Review round 2: a durable nobody ever read (no pull waiting, no delivery ever) was
named behind when its twin was not current, and the plain durable of a module whose
program left was never judged (live: claude-code's plain durables 1,000 to 3,000
behind beside a current `_tools` twin). A never-read durable is now an orphan
whatever its twin; one read once and lagging stays behind. The pass judges both
twins of every heard entry wherever no pass composes them, also on a stream where
only such durables are held.
Builds: design 47 H2 H-A-472 (eleventh addendum)
Review round 1 found that with no `_tools` twin composed, the behind and orphan
checks never looked at one: live, five twins on EVENTS held 1,000 to 58,000
messages and none was named. The pass now reads, on the streams each consuming
module's heard entries name, every held `<consumer>_tools` and judges it as any
consumer: behind by the bound, or an orphan when no pull waits and its twin is
current. The stale comments on HasProgram and the orphan finding's text follow.
Builds: design 47 H2 H-A-472 (tenth addendum)
The grants part stays the one composer of filters while the tool runner makes its
own `_tools` twin: each heard entry, EVENTS included, carries the twin's filters
(the again filter named for it), ack wait, max deliver, max ack pending and
deliver policy. The pass composes no `_tools` consumer any longer, so it makes no
twin nothing reads; the reader grants add the create and delete of the module's
own twin by name. A membership of a module no longer assigned on its node is
withdrawn by a whole pass, so the runner there removes that module's twins. The
part's own hearing loop leaves EVENTS to the hub, now that the membership names it.
Builds: design 47 H2 H-A-472 (fifth, sixth, seventh, eighth and ninth addenda)
A revision's kept grants carry the digest of the inputs they were
composed with; an ask whose inputs differ composes again, so grants-for
never lists a change the candidate did not cause. TestGrantsForKeepsWhatItMay
now proves that a revision dropped beyond the bound has no grants kept
either (review round 1); the Logf comment says the tools program's
journal too.
Builds: design 47 H2 H-A-478 second addendum
grants-for composed every module twice per ask (3 to 4 s against the
runner's 5 s bound, cut three times on 2026-11-22). The tools program
that serves it now keeps each kept revision's composed grants beside the
revision (the newest's and the candidate's), bounded by
grants-for-revisions-per-branch, so a second ask for a kept revision
composes nothing; it reads its inputs at each ask and journals each
stage's time on its standard error.
Builds: design 47 H2 H-A-478, H-A-478 addendum
Review round 1: the day after which a twin is no longer current was a
constant here and again in the bus module. It is now the setting
twin-current-hours of the manifest (24 by default), written into a file
the part reads at its start, the field name the bus module's manifest
names too, so the operator sets it once per module alike; the
durable-orphaned condition names the bound. hasOwnProcess no longer sits
under ownServesOf's doc comment.
Builds: design 47 H2 H-A-472 (2) and its third addendum (2); review round 1 findings (2), (3)
The pass made <node>_<module> beside <node>_<module>_tools for every
consuming module, so a module whose events only the tool runner reads
got a durable nobody pulls, which piled up and came back after each
removal. The program's consumer is now composed only for a module with
an account of its own and a process resource with no bus of its own;
the tool runner's twin stays for every consuming module.
Builds: design 47 H2 H-A-472 addendum and second addendum, G A-397
About 110 durables left by modules that moved into or out of the tool
runner pile up behind their stream's head and raised 'Messages pile up
for a listener' for no real backlog. A durable with no pulls waiting
whose twin on the stream (<node>_<module> and <node>_<module>_tools) has
pulls waiting and delivered within the day is now left out of that check
and named once in mesh-grants.durable-orphaned, cleared when gone.
Builds: design 47 H2 H-A-472 (2), G A-340
"Main's workers are held already" held only while the controller made
them: the Node build module's install queue, added since, never had its
plain worker, so the module pulled nothing. The account consumers now
name each queue's plain worker beside its named one; a held one is left
as it is.
Builds: design 47 H H-A-468 (third addendum)
The live registration history is held with `….registration.*`; the bus
refuses adding `….registration.>` beside it (JetStream 10052, overlapping
subjects), so df5a8895's subject never reached the live stream. A held
subject a composed one covers is now replaced by it, the messages kept and
no other subject dropped, and the assertion says the replacement.
Builds: design 47 H H-A-470 (third addendum)
The builder binds the seat's worker when the bus says it was created and again after a deletion, but no holder was
granted those advisories, so under scoped grants it would never hear them. Each holder of a seat with an accept not
pulled by kind is now granted the subscribe of CREATED and DELETED for that seat's worker. The worker the grants part
creates is SeatQueue's as returned, every shared setting copied, and a test holds it equal to SeatQueue's.
Builds: design 47 H2 H-A-468 (1), addendum, second addendum
The builder binds SEAT_MESH_BUILDS_worker and waits for the broker's part to make it, but the grants part made shared
workers only for seats pulled by kind, so no part made it and the seat's asks piled up unread. AskStreamsOf now keeps
the worker SeatQueue returns for every seat with an accept that is not pulled by kind; the grants principal carries
them, so AccountConsumers grants their create and the consumers pass creates them where the bus lacks them, leaving a
held one as it is and deleting none.
Builds: design 47 H2 H-A-468 (1)
mesh-builds' and mesh-wants' accepted kind step names max-age 24h, max-msgs
10,000, ack-wait 60s, in-progress 20s, max-ack-pending 1,000, max-deliveries 5
and a duplicates window of 2m; the grants part makes each seat's ask stream
through mesh-sdk's buscompose.SeatQueue and says a seat whose settings it
refuses, and its own copy of the derivation goes.
Builds: design 47 H2 §777; H-A-350 DMC-1, H-A-362
The node enroller's and the watchdogs' generated decoders answer the form
they read (mesh-contracts K1); their readers here take it.
Builds: design 47 H2 §780
The kept catalog revisions had no design line and their bound stood in
the code. They are now bounded per branch by the module's setting
grants-for-revisions-per-branch (default 4, oldest dropped, a run's
candidate counted within its branch's), read at the start from the file
the manifest writes, and kept only for the watched branches the part's
branch follower follows: a branch that leaves them is freed at the next
grants-for or catalog-revision event. A kept revision is a cache of the
catalog seat's own truth, never a state of the mesh.
Builds: design 47 G §441, §401; H H-A-195, H-A-193
A grants-for call still read the first page of each revision (up to
768 KB) only to learn its number. The grants holder already hears the
seat mesh-catalog's catalog-revision event, which names each branch's
newest trunk revision and the source revision it was taken in from; the
part now records both from the event and asks the revision verb only for
a revision it does not hold, by its number. Before the first event after
a start the first ask's answer names the newest once. An event below the
newest held is the numbering started again: what is kept of the branch
is dropped.
Builds: design 47 G §441, §401; H H-A-193
grants-for read two whole catalog revisions page by page on every call
(about 1.4 MB each) and answered in 3.4 to 5.0 s, so the pipeline
runner's 5 s read bound cut it and every widening touching the vault
said "could not run". A revision never changes, so the part now keeps
each revision it read whole, by branch, kind, group, heads and number,
the newest four per branch for the process's life (as the runner keeps
its revisionCache); once a first page names a kept revision the pages
after it are not read. The branch's newest is read on by the number its
first page names, and a revision whose pages name another is not kept.
Quick fix (not chosen): raise the runner's widens read bound, a catalog
merge that hides the cost. Thorough fix (chosen): the cache here. The
compose of every account twice is kept whole: narrowing it to the
candidate's modules would stop a refused principal elsewhere from
failing grants-for, which is not obviously safe.
Builds: design 47 G §441, §401; H §774
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
The module mesh-delivery-checker goes whole with its manifest, and its seat's
definition leaves the seat list. The status view reads no tries: build_usage
answers no tries running and counts no step as a try. The grants part grants
main's accept subject to the builder alone. The tests name no checker as an
asker or a seat. The seat's generated packages, the indexer's readers and the
assigner's verdict constants go in the join's cleanup, once those branches
stand on this one (the lead, 2026-11-20).
Builds: design 47 H §764 (the checker (H1): mesh-delivery-checker whole); lead answer H-A-57 (hq f599b761)
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
Every account may call every verb: the grants part reads no definition's
callers (callerOf, the verb and accept callers, admits) and grants every
module's account every verb of every seat (mesh.seat.*.tool.> and its
caller-named form), a sent store to every module invoking a verb of its
claimant's seat, and every queue's named ask to every asker. grants_callers
goes; grants_why_refused answers the bus's own verdict. The indexer keeps
no verdicts: their first view, events, invoke and the engine's read go, and
a want names its build alone (no newest). internal/... and cmd/... tests are
green; the root package's grants_test.go still reads the verdicts (rows at
grants_test.go:303-307, 741, 1184, 3700-3870, 4560-4615) and does not compile.
Builds: design 47 H §745, §750, §752, §764 (every seat contract: the callers lists); lead answers H-A-37, H-A-66, H-A-79, H-A-113 (hq f599b761)
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
ConsumerFor drew its filters from PermissionsFor, which since every account may subscribe to every subject (e1476cfd) is ">": no filter
named an event, so no module got an account consumer on EVENTS. The scoped permissions give them again. The integration subtests that
read each account's own subjects run with the wide grant off, as the unit tests do.
Builds: design 47 G A-434 (G-22-lead-answers-2.md: every account may do everything, the mechanisms stay whole), the grants part's account consumers (ADR 0561 §6)
GrantWaits compared permissions by their text, so a composed subject or allow-responses counted as
missing though the loaded list's '>' already allowed it: all 182 grants waited on '1 permissions'
from 15:43 on 2026-11-16 and no want moved. A subject a loaded one covers is carried now, and
answering is carried where the loaded user may publish every subject.
A grant held (kept or waiting) or left out was read only from the pass object, which no tool opens,
and the assigner's sixth rule waited on it in silence: no grant for any module was written after
15:43 on 2026-11-16 and nothing said why. Each pass now says its grants: wanted, written, deleted,
and every kept, waiting and missing one with its reason.
The list giving every account every subject was held at 23:20:32: the guard reads each seat's holders from the text of
the subjects users subscribe to, and > names no seat, so it saw every seat lose its holder.
The operator, 2026-11-15 23:12 CET: the mesh accounts should have all the permissions, otherwise it becomes a mess;
the operator can do everything, always. Permissions derived per module refused the indexer's chunks, seat asks and
seat verbs again and again while G is delivered. The approval mark's and the desk's denies stay; an enrolling node
keeps its few subjects. The research after G replaces the accounts themselves (mesh, agent, operator).
The validation of mesh-delivery #219 failed twice: the indexer answered seat:mesh-catalog.input, and the bus refused
its chunks, since only an account whose process repeats the verb in serves was granted to send them. The indexer's
and the assigner's programs answer their seat's input verb on the module's own account; their claims say they serve it.
The validation of mesh-delivery #219 failed: the check program on shanks asked seat:mesh-catalog.input, whose answer comes
in chunks, and the bus delivered none, since only an account naming the verb in invokes was granted its chunk subjects.
Inside the mesh any module may call any seat verb (the operator, 2026-11-15).
The bus refused SEAT_MESH_BUILDS's and SEAT_MESH_WANTS's create: the part's account is granted each stream it creates by
name, and the ask streams were not named. They join the declared streams, as the kept-event streams do.
The pipeline runner's step ask to mesh-builds timed out: it publishes on the seat's accept subject through JetStream and
no stream held mesh.seat.mesh-builds.accept.>. Before G the controller made SEAT_<SEAT> for every seat that accepts work;
the seats whose queues came with G (mesh-builds, mesh-wants) had none. The grants part now makes each such stream, from
the seat list, at every pass, as it makes the seats' kept-event streams: only where the bus holds none.
The bus refused the indexer's validation ask and the pipeline runner's step asks: a named ask was granted only from the
seat definition's callers and a composed manifest, and both accounts were kept from the list in force without them. The
operator, 2026-11-15 about 18:10 CET: no fine-grained permissions inside the mesh; every account of a module, its
processes, its node and the tool runner is granted mesh.seat.*.accept.>, composed or kept. A seat's events and verbs keep
their own rules until the research after G decides the whole shape.
The user list in force on novox stayed at 14:53 UTC while the shadow list changed, and the pass said nothing of why;
each refused create of a state's object store waited out its bound, one after another. Recovery mode (lead-47).
A state of the field state had no <state>-objects store: the grants part made and granted object buckets from
the manifest's objects alone, so the indexer's state inputs (A-237) had no inputs-objects, and every validation
failed on a refused STREAM.INFO of OBJ_mesh-delivery-indexer_inputs-objects; the assigner's inputs likewise.
objectsOf now gives each such state its store (its size, 16 MiB when it names none; none for a state only read),
so the grants part creates it and grants it to its owner as the module's own bucket.
And a state asked in requires by the template state: the bus module's process may now make its store too (the
grants part's BusStateStreams names it, as mesh-sdk's does), and a kept binding's objects is among its streams.
Builds: design 47 G §58, §607, A-237
Live on shanks the build runtimes' seat list and catalog-half reads were refused on their caller-named subjects, since the modules' manifests do not invoke them, and the no-caller fallback is answered by nobody.
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
Holders subscribe and callers publish on mesh.seat.<seat>.tool.<verb> (and its call form) beside the machine-named subject, so a verb asked of no machine is answered by any one holder in a queue group.
Builds: lead answer A-393 (design 47 G)
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
Covers busstate.go keeperConsumer for the seat definition subject catalog-revision.<branch>: restoring the old body fails the table test.
Builds: design 47 G (seat definition subjects with parts)
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
mesh-catalog's seat.json gives catalog-revision the subject catalog-revision.<branch> (likewise source-moved.<branch>,
group-ended.<branch>.<group>, owed.<repository>.<number>). watchedEvents widened only a node seat the node-engine holds
(machineEvent), so the consumer and subscribe grant of a module consuming mesh-catalog.catalog-revision (node-tools,
builder, teller, pipeline) stayed on the bare subject and never took the branch-form event. One rule now
(watchedEventSubject, used by the grants and the bus module's consumer): each part is *. A held consumer gains the
filter through AddFilters (grants/consumers.go).
Builds: design 47 G (seat definition subjects with parts)
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h