[objectstack-fleet]

Requested by Elon · project thread

Fixes #26051

Clause-②: no
Metadata-protocol: none

Part of the p1 merge-queue backlog work (#26074). Dispatched by seat domain:devx#1 (#6023), claim 6415558140, card-thread session session_01SMJGPXxFBpbgAgXdnsXKVS.

What changes

scripts/sdui-manifest.record.json carried ONE inTree.registrySourcesSha256 over the src/ of all 27 ui registry packages, so any two ui PRs that touched any registry package conflicted on that line. Now:

• One record per registry package, one file each, one line each: scripts/sdui-manifest.sources/PACKAGE.json = {"package","sources","sha256","manifestSha256"} (the package's own src/ digest, its file count, and the sha256 of the artefact it was verified against). 27 files.
• The shared record (scripts/sdui-manifest.record.json, same path) keeps only what describes the artefact: sha256, components, generatedAt, and inTree = modulesRoot, objectuiWorkspaceVersion, registryRecords (the directory). The retired registryPackages / registrySources / registrySourcesSha256 are refused if they reappear.
• The reached package set is the set of record files: a reached package with no record (new package, new dependency edge into the registry, deleted record) is RED; a record the registry no longer reaches (removed package, dropped edge) is RED.
• sdui.manifest.json is byte-identical (sha256 1f6e1ccb925a…, 106 components).

lint.yml's command line (node scripts/check-sdui-manifest.mjs --self-test && node scripts/check-sdui-manifest.mjs) is unchanged and keeps working. No .github/** or ui/.github/** edit.

Measured first (Ask 1): own src/ per package, not dependency-inclusive

Tree: objectstack 0a6948210c (base), 27 reached packages, 922 non-test sources.

Detection is the same as one digest over the set. Any edit under any reached package's src/ moves exactly that package's digest, so a single PR's edit is caught exactly as before; a dependency's edit is caught through the dependency's own record. Dependency-inclusive digests add no single-PR detection. They would only add conflicts, and at this cost (records a src/ edit would touch):

package edited	own src/ (this PR)	dependency-inclusive
types	1	25
core	1	21
data-objectstack, i18n	1	19
react, react-runtime, sdui-parser	1	18
components	1	17
fields	1	14
plugin-detail	1	10
plugin-form, plugin-grid	1	2
the other 11 plugins	1	1

Merge experiment, real tree, git merge-tree --write-tree (no merge driver routes these paths; git check-attr merge = unspecified, so this is GitHub's answer): each branch appends a comment to one src/ file and runs the generator.

pair	before (single digest, on 0a6948210c)	after (this PR, on a8114011)
plugin-grid/src/FormulaBar.tsx × plugin-form/src/DrawerForm.tsx	exit 1, CONFLICT in scripts/sdui-manifest.record.json	exit 0, clean; the merged tree passes check-sdui-manifest (exit 0)
plugin-grid/src/FormulaBar.tsx × plugin-grid/src/ObjectGrid.tsx	(conflicts)	exit 1, CONFLICT in scripts/sdui-manifest.sources/plugin-grid.json only, by design

The open ui PRs today (registry packages each changes, non-test src/ only; none changes sdui.manifest.json): #25987 i18n, plugin-grid, plugin-list · #25998 plugin-detail · #26040 components, i18n · #26045 i18n · #26027 plugin-dashboard · #26030 plugin-grid · #26033 core, plugin-form, plugin-grid · #26042 plugin-form · #26044 i18n · #26048 i18n, plugin-map · #26050 fields, plugin-grid · #26062 fields. Under the single digest all 12 conflict pairwise (12 serial landings). Under this layout they conflict only within a shared package; the largest group is i18n (5 PRs), so they can land in 5 rounds, e.g. {#25987, #25998, #26027, #26042, #26062} → {#26040, #26030} → {#26045, #26033} → {#26044, #26050} → {#26048}, assuming no source-level conflicts between them.

The check (Ask 2)

checkTreeWithoutPin compares each reached package's digest with its own record and names every stale package (ui/packages/plugin-grid: now 38 file(s) sha256 …, recorded …), every unrecorded package, every record the registry no longer reaches, every record verified against another artefact, and every malformed record, then prints the one command: node scripts/gen-sdui-manifest-node.mjs. The pin mode (.objectui-sha present) is untouched.

Why manifestSha256 is in every package record, and why one line: a PR that changes the artefact re-records every package against the new sha256, so it still conflicts with every in-flight registry PR, which must regenerate on the merged tree (that is today's behaviour, kept for the case where it matters). Two PRs on the same package touch the same line and conflict. One line per file guarantees git cannot merge a digest from one side with an artefact sha256 from the other; a record that is not in the producer's one-line form is RED.

What this does not force, stated rather than implied: two PRs on different packages that each leave the artefact unchanged, where only their combination would change it, land without any run seeing the combination. Under the single digest their textual conflict forced that run. Here the next regeneration of any registry package sees the combination. Dependency-inclusive digests would close this only along import edges (not between siblings that interact through ComponentRegistry), at the fan-out above. No stale or missing registry source passes: every source edit still reds until regenerated, and every package added to or removed from the reached set still reds.

Self-test (Ask 3)

29 rows (was 22), roster and floor updated. New or changed in-tree rows: a staged edit to one package is RED naming that package and NOT naming plugin-form or core; a new package the registry reaches; a new workspace edge (now names ui/packages/unrelated); a removed package (names scripts/sdui-manifest.sources/types.json); a deleted package record; a record verified against another artefact; a record out of its one-line form; a record from the single-digest layout; and the producer re-recording only the moved package (it back-dates the shared record first, so a producer that stamps today's date reds the row).

Ablations, one at a time through scripts/ablation-replace.mjs, each restored with git diff HEAD empty:

guard removed	self-test
the other-artefact comparison (gate)	exit 1, row "a package record verified against another artefact is RED" expected RED got GREEN
the one-line form comparison (gate)	exit 1, row "a package record out of its one-line form is RED"
the unreached-record list (gate)	exit 1, row "a removed registry package is RED naming its record"
the generatedAt carry (producer)	exit 1, row "the producer re-records only the package whose sources moved": the second run changed the shared record too
the per-package record write (producer)	exit 1, the two producer rows

The generator (Ask 4)

It re-records only the packages whose line changed and leaves the others byte-identical; it removes a record the registry no longer reaches; it prints which packages it re-recorded. The shared record moves only with the artefact: generatedAt and the provenance version are carried while the artefact's sha256 is unchanged (the same rule the pin-keyed fields already follow), so a regeneration on another day no longer moves a shared line.

It still builds, and why: the build plus the enumeration is the only oracle that says whether sdui.manifest.json moved. A record written without it would stamp a current digest on a manifest no run checked, which is the narrowing this card forbids. The 8 minutes is a cold turbo cache, not the build of one package. Measured on 0a6948210c, one container, no remote cache (TURBO_TOKEN unset):

run	wall	turbo
cold	8m49s	0 of 63 cached, 8m43s
nothing moved	6s	63 of 63 cached
one edit under plugin-grid/src	33s	61 of 63 cached, 27s
one edit under core/src	2m50s	41 of 63 cached, 2m44s

Readers moved (Ask 5)

• docs/releases-maintenance.md: the regenerate section describes the per-package records, the re-record rule and which PRs still conflict.
• scripts/check-changeset-no-major.mjs: the "measured out" comment, the real-build assertion (neither record is shipped) and the records-only control now cover scripts/sdui-manifest.sources/. Its --self-test passes (456 assertions).
• scripts/ui-import/README.md is not edited: line 140 records what the cutover's step 4 wrote at the time, and that file still exists at that path.
• git grep -n "sdui-manifest.record" outside CHANGELOGs: only the two scripts, the docs page, check-changeset-no-major.mjs and that historical README line. git grep registrySourcesSha256: only the retired-field list and its self-test row.

After this lands: what an in-flight ui PR does (once)

A ui PR whose branch re-recorded the old single digest conflicts once on scripts/sdui-manifest.record.json:

git merge origin/main                                   # conflict in scripts/sdui-manifest.record.json
git checkout --theirs scripts/sdui-manifest.record.json  # take main's layout
node scripts/gen-sdui-manifest-node.mjs                  # re-records only this PR's package(s)
git add scripts/sdui-manifest.record.json scripts/sdui-manifest.sources sdui.manifest.json
git commit

After that, PRs on different registry packages no longer conflict with each other on the record.

Gates

Local, on a8114011 (derived with node scripts/pm/dispatch-gates.mjs --commands): 39 of the 40 derived families exit 0, including node scripts/check-sdui-manifest.mjs --self-test (29 cases) and node scripts/check-sdui-manifest.mjs, node scripts/check-changeset-no-major.mjs --self-test and --base origin/main, pnpm check:nul-bytes, pnpm check:watch-hint-literal, pnpm check:required-contexts, pnpm check:doc-authoring and pnpm --filter @objectstack/lint run check:doc-formula-expressions (after the lint closure build). pnpm check:pm-dispatch-gates was still running when this PR opened; its reading goes into the report comment on #26051, and CI runs it in Lint & Repo Gates.

CI on the PR's check-runs: pending at the time of writing; the seat reads them.

Changeset: none, skip-changeset. The diff touches only root scripts/ (private), docs/ and the records; sdui.manifest.json is byte-identical and neither record is copied into packages/console/dist/ (the check-changeset-no-major real-build assertion above).

Not Tier H: no governed path (git diff --name-only has no .github/, .claude/, docs/adr/, skills/, AGENTS.md, CLAUDE.md).

Acceptance notes

• i18n is the hot package: 5 of the 12 open ui PRs touch its src/, so they still serialize on scripts/sdui-manifest.sources/i18n.json. Whether locale data belongs in the digest at all is a narrowing question (it would let a source change pass unregenerated), not something this PR decides.
• The cold-cache cost of a regeneration (8m43s here) is a fresh-container property; card threads run with no turbo remote cache.

---

Generated by Claude Code

[objectstack-fleet]

席位核对 — ACCEPT at bc64912d

Seat review of #26120 for #26051, run under p1 #26074 (devx seat #6023).

• Scope: scripts/check-sdui-manifest.mjs, scripts/gen-sdui-manifest-node.mjs, scripts/sdui-manifest.record.json, the new scripts/sdui-manifest.sources/*.json (one one-line record per registry package), and the readers that moved with them (scripts/check-changeset-no-major.mjs, docs/releases-maintenance.md). .github/** is untouched, so the lint.yml command line runs unchanged. Not governed.
• Measured as the card asked: own-src records keep single-PR detection the same, while dependency-inclusive digests would touch 21 (core), 25 (types) and 18 (react) records. Every source edit still reds until it is regenerated, and every package added to or removed from the reached set still reds.
• Proven twice on real merges: after fix(plugin-list,plugin-grid): list empty state tells no-access and "mine" apart from no records #25987 the gate named exactly i18n, plugin-grid and plugin-list, and the generator re-recorded only those 3 of 27. After fix(types): the @object-ui/types/zod mirrors keep an unknown key instead of silently dropping it #26019 it re-recorded only types.json.
• Residual: ruled by the maintainer. Two PRs on different packages that each leave the artefact unchanged, but would change it together, are no longer forced to regenerate together. The next regeneration finds it. The seat put this to the maintainer, and he chose 「接受残余,落地」 (2026-10-26 13:28Z, devx seat thread). The structural fix, where CI regenerates and compares so the merge queue checks the combination, is filed as devx(sdui): CI regenerates the sdui manifest and compares it to the committed artefact, so ui PRs conflict only on real contract changes and the merge queue catches combined effects #26137 (Tier H, after this PR).
• skip-changeset: private root scripts and docs only.

Landing: pr_ready + automerge_enable once every check on bc64912d is green, under the maintainer's standing grant. He may move it up the queue by hand. Every ui registry PR that lands first costs this PR one more re-merge. Fixes #26051.

Card: #26051

---

Generated by Claude Code