A module held because the recovery module could not be asked kept that hold
across the engine's start as if it waited for the operator's answer, while
its have carried a wait, so the assigner never raised that ask: the hold was
for good (the indexer on the control-node). The engine now judges such a hold
again at its start and at every change of the placements: with no holder of
mesh-recovery placed it moves the module without a point (H-A-177), with one
placed it asks the point once more. A point a holder answered and failed is
held as a failed backup, with no wait, so the assigner raises the operator's
ask (back up again or skip); an older hold of that kind written with a wait
is written again without one at the start.
Builds: design 47 H-A-180, H-A-177, ADR 0568 §4, G-14 (a snapshot-kept that ends failed holds as a failed backup), A-453
Findings are keyed by scope, id and kind across the mesh, so the module
alone as the id let two machines clear each other's restore-point-missing:
its id is now <module>.<node>, as a held module's is. Before the watch of
the assignments has caught up the point is asked as before and a missing
answer holds the move (H-A-178, built in d217056c); the composer's Watch
test now shows RecoveryPlaced answers read only once the assignments are
caught up.
Builds: design 47 H §456, H-A-177, H-A-178
The node-engine held every data-holding module's take-in for good: it asked
the seat mesh-recovery for the restore point, and nothing answers because
no holder of the seat is placed on the mesh (H §456). It now asks only where
the assigner's placements it holds name a holder (and while they are not
read yet); where none is placed it moves the module without a point and
raises the finding restore-point-missing (warning, scope machine, id the
module), cleared at the module's next take-in that takes its point. A holder
placed that does not answer still holds the move.
Builds: design 47 H §456, H-A-177
The apply's fetch tells the composer of bytes that are not of the declared
digest (fetch_first.go), but no test held it: with that call disabled every
test of internal/apply passed. TestAStoreThatStaysAwayCostsOneBoundedWait now
has a store that serves other bytes to two modules from one source, and the
composer is told exactly once, with the declared digest.
Builds: design 47 H H-A-150
A fetch nothing answered for (a bus or store that did not answer, a gateway
saying the store is away) proves nothing about the digest, so it is tried
again as before and the module is not held for it; only an answer of the
store's (a 404, bytes not of the digest) forgets the resolution. The apply's
fetch of a bundle meets the same wrong digest as the take-in's, so it tells
the composer too (apply.WithFetchFailed, set in the node agent's apply).
Builds: design 47 H H-A-150, H-A-146, G A-430a
The engine kept a want's version resolved to a digest for its process's life
even when the artifact store's fetch of that digest failed (404, or bytes not
of the digest), so a fix of the store's answer never reached the node without
a restart. A failed fetch now forgets every resolution naming that digest, says
"resolution of <module> <version> forgotten: <why>" in the journal and holds
the module; the seats' verb versions is asked again at the module's next want,
at the next compose by hand, and after resolveAgain (the bound a failed lookup
already has). A resolution whose fetch succeeded is kept.
Builds: design 47 H H-A-146, G A-430a
The node-engine consumes the entries client's asking again (H-A-60, H-A-110): the setting entry-asks (default 5; module.json
config and file resource, read at the engine's start) goes to Entries.Asks; Entries.AskAgain is called at each link's start and
on each catalog-revision the composer hears (the shadow's cache too); the finding entry-unanswered joins the engine's finding
kinds, written with a journal line when an entry ends unanswered and cleared when it is answered or no longer wanted; the
shadow's wait and the have's hold end on it instead of waiting without bound.
Builds: design 47 H, H-A-60, H-A-110
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
The seat mesh-catalog keeps no state verdicts (H-A-79): frombus taken, the composer's parse and the composer's
reading of a candidate drop the verdict branch and decode the manifest alone, a module in a catalog revision being
accepted by being there; the composer no longer reads the request field Case.Verdicts (H-A-112). Tests that held a
module by a verdict hold it by the decoder, or go.
Builds: design 47 H, H-A-79, H-A-111, H-A-112
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
A want carries no composed digest, no keeping node and no newest build (design 47 H 750,
H-A-37): the agent takes the digest from the composition the node made for the want's
build, the shadow's resolution around a missing entry (A-405) and KeptBuildOf's keeping node
go, and the artifact is fetched from the store. Tests of the removed behaviour go.
Builds: design 47 H §750; H-A-37
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
Unit test functions: 1,719 before H1, 1,717 after (three fewer by the merges, one added for the
claims read); integration tests: 10 before, 11 after (budget 50). The module stays over the
unit budget of 500.
Builds: design 47 H §774
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
A want of rank 2 whose composition is made first is held in a wait of step placing on the
module of rank 1 of its pipeline; rank 1 is applied first, rank 2 after it, and the wait ends.
Builds: design 47 H §771, §774; H-A-59
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
One row per unapplied changed want: module, pipeline, rank, position, why, waits.
Served beside process-builds and claimed in module.json.
Builds: design 47 H §771; H-A-46
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
settle applies the changed wants in the order of design 47 H 771: the node's own composition,
a reverted module, the holder of node-backup and the bus module (by the claims of the manifests
the composer holds), each pipeline's wants by rank (older source first), the rest by name, the
engine's own module last. A want of a higher rank waits on this machine alone for the lower
ranks of its pipeline, held in its have as a wait of step placing and judged again at the next
settle. The want's pipeline and rank are the mesh-delivery contract's stand-in until its commit.
Builds: design 47 H §771; H-A-46, H-A-59, H-A-67, H-A-68
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
The record's capabilities no longer hold want-parts; manifest-reads holds the word
want:version, stated at the first reading and said by manifest-reads-changed.
Builds: design 47 H §772, §764; H-A-45
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
The node-engine no longer reads a want's parts: the parts check (G 524) and its hold
parts-differ are removed, the digests of a build are resolved through the registry
seats from the want's version, and a candidate's added module is resolved from its
version alone.
Builds: design 47 H §750, §764; H-A-37, H-A-38
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
Subscribe to meshcatalog.SubjectCatalogRevisionAll (no string of its own); module.json consumes
mesh-catalog.catalog-revision so the node account is granted it; the two new tests are folded into
TestTrunkAskBackoffAndCatalogRevision and wait on channels, not time.
Builds: design 47 G-22 A-487a, A-487c, principle 3
The subscription is a wake-up; the backoff bounds the asks without it. The refusal is said with its subject and error
(principle 3), the composer goes on, and the subscription is tried again on laterAgain's steps until it holds.
Builds: design 47 G-22 A-487a, principle 3
A trunk build answered none waits laterAgain's step for its module (5 s growing to 5 min); the seat mesh-catalog's
event catalog-revision, heard through Env.Hear, ends the wait of every module at once; an answer starts the steps
again. A pass is woken at the step's end.
Builds: design 47 G-22 A-487a, A-461e, G-12 line 317
Entry of the running build, else the composition this node runs of it, else the module's trunk manifest through the seat mesh-catalog's verb module;
said once per module and source; words naming the three where none answers. Read there: a held module's offer (frombus Scope.RunningManifest),
FormChanges (A-483), the store holder and the stand of a running module without an entry (A-461d, A-480a), and the tool composition.
Builds: design 47 G A-484, A-480, A-480a, A-483, A-461d, A-482
compose fills Composed.Templates from the resolved manifest's TemplatesDigest; partsDifference adds them to the digests fetched.
Builds: design 47 G A-481, A-481a
A module the node holds keeps running what it runs, so a module that requires what it provides is no longer held for
"nothing provides": the held offer's provisions are the running composition's build's manifest, else the manifest the
module's running event names.
Builds: design 47 G A-480 (G-22-lead-answers-2.md)
The trunk build is asked, and a placed module laid, only for the store's
holder (a store seat's holding on that node) and the providers this node's
modules are bound to; never every module placed elsewhere. The asks run beside
the pass, three at a time, and an answer asks a pass; a module that waits is
only the one needing that address.
Builds: design 47 G-22 A-461, A-461d, A-461e
lacksEntry holds for a stopped run and for one naming no build, and the loop
for another node's placed modules lays every one not yet laid where it holds.
A trunk build askTrunkBuilds could not get is said with its reason and asked
again on the next pass.
Builds: design 47 G-22 A-461d, A-461e
Declared.Held carries the modules composeWithoutStoreFetchers leaves out
(A-452: each stays on the build it runs). The store-address comment and code
say A-461e: no address read without error holds only the fetchers, a fault
reading the inputs still refuses the node.
Builds: design 47 G-22 A-452, A-461, A-461e
Being heard running is no condition: such a module counts as one whose
running build has no entry (wantedBuilds, askTrunkBuilds, layModules,
lacksEntry), so the store holder is found from the build its placement names or
its trunk build.
Builds: design 47 G-22 A-461d, A-461b
Typed resolution.NoStoreError from artifactsInto (image, source, tools bundle,
runtime program, templates archive), laid on its module by
composeWithoutStoreFetchers with A-452's "nothing provides" words and composed
again, instead of a whole-node refusal. Replaces storeHeld and
modulesNeedingTheStore, whose shape list missed bundles and held stand-ins.
Builds: design 47 G-22 A-452, A-461, A-461b, A-459b
A hold that changes the problem but leaves the machine not composing keeps that module held and blamed,
and the search goes on against the retry's problem within the same blame_tries, until the machine
composes or no hold changes it: every module the change refuses is blamed, never only the first.
Builds: design 47 G §721, §722, A-470c, A-470v
The held composition's own blames and its search come first; the module whose build can never be made is
then blamed beside them. A held composition that still fails and blames nobody stays on the node's own
account, so the module is not blamed in its place.
Builds: design 47 G §721, §722, A-470j, A-470k
ComposeSent no longer runs the node-engine's own culprit search (fixed at eight modules), and the composer's
own loop, which reran that search at every try, is gone: ComposeSent holds the case's changed modules on the
machine one at a time, one composition each, at most blame_tries, and blames the first whose hold ends the
problem; past the bound the untried are named. Without blame_tries there is no search. A module the node
holds composing without the change is held as there, and two owners of one path are blamed outside the search.
Builds: design 47 G §721, A-468i, A-468l, A-469j, A-469k, A-470c
A module awaited on a machine whose revision gives it no fingerprint (the case's none) is held as its
node-engine holds it, the machine judged from that composition, and the module blamed with the revision's
none words, so only its own pipelines fail. An awaited build with a fingerprint keeps the machine not judged.
Builds: design 47 G §721, §722, A-470j, A-470k
The composer program has no settings, so the assigner bounds the search through the request: only the candidate's
changed modules placed on the machine are held, one at a time, the search ending at the first whose hold ends the
problem; past the bound the untried are named. A request with no blame_tries gets no search.
Builds: design 47 G §721, §733; A-468h, A-468i, A-469i, A-469j, A-469k