Repository navigation
Finish the bake/gate split: the PR lane and the satellite lane still bake through a mesh #2064
Description
#1763 split BAKE from GATE and moved the platform's own lane (main-cd.yml's publish-bake) onto both halves — PR #2059. Two lanes are deliberately still on the fused, mesh-driven bake, each for a different reason.
1. The PR lane (dotnet-test.yml, doc-gate)
It still runs mw-plugin-test <stage> --allow … --bake-output "$BAKE" for the Doc stage and again for the samples stage. Its own comment states the purpose:
--bake-outputhere is a SMOKE TEST of the bake stage, NOT a delivery lane (#1725). … it makes a bake-stage regression fail on the PR that causes it rather than in CD.
That is exactly right, and it is now aimed at the wrong code path: CD bakes with compile, so a regression in the compiler-driven bake reaches CD without ever reddening a PR.
Blocked on #2063. The tree baker discovers NodeTypes the gate's own discovery skips (BOM'd .json), so switching the lane today newly gates types absent from .github/samples-gate.allow and reds every PR for a reason that is not the change. Fix #2063 first, let the ratchet absorb what it reveals, then switch the lane to:
mw-plugin-test compile "$STAGE" --output "$BAKE" --allow .github/doc-gate.allow --source-sha "$GITHUB_SHA"
mw-plugin-test "$STAGE" --seed "$BAKE" --allow .github/doc-gate.allow
Roughly cost-neutral: the gate adopts instead of compiling, so the tree is compiled once either way.
2. The satellite lane (node-repo-publish-bake.yml)
Untouched on purpose, and #1763's own comment says why:
A satellite calls the reusable workflow with its own pinned image, which may predate the verb — that lane needs a version floor or a fallback before it can switch, and switching it blind is how a satellite's bake goes dark without failing.
So it needs, in order:
- a way to ask an image whether it has the
compileverb (a probe step, or a floor on the platform version the pin resolves to —mw-plugin-test framework-identityis the existing shape for "ask the image a question"); - a RED failure when the pinned image is below the floor — never a silent fallback to
--bake-output, because a satellite bake that quietly keeps using the mesh is precisely the thing that stayed invisible for months; - then the same two-step as the platform lane.
Why both are worth finishing
Until they are, "the estate bakes without a mesh" is true of one lane out of three, and the two remaining ones are the ones that gate PRs and that feed the satellites' portals.
Activity
Merge pull request #2250 from Systemorph/fix/2063-gate-discovery-equa…
rbuergi commentedon Aug 25, 2026
Stage 1 landed in #2250 (merged). Narrowing to the satellite lane.
Done — the PR lane now bakes with the compiler.
.github/workflows/dotnet-test.ymlno longer uses--bake-output; the plugins, doc and samples gates run the two-stepcompile … --output+<gate> … --seed, mirroringmain-cd.yml:1232/1272. The acceptance is not "the job is green": the log is grepped for the compile-fallback family (no prebuilt bundle,carried no assemblies — compiling instead,could resolve NO artifact,no prebuilt assemblies adopted) and fails on a hit, and the--seedpostcondition assertsdeclared = adoptedwithadopted > 0— the> 0being load-bearing, since the postcondition passes vacuously on an empty bake.Also done — #2063, which was the prerequisite: gate discovery now matches the baker's parser tolerances (BOM and otherwise), so
PensionFundreports 5 types where it previously reported 0. The gate had been passing over those types entirely.What remains here: the satellite lane —
.github/workflows/node-repo-publish-bake.ymlstill bakes through a mesh. Three parts, per the plan:compile --helpagainst the pinned image) that fails RED naming the pin, resolved digest and minimum digest — 🚨 never a fallback to--bake-output, since a satellite bake quietly reverting to the mesh is exactly what stayed invisible for months;:487-551, keeping both steps insidepublish-bake(node-repo-gate.ymlhas no--seedinput — do not add one);publish-bakein the estate isneeds:-gated behind gates that can go red and therefore renders a greyskipped, which is indistinguishable from a pass. Education demonstrated this in production: nine consecutive runs, ~10 hours, zero bundles published, nothing red.🚨 Sequencing constraint that has not changed: adopting the renamed reusable workflow renames the satellites' required-status-check contexts to
<caller job> / <name>, and MeshWeaver.Plugins is the one repo withstrict: true— a mismatch makes it unmergeable. Platform first, then Plugins (with its protection updated in the same change), then Education → Reinsurance → SocialMedia. Manufacturing is excluded by maintainer decision (2026-08-25).Two findings from #2250 worth carrying into the satellite work:
.github/samples-gate.allow'sPensionFund idempotenceentry is intermittent debt, not stale — the gate reported it "now passing" on one run and a "new failure" on the next, because whether it fires depends on whether the Currency NodeType has compiled. Do not remove it on a single green run; the underlying defect is Second install rewrites default-valued nodes: default-suppression makes the stored node differ from the authored file, so the unchanged-skip misses #2271.Merge pull request #2456 from Systemorph/ci/node-repo-bake-without-mesh
rbuergi commentedon Aug 28, 2026
Closing — both lanes named by this issue no longer bake through a mesh.
dotnet-test.ymlto the two-stepcompile … --output+<gate> … --seed(recorded in the 2026-08-25 status comment above).node-repo-publish-bake.yml— the sharedworkflow_calllane every node repo (Plugins, Education, Reinsurance, SocialMedia) uses. Verified at HEAD: the workflow's bake step is annotated "THE BAKE — compiler-driven, NO MESH ANYWHERE IN IT (Finish the bake/gate split: the PR lane and the satellite lane still bake through a mesh #2064)" and runsmw-plugin-test … compile /repo --outputfollowed by the gate with--seed /bake; the fused--bake-outputis documented as forbidden in the same file.bake-run-id) — so the published bundles are the very bytes that were gated.Still in flight but not a mesh-bake: MeshWeaver.Plugins#745 extends the bake-reuse to the plugins repo's own lane (an efficiency change; the producer there is already the compiler).
Related open defects that are NOT this issue: #2543 (per-node hubs stop answering mid-bake — a wedge inside the consuming gate mesh, not a producer/consumer split question).