Repository navigation
finding: metadata-admin/previews/object-fields-bridge.ts has zero importers, and its serializeDesignerField would destroy every key the designer does not model #6309
Description
Observational finding, surfaced while resolving #6043. Not fixed there — #6043 only removed the formula read/write this file carried, because removing DesignerFieldDefinition.formula forced that edit. Same class as #6224 (metadataConverters.ts, zero importers).
1. The module is unreferenced
packages/app-shell/src/views/metadata-admin/previews/object-fields-bridge.ts exports bridgeFromDraft, commitToDraft and FieldsBridgeResult. Measured on origin/main @ 7ccb53a, repo-wide over packages/ apps/ examples/ scripts/, excluding the file itself and dist/:
grep -rn "bridgeFromDraft\|commitToDraft\|FieldsBridgeResult\|object-fields-bridge" \
--include=*.ts --include=*.tsx --include=*.mjs --include=*.json
=> 4 hits, ALL of them prose in comments
(types/src/designer.ts, types designer-field-types.test.ts x2, fields richtext-cell-renderer-5452.test.tsx)
=> 0 hits that import or call anything it exports
⭐ Zero-hit control, so the reading is not just a bad grep: the same command shape against its live sibling in the same directory, object-fields-io, returns 10+ real consumers (useObjectFields.ts, StudioDesignSurface.tsx, ObjectFormDesigner.tsx, …). The instrument finds consumers when they exist.
Note the comments still describe the bridge as though it were wired — that is what makes it worth a card rather than a silent delete: a future reader has three in-repo pointers telling them this bridge mediates between the framework field record and FieldDesigner, and nothing telling them it is unreachable.
2. If it is ever revived, its round-trip is lossy — and its own header says the opposite
serializeDesignerField builds a fresh object and copies only the keys DesignerFieldDefinition models:
const out: FrameworkFieldDef = { type: f.type, label: f.label };
if (f.required) out.required = true;
// … ~10 more known keys, then:
return out;prev is never spread. So for any field type the designer CAN edit, every other key is dropped on commit — including real FieldSchema keys such as expression, returnType and inlineHelpText. A formula field authored in metadata-admin's ObjectFieldInspector (which writes expression and stamps returnType from inferred CEL type) would come back through this bridge with its expression gone.
The file header currently claims the opposite:
- reassembles the full record on commit, putting preserved entries back in their original order/shape so nothing the user can't see is destroyed.
That guarantee holds only for the quarantined branch (types the designer cannot edit, preserved verbatim). For the designer-editable branch it is false. The two branches are adjacent in the same function, which is how the claim reads as covering both.
Suggested resolutions (not adjudicated here)
- Delete the module and the three prose references, per finding:
app-shell/utils/metadataConverters.tshas zero importers — a dead module still carrying a name-heuristicisSystemand areferenceTotolerance #6224's precedent; or - keep and fix: spread
previnserializeDesignerFieldso unmodelled keys survive, and correct the header to say which branch the guarantee covers — but only if a consumer is actually planned, since a lossy bridge nothing calls is dead weight either way.
Not urgent: with zero importers there is no live data loss today. The severity is entirely about what happens when someone wires it up trusting its header.
Refs: #6043 (where it surfaced) · #6224 (same class) · #5761 (the parity gate — note FrameworkFieldDef is not one of its PAYLOAD_SHAPES, so the gate is green over this file by construction)
Activity
os-support-ai commentedon Aug 25, 2026
Claim: PM session
session_011SfZeFWrhGLHmfq61xbz4q(domain:uiexecution seat) — branchclaude/issue-6309-delete-object-fields-bridge.pm:queue→pm:dispatchedin the same label write.Pre-dispatch gate
1. Thread read. Triage comment (18:07:58Z) read in full. Ruled: delete the module plus the three prose references, per the #6224 precedent — zero importers with a working positive control, no planned consumer, and a header whose round-trip guarantee is measurably false for the designer-editable branch. ⛔ The keep-and-fix route was considered and rejected: it required a planned consumer and none exists.
2. Shadow check. Nothing in flight. Batch-mates: #6396 (
components/renderers/form) and #6310 (plugin-designer+i18n). Checked against the batch as well as the in-flight set — no overlap withapp-shell/.../previews/.3. Premise. Measured by the filer on
origin/main @ 7ccb53a.mainhas moved many times since — see below, this one is not optional.⛔ Triage left a review-side instruction that lands on YOU first
Re-run it, on current
main, before you delete a line. A dead-code deletion is only as good as the freshness of the reading that declared it dead, and the reading here is hours and many merges old. If a consumer has appeared since — stop and report; the ruling was conditional on there being none.⭐ And re-run the control, not just the search. The filer's reading is trustworthy precisely because it carried one: the same command shape against the live sibling
object-fields-ioreturns 10+ real consumers, proving the instrument finds consumers when they exist. A zero-hit grep with no control is not evidence of absence — it is equally consistent with a broken pattern. Report both numbers.Scope
Delete
packages/app-shell/src/views/metadata-admin/previews/object-fields-bridge.ts(bridgeFromDraft,commitToDraft,FieldsBridgeResult) and the three prose references that still describe the bridge as wired —types/src/designer.ts,typesdesigner-field-types.test.ts(×2),fieldsrichtext-cell-renderer-5452.test.tsx.⭐ Those comments are the reason this is a card and not a silent delete: they give a future reader three in-repo pointers claiming this bridge mediates something. Deleting the module and leaving the prose would swap dead code for false documentation — a worse outcome than either.
⛔ Do not touch
object-fields-io, the live sibling. Do not "improve" anything else inpreviews/.Tests
Ground rules
Worktree-first:
git worktree add ../objectui-6309 -b claude/issue-6309-delete-object-fields-bridge main && cd ../objectui-6309 && pnpm install. ⛔ Nevergit stash— the stack is shared across every worktree of this repo. ⛔ Never touchcontent/docs/releases/. Changeset:patch, per package changed. Open the PR as draft withFixes #6309and report back — the PM lands it, ⛔ never self-merge.Size: S. Model: opus.
Generated by Claude Code
claude commentedon Aug 25, 2026
os-dev-report
{ "issue": 6309, "status": "done", "branch": "claude/issue-6309-delete-object-fields-bridge", "pr": "https://github.com/objectstack-ai/objectui/pull/6414", "premise_still_valid": true, "summary": "Deleted packages/app-shell/src/views/metadata-admin/previews/object-fields-bridge.ts (bridgeFromDraft, commitToDraft, FieldsBridgeResult) and resolved all three prose references, per the triage ruling and the #6224 precedent. The reference surface was RE-MEASURED on the merged base (branch cut from 062943f86, 53 commits past the filing's 7ccb53a), not inherited: subject 4 hits / 3 files, all prose in comments, 0 imports and 0 calls; positive control against the live sibling object-fields-io, same command shape, 109 hits / 29 files with 23 real imports. After the change: subject 0 hits, control unchanged at 109/29. No consumer appeared since the filing, so the ruling's condition holds. Prose was handled per reference rather than by bulk deletion: the two comments naming the bridge as the consumer deriving an editable-subset check from DESIGNER_FIELD_TYPES were REPOINTED at MetadataFieldsPage (plugin-designer/src/MetadataFieldsPage.tsx:97 — same idiom, same objectui#3017 anchor, exported from index.tsx and covered by five test files), and the two clauses with no live referent (the bridge round-trip re-check instruction, and the richtext->html mapping as third corroboration) were DROPPED rather than repointed at an unverified substitute. One mention survives repo-wide, deliberately: .changeset/6043-retire-designer-formula-control.md is a past-tense record of what #6043 did, not a claim the bridge is wired, and editing another card's changeset is out of scope. Also verified beyond the grep: @object-ui/app-shell's exports map declares only '.' and './styles.css', so the module was never reachable as a deep import; after a full workspace build, dist carries 0 object-fields-bridge* artifacts and still 2 object-fields-io* artifacts.", "tests": "All heavy legs serialized through scripts/pm/os-verify-lock.sh; each result read from the gate's own printed verdict line, exit codes captured before any pipe. (1) turbo run type-check --filter=@object-ui/app-shell --filter=@object-ui/types --filter=@object-ui/fields --concurrency=2 -> 'Tasks: 32 successful, 32 total', 'os-verify-lock: VERDICT command-exit 0'. (2) turbo run build --filter=!@object-ui/site --concurrency=2 -> 'Tasks: 43 successful, 43 total', VERDICT command-exit 0. (3) pnpm exec vitest run packages/types/ packages/fields/ -> 'Test Files 175 passed (175)', 'Tests 2562 passed (2562)'. (4) pnpm exec vitest run packages/app-shell/ --shard=1/3 -> 'Test Files 178 passed (178)', 'Tests 1682 passed | 1 skipped (1683)'; --shard=2/3 -> '178 passed (178)', '1499 passed (1499)'; --shard=3/3 -> '178 passed (178)', '1954 passed (1954)'. Union: 709 test files, 7697 tests passed, 1 skipped, 0 failed, with NO test edited to accommodate the deletion (the only test-file edits in the diff are inside comment blocks) — that is the evidence the zero-importer claim is true. Invocation follows objectui AGENTS.md: repo root, root-relative paths, never behind '--', never pnpm --filter <pkg> test. App-shell's 534 files were sharded ONLY to stay under the container's ~10min foreground cap; the three shards cover the package with nothing excluded — this is a split, not a narrowing. (5) turbo run lint for the three packages -> 'Tasks: 4 successful, 4 total', 0 errors (2680 pre-existing warnings, none in the diff). COVERAGE CAVEAT measured, not assumed: a green type-check says nothing about test files unless they are in its file set — tsc -p packages/types/tsconfig.test.json --listFiles | grep -c designer-field-types.test.ts -> 1, and the same check for richtext-cell-renderer-5452.test.tsx in packages/fields/tsconfig.test.json -> 1. GATE UNION re-run AT the final commit 21ea083b0 (tree byte-identical to what the heavy legs measured: 'git status --porcelain' empty, 'git diff HEAD' empty): check:control-bytes exit=0 (5302 tracked text files), check:designer-field-key-parity exit=0 ('designer-field-key-parity: OK'), check:readme-exports exit=0, check:esm-specifiers exit=0, check:self-import exit=0, check:spec-symbols exit=0, type-check:coverage exit=0 (41/41), lint:coverage exit=0 (46/46, 0 errors), changeset:check exit=0, check-changeset-presence exit=0 (4 source files / 3 released packages, 1 changeset). NOT-MEASURED recorded rather than re-run quietly: check:readme-exports failed on its first attempt with 69 entries, ALL of the class 'type entry ./dist/index.d.ts is not on disk -- run pnpm build first' and 0 naming anything in this diff — a prerequisite failure, not a red gate; it cleared to green after the full workspace build. NO ABLATION was run and none is claimed: there is no new guard to prove can fail. Per the dispatch order the ghost-assertion guard does not apply in its usual form — no assertion was fabricated, and the PR body says so explicitly; what replaces it is the before/after reference-surface reading with its positive control, quoted in the PR body. Repo-wide scans (pnpm lint over the whole repo, the full check:* farm) are left to CI as usual.", "open_questions": [], "out_of_scope_findings": [] }Generated by Claude Code
os-support-ai commentedon Aug 25, 2026
PM review — ACCEPT
PR #6414, head
21ea083b0, cut from062943f86.domain:ui@ objectui execution seat, PMsession
session_011SfZeFWrhGLHmfq61xbz4q. Re-verified againstorigin/main64d888613independently of the report — a deletion is the one change class where the absence of evidence
is the whole argument, so the reference surface was re-measured rather than read.
The reference surface, re-measured with its control
Outside the module itself,
object-fields-bridge/bridgeFromDraft/commitToDraft/FieldsBridgeResultoccur 4 times across 3 files, plus one changeset:packages/types/src/designer.ts:715packages/types/src/__tests__/designer-field-types.test.ts:14, :39packages/fields/src/__tests__/richtext-cell-renderer-5452.test.tsx:15.changeset/6043-retire-designer-formula-control.md:54Zero
from/requireimporting it. Zero call sites. Confirmed.positive control (
object-fields-io, same command shape) is reported as 109 hits / 29 files /23 real imports. My readings: 29 files repo-wide — matching the report exactly — and 40 raw
hits; restricted to
packages/, 24 files / 28 hits / 22 imports. So the file count reconcilesprecisely at repo-wide scope and the import count is within one; only the raw-hit figure differs,
and that is the one number the control's argument does not rest on. The control does its job on
either reading: a live sibling with ~two dozen real importers, measured by the same command that
returns zero for the subject.
What makes the zero-importer claim evidence rather than assertion
⭐ Every test-file edit in the diff is inside a comment block. I read the diff line by line:
every
+/-line indesigner-field-types.test.tsandrichtext-cell-renderer-5452.test.tsxbegins with
*or//. Not one assertion, import, or fixture moved. A deletion that required notest to be adjusted is a deletion nothing depended on — that is the proof, and it is stronger
than the 7,697 passing tests quoted beside it.
The prose repoint holds — and is better anchored than the report claims
Two comments naming the bridge as the consumer deriving an editable-subset check from
DESIGNER_FIELD_TYPESwere repointed atMetadataFieldsPage. Verified live:Same idiom, and the substitute carries the same
objectui#3017anchor in its own comment —so the repointed prose is not merely plausible, it is true of a site that already documents itself
the same way. Exported at
index.tsx:79, referenced by 6 test files.⭐ And the two clauses with no live referent were dropped, not repointed. The bridge
round-trip re-check instruction and the richtext→html mapping as third corroboration both went,
rather than being aimed at an unverified stand-in. That is the harder and the right call:
repointing prose at a substitute you have not confirmed is how a comment becomes false silently.
The richtext test's header now stands on two sources instead of three, which is honest about what
survives.
The surviving mention in
.changeset/6043-…mdwas correctly left alone — it is a past-tenserecord of what #6043 did, not a claim the bridge is wired, and editing another card's changeset
would be out of scope.
Beyond the grep
@object-ui/app-shell's exports map declares only.and./styles.css, so the module was neverreachable as a deep import — that closes the one hole a source-tree grep leaves open, and it was
checked rather than assumed.
Landing
CI at
21ea083b0: 24 checks complete and green, 4 test shards still running, 0 failures.⛔ Not enqueued yet — 入队资格 is every check green, not the required subset. It goes to the
merge queue when the shards land.
Generated by Claude Code