main-cd: plugins-bake fails with CS0234 since MeshWeaver.Maps left the platform — no plugins publication sealed since 2026-09-01T06:43Z #3069

Activity

1. rbuergi commented on Sep 2, 2026

   

   rbuergi

   on Sep 2, 2026

   Contributor

   More actions

   The predicted red, confirmed — and explicitly NOT a verdict on #3071.

   CD run 33606528599 reached the seal at ~09:0x and failed exactly as forecast:

   Bake (compiler --output), then gate a mesh against it (--seed)   FAILED
   CS0234 Error (line 33): The type or namespace name 'Maps' does not exist in the namespace 'MeshWeaver'
   CS0234 Error (line 53): …
   CS0234 Error (line 79): …
   

   Ten occurrences across the affected galleries, then exit 1.

   🚨 Read the built commit, not the run's head. That run built 1b044e5, which predates #3071 (19035c1ea, merged 08:31:56). Its bake therefore ran the OLD lane — compiling against the tester image's /app, which has no MeshWeaver.Maps — so this failure is the bug reproducing, not the fix failing. A run's head_sha is the trigger ref and is never a claim about what it built; conflating the two is what cost #3022 a reopen earlier today.

   The acceptance is still outstanding. It is the first CD run that publishes a commit at or after 19035c1ea, where #3071 makes the bake and its gate compile against the portal image's /app (the reference set) with the tester supplying only the CLI.

   Two attempts have not produced it yet, both for reasons unrelated to the fix:

   • run 33609421101 (push path, 19035c1) — cancelled, evicted from the single pending concurrency slot by the next main run;
   • run 33609653563 (reconcile, 19035c1) — success with everything skipped: the required check was absent/none (commit is 5 min old), so the gate correctly declined to publish a commit CI had not yet vouched for. It emitted the annotation "main 19035c1 does NOT have a complete image set" and still reported green. That is main-cd's reconcile can only ship HEAD, so it starves exactly when the repo is busy #3077.

   main's required check has since completed, and no PRs are queued behind it, so the next push-path or reconcile run should publish and produce the verdict.
2. rbuergi commented on Sep 2, 2026

   

   rbuergi

   on Sep 2, 2026

   Contributor

   More actions

   Fixed — the acceptance run sealed green. #3071's lane works.

   CD run 33610520524, which built 17ee6d9f (after #3071's 19035c1ea), completed success with the whole chain green:

   success  Promote: tag the full set (all-or-nothing)
   success  Bake platform content in the shipped image + publish (linux-x64)
   success  Plugins: bake + seal the publication for this identity
   

   The decisive steps inside the seal are the ones #3071 introduced, and the one that used to fail:

   success  Resolve the platform image (the reference set, the identity, the …)
   success  Pull the tester and the platform image and extract their /app
   success  Resolve the framework identity this bake will target — the PORTAL …
   success  Compose the gate host (the portal's /app + the tester CLI)
   success  Bake (compiler --output), then gate a mesh against it (--seed)   ← was CS0234
   success  Publish bundles to every portal target
   

   And it genuinely published rather than no-opping:

   sealed: …/prebuilt-bundles/sd245cf578b98012064a52e3b3bbb19f1/plugins/_complete
           (36 bundle(s), source 3993c83fa56c87e36428bef7ab41e1d42e3fa802)
   bake published: identity=sd245cf578b98012064a52e3b3bbb19f1 arch=linux-x64
           source=plugins bundles=36 targets-published=2
   

   Nothing to bake (this content is already sealed for this identity) was skipped, so this was a real bake, not a re-assertion of an existing seal.

   Why it works

   MeshWeaver.Maps is a MeshWeaver.Plugins project referenced by Memex.Portal.Gui, so it exists in the portal image and nowhere else. The old lane compiled the content inside the tester image, whose /app never carried it — hence CS0234 on the galleries. #3071 composes the gate host from the portal's /app (the reference set) with the tester supplying only the CLI, so the assembly is present at compile time.

   One thing this does NOT mean

   The seal proves the bake now compiles against the image that carries Maps. It does not mean Maps is riding as a module. platform-shipped.txt still lists MeshWeaver.Maps, while core dropped the projects in #2941 — the halfway state that file's own header forbids ("Step 3 must land with step 2, not after it"). That remains MeshWeaver.Plugins#1061 step 3, and the 409 hazard its body describes is still live for whoever takes it.

   Also worth noting for the next reader: the earlier red on run 33606528599 was not a verdict on #3071 — that run built 1b044e5, which predates it, and therefore ran the old lane. A CD run's head_sha is the trigger ref, never a claim about what it built.
3. rbuergi

   closed this as completedon Sep 2, 2026