The mesh knows no pull request (design 47 G, G-1 §38-39). Removed:
- the states `pulls` and `kept` and all code that reads or writes them (pulls.go, the held values a proposal holds, the
listing of held values), SweepPulls, settlePull, HeardPull and the forge call;
- the consumes of git.pull-merged and git.pull-closed, the invoke of seat:git.pull;
- the config verb's `pull`, `head`, `declares` and `pulls` arguments and its actions `pull-declares` and `held`;
- the indexer's config candidates, which only that path used: the consume of mesh-delivery-indexer.views.config-candidates
and the invoke of indexer_read_config_candidates (the indexer's state goes in the same group).
A stored proposal that names a pull request is judged not one this module makes, so nothing is set on it.
Builds: ADR 0681 §1 (no pulls or kept, no git events, no seat:git.pull, no `pull` argument, no pull-declares or held), §3, §5, §6;
lead answer A-490l (a stored proposal naming a pull request is refused on its approval).
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
mesh-config sent about 262 requests a second and received about 1,540 since its start: every sources() call (a
resolution pass, and a `delivered`, `manifest-reads-changed` or `manifest-accepted` event, each of which read the sources
whether or not a window or a retirement was open) read every key of its own states by one get each (pulls 1,176 keys,
resolved 447, declared 168, values 101...: about 1,200 requests for an event, 2,800 for a pass), serialised by one lock, so
a delivery's burst of events kept it asking the bus back to back.
- The module's own states are read from mesh-sdk's WatchedStates (one watch per state, its own writes taken into the copy)
instead of a get per key (ownstates.go, main.go runtimeStore).
- An event judges what it names: `delivered` and `manifest-reads-changed` read no source when no rename window is open,
`manifest-accepted` reads none when every row is already marked or read again (renames.go windowOpen, retiring.go).
- Open windows and retirements are judged again after each pass, since an entry that answers after the judgement was
made moves the pass, not the event; the slow reads had hidden that race.
- No timer polls: the five-minute resolution and sweep of every pull request on the forge, and the minute's read of the
router's record, are gone. The sweep and Decide run once at start (an event said while the module was down is not said
again), then on git.pull-merged/-closed and operator-channel.decided.mesh-config; an ask ended with no choice also
takes the retirements from the router's record. One timer is left, a deadline: for the earliest waiting proposal
(ProposalLasts + 1 h) or unanswered retirement (HeldFinding), set again when either is written (NextDeadline).
Wake-ups kept, each by its record: `delivered` and `manifest-reads-changed` close a rename window (design 32 *A config
value renamed*, ADR 0609 §4); `manifest-accepted` judges one module's retirements (design 32 *A value no longer read*);
`git.pull-merged` and `-closed` settle one pull request's held values (ADR 0502 §7); `decided.mesh-config` takes one ask's
warrant (ADR 0259); a changed view (assignment, want, predefined list, config candidate, catalog revision, entry) moves a
resolution that writes only what moved (ADR 0565 §3).
Builds: ADR 0681 §1 (no polling of the forge: the five-minute look ends); ADR 0580 §1; ADR 0426 §3; design 32 (*A config value renamed*, *A value no longer read*);
lead answer A-490l (hq main, 03-DESIGN/01-to-be/47-milestones/G-22-lead-answers-2.md) for the wake-ups and the one deadline timer.
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
mesh-build-go and mesh-build-image follow the seat list as mesh-build-node does, so each consumes mesh-catalog.seat-definition-changed and
-removed, and all three invoke seat:mesh-catalog.seat-definitions, which the reader calls.
Builds: design 47 G A-488a, A-488b (hq 57c14047)
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
Its runtime follows the seat list (seatlist.Start subscribes to mesh-catalog.seat-definition-changed and -removed); with no consumes
there was no consumer shanks_mesh-build-node_tools, the seat list was never read and the asks of a named asker were given back.
Builds: design 47 G A-488a (hq 25df48ad)
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
Its processes run as the operator account, so its own secrets are owned by it (design 20 "Whose a process's own secrets are").
Builds: ADR 0680 §1, §4; design 20 "Whose a process's own secrets are"
The single-module proof of design 47 G (the operator, 2026-11-17), after pack49:
one change to one leaf module, followed to delivered on g14 and shanks with no hand act.
The single-module proof of design 47 G (the operator, 2026-11-17): one change to one module,
followed through its pipeline to delivered on g14 and shanks.
The toolchain a build module reports (guard key and toolchain-reported event) names the
language version alone (mesh-delivery-sdk steprun Guard); the build module's own
Toolchain func keeps the installed text. Delivered with or after the sdk change.
Builds: design 47 G A-471 (G-22-lead-answers-2.md)
The level took 67 s locally and 74 s in the pipeline, over IntegrationRun's 60 s. The eight slowest tests share no state and run in
parallel: 31 s locally with the fixture's store and bus. The real-store test makes its work directory under the test temp
directory, not /tmp.
Builds: ADR 0445 §3 (the integration level's 60 s bound)
A pipeline's test step places no repository beside the tree and a module's tests read nothing outside their module (lead answer
A-474), so the test no longer needs MESH_CHECK_BESIDE or the node-engine's tree: it fetches a small program of the composer's
contract (testdata/composer) and still proves the build once, the kept composer on the second ask, and the verdict by case.
Builds: design 47 G lead answer A-474 (principles 10 and 9)
The bridge kept a directory for the pre-G indexer's package publishing; its own commit says it goes once G's indexer runs, and
the test of the delivery SDK's step code refuses it. The unit level took 6.3 s over its bound of 5 s in the take-in's pipeline:
the digest tests ran their go commands one after another; the marker table is a subtest per case in a fixture of its own, the
list half a test of its own, and the six slowest independent tests run in parallel (4.6 s locally).
Builds: design 47 G §464, §680 (the build modules keep no copy of the delivery SDK step code); ADR 0445 §3 (the unit level's 5 s bound)
The forge bridge's merged moves the repository's head and says source-moved; at its start each head is compared with the forge's branch head; gate_source_follow does it for one repository by hand.
Builds: design 47 G-22 A-464, A-464b
bundleKeepDays read the config value snapshot-keep-days while allTools
listed the tools, before stdio.Serve; a config read waits for the runtime's
channel, which opens only once the bundle serves, so the bundle never
answered initialize and node-tools served none of its tools (novox
2026-11-16 22:21 and 22:26, after the first build from main since A-112).
The setting is now read when nats_snapshots is called.
A build seat's work queue (SEAT_NODE_BUILD_GO) held about 370 asks of
pipelines nobody waits for, ahead of the one pipeline wanted, and no tool
could empty it short of removing the stream. nats_stream_purge removes the
messages below a given sequence, keeping the rest, the stream and its
consumers; it needs why, reads first, offers dry_run, refuses a purge past
the stream's end and the module-kept data nats_drop owns, and says
stream-purged as its audit event, like the three removal tools.
Kind: a version written into a workspace's modfile without what that version's graph needs, judged against files the build does not use.
Builds: design 47 G (ADR 0573 §2, §4, as written in versions.go: asked versions hold whatever go.mod names)
A-379 replaced A-376 for the start, and the seat mesh-catalog offers neither the states modules and manifests as views
nor verbs of those names, so the validator refused the manifest's two consumes and two invokes.
Builds: design 47 G A-379, A-445 (G-22-lead-answers-2.md)
The operator channel kept at most three asks of one asker offered and held its oldest back behind newer ones, so asks
waited unseen (the pipeline's proposal behind two mailu asks). It keeps no count any more and holds nothing back: every
ask is offered when it is raised, and lapses at its expiry or ends at its answer. An ask held back before the change is
offered again at the router's next tick; ask-again offers one so and holds nothing back in its place.
Builds: design 47 G A-424
The builder's publish again of mesh-delivery-pipeline 0.1.6 sends the version and the digest with no parts, as A-73
has it; the build stored the bundle untagged, and the holder answered 'publish: no parts' unless the tag existed
already. The holder now checks the digest against the store under the name (the artifact store, A-50), asks the
builder's reservation, tags it with the version and says published; a digest it holds none of is refused as not
stored, naming the digest.
Builds: design 47 G A-417, A-73, A-50, §588
The forge's git verbs may serve after the gate does; one pass with every head read refused wrote nothing and was never retried.
Builds: design 47 G A-401