finding(docs): AGENTS.md section 4's node interface diverges from the real BaseSchema — two members that do not exist, one unexported element type, three declared differently #8092

Filed by the domain:skills @ objectui execution seat while implementing PR B of objectui#7434 (PR #8090), unassigned, finding only — grading and domain:* are triage's. ⛔ Not repaired in that PR: ruling item 3 of that dispatch stated the fence is illustrative prose and is not to be expanded, and asked for this measurement to be reported instead.

⚠️ Angle-bracket shapes are spelled out in words throughout this card. GitHub's body sanitizer eats tag-shaped fragments even inside backticks and fenced blocks, and this card is entirely about type declarations — a literal rendering would collapse exactly the pairs it exists to compare.

The observation

Root AGENTS.md section 4, "The JSON Protocol (the DNA)", introduces its fenced block with:

Every node in the UI tree follows this shape (@object-ui/types):

PR #8090 renames the interface in that block from the unexported UIComponent to the ruled name BaseSchema (ruling 5556352576). That rename makes the block's claim stronger than it was: it now names a real, exported interface, and the sentence above says the tree follows this shape. The member list does not follow it.

Measured on packages/types/src/base.ts at origin/main 295804a6, against the nine members the fence lists:

Two further readings:

• The fence's events member names ActionDef as its element type. @object-ui/types does not export that name either — the exported one is ActionSchema (packages/types/src/crud.ts:40, which extends BaseSchema). So the block introduced one unexported identifier while the card was in flight to remove another.
• Fourteen members of the real interface do not appear in the fence at all: name, label, description, placeholder, style, data, body, visible, visibleWhen, visibleOn, hiddenOn, disabledOn, testId, ariaLabel.

Re-derive both halves:

sed -n '72,84p' AGENTS.md
awk '/^export interface BaseSchema \{/,/^\}/' packages/types/src/base.ts | grep -E '^  [a-zA-Z]+\??:'

Why it is worth a card rather than a silent edit

⭐ The failure direction is the one objectui#7434 exists to stop. That card's whole argument is that authors reach for whatever the instruction file last said, and that the sites regrow when the vocabulary is wrong rather than the instances. An author who copies this block now gets a name that exists (BaseSchema — that is PR #8090's contribution) attached to a member list that partly does not (props, events, ActionDef), and two members whose declared types would not type-check as written. That is a narrower error than before, and therefore a harder one to notice: the import succeeds and the mismatch surfaces later, at the member.

⚠️ Nothing catches it. AGENTS.md is read by no doc gate: check-doc-fence-languages.mjs sets its root pages to README.md only, and check-doc-snippet-types.mjs's own coverage ledger names the non-README.md root pages, AGENTS.md among them, as unscanned. So the block is never compiled, and no run has an opinion about it.

⛔ Not a mechanical repair — the decision this card asks for

There are at least three defensible answers, and picking one is a judgment about what an instruction file is for:

1. Leave it illustrative and say so — one sentence above the fence stating that the block is a teaching sketch, not the interface, with a pointer to packages/types/src/base.ts. Cheapest, and honest about what the block is; a reader who wants the real list has somewhere to go.
2. Make it true and keep it short — drop props, events and the ActionDef reference, correct the three declared types, and leave the fourteen absent members absent under an explicit "abridged" note. The block stays readable and every line in it is accurate.
3. Track the real interface — expand to all twenty-one members. ⛔ This seat would argue against it without a gate to hold it: an unscanned copy of a real interface is a drift generator, which is the class of defect this whole card family is about.

⛔ This seat is not choosing. It also does not think this is urgent: the block has read this way for a long time, and PR #8090 deliberately changed only the name it was dispatched to change.

Related, checked, not duplicates

Dedup run with a self-validating query — a repo-scoped semantic search for the vocabulary terms returned 3 on-topic results, so the absence of an exact match is a reading and not a broken channel. packages/types/README.md and content/docs/api/schema-reference.md were both repaired by PR #8064 and are not affected by this.

• objectui#7434 — the parent card. It rules the name; this card is about the member list underneath it. Its two prescribed PRs (docs(types,api): name the component node BaseSchema and pin the vocabulary #8064, docs(agents): name the component node BaseSchema in AGENTS.md section 4 #8090) neither touch nor mention the fence's members.
• objectui#8065 — the published skills/objectui tree teaches the same unexported name. Adjacent and same family, but it is about a third surface carrying the old name, not about a member list being wrong. ⚠️ Worth reading together: that card's two files carry the same fenced shape, so whatever is decided here probably applies there too.
• objectui#4631 — three declared surfaces of a component type disagreeing. Adjacent, about reconciliation between real declarations; this one is a doc block versus one real declaration.

Generated by Claude Code for the domain:skills @ objectui execution seat, session session_019RfFHiRCSs3JXLK4cwcfox.

Generated by Claude Code