G's grants already carry a state's max-mib onto its bucket (Bucket.MaxBytes, §469); main's second copy of that cap in
BucketStreamConfig and Composed is dropped for G's, and main's tests (the declared cap row, the held bucket raised)
are kept. The checker keeps G's version 0.1 (§368) with main's compositions max-mib 256.
Builds: design 47 G §469, §368; A-277 (since mesh-delivery #206 the grants part carries max-mib)
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
The delivery repository's contracts and SDK live under modules/ as
mesh-delivery-contracts and mesh-delivery-sdk, their Go module paths
renamed to match the folders (the TypeScript names already did), and
every importer in mesh-delivery, mesh-catalog, mesh-host, mesh-tools
and mesh-sdk follows in the same group: a design as it should be, one
path per module, checked by the group's check building every importer.
Builds: design 47 G §427, §88
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
The checker's fixed names of the build modules (buildmodules.go in its program, rules and check) go: the
Examiner finds each build module by the language its seat's acting rule builds (buildrun.BuildSeats over
the seat list's definitions and the holders the indexer's modules name by their manifests' claims), the facts
carry the language where a level counts by it, and the program's manifest reads the steps and toolchains of
the seats and hears their toolchain-reported. The tools read the steps of each holder and follow the seat
list as the view does. Timers stay (A-118).
Builds: design 47 G A-82, A-118, §204
Claude-Session: https://claude.ai/code/session_01NcbgfGKaxw984XyAPfLm4h
A check that failed for a part of the mesh not heard was judged again at
every change of the nodes. It now says could not run, naming the part,
and is kept in the checker's own state could-not-run; only the running of
that part asks it again, once per heads and part at one fingerprint
(PartRunning). Not asked again within an hour it is the finding
check-could-not-run, which clears when the heads are checked again,
replaced, or their pull request closes. checker_status reads the state.
The node-engine seat's running event is wired once its definition lands.
Claude-Session: https://claude.ai/code/session_01JG73KEq4zoDfM9WtYpcQgZ
Builds: design 47 G §253, §315
A large node's request is 1.2 MB, over the bus's 1 MiB message limit, so the compose ask failed once keeping no longer did.
The checker packs the request (a JSON string "gzip+base64:..." in the part's `any` request, no contract change) and
buildstep.ComposeJob.RequestBytes reads either form. A kept request of another version is refused naming it, and listing
no longer reads pieces. An older mesh-build-go refuses a packed ask loudly (it holds no request object): the check fails
naming it, so the catalog change is delivered with this one as a group.
Quick fix (chosen as the recovery): pack the request, ~290 KB for the largest node. Thorough fix (milestone G): the request
named by digest and served on a seat verb, as RunInput is.
Claude-Session: https://claude.ai/code/session_01NcbgfFfQ4Z9UE5jK32PmE5
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 (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/check checking 918d21b7dac5 on 6bc232eb78b9 | novox/mesh-delivery#151 in group feat/registry-milestone
The request sent to the composer is kept gzip-compressed in the state compositions (ttl 3 days) under the check's key and the node, and the tool checker_composition_requests reads one by check and node or lists a check's. A request that cannot be kept fails the node's composition, naming why.
Claude-Session: https://claude.ai/code/session_01NcbgfFSnn24M3GcqttoU7i
B0's shape wins: the running checker asks and records facts, the checking program
decides. The registry's types win: the rules and the program read the generated
checker contract and the domain SDK, and no mesh-sdk delivery import is left.
The registry's series, renames and unparsed seat definitions become facts with
their refusals as rules; the tagged versions ride on the case. The program reads
the heads' seat definitions through sdk/seatdefs and the running checker lays them
over its reader's list by seat, reading no file of the heads: the reader's list
is larger than a run input's bound.
Claude-Session: https://claude.ai/code/session_01NcbgfFSnn24M3GcqttoU7i
mesh/check checking 9c9ed7a1d8f1 on 2d201f781ced | novox/mesh-delivery#162 in group feat/b0-checking-program
The operator's answer: mesh-build-go reading the checker's state inputs broke ADR 0596 §2. The run ask's input now names
the verb seat:mesh-delivery-checker.input; the checker's tools serve it from the state inputs (Reader.Input, checked
against its digest), and the state is the checker's own again, read by no other module.
Claude-Session: https://claude.ai/code/session_01NcbgfFSnn24M3GcqttoU7i
The running checker asks the checking program twice on the build seat's job run: first the
heads' seat list, which the manifest validator's requests (change and trunk) and the
node-engine composer's carry; then the verdict of the facts it gathered. The program is built
at the group's mesh-delivery head, else at mesh-delivery's trunk, and its input is kept in the
checker's new state inputs as the exact bytes of its digest, read with checker_inputs. A run
ask no build node begins within 5 minutes is put once more, then fails the check naming the
job; a program that does not build or answer fails the check naming it, and the verdict
records who judged it (JudgedBy). A second step-seen hook on a context is added to the
first, so a try still sees its steps.
Claude-Session: https://claude.ai/code/session_01NcbgfFSnn24M3GcqttoU7i
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
The checker says try.started (members at their commits) and try.ended for a whole try; the indexer writes and
deletes the candidate from them and finds the open tries again at its start through checker_tries, never reading
the checker's state. The try's digests are judged inside the one pass that takes the seats; the definitions of every
contracts module of each member are read; the config declarations leave a try's candidate alone.
Claude-Session: https://claude.ai/code/session_01NcbgfFSnn24M3GcqttoU7i
A whole try has no pull request, yet compose-candidate and the manifest validator read the indexer's
candidate of a group. The indexer reads the checker's tries, writes candidates.<try id> (number 0 per member) from the
previews, judges its manifests with the seat definitions the members carry, leaves try-* alone in the open message,
deletes it when the try ends, and has a tool that reads and one that drops it.
Claude-Session: https://claude.ai/code/session_01NcbgfFSnn24M3GcqttoU7i
A try names several repositories, each the exact commit asked. A whole try (not narrowed to a level or
tests) runs the group check's own path: the manifest validator, the module graph and the composition on every
placed node, reading the indexer's candidate of the try's group. A try with no such candidate is an error that
says so, never a pass. A try's candidate is no check: the checker leaves it alone and owes the forge nothing.
Claude-Session: https://claude.ai/code/session_01NcbgfFSnn24M3GcqttoU7i
mesh-delivery's sdk/firstview moved whole into mesh-sdk go/views (J3): the builder, checker, watchman, teller and view
follow each owner's state with views.ProcessReader.Follow(owner, state, tool, stream), naming the owner's read tool as
their invokes name it and its views family as the owner's manifest's `streams` names it ("indexer-views",
"assigner-views"); a refused Follow stops the start. Their beds serve the owners with viewstest.Serve and grant with
viewstest.Grants by the same names. The view's tools ask the owners' read tools by their literal names. The assigner
names its own views family and the indexer's itself (views.AssignerStream and IndexerStream are gone).
Every views.StatusTools call takes handacts.Recorder{Ask: stdio.Ask}: a first view asked again by hand is a hand act,
recorded through seat:mesh-hand-acts.hand-act, which the builder's, checker's, watchman's, teller's and view's
manifests now invoke (a row in each manifest test, each failed first).
Row (A2j's written-time switch): TestRunNeedsItsCredential "a want stood since the assigner's bus wrote it": the
watchman's want time (createdOr) is the time the owner's bus wrote the key, through ProcessReader's first view (failed
with createdOr ignoring the entry's time). sdk/go.mod: nats.go is indirect now.
Claude-Session: https://claude.ai/code/session_01JG73KJNrH5EgFXMKKYSrig
ADR 0615 §1 and ADR 0596: the checker invokes the six read tools, consumes their views families and
reads none of the six buckets; this ends its direct read of the packages state (the lead's ruling 2).
It keeps first-views and findings and announces both as its own. The lead's ruling: an owner that is
down never holds the start; the checks, the tries and the hand asks start once every first view is
answered, the placement is followed into the toolchain gate once its first view is. firstview.Reader
opens a followed state's picture as a process.BucketOpener (the builder uses it too). Tools
checker_first_views and checker_first_view_again.
Claude-Session: https://claude.ai/code/session_01JG73KJNrH5EgFXMKKYSrig
A failed candidate build no longer needs mesh-delivery.recheck, an exception verb (ADR 0421 §12): the checker hears the
builder's keys the failed verdict names and, once each is written done (builder_build_again), retakes the check of the
same heads as it does after a failure of the nodes, at its start too. When the tests have already failed the check
composes at once and names the builds it did not wait on, so its red status is never held back by a build.
Claude-Session: https://claude.ai/code/session_01JG73KG7AfSdh6FP8VrBqCn
Worked out from the builder's builds state at each read, stored nowhere: of each running check, the candidate's builds
that are neither done nor failed.
Claude-Session: https://claude.ai/code/session_01JG73KG7AfSdh6FP8VrBqCn
The operator, 2026-11-02: "A failed check shows only its first fault". Group feat/1126-removal's
check failed four times, each verdict naming one fault, so three faults took three runs of 20-40 minutes.
The checker no longer returns at a member's first failure: every member is asked, every level's
failures and every member's refused step or conflict are joined in judging order, the manifests are
judged and the composition is read also when the tests failed, and a failed composition is added after
them. The status line says how many failed and names the first, cut so "; and N more" stays whole.
checker_verdicts and checker_pull answer the whole list (failed). Needs mesh-sdk's buildstep change
(every level runs after a failure) pinned at the group's head.
Claude-Session: https://claude.ai/code/session_01NcbgfJ9Gum47zkFKWRaovL
buildrun leaves mesh-sdk for the builder's own Go module (ADR 0593 §1); the
checker, assigner, indexer and watchman require it by a replace to the sibling
directory. Issue 1930's bound was already on mesh-sdk's main and moves with it.
The builder copies the imports a build module answers to the done key and the
log's end entry, where the indexer keeps them (§2; D7).
Claude-Session: https://claude.ai/code/session_01NcbgfJ9Gum47zkFKWRaovL
Review of fix/lost-check: an ask no node took yet has no step key, so
asking again then would run the check twice (issue 1791's kind). The
checker keeps the asks each running check put in a new state asks, and the
tool refuses unless every one of them has a step key. Asking again under the
same ask id is not done: the stream drops a second put of an id within its
de-duplication window and keeps both past it, so the waiting copy is never
replaced.
Tests: the hand row of the lost-ask table refuses while its ask waits;
counts unchanged.
mesh-catalog #553's verdict stayed running after its step mvsgthyjed5786
passed on shanks at 19:09:31Z: the watch of that one step key, opened at
18:46:05Z, heard neither its start nor its end and never closed, so the
wait had no bound running and nothing would ever judge it.
- HeardEnds stands between the asker and each ask's watch: at every step
end of the same build module (one watch per steps state), at a
reconnect, the asks waited on read their keys now and are judged; a
watch that closes is opened again, at most 3 times.
- A step the build module's next process ended (RestartReason, a delivery
replacing it) counts as a lost ask: asked again under the claim, at most
LostAskTries, then failed as not heard, judged again when the nodes change.
- checker_ask_again {group, why} ends a running verdict as asked again by
hand; the checker hears its verdicts, ends the wait and claims it anew.
Refused while a step of its pull requests runs, or for a verdict that
is not running.
- checker_status names such checks under lost, and each step's what.
Tests: one table replaces the given-up test (rows for given up, replaced,
deaf and nudged, deaf and asked by hand); counts unchanged (147 unit, 10
integration in the checker).
Only tries a build node holds count toward the 4 at once; tries waiting for a node are bounded
apart at 8, and the refusal names both counts (issue 1890: at 19:53 four tries queued behind 44
checks for up to 47 minutes refused every further try).
The checker wrote a try's asks only at its end, so a 31-minute try could not be told queued from
running (issue 1890). The try's record now names each ask as it is put, then the node that took
it, its state and the phase its step says, written on the step key's own changes. build_usage
names a running try's asks and takes a build module's capacity.<node> (configured bounded by its
node's build CPU, issue 1806's follow-up) when it is current with mesh-config.
mesh-grants moves to the same mesh-sdk (3203358c), so its shadow user list grants the checker and the builder the
DEAD_LETTERS reads the controller's list will, and the take-over is not blocked by a difference. The checker and the
builder hear DEAD_LETTERS through buildrun.DeadLetterHearing: in the background, again at every reconnect and every ask
waited for while deaf, and keep whether they hear it in their new state hearing, which checker_status and builder_stuck
say as dead_letters. checker_try says that a try ended by the not-taken bound leaves its ask in the stream.
Not done as asked: a module health tool. The catalogue refuses a tool check on a module with no runtime of its own and no
other check beside it (ADR 0227 rule 8, ADR 0366 §1), so the deaf state is said in the status tools and the log.
A check ask handed out five times during a bus reload was given up into DEAD_LETTERS and no build node wrote a step for it,
so try-1793017340003993788 stayed running for ever, held one of the four try places, and checker_try of its head kept
answering its id. The checker now hears the copies DEAD_LETTERS keeps of the build seats' given-up asks (an ordered
consumer from its start; the grant is mesh-sdk/mesh-controller DeadLetterReads, branch fix/734-lost-ask) and ends the
step it waits for: a try ends as error saying why, a check is asked again under its claim, up to three times, then fails
as not heard. A try whose ask no node takes within the step bound ends the same way (a check's backlog stays no fault,
ADR 0525 §9). checker_try_end ends a try by hand, freeing its place and stopping its judging.
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.
mesh/check checking 4ffe5f4d5243 on 85aca46348a0 | novox/mesh-delivery#25 in group feat/fastcheck
A new pull request waited 25 to 45 minutes behind every open pull request's re-judge after a
merge. The semaphore becomes a ranked admission (ADR 0536 §4); checker_status names the
waiting checks by kind.
mesh/check pass e76202876032 on b714eab995a5 | unit 422 run, 0 kept | integration 467 run, 0 kept | over bound: integration (slowest integration mesh-…
The checker's own places are released when a step ends or its wait passes the step bound; checker_status read every
'running' key in the build modules' steps buckets as a step in flight, so keys left by stopped, restarted or unassigned
build modules looked like full nodes. It now lists them under left_running with why (past the bound, module not on the
node, module not running), and leaves hq-only groups out of the unjudged candidates.
mesh-controller's try asked build-agent, which leaves the checking at R4e. A try is now a key in the checker's state
tries: the checker judges the head merged onto its trunk as it judges a pull request's head, through the same
judge, places and holds, writes its end in the key and owes the forge nothing (no status, no preview, no verdict).
One try per head, at most four at once; hq and a repository whose trunk the gate does not name are refused with
"no level of <repo> is checked by the checker".