Repository Dependency Direction

Systemorph/MeshWeaver is the platform. MeshWeaver.Plugins and the four satellites consume it. The dependency runs one way, and every edge that runs the other way is a way core's own build or release can be broken by a repository it should not know about.

Plugins depending on core is CORRECT and must not be "fixed": $(MeshWeaverRoot) project references, the digest-pinned platform image the in-mesh gates compile against, and a plugin repo calling core's reusable node-repo-*.yml / plugin-build.yml workflows are all the intended shape. This page is only about the other direction.

Measured against Systemorph/MeshWeaver@cea28557b. Re-measure before acting on it; line numbers move. See also Carving Projects Out Of Core, which covers the moves that CREATE these edges.

The inventory

A · The image lanes build plugin source — structural, not a defect to delete

Memex.Portal.Distributed (the portal host) and Memex.Database.Migration (the migration worker) LIVE in MeshWeaver.Plugins since #2293. Core's image lanes therefore check that repo out and dotnet publish from it. Without the checkout there is no project to publish — this is not incidental coupling that can be tidied away.

🚨 The blast radius is the FLEET, not this repo. A broken plugins tree does not merely red a core job. The release wave itself is memex's — the build fact still reaches Hosting/PlatformBuilds and memex still dispatches meshweaver-framework-released to every subscribed repository — but each satellite's bake seeds the Plugins publication for the released identity, so with plugins-bake red every one of them fails upstream 'plugins' has no SEALED publication and the fleet adopts nothing, reporting FrameworkDeclined at its next fetch. That is the cost of the edge, stated plainly.

🚨 ref: main is the dishonest part, not the checkout. A live checkout of a sibling's moving default branch means "which plugins commit shipped with which core commit" is decided by whoever merged last, and the same core commit can build differently on two runs. The variable MW_PLUGINS_REF makes that coupling steerable during an incident; it does not make it a pin.

The honest seams, in order of increasing cost:

1. Pin, do not follow. Core records the plugins commit it builds against — a lockfile in core, or MW_PLUGINS_REF set to a sha rather than main — and a plugins release proposes a bump the same way a package upgrade does. The image set already keys on the PAIR (check-image-set.sh takes the plugins short-sha), so the identity machinery is in place; what is missing is that the pair is chosen rather than observed.
2. Move the image build to the repo that owns the host. Plugins already builds and tests these projects; the lane that publishes them could live there and consume core as a released artifact. This inverts nothing else and removes every edge in the table.
3. Move the hosts back to core. Correct by construction and the largest change; it also gives back the NuGet publication a source move costs (see the sibling page).

Until one of those lands, the coupling stands and should be stated, not discovered.

B · A gate whose subject is core, living in the plugin repo — a defect, fixed

DocumentationLinkIntegrityTest and DocumentationEmbedIntegrityTest pinned this repository's doc tree (src/MeshWeaver.Documentation/Data) from MeshWeaver.Plugins/src/MeshWeaver.AI.Test/, because they check the union of the Doc, Agent and Skill partitions and the latter two are embedded in MeshWeaver.AI, which ships from there.

The consequence: a core pull request that broke a doc link or a @@ embed went green here and turned a DIFFERENT, private repository red — hours later, on a change it did not make. A required gate sitting behind a sibling checkout is the sharpest form of an inverted dependency, because the repo that owns the subject has no way to run it.

The fix is a SPLIT, never a delete: core now carries the Doc-only half in test/MeshWeaver.Documentation.Test/, and the plugin repo keeps the union — the repo that holds BOTH assemblies validates cross-partition links, which is the correct direction. Nothing became unpinned; core simply stopped depending on someone else to notice.

🚨 A gate that moves and leaves its subject unpinned is not fixed, it is unpinned — with nothing turning red to say so. When a gate cannot move whole, split it along the assembly boundary and say in both copies what the other one covers.

C · A core PR gate that reads the fleet — deliberate, and worth naming

dotnet-test.yml:1768 (shared-rules) reads AGENTS.md from all seven repos and is a needs: of collect-results (:1955), this repo's only required status check. A shared-rule block that drifts in a satellite turns core red and blocks core merges until it is fixed there.

That is chosen rather than accidental, and the reasoning holds: a per-repo self-check cannot detect the missed-repo case, because a repo with no pull request never runs its own gate — "six of seven were updated" and "all seven were updated" produce identical evidence. The scheduled half (shared-rules.yml) exists for the weeks nobody opens a pull request here.

It is listed anyway, because anyone auditing "does core's build touch plugins?" must be able to find it rather than rediscover it.

dotnet-test.yml (cross-repo-pair) is the SECOND edge of this class, added for #2689. It resolves the pull request a surface-removing change declares (Pairs-with: Systemorph/MeshWeaver.Plugins#904) and refuses to let core merge until that counterpart is merged into its repo's default branch — see The Cross-Repo Pair Gate. It exists because a public type leaving core's src/ reddened MeshWeaver.Plugins' trunk for two hours on a change none of that repo's pull requests had made, and core's CI cannot see it: core does not build the plugin repos, so the coupling can surface nowhere except CD, after the fact.

🚨 The distinction that makes this class permitted while a checkout is not: a checkout puts another repository's SOURCE into core's build; an API read puts only a FACT about it into a verdict. Core still compiles, tests and ships with no sibling on disk. Every edge of this class is ASSERTED as well as documented — PlatformNeverDependsOnPluginsGuard.ApiReadLedger enumerates them and fails in both directions, so a new one has to be a decision rather than a diff nobody noticed.

There is deliberately NO third edge. A dispatcher that asked MeshWeaver.Plugins to build against a core pull request (dependent-suites.yml, #3103) existed for a few hours on 2026-09-03 and was withdrawn by the maintainer: "none of the top-level repos should have any dependency to anyone else." Even an event-plus-read is a dependency — a red on a core pull request that only a plugin repository can turn green couples the two repositories' trunks. The rule is the strict one: core reads facts about sibling repositories (the two ledger entries above) and sends them nothing; a plugin repository discovers a platform break in ITS OWN CI when ITS platform-ref moves, and fixes it there. PlatformNeverDependsOnPluginsGuard.ApiReadLedger lists exactly the two permitted reads.

What is NOT an inverted edge

Checked and cleared — do not "fix" these:

• plugin-build.yml, node-repo-{validate,compile-check,gate,tag-modules,publish-bake,module-pack}.yml — workflow_call only. A plugin repo calling core's reusable workflow is the intended shape. plugin-build.yml carries an explicit note that its cross-repo checkout was REMOVED and must not be re-added on the strength of a grep.
• No dispatcher. Core sends repository_dispatch to no repository. notify-dependents.yml was withdrawn on 2026-09-03 (maintainer: "core publishes an event and finishes"): the release wave is emitted by memex from the build fact core POSTs into Hosting/PlatformBuilds, to the repositories the Hosting/Deployment records name as registry sources. PlatformReleaseNotifyGuard.CoreDispatchesToNoRepository refuses a /dispatches POST in ANY workflow under .github/workflows — the reusable lanes included: node-repo-publish-bake.yml no longer wakes dependents itself, it ENDS by registering the publication with memex (register-publication), and memex emits meshweaver-upstream-published.

The contract (maintainer, 2026-09-03: "end of github pipeline must call memex, which must register release and publish event") is three sentences:

1. Every publishing pipeline ENDS with one call to memex. Core's CD, after the image set is promoted, POSTs the signed platform build (event: platform-build) into the control instance's Hosting/PlatformBuilds inbox (notify-platform-update). Every node repository's node-repo-publish-bake.yml run, after its bundles are sealed for an identity, POSTs the signed publication record (event: bundle-publication — source, identity, commit, tester + portal image) into the same inbox (register-publication, its last job). Nothing runs after that call, and no pipeline sends a repository_dispatch to another repository.
2. memex REGISTERS the release as a durable node — Hosting/PlatformBuilds/<version> for a platform build, Hosting/Publications/<identity>/<source> for a bundle publication — the source of truth for "what is published for which identity" (what the self-update availability check reads).
3. memex PUBLISHES the event from that registration: FrameworkReleaseBroadcaster sends meshweaver-framework-released (platform) or meshweaver-upstream-published (bundle publication, client_payload.version = the identity) to the subscribed repositories — the repositories the control instance's Hosting/Deployment records name as registry sources. The subscribers' CI receives it, resolves both images from the version, builds and publishes for that identity — and ends by calling memex (1).

 pipeline (core CD | a node repo's publish-bake)        memex (control instance)              subscriber CI
 ───────────────────────────────────────────────        ────────────────────────              ─────────────
 promote / seal ✅                                       WebhookInbox Hosting/PlatformBuilds
   └─ ONE signed POST ──(platform-build |──────────────▶│ verify HMAC
      bundle-publication)… and FINISH                    ├─ REGISTER  Hosting/PlatformBuilds/<version>
                                                         │            Hosting/Publications/<identity>/<source>
                                                         ├─ subscribers = Hosting/Deployment records'
                                                         │              pluginRepos[].isRegistrySource
                                                         └─ PUBLISH   repository_dispatch ─────────────▶ on: repository_dispatch:
                                                            meshweaver-framework-released |               types: [meshweaver-framework-released,
                                                            meshweaver-upstream-published                        meshweaver-upstream-published]
                                                                                                          → bake for the version → seal → POST memex

Where the pieces are: the POST steps in main-cd.yml and node-repo-publish-bake.yml (this repo); the inbox watcher, registration and broadcast in the Hosting module's PlatformBuildInboxWatcher (MeshWeaver.Plugins, Hosting/Deployment/Source); the broadcaster in src/MeshWeaver.GitSync. PlatformReleaseNotifyGuard.CoreDispatchesToNoRepository refuses a dispatch SENDER in any workflow under .github/workflows — there is no ledger — and UpstreamBuildGateGuard.TheLaneEndsByRegisteringWithMemex_AndDispatchesToNobody pins the lane's call.

• scripts/check-type-forwards.py --sibling — the flag can resolve a departed type against a sibling checkout, and CI deliberately does NOT pass it (dotnet-test.yml): without the flag the gate is the conservative one, and passing it would make a core verdict depend on another repo's moving HEAD.
• The ~100 prose mentions in comments, and the repository NAME used as test data (ReleaseFactTest, FrameworkReleaseBroadcasterTest, the self-update suites). A string is not a dependency; none of these reads the repo.
• Every $(MeshWeaverRoot) reference from Plugins into core — twelve projects reference MeshWeaver.Documentation.csproj alone, three of them production. That is the arrow pointing the right way.

The rule

A gate belongs in the repository that owns what it measures. When you move a project out of core, the gates that measure the code STAY with the code — but a gate that measures core content while merely needing a plugin assembly to do it belongs here, split along the assembly boundary. And when core genuinely must build from a plugin repo, it pins a commit; it does not follow a branch.