Compare commits

..
11 Commits
Author SHA1 Message Date
jschoubben 268bbdb6bd Design 47 H: H-A-137, a root module's generated-from keeps the seats directory and its generated packages 2026-11-20 13:00:31 +01:00
jschoubben cf872d91d4 Design 47 H: H-A-136 reads the fetched commit's manifests; the key question goes to H2
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
2026-11-20 12:43:31 +01:00
jschoubben aca06fc312 Design 47 H: H-A-136, what a step's workspace holds (generated-from directories, same-repository zero-version packages)
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
2026-11-20 12:36:37 +01:00
jschoubben 40be107a7e Design 47 H: H-A-134 names delivery from new messages only
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
2026-11-20 12:20:18 +01:00
jschoubben 8d30cc4237 Design 47 H: H-A-134 names the ack wait and max deliver of the indexer's own durables
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
2026-11-20 12:13:51 +01:00
jschoubben 4b5066bc66 Design 47 H: H-A-135 names the earlier branch.* objects too
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
2026-11-20 12:10:29 +01:00
jschoubben 73685d41b6 Design 47 H: H-A-134 the indexer makes its own per-branch durables until H4b; H-A-135 old trunk revisions and current mark go
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
2026-11-20 12:06:31 +01:00
jschoubben 9f231e07a2 Design 47 H: H-A-133, the follower on a module's own runtime hears when its bundle subscribes
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
2026-11-20 11:31:58 +01:00
jschoubben 934f2c2bc0 Design 47 H: H-A-132, the tool runner binds a module's consumer once
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
2026-11-20 11:20:27 +01:00
jschoubben be0bf71651 Design 47 H: H-A-130 the start of a take-in read by the verb; H-A-131 a module-less repository and the gate
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
2026-11-20 10:32:23 +01:00
jschoubben 58ea7aece1 Design 47 H: H-A-129, the indexer reads a branch's source by the verb when its consumer is not bound
Claude-Session: https://claude.ai/code/session_01JG73KHUnBze48rdua8u2sy
2026-11-20 10:30:14 +01:00
@@ -568,3 +568,68 @@ Where two preparers disagreed, the chosen one is named.
manifest on any node (the keeper's accepted set holds modules, not nodes; the developer's reading of 06:23), since a
module may name more than one durable with a suffix. The tool refuses those and removes every other. P3 (a wrong
removal of a live consumer is a silent failure), P4.
- **H-A-129 (the indexer's first take-in at a clean state, 10:30)** Met live at the clean state: the indexer takes the
trunk in only through its per-branch consumer of `source-moved` (§743), the grants part makes that consumer only from
the modules view (H-A-30), and the indexer answers the modules view only after its first take-in: nothing ever takes
the trunk in. So at its start, and again whenever a watched branch's consumer cannot be bound, the indexer reads
that branch's current source revision through the seat verb `mesh-delivery.source` (a call where an answer is needed,
P1) and takes it in as a heard `source-moved` of that revision would, when its view holds no revision of the branch
at or above it; it keeps binding the consumer on `registration-changed` as before. No timer. A fault that reports
nothing wrong while nothing moves is P3's failure. Fixed under H1 (playbook 08) before step 5.9 can finish.
- **H-A-130 (the start of a take-in read by the verb, 10:33)** The seat's `source` answer carries no `start`. The
take-in from it uses start true exactly when the indexer's view holds no revision of that branch, false otherwise
(movers against the newest kept): the first take-in's entries are the base later ones are held against (H-A-19), and
a first take-in with movers would pin and start pipelines for every module (P3). The answer names every repository of
the revision; all are taken in as heads (the movers are the fingerprint changes alone, so the outcome is the event's).
Adding `start` to the seat's answer is not built here (P7).
- **H-A-131 (a repository with no module is merged by the gate too, 10:34)** Met live: the new gate refuses hq's pull
request 796 ("its base main is no watched branch") because a watched branch is a registration's and hq registers no
module; the forge's own tools refuse a merge by design (ADR 0266). Until hq has a registration of its own (a module
whose unit step is its checks, P5: everything ships the same way; built after the clean state), hq's lead answers
stay on their branches, pushed, and are cited by branch. Decided as a design detail; told to the operator.
- **H-A-132 (one pull loop per module consumer in the tool runner, 11:24)** Met live: in a module's own runtime the
tool runner binds the module's durable twice, once for its branch follower (the first views) and once for the
bundles' events, so each event reaches one of the two; the catalog-revision of main.1 reached the follower and never
the bundle, whose trunk view stayed at revision 0 and whose `revision` verb answered none-yet to every reader. The
tool runner binds a module's consumer once and hands each event to both the branch follower and the bundles in
process; no second durable (a consumer the grants part would have to make, H-A-30; P7, P12 a module writes only its
own). Fixed under H1 (playbook 08, mesh-tools) before the clean state goes on. A durable bound twice in one process is
a kind: the reviewer reads every other bind in the runtime for it.
- **H-A-133 (the follower on a module's own runtime hears when its bundle subscribes, 11:33)** With H-A-132 the branch
follower of a module's own runtime is handed the events the module's consumer delivers to its bundle. It therefore
hears from when the bundle subscribes, waits while the bundle reloads, and loses an event the bundle fails five
times (the durable's limit), said in the journal. That is accepted: the follower exists to serve that bundle's
first views, so a bundle that never subscribes needs no follower (P7), and acknowledging before the bundle took the
event would lose the bundle's (P3). A lost event's picture catches up on the next one of its family; a re-ask on
that loss is a finding for H4. The machine's runtime keeps its own binding.
- **H-A-134 (the indexer makes its own per-branch durables, 12:08)** Met live: the grants part composes no per-branch
durable `<node>_<indexer>_<branch>` on the source-moved stream (H-A-30 named it the maker; the grants part's events
work is H4b), so the indexer never hears a source revision after its start and takes a merge in only at a restart.
Until H4b, the indexer makes that durable itself when its bind finds none, with the configuration the mesh gives a
module's seat-event consumer (mesh-sdk's SeatEventConsumerConfig, one message at a time, filtered to the branch, the ack wait of 30 s and
at most 5 deliveries the grants part gives a module's consumers, so that H-A-133's bound holds here too, and
delivering from new messages only: the start's read by the verb covers the current revision, H-A-129), and
the grants part matches it as it matches a module's consumers. It is the indexer's own kept consumer (P12: a module
writes only its own); H-A-30 is read so. P1, P3.
- **H-A-135 (the earlier model's trunk revisions and current mark go at the first branch take-in, 12:08)** Met live:
the catalog-revisions object store keeps the earlier model's `trunk.<n>` objects and its `current` mark on the newest
of them beside the branch-keyed revisions (`main.1`), and the runner's first view hears "revision 528" while its
picture holds 1, forever behind. At a branch's take-in the indexer drops every `trunk.*` and `branch.*` object of the earlier forms no record pins and
the old `current` mark (the branch's newest revision is the current one, H §743); said in the journal, counted. The
stale readers' first views then agree. P3, P7.
- **H-A-136 (what a step's workspace holds, 12:38)** Met live at the first pipeline: the Go build module fetches for a
step only the paths the artifact names plus module.json, so (1) the seat generator's out-of-step check for a generated
package finds no seat definitions (`generated-from` points outside the fetched paths) and marks every file out of
step, and (2) a requirement at the zero version of a package of the same repository (ADR 0573 §2) is "no module of
that path in the fetched tree". A step's workspace holds, beside the artifact's paths and module.json, every
`generated-from` directory an artifact of the module names, and the directory (with its go.mod) of every package of
the same repository the artifact's go.mod requires at the zero version, found by the module.json files of the fetched commit (the same manifests the catalog revision holds); nothing else is
fetched. Whether a change in such a directory moves the artifact's key (ADR 0499 clause 4) is a finding for H2. The published version of a package of another repository is
used as before (ADR 0573 §4). P3, P7.
- **H-A-137 (a root module's `generated-from "."`, 13:05)** Met in the review of H-A-136's build: a module whose
manifest is the repository's root (mesh-contracts) names `generated-from "."` for its generated packages, and "the
generated-from directory" read literally is the whole repository, against "nothing else is fetched". For a root
module the step's workspace keeps, beside the artifact's paths and module.json, the generator's inputs and outputs:
the `seats/` directory the generator reads and the paths of every artifact of the module that names
`generated-from` (the generator's check reads every generated package, and one left out counts as missing). Nothing
else of the root. A module in its own directory keeps that directory, as H-A-136 says. P7, P3.