The seat verb try gains members (ADR 0637 section 6); the try's group name and the missing-preview reason are one definition of mesh-delivery-sdk (ADR 0637 section 8).
Claude-Session: https://claude.ai/code/session_01NcbgfFSnn24M3GcqttoU7i
mesh/check being carried to c78941e80b59 on 4596fbe1b7ac | from ae31b66a9e33, mesh-catalog fd4931a75261 | novox/mesh-delivery#153 in group feat/take-in
The group's check failed at 7b7096f3: the group's own manifest validator refuses the indexer's stream "take-in", outside
the bus's predefined forms. The lead's ruling: the indexer says take-in.pending and take-in.taken as its ordinary events
(its emits, mesh.mod.<module>.event.take-in.<kind>), as manifest-accepted is said, and keeps no stream. The wait's key is
in the body only: an emitted event names one event (no wildcard), and a key such as g.<group> is no event name. The
watchman consumes the two events by name on the shared events stream; it needs no replay, since it seeds its waits from
indexer_pending at its start. indexer_pending says the last event of each wait as take-in.<kind>.
Run locally, as the check runs it: cmd/mesh-judge over every manifest of mesh-catalog, mesh-delivery, mesh-host,
mesh-tools, mesh-networking and mesh-media-catalog, with this branch's and mesh-catalog's feat/take-in manifests in the
change case and origin/main's in the trunk case: 171 of 171 accepted in both. The same run with the indexer's manifest at
7b7096f3 refuses it with the check's words.
Claude-Session: https://claude.ai/code/session_01JG73KG7AfSdh6FP8VrBqCn
mesh/check failed (2): mesh-build-go: the indexer's judge at the group's mesh-delivery head 7b7096f37ea76e1…; and 1 more | 7b7096f37ea7 on ae31b66a9e33
- A take-in releases its claims (compare-and-set, only while the claim is still its own) on a lost claim, on a claim it
cannot read back, and on a merge it fails to take in: a claim left behind refused every later take-in until a
restart. T5 row with a store that fails once, both ways.
- The start step waits for nothing: with the build it runs unknown it says so and the forge consumer starts; a merged
message left unacknowledged asks the build again (clause 1). T10 row: the start answers at once and takes nothing in.
- A toolchain report evaluates pending again (ToolchainReported), since the build a revision names moves with the
toolchain. T7 row: the revision is taken in once the report makes it name the running build.
- The group's end is told by its whole subject (process.ModuleStreamSubject), so a repository named forge/group is
never taken for one; row in the indexer's handle test.
- The watchman asks the indexer's waits through indexer_pending at its start (invokes), before its durable consumer,
which does not hear again what it acknowledged before a restart; the comment that said the whole stream is heard is
corrected. W2 row: a wait seeded within its bound is raised past it.
- bridge_groups names each group's last end ask, its who, why and when, and the bridge's answer (state end-asks).
- The check's words for the indexer's rules: the trunk's manifest validator.
- The bridge's end drops the merged-state test its merge commit already implies.
- Test budget: groupask_test folded into groups_test, and the K1 mesh/check test into the owed status one: 804 unit
tests (was 806).
Claude-Session: https://claude.ai/code/session_01JG73KG7AfSdh6FP8VrBqCn
The operator, 2026-11-04: logic one module owns lives inside that module as a package of its own, so the take-in step
changes no mesh-sdk. TakeIn, its keys and event names, PendingWait and PendingAnswer move to
modules/mesh-delivery-indexer/takein (its merges a takein.Merge of its own); Forge whole, with its group fields,
ForgeGroupEnd and GroupMerge, moves to modules/mesh-delivery-forge-bridge/forgemsg. Both are nested Go modules that their
readers take by replace, as buildrun is. Every importer moves in this commit: the indexer, the bridge, the watchman and
the view. Forge's private checks (the repository name, the version) are copied small; mesh-sdk main still holds an unused
Forge until its own fold removes it.
The indexer, bridge, watchman, view and checker pin mesh-sdk 5a38a12 (origin/main), no feat/take-in commit.
The bridge's integration bus now grants the state end-asks that bridge_group_end_again added: without it every test of
the package failed on a permissions violation. Test budget: two new test functions (TestForge, TestTakeInEvents) and the
two the join added are paid by merging four pairs in internal/bridge; the repository stays at 806 unit tests.
Claude-Session: https://claude.ai/code/session_01JG73KG7AfSdh6FP8VrBqCn
Bind runs once at a part's start, so the finding's words no longer promise it clears when the grants part makes
the consumer. The assigner's comment on firstViewBound named five seconds for views.Bound, which is two minutes.
Claude-Session: https://claude.ai/code/session_01JG73KJNrH5EgFXMKKYSrig
The indexer's, the assigner's, the builder's and the watchman's consumers that were not there or could not be
heard only said so in the log, so the part stayed healthy while it heard nothing (ADR 0615 §5). sdk/heard is
the one shared helper: the finding family-not-heard names the observer and the family, written when the
consumer does not bind and deleted when it does.
Claude-Session: https://claude.ai/code/session_01JG73KJNrH5EgFXMKKYSrig
The bus module says created once per entry, so every later accepted manifest of the module raised a false
bus-state-overdue five minutes on. The watchman keeps a mark per entry said created, on its pulse state with no
time to live and read back at a restart; a refused or an entry no longer asked drops it.
Claude-Session: https://claude.ai/code/session_01JG73KJNrH5EgFXMKKYSrig
Conflict: modules/mesh-delivery-watchman/manifest_test.go, both sides asserting m.Consumes,
and both adding a Consumes field to the test's manifest struct (kept once).
Git also auto-merged module.json into two "consumes" keys (A2g's first views and A5's
manifest-accepted and mesh-broker events): joined into one list, and the test asserts all five.
Claude-Session: https://claude.ai/code/session_01JG73KJNrH5EgFXMKKYSrig
ADR 0615 §1: the watchman invokes assigner_read_wants and indexer_read_modules, consumes their views
families and never reads either bucket; it keeps first-views and writes an unanswered first view as a
finding of its own. The lead's ruling: an owner that is down never holds the start; each state is read
on a goroutine once its first view is answered, and what it holds (seeded pulses, stuck keys, the
findings it is the source of) is reconciled then (watch.Awaits, ReadLate), never dropped at the start's
end. Tools watchman_first_views and watchman_first_view_again.
Claude-Session: https://claude.ai/code/session_01JG73KJNrH5EgFXMKKYSrig
The seat's definition (mesh-contracts meshbroker) names local-name and storage-type and puts
the events on mesh.seat.mesh-broker.event.<event>; design 25 § 12.3 keeps them on the bus
module's own stream. Both are read until that is ruled.
Claude-Session: https://claude.ai/code/session_01JG73KJNrH5EgFXMKKYSrig
It consumes the indexer's manifest-accepted and the bus module's created and refused on its
durable consumers. The five-minute bound is a row of its signals table (a pulse key whose
expiry is the finding, read back at a restart), not a new state: the prepare's state
overdue would have kept the same keys with no expiry to act on.
Claude-Session: https://claude.ai/code/session_01JG73KJNrH5EgFXMKKYSrig
The check of 2026-11-02 22:35 on a loaded shanks failed two buildrun tests on wall-clock bounds of 60 ms, the
second failure of this kind tonight (the first: mesh-sdk's busretry tests, f02f142).
Quick fix: give the two buildrun asker tests bounds of seconds.
Thorough fix (chosen, because the kind recurs across every module of the repository): every test that raced a
wall clock now waits on an event, steps a clock it is given, or has a bound only a failing test reaches.
- buildrun asker tests: the steps' own times stand for the clock (bound an hour, starts and phases written in the
past), no goroutine paces phases with sleeps.
- builder: the decline bound reads Builder.Now; the decline test steps it; the stream-wait test waits for five
check reads instead of 100 ms.
- checker: the lost-ask try bounds an untaken ask at 200 ms only where that is what ends it; the program says when a
merge turn changes, and the integration test waits on that line instead of 300 ms.
- teller: the bed makes the teller's own ends bucket (a 10 s stall at start the tests raced by 0.5 s), waits for its
warrant subscription instead of 500 ms pauses, and the package-level now is restored with t.Cleanup.
- watchman: the recount, run in its own goroutine, is waited for instead of 10 ms.
- polls of 2-5 s become 10 s (unit) and 20 s (integration).
Under a throttled cgroup (half a core shared with two busy loops) beside the node-engine's suite, the buildrun
tests failed 5 of 20 runs before and passed 30 of 30 after; no test added.
Claude-Session: https://claude.ai/code/session_01NcbgfJ9Gum47zkFKWRaovL
The watchman held a node a control node, urgent after half an hour
silent, while its have named mesh-controller; the controller ended by
its unassignment (ADR 0590), so no node was one any more and the bus
node's silence stayed a warning. The control node is now the node whose
have holds the bus module, nats, which the watchman already hears by
that name.
Claude-Session: https://claude.ai/code/session_01NcbgfJ9Gum47zkFKWRaovL
ADR 0590 supersedes the lease write: mesh-grants says no lease event any
more, so the watchman's lease-altered and removal-failed findings and
their row had nothing to hear. The ended controller's consumers are
removed as hand acts, so their deletion is no consumer lost; only the
one on DEAD_LETTER_NOTICES is still read, since its removal hands the
notices to the watchman. The builder hears artifacts-gone from the
artifact sweep seat alone, and mesh-enrolment's tokens no longer report
an issuing gate that nothing can turn off.
Claude-Session: https://claude.ai/code/session_01NcbgfJ9Gum47zkFKWRaovL
The controller ended by its unassignment, so its lease key is absent for
good and mesh-sdk reads that as not handed away: mesh-enrolment's issue
stayed off and the watchman never took DEAD_LETTER_NOTICES, whose notices
age out after a week. Both now act without reading the lease; the
watchman still waits, and says so, while the controller's consumer on the
notices stands. Neither manifest reads the lease any more.
Claude-Session: https://claude.ai/code/session_01NcbgfJ9Gum47zkFKWRaovL
- redeem writes the node into placement before it asks the vault for the account and the certificate of the serving key
presented (the vault reads the node's address from placement); a retry by the same keys still finishes (blocker 3).
- loadedVia reads mesh-grants.user-list's pass as the part's whole Result, the list state under user_list; the fixture
testdata/grants-pass.json was written by mesh-grants' own Result, not by hand (blocker 4).
- enrolment-expired is delivery.EnrolmentExpiry with IssuedID, through its codec; a key that expired before its content was
seen names no issued event, so nothing is said and its token is dropped at once (blocker 6).
- issue builds the joining node's first tunnel from placement (the hub, its endpoint, the node's address) and the hub's
record (its tunnel key) and passes it to enrolment-token; no hub, or a hub with no key or endpoint, is a finding of issue
(ADR 0587 §3, insight of 2026-11-02).
- The assigner takes an add-node only from mesh-enrolment's own bus user (stdio.Callers), never person.mesh-enrolment.
- The watchman raises machine.<node>.node-account-fallback from a node's health report's fall back and clears it when the
report no longer carries one (ADR 0583 §6, blocker 7).
- mesh-enrolment's ten unit test functions are two table tests now (blocker 8): the repository is back at 816.
(issue 774)
Claude-Session: https://claude.ai/code/session_01NcbgfJ9Gum47zkFKWRaovL
ADR 0582 clauses 4 and 6: mesh-grants says lease-altered and removal-failed "as a finding the watchman raises".
The watchman hears the seat mesh-grants' three events on core subjects (mesh-sdk WatchmanHears, its grant and
its subscriptions from one list, rather than a consume and a durable consumer), raises lease-altered naming
what changed and clears it at the next lease-handed, and raises removal-failed naming what failed and clears it
once what it names reads gone: the controller's consumer of DEAD_LETTER_NOTICES by its deletion advisory
(heard already), the stream ASSIGNMENTS by its own. The finding DEAD_LETTER_NOTICES.controller stays as it is.
Test: a row in TestTheWatchmansFindings; no new function.
(issue 774, 774/S22)
Claude-Session: https://claude.ai/code/session_01NcbgfJ9Gum47zkFKWRaovL
Issue 1585: the indexer keeps an accepted manifest accepted when a stricter judge re-judges it, its problems as
warnings carried in its value; a merge's new manifests are judged in full. The watchman raises manifest-outgrown as
a warning until mended. The checker fails only on a manifest the change adds or changes and names the rest as
warnings. Every module pins mesh-sdk 19c2042.
Claude-Session: https://claude.ai/code/session_01NcbgfJ9Gum47zkFKWRaovL
Issue 1585, S10: the watchman watches the indexer's modules and finds manifest-refused for a module whose value
carries a refusal, ending it once accepted. Its mesh-sdk pin moves to the group head, so it takes health.Finding's
Ref (issue 1952 renamed Token).
Claude-Session: https://claude.ai/code/session_01NcbgfJ9Gum47zkFKWRaovL
At a restart a standing finding is observed again by the replayed failed have: kept, not confirmed again, and its urgent
signal runs on from the finding's Since; an urgent one stays urgent with no signal restarted.
Quick: skip the confirm when a finding stands. Thorough, chosen: the same, plus urgency read from the finding itself.
Claude-Session: https://claude.ai/code/session_01NcbgfJ9Gum47zkFKWRaovL
A have that turned failed once and stayed unchanged was written once, so the old count of two writes never fired and the
finding was never written (fail2ban on novox). Confirm by time, with a signal, also at a restart.
Quick: raise the count to one write (loses the confirmation). Thorough, chosen: confirm by a time bound, as ADR 0564 s4's
signals are keys that expire, so no write count is needed.
Claude-Session: https://claude.ai/code/session_01NcbgfJ9Gum47zkFKWRaovL
A key that ages out of its bucket (the build seats' asks pass the stream's 168 h) leaves no marker a watch hears,
and one deleted while the watchman was down is not heard either: its stuck key stood unmoved for ever, and 26
stuck-key conditions of asks from 2026-10-21 were open on 10-30.
Quick fix: mark stuck keys older than the stream's maximum age moved.
Thorough fix, chosen because it covers every bucket the watchman mirrors and any way a key goes: at the start's
end, a stuck key not moved whose bucket was read whole at this start (each build seat's steps, the assigner's
wants) and whose key was not read standing is written moved, which clears its finding, and the closing is said
in the log. A bucket that could not be read closes nothing. A key that ages out while the watchman runs is
closed at its next start: a KV age-out emits no event, and no timer is added.
S1 is urgent after half an hour for a control node (the node running mesh-controller). S9 also hears the bus module's
slow consumers and closed connections, each closure counted once. S21 is one finding per heard machine from its failed
haves and its provisioner clients' reports (PROV_APPLIED). DEAD_LETTERS is counted again at a letter's age and every
five minutes while letters are held, on the watchman's own pulse. needs-operator waits for L6: its inputs are not on the
bus. Pins mesh-sdk b6ef01d.
The signals rows S1, S8, S9, S10, S11, S21 and the haves' unhealthy and needs-operator are pulses whose expiry is a finding;
stuck keys and presence are its own findings; it keeps the dead letters from the lease write on its own consumer, lists and
shows them, and refuses deliver and drop until the hand-act log's home (issue 1953). Tools read every state it keeps and
start resync. Pins mesh-sdk 2a04bf0.
jschoubben