drops every such
object no record pins, says the count in the journal, and the revisions tool marks
each branch's newest revision current.
Builds: design 47 H §743, H-A-135
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
The grants part composes no per-branch durable on the source-moved stream, so the
indexer never heard a source revision after its start. Its bind now makes the
durable itself, as a module's seat-event consumer is made (SeatEventConsumerConfig,
named <node>_<indexer>_<branch>, filtered to the branch, one message at a time),
and binds an existing one as it stands.
Builds: design 47 H §743, H-A-134
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
At a clean state the grants part makes the per-branch consumers of source-moved only from the modules view, which the
indexer answers only after its first take-in, so nothing ever took the trunk in. At its start, and again at each
failed bind, the indexer now reads the branch's current source revision through mesh-delivery.source and takes it in
as a heard source-moved of it would, when its view keeps no revision of the branch at or above it: start true exactly
when it keeps none, every repository the answer names a head. No timer; binding on registration-changed is unchanged.
Builds: design 47 H §743, §768, H-A-129, H-A-130
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
Review round 1, H-A-121 (2): beside the one-item form, all true drops
every pin of the state pins, each recorded through mesh-hand-acts and
asked of the indexer as the one-item form is; it answers the count
dropped and the pipelines the pins named.
Builds: design 47 H §765, H-A-121, H-A-124
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
Review round 1, H-A-121 (2): beside the one-item form, all true drops
every want key, held or not, and every value of an earlier form the
decoder refuses, each act recorded through mesh-hand-acts first; it
answers what it dropped, counted per kind (held, current, undecodable).
Builds: design 47 H §765, H-A-121, H-A-123
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
Review round 1, must 6: the stream's definition names its keep, as the
manifest template states a stream's age (max-age-days).
Builds: design 47 H §739, H-A-23
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
packageEntries, manifestBytes, writtenModules, WriteHeard, heldOldSections, keepOldSections,
the overlay's written and changed, moduleOfDir with moduleAt, EncodeSeriesPackage,
InputDigest, HoldSeats and the barBlock shape's functions are referenced by no file of the
repository; with HoldSeats gone the seat list's lock is a plain mutex WithSeats takes.
barShows and barOption stay: a judge test reads them.
Builds: design 47 H §764; lead answer H-A-48
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
TestPullOrderAndCapacityOnTheBus still asserted deliveries first, red since 773f89e9 made
the build seats' queue kinds take turns (not since 83d45421: it fails at 83d45421^ too).
It now asserts a delivery, then the check in its turn, then the other delivery; it fails
on a pull that takes deliveries first.
Builds: design 47 H H-A-41, H-A-117, H-A-120; G A-440
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
Before the state held is bound, hearing a source-moved read it to hold the event behind
earlier ones, failed and left every message unacknowledged, so nothing was taken in (two of
the indexer's integration tests, which make no held state). A state not there yet holds
nothing: the event is taken in; holding it still needs the state.
Builds: design 47 H H-A-114, H-A-118
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
The teller reads pipeline-ended's H1 form, saying each mover's widening once for its module,
watched branch and commit; its hearing of record-opened, whose forge-made record H §763
retires, goes with its consume in the manifest.
Builds: design 47 H H-A-119, H-A-72, H-A-51, §763
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
The seat mesh-pipeline-runner's MoverEnded gains commit and widens (regenerated, with a
fixture row), and the runner writes them from its pipeline value when it says
pipeline-ended, so the teller says a widening from the event alone.
Builds: design 47 H H-A-119, H-A-72, H-A-51
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
A source-moved the indexer cannot take in is written into the state held, the event whole
(call, heads, start beside branch, source, why, since), and acknowledged once written; a later
source-moved of a held branch is held behind it, and the branch's key keeps them in order.
The indexer takes them in again from that state alone, in their order, on its own room and
at its start; one neither taken in nor written stays unacknowledged. The whole test's held
case reads the new form and that nothing is left unacknowledged.
Builds: design 47 H §744, H-A-114, H-A-118
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
The passer's catalog entries are the followed branches' own, so each catalog-revision asks
again an entry whose ask failed to reach the holder, and so does each reopening of the
assigner's link to the bus; the manifest's setting entry-asks (default 5) bounds it, and past
it the assigner's finding writer writes entry-unanswered (and entry-not-found) through the
entries client.
Builds: design 47 H H-A-60
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
The view's tools already read indexer_pending's {held: [{branch, source, why, since}]} through
view.DecodeHeldTakeIns at the joined base; the take-in package's earlier PendingAnswer,
PendingWait and DecodePendingAnswer, read by nothing since, go with their test.
Builds: design 47 H H-A-102, H-A-97
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
The seat mesh-delivery-checker's generated Go seat, client and TypeScript package go from
mesh-delivery-contracts (and their entries in its manifest), with the old copies under
contracts/go; the indexer, the forge bridge and the assigner require them no more, and the
assigner judges with its own internal/assignment verdict constants. The git, node-build-go
and node-build-node fixtures, the buildstep and steprun tests and the indexer's and
assigner's whole tests name the seat mesh-wants (the asker of every check-kind run in H1)
or the runner where they named the checker; the indexer's unused CheckerModule and the
judge's stream exemption for the checker go. The step kinds check and recheck stay with
their queues (H-A-117).
Builds: design 47 H H-A-57, H-A-117
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
The assigner branch was built before SDK-C took Case.Verdicts out of the composer's request;
its test still read the field, so the joined line did not compile.
Builds: design 47 H H-A-112
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
The assigner's whole tests no longer assume wants follow the trunk's
fingerprint: a runner's pipeline-ended names the movers and versions, the
wants are written tier by tier on running, and pipeline-wanted says running
at the end. Test 113's cases: a provider and its consumer in tiers, the
acting part's build and the old holds ignored, a failed pipeline's version
and an older end heard late writing nothing, a placement's build reverting
the want, tier-late after tier-bound with no later tier written, and a node
not stating want:version held awaiting its hand placement. Test 117's cases
run the step validate whole: a refused manifest, a pass and a cycle. The
first views answer every account (no callers), and tier-bound's unit is a
variable the whole test shortens.
Builds: design 47 H §774, §750, §751, §772, §745, §752; acceptance tests 113, 117; lead answers H-A-10, H-A-37, H-A-40, H-A-41, H-A-56, H-A-73, H-A-75, H-A-99, H-A-108
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
The composer's Case.Verdicts goes (its one writer was the checker); the
validate step judges the manifests before the composition step, so caseOf
answers the manifests and none words alone.
Builds: design 47 H §745; lead answers H-A-112, H-A-108
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
The seat mesh-catalog's module value and entry inputs lose src and fp-src
(C1d), so a trunk entry's release is the version-released whose parts are
what its build.ended made, the one way left of the two the assigner had.
Builds: design 47 H §764; lead answer H-A-48
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
The assigner takes the seat mesh-wants' step validate beside the composition
step. Every mover's validate step of one take-in reads one run, the key
<pipeline>.validate of composition-runs with no machine, taken up at the
start like the composer's runs: a cycle the revision's entries make
(graph.RefusedCycle) ends it at once; else the manifest validator built from
the revision's entry of the module holding mesh-catalog runs on the Go build
seat, its input kept and served on mesh-wants' input, over the revision's
manifests and the branch's previous revision's. The step ends failed cycle,
failed manifest-refused naming every problem the revision causes, or passed.
Builds: design 47 H §745; H-A-34, H-A-51, H-A-108
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
The words of each node-engine's manifest-reads-changed are written into the
assigner's state engine-reads, key the node, value {words, at}, before the
event is acknowledged, and read back from the state's watch at the start, so
a restart loses no node's words. The verb manifest-reads is no longer asked:
the keeper's asks, the unanswered-call finding and the node-record fallback go,
with the invoke in module.json. assigner_engine_reads lists the state.
Builds: design 47 H §772; H-A-104 (1), H-A-45
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
The pass reads each node's placement before the rule: a placement naming a
build has its want name that build, by the version released at it (the kept
end's when it is that build, else the release whose parts the build made),
with no pipeline and no rank; with no released version known no want is
written, said once. A placement's build is no longer made a held put:
placementPuts and its test go, and a placement naming no build any more
gives the node back to the rule.
Builds: design 47 H §750, H-A-5, H-A-37, H-A-39, H-A-99
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy