mesh-host #219's check failed: the request kept for g14 was 300127 bytes compressed, over the 262144 the state takes.
The key of the check and node now holds a pointer (digest, size, pieces) and each piece of the gzip bytes is a value of
its own under pieces/<digest>/<i>, at most 128 KiB; checker_composition_requests follows the pointer and checks the digest.
Quick fix: a larger cap or a better compression: still a limit, met again as the mesh grows. Thorough fix (chosen, and
bounded only by the bucket): pieces. Not an object store: the checker is a foundation part, whose data stays in the old
sections (ADR 0619 §2), and the tool runner serves no old-section object read to its tools. No manifest change.
Claude-Session: https://claude.ai/code/session_01NcbgfFfQ4Z9UE5jK32PmE5
mesh/check failed (6): integration: mesh-delivery-indexer/cmd/mesh-delivery-indexer.TestIndexerWritesModule…; and 5 more | 3a63813b3d75 on fb69c8dee691
An indexer that started while the bus's storage was down read no published package,
took every import for unpublished and wrote the version line over ~130 modules'
fingerprints, which a later pass kept through the build's shortcut. A failed or empty
read of the packages state, where held modules were computed from published ones,
now fails the pass and says why; a version line clears the import form it was
computed beside; indexer_fingerprints_again computes the import form again by hand,
acting first and recording the hand act after, with a bound.
Claude-Session: https://claude.ai/code/session_01NcbgfFSnn24M3GcqttoU7i
The verbs ask mesh-delivery-indexer.modules and .manifests through the tool runner
since the asks moved off the indexer's state, but the test's stand-in answered every
ask {ok:true}, which is no answer: propose failed on main with 'unexpected end of
JSON input'.
Claude-Session: https://claude.ai/code/session_01NcbgfFSnn24M3GcqttoU7i
mesh/check checking a78c5700fdf4 on fb69c8dee691 | novox/mesh-delivery#188 in group fix/1966-requires-state
The grants part wrote a binding into a membership on the bus module's
created, but granted its subjects to no account but the ledger's, so
distribution's program and the tool runner carrying it were refused their
own key-value, object store and stream entries (issue 1966, ADR 0595 §1,
design 25 § 12.2). The kept bindings now reach Declared in the same pass that
writes them into the membership, and one function, boundOf, grants each to
the module's own account and the tool runner, as its old section is granted;
a process's bus.state names a key-value entry as an old state. The composed
membership carries the bindings, so grants_membership's two sides agree.
Claude-Session: https://claude.ai/code/session_01NcbgfFSnn24M3GcqttoU7i
mesh/check failed (2): integration: mesh-delivery-assigner/cmd/mesh-delivery-assigner.TestAProposalIsAnswer…; and 1 more | af3b678a92f0 on 042b8d1fc160
holdLost builds a pending take-in naming neither WaitsFor nor Unknown, which TakeIn's check refused, so the hold stayed
in the journal and the watchman never saw it.
Quick fix: set Unknown on holdLost's event. Thorough fix (chosen): the check accepts a pending revision that names Lost,
the field made for it; one naming nothing is still refused. Tests: a row accepting Lost, the refusal row kept, and the
lost-hold scenario encodes its said event.
Claude-Session: https://claude.ai/code/session_01NcbgfFSnn24M3GcqttoU7i
Quick fix: none smaller than this (the bridge fix alone leaves a covered merge needing the bridge).
Thorough fix (chosen): loadNewest asks takenAlready (already used for group members) before loading the newest merge of
a repository; at or behind taken it is marked Covered, said once in the log, and never read. A merge whose ancestry
cannot be asked is read as before. Test: a row pair behind/ahead of taken with a bridge that cannot answer.
Claude-Session: https://claude.ai/code/session_01NcbgfFSnn24M3GcqttoU7i
Both ForgeMade and Tags took the tool's wrapper for the pull request / the tag list, so every forge-made merge read
"#0 is not merged" and bridge_tags failed, leaving builds unnumbered.
Quick fix: unwrap "result" in forgeAPI (the one Go reader of gitea_api).
Thorough fix (chosen, the kind was met twice and hid behind a drifted fake): bridge.API answers what the tool answers,
forgeBody is the one reader of that shape and refuses an answer without a result, naming it; ForgeMade refuses number
0 as "no pull request"; the unit fakes answer the tool's real wrapped shape, with rows for the bare body.
Claude-Session: https://claude.ai/code/session_01NcbgfFSnn24M3GcqttoU7i
The operator, 2026-11-09 09:07: raise only. The grants part's assert raises it, says it in its asserted event and the
journal with the old and new cap, and grants_assert asks it by hand.
Claude-Session: https://claude.ai/code/session_01NcbgfFSnn24M3GcqttoU7i
The log grew about 3 MiB an hour on 2026-11-09 (951.9 MiB at 07:47 UTC, 954.9 at 08:42)
against the bus's default cap of 1 GiB, about 23 hours before builds stop. ADR 0571 §1
keeps it 30 days, so the cap rises to 3072 MiB instead of the age falling.
Claude-Session: https://claude.ai/code/session_01NcbgfFSnn24M3GcqttoU7i
mesh/check checking 173a7654ba61 on e39ec444bd39 | novox/mesh-delivery#184 in group fix/1966-views-durable
A consumer is made from now, so an event said between a first view and the consumer is never delivered; the hub says when the bus made it (EventHub.OnBound) and each follower asks again if behind (Follower.ConsumerMade).
Claude-Session: https://claude.ai/code/session_01NcbgfFSnn24M3GcqttoU7i
Neither runtime of the assigner gave catalogue a seat list, so propose judged
every mesh seat as held once (mesh-dns-resolver, replicated, refused every
proposal) and seat-holders marked every seat outside the list. Both now follow
the list as the grants part does (ADR 0636 §1, §4, §7): a proposal laid again
is not ended and seat-holders is not written until the list is read, and the
two seat list tools read and read it again.
Claude-Session: https://claude.ai/code/session_01NcbgfFfQ4Z9UE5jK32PmE5
A mesh seat held on two nodes was refused as held once whenever catalogue did
not name it, and a host that has not read the seat list holds only the
bootstrapped definitions: mesh-dns-resolver, a replicated seat, refused every
proposal (ADR 0636 §1).
Claude-Session: https://claude.ai/code/session_01NcbgfFfQ4Z9UE5jK32PmE5
The markers and the values of the lists taken away before the purge still fill
the state user-list. Before its first pass the process purges every
sealed-grants key whose last operation is a delete or purge marker, and every
user's key of a list the index of kept lists does not name; never a kept
list's key nor another key of the state, and nothing when the index cannot be
read. It logs the counts and the state's bytes before and after.
Claude-Session: https://claude.ai/code/session_01NcbgfFfQ4Z9UE5jK32PmE5
A key-value delete only adds a marker: every value of the per-user keys of the
lists older than the two kept stayed in the state user-list's stream, which
reached 94% of its 64 MiB cap on 2026-11-08 (5,968 keys, 13,228 values). At the
cap the bus refuses grants' next write of the sealed list and deliveries lock.
The part now takes those keys away through a purger on the process's own
JetStream connection, a publish to the key's own subject the account already
writes on; the tools, with no purger, still delete.
Claude-Session: https://claude.ai/code/session_01NcbgfFfQ4Z9UE5jK32PmE5
With exactly one filter in FilterSubject, nats.go creates on $JS.API.CONSUMER.CREATE.<stream>.<name>.<filter>,
a subject the grants part is never granted (CreateConsumerGrants keeps filters out of grants), so the creates
of ASKS.novox_mesh-delivery-indexer, ASKS.shanks_mesh-delivery-builder and FORGE.shanks_mesh-delivery-builder
were refused and timed out (2026-11-08, issue 1966). Every filter now goes in FilterSubjects, which the
client creates on the plain subject.
Claude-Session: https://claude.ai/code/session_01NcbgfFfQ4Z9UE5jK32PmE5
Since 16:39 the running assigner cannot answer its wants page within a call's
bound, and the failed wants read held the whole user list, so no new build got
its credential. As the assignments view already does, an unanswered wants view
now composes what runs from each have's fingerprint, is said as a gap that is no
failed read, and keeps the pass to one that only adds (keepsLoaded), so the list
is held only where it would take from a loaded user.
Claude-Session: https://claude.ai/code/session_01NcbgfFfQ4Z9UE5jK32PmE5
The channel refuses a raised condition whose explanation does not open with
"Needs you: " and its needs, so seven of grants' conditions, among them
first-view-unanswered, never reached the operator. One helper now writes both,
and every condition the unit tests say passes the channel's own check.
Claude-Session: https://claude.ai/code/session_01NcbgfFfQ4Z9UE5jK32PmE5
The keeper on novox (nats fea2c02b) has been refused $JS.API.INFO every 30 s since
21:58 UTC: the client library's CreateKeyValue asks the account's information before
it creates a bucket, so each create of KV_distribution_* and KV_gitea_* timed out and
the keeper never finished a fill pass (health failed). Read-only: the account's limits
and usage, never a stream's data.
Claude-Session: https://claude.ai/code/session_01NcbgfFfQ4Z9UE5jK32PmE5
mesh/check failed (2): integration: mesh-delivery-assigner/cmd/mesh-delivery-assigner.TestAProposalIsAnswer…; and 1 more | 5452fa92a731 on f9cadbfdbb93
A running build older than the trunk lost reads it still uses (the vault's read of the assigner's placement), failing every node-engine pass. Memberships now come from the union of the have's build, the want's and the trunk's manifest.
Claude-Session: https://claude.ai/code/session_01NcbgfFfQ4Z9UE5jK32PmE5
mesh/check failed (3): mesh-build-go: the go.mod of novox/mesh-sdk at 396511b7a276, beside the change, cann…; and 2 more | 0a82e3351804 on 9c6bf236b6b1
The second instance of the kind: checker_status and checker_pull read the indexer's
candidates and modules from its bus state, which the checker's manifest does not
declare under reads, so the runtime refuses them ("keeps and reads no state
mesh-delivery-indexer.candidates"). They now ask indexer_read_candidates and
indexer_read_modules, the indexer's read tools the manifest already invokes: the
keys with no argument, one key's value with key (ADR 0596). The assigner's put
read its newest fingerprint from the same refused state and now asks the
indexer's first view of modules, as the seat verbs do. A sweep of every module
finds no other read of another module's state that its manifest does not declare;
the checker's tools carry a test that refuses one.
Claude-Session: https://claude.ai/code/session_01NcbgfFfQ4Z9UE5jK32PmE5
mesh-module-assigner.propose and mesh-wants' upgrade and bus read the indexer's
modules and manifests straight from its bus state, which the runtime refuses
since those reads moved to the indexer's first-view tools (ADR 0615 §1, ADR 0621):
"mesh-delivery-assigner keeps and reads no state called mesh-delivery-indexer.modules".
Another module's data is asked through its tools, never read from its state
(ADR 0596). The verbs now ask mesh-delivery-indexer.modules and .manifests through
the tool runner, every page, at each call; the manifest already invokes both and
the indexer answers them to the assigner.
Claude-Session: https://claude.ai/code/session_01NcbgfFfQ4Z9UE5jK32PmE5
mesh/check failed (3): mesh-build-go: the go.mod of novox/mesh-sdk at 396511b7a276, beside the change, cann…; and 2 more | 00232c71977b on 9c6bf236b6b1
Every recompute wrote every key of its modules state again with identical bytes: recomputeLocked
compared the package versions with reflect.DeepEqual across two map types (the contract's
map[string]string and sdkstep.PackageVersions), which never match. Each write was a views
event, and a reader of the views took each as a gap and asked the whole first view again.
The comparison is by maps.Equal now, and write itself skips an answer whose bytes are the
bytes held, for every caller and every state it writes; the merge's written list counts only
real writes.
Claude-Session: https://claude.ai/code/session_01NcbgfFfQ4Z9UE5jK32PmE5
mesh/check failed (2): mesh-delivery-checker: no published of the tagged version of go-sdk of mesh-delivery…; and 1 more | ee6fd92b4d1a on 7f412185525e
Every reader of the seat list holds the bootstrapped seats from its start
(ADR 0636 §2), so a manifest naming one is read by every node-engine
whatever its statement names. graph.MissingReads, in the Delivery domain's
SDK, is the one rule both holds judge with; mesh-sdk is not changed.
ace's engine states no seat, so the want of the node-engine's own build
0bd8cbad (needing mesh-catalog) and of restic (needing node-backup) were
held on ace: the engine that would state them could not be delivered. The
indexer's hold judges only mesh-wide reads today, so its verdicts do not
change; it takes the rule so the two stay one.
Claude-Session: https://claude.ai/code/session_01NcbgfFfQ4Z9UE5jK32PmE5
The grants part built the node-engine's declaration with its consumes alone,
so a node's account was granted none of the engine manifest's invokes: ace's
engine could not publish mesh.seat.mesh-catalog.call.seat-definitions.node~ace,
read no seat list and composed nothing new. And its consumes of
mesh-catalog's seat-definition events were granted on the emitter's subject
only, where the engine subscribes the seat's (ADR 0636 §4).
The engine's declaration now carries its invokes and the seats it watches,
split by the same consumedOf a placed module's consumes are; the node's
account is granted the invokes as a module's are (in its own caller token
by callerNamed) and the watched events on each seat's subject by the same
watchedEvents a module's use. Its consumes stay granted on the emitter's
subjects too, so nothing granted before is taken away.
Claude-Session: https://claude.ai/code/session_01NcbgfFfQ4Z9UE5jK32PmE5
mesh/check failed: mesh-build-go: the change's manifests cannot be judged: the seat list is not read (the seat list is … | 2e9f23f34828 on 68581bed7e36
The review of 0f5c186b: counting each record's packages made 94 sub-package
paths, commands among them, known names, which design 32 forbids; the seven
names refused live are each their record's module (issue 1966).
Claude-Session: https://claude.ai/code/session_01NcbgfFfQ4Z9UE5jK32PmE5
The indexer's judge refused manifests whose `packages` name a Go package that no
manifest declares as a package artifact: such a package has no series key, and
knownPackages took series keys only, though each module the trunk published has
a per-module record in the packages state naming its Go module and packages.
knownPackages (and the checker's PackageNames) now also count each record's
module and packages[].package; an unreadable state still judges no name.
Quick vs thorough: this is the quick fix, the smallest change that lets the
trunk's own records name what the manifests name. The thorough fix is that a
build module builds Go package artifacts, so every Go package gets a series
(the registry milestone); then the records stop being a source of names. Done
now because the 15 refused manifests hold the bootstrap of the registry delivery.
Fault: the new assigner composes a membership that drops the placement seats and adds its state view-marks. The partial pass
kept the published one whole, so the assigner's tools never got view-marks, its assignments view never answered, and the
pass never left its partial state (third bootstrap loop of the kind after #170 and #171).
Kind: the bootstrap hold freezes whole entries, so a grant needed to end the hold is never given.
Quick fix: special-case the assigner (or let additions through for it). Rejected: the next module that gains a state meets the same loop.
Thorough fix: while a pass keeps what the bus loaded, publish the union of the membership last published and the one composed
(every list, deduplicated, published entries first; the private network, contributions and bindings the pass's), published
when it differs; the pass says what it kept beyond the composed one. The user list already unions (KeepLoaded), additions pass there.
Claude-Session: https://claude.ai/code/session_01NcbgfFfQ4Z9UE5jK32PmE5
Review round 1 of issue 1966's keep-what-is-loaded fix.
- KeepLoaded intersected the denies, and the composer gives every composed user an empty
deny, so a user in force with a wide publish and a deny was sealed with no deny. The
union of every deny any list says is kept now, and Losses counts a deny dropped.
- A pass composing part of the mesh (the seat list unread, or placed by the haves) no
longer narrows a published membership: one whose published value serves, hears or gives
more than the pass composes is kept as published and said so (not_published), so a tool
runner or a seat holder goes on serving what it serves while keeping its login.
Claude-Session: https://claude.ai/code/session_01NcbgfFfQ4Z9UE5jK32PmE5
Issue 1966: the list sealed 2026-11-07 23:09, composed while the seat list was unread
(only the bootstrapped seats compose), took the node seats from the tool runners, the
vault's reads from the node-engines and the seat queues from the build modules, and left
out ledger.novox.mesh-knowledge.
- While the seat list is unread (the gate's MayAct with no seat) or a pass is placed by
the haves, each user of the loaded list and of the list in force is sealed with the
union of what the pass composes and what those lists give it (publish, subscribe,
responses; a deny only where every list saying one denies it), kept the way a refused
principal is; the users so kept are the pass's kept_loaded.
- In such a pass a list that still takes a permission from, or leaves out, a loaded user
is held (user-list-held), naming each user and how many permissions it loses.
- A held module keeps its ledger account, ledger.<node>.<module>, with its other users:
heldModuleUsers matched <node>.<module>[.<process>] only, which dropped the ledger user.
Claude-Session: https://claude.ai/code/session_01NcbgfFfQ4Z9UE5jK32PmE5
Review round 1 of issue 1966's fix: the membership check alone did not pin the removed
phase; the row now reads the inputs and requires mod-y off b while its have is removed.
Claude-Session: https://claude.ai/code/session_01NcbgfFfQ4Z9UE5jK32PmE5
Issue 1966: the seat list is read through mesh-catalog's seat-definitions, which got no
responders only because nothing was placed. A row shows that, the assignments view
unanswered, the indexer placed by its have (its real manifest) is composed to hear
mesh.seat.mesh-catalog.* and its membership serving seat-definitions is published, while
the seat list stays unread and other modules still wait for it.
Claude-Session: https://claude.ai/code/session_01NcbgfFfQ4Z9UE5jK32PmE5
Bootstrap only (issue 1966, notes 63 and 64; the operator chose it 2026-11-07 22:19,
bending ADR 0567 §9 for the bootstrap). An assigner that serves no first views answers
no assignments, so the grants part placed nothing and wrote no grant, membership or
bucket for anything, not even for the assigner that would serve the view.
When the assignments read fails and the haves read, each module is placed on the node
whose have says it runs (any phase but removed), the bus node found among them, and what
runs composed from each have's fingerprint instead of the assigner's wants. Said as a gap
naming the fallback, and kept as Inputs.ByHaves: such a pass retires no account and deletes
no withheld hearing (ReadFailures), and seals no list that drops a user of the list in
force or of the loaded list (holdsBack). Once the view answers, the fallback is unused.
Quick fix; the thorough one stays the module-graph edge (the assigner's views before the
grants part that reads them) in the delivery plan.
Claude-Session: https://claude.ai/code/session_01NcbgfFfQ4Z9UE5jK32PmE5
mesh/check pass 71a56d4606c9 on b73315ffe05c | carried from a37ee99b41d5 | unit 495 run, 947 kept | integration 494 run, 1014 kept | composition passe…
Both were published with the JetStream acknowledgement awaited, a request and response, for ten seconds each, and
manifest-accepted once more on failure. On a bus refusing the subject (the old grants do not grant the new indexer
mesh.seat.mesh-catalog.event.manifest-accepted) no acknowledgement ever came, so each judging pass took about twenty
seconds a manifest and the indexer never reached the forge's messages nor the seat definitions: the take-in stood
still. Now both are published plainly on their subject with the dedup id in Nats-Msg-Id, as an event is (the
operator, 2026-11-07: everything is event-driven); the stream that keeps the subject drops a duplicate by its id, and a
refusal is said on the connection's error handler, as the journal already shows. An error of the publish itself (no
connection) still goes to the caller, which retries once and writes the unsaid finding.
Quick vs thorough: thorough for these two events (every publish of a seat event by the indexer is now an event);
sayTakeIn keeps its acknowledged publish, said in the issue for 47/G.
Claude-Session: https://claude.ai/code/session_01NcbgfFSnn24M3GcqttoU7i
mesh/check pass e01d265decb3 on b73315ffe05c | carried from a37ee99b41d5 | unit 495 run, 947 kept | integration 494 run, 1014 kept | over bound: integ…
The running indexer (built at the registry merge) publishes every Go module of a repository it published before in
one forge listing; seatfiles, published at 6bc232eb and removed since, is not in the tree, so the listing fails and
every take-in of mesh-delivery is 'computed and not written' (journal 19:56:29: 'the directory
modules/mesh-delivery-indexer/seatfiles is not in novox/mesh-delivery'). mesh-delivery #165 fixed that in the new
indexer, which cannot run yet. The module comes back as it stood at 6bc232eb, imported by nothing, until the new
indexer runs; the follow-up removes it. The operator's bootstrap of 2026-11-07.
Claude-Session: https://claude.ai/code/session_01NcbgfFSnn24M3GcqttoU7i
mesh/check checking e0f510facb17 on b73315ffe05c | novox/mesh-delivery#166 in group fix/1966-bootstrap-no-packages
The running builder predates Go package artifacts ("refused kind: package of language go ... no build module
builds it"), so it can build neither the new builder, the grants module, the Go build module nor the checker:
each new part is needed before the others can be built or delivered. For the bootstrap of the registry
delivery, at the operator's word (19:4x, "do the temporary change"), the package artifacts leave the manifests of
the parts that run; the contracts and SDK modules, which run on no node, keep theirs. They come back in one change
once the new builder runs. Quick vs thorough: quick; the thorough fix is 47/G's tier order (issue 1921 note 80).
Claude-Session: https://claude.ai/code/session_01NcbgfFSnn24M3GcqttoU7i
publish asked the forge for every published module of the repository in one call; when the merge removed one (seatfiles), the whole call failed and a revision's own new Go modules stayed unpublished, so the take-in waited for a version only its own take-in would publish. packagesAt now reads the modules one by one when the tree lacks a directory and leaves the missing one out, said.
mesh/check checking 4735fc0b5fdb on 8ad2bf6951d5 | novox/mesh-delivery#151 in group feat/registry-milestone
Conflicts in the checker's judge.go, builds.go, composition_test.go and try_candidates.go:
the group's contract types and kept composition path stay, #164's behaviour is taken whole.
WaitsOnPreview reads the contracts' meshcatalogclient.NoneNoPreview instead of a copied
constant; awaitBuilds names its failures failedBuilds, beside the builds package.
Claude-Session: https://claude.ai/code/session_01NcbgfFSnn24M3GcqttoU7i
jschoubben
mesh-tools