The checker stopped reading a level's test results at the first that did
not pass and kept only its first failure, so a check named one fault per
round. It now reads and counts every result and writes every failed test
with its first line into the verdict's Failures and the try's failed
list (at most 50, then how many more); Failed keeps the first one's
words whole (issue 1901). Pins mesh-sdk fix/1921-every-failure.
mesh/check pass 769dab43a24d on 2f697216eee4 | unit 168 run, 488 kept | integration 178 run, 554 kept | composition passed on ace
The held record is the memory that a fingerprint failed (ADR 0408 section 4); a revert returns the
trunk to exactly it, and without the key rule 3 would move the canary onto the failed build with
nothing asked. A stale hold ask now clears only the condition (raised again on a revisit), and the
hold is asked again under a new asking. Also read the single want key for a record ask.
Review round 1: a record ask is about the composition of any fingerprint a node may still want
(trunk, want, have), not only the trunk's, so deleting it looped with the wants rule that asks it
again and the operator was never asked. It is now stale only when the module left the trunk or no
read composition of those fingerprints has its digest. A hold or restore ask taken away as stale
also takes its held record and clears its held condition.
An ask outlived its fingerprint: a lift left the restore asks of its fingerprint on every node,
and no pass removed an ask of a fingerprint that is no longer the module's trunk fingerprint,
so the operator kept being asked about a module that runs the trunk healthy. Each pass now
takes those asks away (the teller then withdraws the router ask), and a lift takes the restore
asks of its fingerprint. Nothing is taken away while the trunk or a composition is unread.
mesh/check pass aea22bf793b4 on 7ddbf40715ff | unit 168 run, 479 kept | integration 178 run, 545 kept | composition passed on novox
The part reads seats, buckets, logs and streams, never a config value, and
the merge gate already judged the fallbacks; reading the controller's
predefined mirror would tie the part to the leaving controller (ADR 0547
clause 1). ParseManifestForInputs takes a fallback as it declares itself.
Issue: 1945
A manifest falling back to a predefined key (claude-code, dunst) was parsed
with no list, refused silently, and its module counted as placed nowhere, so
three memberships would lose claude-code's reach addresses and were held back.
The part now reads the list the controller mirrors (mesh-controller.predefined,
key list), follows it, and lists every manifest it cannot read in its gaps,
its pass (refused) and its log.
Issue: 1945
mesh/check checking 12f16bbebd40 on 13380e3d0e52 | novox/mesh-delivery#58 in group fix/1790-grants-create-buckets
A module's new state reached the bus only when the controller started (issue 1790, met three
times: puts, queues, mesh-vault's bus-accounts). The operator allowed one more grant in the
controller's composer on 2026-10-26 08:54, creation only: the part's process now creates each
bucket, log or stream a placed module's manifest declares and the bus lacks, with the same
configuration the controller asserts (mesh-sdk BucketStreamConfig, Log.Config,
ModuleStreamConfig), on the event that brought the manifest, never on a timer. It says asserted
and logs one line per module; a stream the bus holds is never updated or deleted, its difference
only said; a create refused is said and tried at the next event; one created at this manifest
and gone since is made again only by hand. grants_assert asks the process to do the same for one
module now. The user list, the grants and any change to a held stream stay in shadow. mesh-sdk
moves to fix/1790-grants-create-buckets, whose composer grants this, so the shadow list is
composed as the controller's.
Issue: 1790
The guard that holds back a membership losing anything held back five since 01:24: the frozen
controller's memberships still serve the delivery seat's mark-ready, which mesh-sdk d506f21
retired on purpose (ADR 0421), so every membership composed with today's seat catalogue looked
like a loss. A seat verb the catalogue defines no longer is now dropped from the comparison and
named in the pass (retired); any other loss, a module's own seat's verbs included, is still
refused. The log named only the first held membership: it now says each one, with what it would
lose, whenever the set changes, and grants_membership with neither node nor module lists them.
Issue: 1945
mesh/check pass 9132b6a296b9 on 13380e3d0e52 | unit 168 run, 473 kept | integration 178 run, 539 kept | composition passed in 36ms on novox
ADR 0547 clause 9: the vault writes the hash of each bus account it holds into bus-accounts.<user>. The grants part
watches that state (no timer), carries the hash with each shadow user in memory only, and its user-list tool says per
user whether a hash is held and how its line would differ from the loaded one (same, adds, left out for want of a
credential, loses its login); never the hash, since bcrypt salts make two hashes incomparable. Take-over names the
users with no hash.
Issue: 1945
mesh/check pass 3e8bd9b0def2 on bcb5aafdbe80 | unit 168 run, 472 kept | integration 178 run, 538 kept | composition passed on shanks
mesh-delivery #51's re-ask never fired for mesh-build-node.4e7b118e: messages the
SEAT_NODE_BUILD_GO workers had given up on stay on the work queue, and each counted
as waiting ahead, so the ask was never asked again (note 62). A message ahead now
waits only when no node wrote its step and no ask of its kind asked after it was
taken; the old message of an ask asked again is deleted from its stream; and
builder_seat_queue (answered by the builder through its queues state, as only its
account may read the stream) shows each message and whether it still waits.
builder_stuck, checker_kept, checker_verdicts, checker_status, checker_previews, checker_pull and checker_try listed a
bucket of thousands of keys and read each value (builder_stuck also its history) one round trip at a time before
filtering, so they timed out at 30 s. Now keys are narrowed by what the key carries first, history is read only for rows
kept, values are read sixteen at a time, at most 1000 a call with the answer saying it was cut, steps newest first, a new
try reads only the newest 64 tries, and each tool logs the values read and the time taken. Same kind as the view's fix
(PR 52); the helper is copied, not put in mesh-sdk, as a few lines do not earn a group delivery.
Composed without a root-free judgement, a membership would drop a kind's verified-sender and the
operator's channel approvals would stop being honoured; any loss now stays a shadow difference,
named. A tool no longer waits on a pass's round trip. compose-now says the controller serves no
verb to write a module's grant again.
Since release 6 nobody publishes memberships, so a module placed since cannot serve its seat
or reach a new state (the operator, 2026-10-25 22:31). The part publishes each membership on
the assignments stream only when it differs from the one last published, composes from the
manifest the wanted build names, and says a refused publish and tries again at the next event.
Pins mesh-sdk 881da7f, whose composer grants mesh-grants that publish.
mesh/check pass 41f25d3129fd on 73be3df28cd4 | unit 21 run, 579 kept | integration 21 run, 655 kept | composition passed on shanks
On 2026-10-25 node_haves and assigner_wants timed out at 30 s four times, even narrowed to one node and one module,
while assigner_want answered at once. Each listed every key of its bucket and then read every value one after the
other, filtering only afterwards; every value is a round trip from the tool through the runtime to the bus, which
opens the bucket and gets the key, so a few hundred keys outlasted the call. Now a node and a module read that one
key directly, a node or a module alone filters the keys before any value is read, and the values are read sixteen
at a time, at most 1000 a call, the answer saying when it was cut. indexer_modules, indexer_candidates and
builder_builds read their values the same parallel way. The kind: a read tool that reads a whole bucket key by key;
every list tool of the view is covered.
mesh/check pass 4628ba26b41a on 73be3df28cd4 | unit 167 run, 443 kept | integration 177 run, 509 kept | composition passed on 3 machines
The open helpers were copied into seven parts; code shared between
modules lives in mesh-sdk, so they are mesh-sdk process's now and the
parts are pinned at fix/1945-optional-buckets (e0d1de04), moved to its
main once it merges, one delivery group with mesh-sdk.
Met twice on 2026-10-25: the builder of #48 exited on a refused read of
the gate's source and crash-looped, and its new reasks bucket would have
done the same. A grant reaches the bus with a delivery, so a refusal or a
timeout at start is often only a grant not there yet; exiting for it
stopped the whole part until a hand step on the bus.
Every startup open in the parts is now essential or optional. Essential
ones (a part's own state and the input it exists for) still stop the
start, with an error naming the bucket and its grant. Optional ones are
said and left: watches the bus refuses too. The builder builds trunk keys
without the gate's source, which it opens again at the next candidate
that asks a trunk head; toolchains and previews open again at their next
event, as a missing bucket did. builder_ask_again says that asking a
running step again builds it twice.
A new assignment whose trunk build was not done fell through the rules
with nothing said: no want, no neither-list entry, no log line, and the
rule holding it back was computed and dropped. The pass now logs the rule
when a node's reason appears or changes and when it has a want again, one
line per change. A node whose want is removed and that does not run the
module gets the reason in that removed want's words, so assigner_want
answers why; a node with no key gets no removed want, since its node agent
would then remove the module and stop holding its resources back.
A build ask its seat's stream gave up on, dropped or ended with no step was
waited for with no end, and on restart taken over and waited for again: no
build node would ever answer it, and nothing could ask it again (trunk key
of mesh-build-node at 4e7b118e, building since 19:06; seven keys at 18:44).
At the step bound, a context's bound and no timer of its own, the builder
reads what the bus says of an ask with no step: gone from its stream, or
first in line while a node took a later ask. Such an ask is asked again
under a new id the key names, logged and said as build.asked-again, at most
three askings, then the key fails with why. The bus's max-deliveries
advisory would tell it sooner, but the builder is granted neither it nor the
workers' information, and the controller is not changed here.
builder_stuck reads every waiting key with its asks, askings and steps;
builder_ask_again writes the new reasks state, which the builder watches
and answers.
The grants part must compose with the controller's own composer before it can take over the
list; in shadow it proves the two agree, names every input the bus does not carry yet, and
takes no effect. It claims no seat (mesh-sdk's seat list does not name mesh-grants for the
running composer), serves module tools, and keeps running when a read or write is refused.
mesh/check failed: integration: mesh-delivery-assigner/cmd/mesh-delivery-assigner.TestTheAssignerStartsAndSaysWhatItKeeps: the assigner said {V:1 Modu…
mesh/check pass ac82187b852c on 82180bffd2c6 | unit 167 run, 432 kept | integration 177 run, 498 kept | composition passed on 2 machines
The test predates the fix that keeps a held fingerprint while builders disagree (#47, 3e7f852) and still asserted the
blanking that fix removed, so it fails on main 4efe6cf as well. It now asserts the fingerprint is kept after n2's
assignment, and that a merge moving avahi during the disagreement gets none until the builders agree.
mesh/check failed: integration: mesh-delivery-indexer/cmd/mesh-delivery-indexer.TestIndexerCountsOnlyTheAssignedBuildNodesToolchains: within 5 s of n2…
Candidate.Base was never written non-zero (review round 2): the builder's base lookup would have found no trunk for any
live candidate and built no preview at all. The indexer now writes, whenever it writes a candidate's fingerprints, the gate's
source revision whose trunk heads it read the previews onto; that head of the pull request's repository is the preview's
trunk. No mesh-sdk change: the field and its meaning were there. The builder's test now takes the candidate the indexer's
own test writes (testdata/candidate-written.json, kept equal to it), not a hand-made one.
The builder took a candidate build's merge base from the module's value commit. Now that a value keeps the commit its
fingerprint first appeared at, that commit can be many merges behind the trunk, so a preview would be built onto a stale
trunk (review round 1). The base is now the head of the pull request's repository in the gate's source revision the
candidate names (Candidate.Base): the builder reads mesh-delivery-gate.source, watched with its history (a direct get of
one revision is not what a read grants), and a candidate heard before its revision is built once the revision is heard.
jschoubben
mesh-tools