objectstack-ai
diff --git a/‎.changeset/meta-unrecognised-type-refused.md‎
Lines changed: 149 additions & 0 deletions b/‎.changeset/meta-unrecognised-type-refused.md‎
Lines changed: 149 additions & 0 deletions
@@ -0,0 +1,149 @@
+---
+"@objectstack/spec": minor
+"@objectstack/metadata-protocol": minor
+---
+
+fix(metadata): `PUT /meta/:type` refuses a type name the platform does not have, instead of minting a namespace for it (#8421)
+
+<!-- adr-0087: not-required (no-migration-prescription) This narrows when an HTTP endpoint refuses. No authorable key, no stored shape and no spelling changes: `DEFAULT_METADATA_TYPE_REGISTRY` and the URL-spelling map are untouched, so `os migrate meta` has nothing to rewrite. Rows already at rest under an unrecognised type keep their shape, stay readable and stay deletable — the refusal is on the mint path only. #8586's own retirement entry (protocol 18) already carries the declared-kind half of this ruling. -->
+
+
+**BREAKING** accept-set narrowing on a published HTTP surface, landing after the
+v17.0.0 cut (the lockstep launch-window convention ships it as `minor`). A write
+that answered `200 {"success":true}` now answers `400 INVALID_REQUEST`:
+
+```
+PUT /api/v1/meta/fieldz/showcase_task.title
+  before → 200, sys_metadata row persisted with type='fieldz'
+  after  → 400 INVALID_REQUEST, nothing persisted
+```
+
+`fieldz` — or any typo — was neither a declared metadata type nor a known plural
+spelling of one, so the boundary classified it as PLUGIN-registered, which every
+authorization gate is permissive toward by construction. The row was persisted
+under a type nothing reads and nothing serves, and the caller was told it had
+succeeded. That silence is the real cost: a metadata-type typo, from a human or
+from generated code, produced `success: true` and no indication the type is not
+real.
+
+**Why this is only now safe to refuse.** #7894 closed the sibling case (a plural
+spelling of a type the platform DECLARES) and left this one open on purpose: a
+static predicate cannot tell `fieldz` from a plugin kind, and the live-registry
+alternative was measured to be worse than the defect — the live type set is
+ITEM-POPULATED, so it omits every legitimate kind that has no items yet, which
+is the state each kind is in immediately before its first create. What changed
+is the platform, not the boundary's information: #8586 retired
+`MetadataPluginConfig.additionalTypes` and with it the last channel by which a
+plugin could DECLARE a metadata kind, so an unrecognised name can no longer be a
+declaration this refusal has not heard about (maintainer ruling 2026-08-14).
+
+**What still passes, pinned in both directions.** Every declared type in
+`DEFAULT_METADATA_TYPE_REGISTRY`, in canonical and REST-plural spelling; every
+manifest spelling and the singular each folds to; and the six plugin kinds that
+have no static registry entry at all — `theme`, `webhook`, `connector`,
+`sharing_rule`, `analytics_cube`, `rag_pipeline`. `PUT /meta/theme/dark` on a
+deployment with zero themes is explicitly covered, because that first create is
+exactly what a live-registry check would have broken.
+
+**The refusal is scoped to the door that mints.** Reads still ANSWER: a running
+kernel legitimately holds live type keys the static contract does not — `data`,
+`kind` and `package` all enter the registry during an ordinary `registerApp`,
+and `GET /api/v1/meta/types` lists that live set — so refusing unrecognised
+names on the read path would answer 400 for types the same service advertises.
+`DELETE` is untouched for the mirror-image reason: rows minted under an
+unrecognised type before this change are real, nothing rewrites them on upgrade,
+and refusing their deletion would turn the accumulation this fixes into an
+accumulation nobody can clear.
+
+**…but one published ADVERTISEMENT narrows with it, and that is a second
+behaviour change worth reading on its own.** `GET /api/v1/meta/types` keeps
+listing every live type, and every entry keeps every field — what changes is the
+VALUE of one boolean:
+
+```
+GET /api/v1/meta/types  →  entries[] where type ∈ {policy, data, package, kind}
+  before → allowRuntimeCreate: true
+  after  → allowRuntimeCreate: false
+```
+
+The listing synthesised `allowRuntimeCreate: true` for every live type with no
+static registry entry, on the same expired premise as the write door: a name the
+registry does not carry might be a kind some plugin declared. It now derives that
+flag from the SAME predicate the mint door enforces, so the two endpoints agree
+by construction instead of via two rules maintained apart. Nothing ever honoured
+a runtime create on those four — they are internal bookkeeping (seed datasets,
+package rows, kind descriptors) — so the advertisement was a promise the platform
+did not keep, which is the same defect this card is about, relocated to the read
+door. Direct precedent: `api` declared `allowRuntimeCreate: true`, the runtime
+never honoured it, and the 2026-08-07 ruling removed the declaration rather than
+converging the read path onto it.
+
+⛔ The six plugin kinds with no registry entry — `theme`, `webhook`, `connector`,
+`sharing_rule`, `analytics_cube`, `rag_pipeline` — are **not** affected: they are
+in the static spelling contract, stay advertised `allowRuntimeCreate: true`, and
+stay mintable. A UI reading this field (Setup → Metadata, the Studio designers)
+therefore loses create affordances on exactly the four types whose creates were
+already refused, and keeps them everywhere else.
+
+**The premise behind both halves is a CURRENT posture, not a closed door.**
+Maintainer ruling, 2026-08-15, verbatim and untranslated:
+暂时不考虑让插件申明新的元数据类型 — plugins do not declare new metadata types
+*for now*. That word is recorded deliberately: plugin-declared kinds were
+considered and deferred, not ruled out. If they are ever wanted, the two sites
+that encode the deferral name it and its date in place —
+`getMetaTypes()`'s synthesis and `isRuntimeCreateAllowed` in
+`@objectstack/metadata-protocol` — so the decision is findable rather than
+re-derived from the code's silence.
+
+**Two shapes reaching the mint door are exempt, and each is a fact about the
+request rather than a claim the caller makes.**
+
+1. *The COMPOUND arity carries an OBJECT name in the `:type` segment.*
+   `PUT /api/v1/meta/lead/views/all_leads` is `type='lead'`,
+   `name='views/all_leads'` — one operation reaching one save, the shape both
+   the runtime dispatcher and the REST route document verbatim. `lead` is an
+   object, i.e. runtime data no static contract can enumerate, so a type verdict
+   applied there would refuse every object name that is not coincidentally a
+   metadata type. The ruling is about metadata TYPE names like `fieldz`.
+   ⚠️ Residue, stated rather than hidden: `PUT /meta/fieldz/a/b` is therefore
+   still accepted, because at that arity `fieldz` is a claim about an object and
+   the only way to check it is the live-registry lookup this card ruled out.
+2. *A namespace that already exists is not being minted.* `duplicatePackage`
+   re-saves every row of a package under a new name, taking each type from the
+   stored row — measured: a package holding one pre-existing residue row
+   answered `{success: false, copiedCount: 0, failedCount: 1}`, i.e. could not
+   be duplicated at all. That contradicts the `DELETE` reasoning above, so the
+   store (never the request) exempts a type that already has rows. The probe
+   runs only once the refusal has already fired, and a store that cannot answer
+   refuses — a fresh deployment has no residue to protect.
+   `migrate meta --stored` was read as a third victim and measured NOT to be
+   one: an unrecognised type has no manifest collection, hence no ADR-0087
+   chain, hence no notice, so such a row is reported `canonical` and the mint
+   door is never reached.
+
+**What breaks.** A caller creating metadata at runtime, at the simple arity,
+under a type name that is in neither half of the static spelling contract and
+has no rows already. That set is **not** empty in this repo — measured on
+`objectql`, `runtime` and `rest`, three in-tree fixtures minted `trigger` (a kind
+ADR-0088 retired outright), `policy`, and a synthetic `my_plugin_kind`. All three
+are corrected here rather than exempted, and each for its own reason: the
+`trigger` specimens were debt independent of any ruling (a retired kind cannot
+demonstrate a live tier, and they were green only through the hole this card
+closes), `policy` becomes a refusal case of its own, and #7894's control keeps
+its `metaUrlSpellingRefusal` claim while its boundary expectation follows the
+narrowing. An out-of-tree plugin that made its kind live by registering an item
+of it, and then accepted runtime writes to that kind through `/meta`, needs its
+spelling in the contract; there is no declared-kind channel to register one
+through today — that is the trade #8586's retirement made, and the `暂时` above
+is what makes it revisitable.
+
+`@objectstack/spec` gains one export, `unrecognisedMetaTypeRefusal`, alongside
+the #7894 verdict it deliberately does not merge with: one says *you spelled a
+declared type wrongly* and can name the replacement, the other says *there is no
+such type* and never guesses. The residue pin #7894 left behind
+(`metadata-url-spelling.test.ts`, the case that asserted `fieldz` was refused by
+nobody) is **flipped, not deleted**. ⚠️ #7894's positive control keeps its own
+claim intact — `metaUrlSpellingRefusal` still cannot refuse a kind that is a
+misspelling of nothing, which is what makes that control true by construction —
+but the BOUNDARY it drives now refuses six of the twelve names it exercises,
+and that case says so in place rather than leaving it to inference.