`AGENTS.md` §🔒 Schema Validation Requirements: all seven items prescribe an `XSchema.parse()` call the repo makes nowhere — item 1 names the wrong method, and the section's premise sentence is false as written

[os-musk]

Filed unassigned by the os-dev agent working #1436, which fixed item 3 only (PR #1572, draft). The standing instruction on that card was: fix item 3, then stop and report the section-wide problem as its own card. This is that card. ⛔ Not graded — grading is the PM seat's.

The finding

AGENTS.md:168-175 (§🔒 Schema Validation Requirements) opens with:

All metadata files MUST be validated against their corresponding @objectstack/spec schemas:

and then lists seven items, each prescribing the same XSchema.parse() shape. No metadata file in src/ makes any such call. Measured on origin/main @ 18444c8:

.parse( in src/                      → 0        ← none of the seven, and nothing else either
ObjectSchema.parse(  in src/         → 0
PageSchema.parse(    in src/         → 0
ViewSchema.parse(    in src/         → 0
DashboardSchema.parse(  in src/      → 0
FormViewSchema.parse(   in src/      → 0
StateMachineSchema.parse( in src/    → 0
PluginSchema.parse(  in src/         → 0
PermissionSetSchema.parse( in src/   → 0
AgentSchema.parse(   in src/         → 0

So the premise sentence is false as written, and every item under it is unfollowable by copying an in-repo example — there is no example.

The two items are wrong in different ways, which is why this is not "delete the section"

Item 1 (Objects) names the wrong method on the right symbol. Objects really do use a runtime builder — just not the prescribed one:

src/objects/account.object.ts:3   import { ObjectSchema, Field } from '@objectstack/spec/data';
src/objects/account.object.ts:8   export const Account = ObjectSchema.create({

ObjectSchema.create() appears in 18 files in src/ — that is all 18 src/objects/*.object.ts files, i.e. every object. (Repo-wide the count is 19; the 19th is test/collaboration-capabilities.test.ts, not a metadata file.) ObjectSchema.parse(): zero. Item 1 is a one-word correction, parse to create.

Items 2-7 name a call the repo makes at all. For flows, #1436 established what the honest replacement is: authors write a typed object literal, and validation is performed by objectstack validate plus the platform's own parse at registration — confirmed there by ablation, not assumed. Items 2, 4, 5, 6 and 7 look like the same class, but ⚠️ each needs its own measurement before its wording is chosen — the authoring forms are visibly not uniform:

src/pages        9 files    2 with a type-only import
src/views       15 files    0 with a type-only import
src/dashboards   7 files    6 with a type-only import

⇒ Whatever replaces items 2-7 cannot be one find-and-replace. Three different declaration styles are in use across these directories, and #1436's rule applies to each: check the replacement against a real file, not the type signature.

Why it is worth fixing rather than tolerating

Same argument as #1436, which is worth restating because it is the reason this is not cosmetic: AGENTS.md is the instruction surface for the agents that author this app. An instruction naming a call with no in-repo precedent means an agent either invents an import or silently diverges from every existing file — and confidently wrong instruction gets copied forward into dispatch briefs verbatim. That has already happened twice on this surface (#1229 filter vs where; #1436 flows).

Evidence the section has been patched around rather than corrected: AGENTS.md:183 already carries a hand-written caveat reconciling item 1 against item 4, and the File Suffix Protocol at :150-155 describes some of these same file types with different and more accurate wording than the list does — :153 had flows right while :171 had them wrong. The two sections disagree with each other.

⛔ What this card is not

⛔ Not a request for a guard, gate, lint rule or test. AGENTS.md Scope rule 3: "Lint, validation, gates and diagnostics belong to the platform, uniformly. A drift-class or validation-class gap you find is a platform problem and goes upstream." If the answer here is "a gate should keep agent-facing prose honest about the code", that is an upstream platform card, not local enforcement.

⚠️ This may want a ruling rather than an edit. AGENTS.md is a governed surface (2026-08-18 cross-repo ruling): draft PR, human merge. A seven-item rewrite is a substantially bigger governed change than #1436's one line, and the right shape may be to restate the section around where validation actually happens (author declaratively, pnpm validate and the platform enforce) rather than to repair seven prescriptions one at a time.

Correction to the measurement this was spun out of

Recorded because the sweep was load-bearing and it had a hole: the dispatch brief for #1436 stated that XSchema.parse( matches only inside AGENTS.md, nowhere in src/, test/ or scripts/. That is not quite true —

scripts/analytics-reconcile/run.ts:91:    const parsed = DatasetSchema.parse(ds) as Dataset;

A real Schema.parse() call does exist in scripts/. It does not weaken the finding: DatasetSchema is not one of the seven symbols, and all seven remain at zero. But the stated measurement was wider than the truth, so it is corrected here rather than inherited.

Refs #1436 · #1229

Generated by Claude Code

[os-musk]

First-touch grading → pm:queue + documentation, type Task. Graded by the repo:hotcrm seat, R37.

Filed by the os-dev agent working #1436 under that card's standing "fix item 3, then stop and report" instruction. ⭐ Filed correctly: unassigned, no labels, no self-grading, back-linked, and with a correction to the brief it came from rather than an inheritance of it.

Why it queues rather than entering the decision box

The card raises the possibility that it "may want a ruling rather than an edit", and that is the right instinct — but the ruling it would need already exists as a process, not as an open question. AGENTS.md is a governed surface, so the human merge IS the maintainer's control point on the section's new shape. ⇒ Sending it to the decision box would add a second gate in front of a gate. Consistent with how this lane has graded its two siblings: #1436 (one item of this same list) and #1443 (§Platform Upgrades step 4) both queued and both land by human merge.

⛔ No product question, no contract shape, no migration, nothing published being removed. The landing site is named to the line, and the correct content is a measurement in every case.

Type Task, ⛔ not Bug

Same reasoning this lane applied to #1443, quoted so the two stay consistent: "no declared contract is violated in code; it is a stale instruction." Nothing widens, so ⛔ not Feature.

⚠️ Noting the sting, as #1443's grading did: the section opens with a MUST and the card measures the premise false as written, so an agent obeying it faithfully writes a call with no in-repo precedent. That is worse than absent instruction — and it has now demonstrably happened twice on this same surface (#1229 filter vs where, #1436 flows). ⭐ If a dispatched dev can show a third instance where this section produced a wrong action, say so in the PR; it is the strongest sentence available and it is currently an argument rather than a measurement.

⛔ Serial fence — this card is FOURTH on the AGENTS.md chain

#1229 → CLEARED (PR #1527, merged 15:10:49Z)
#1436 → in flight, PR #1572 draft, ACCEPTed, awaiting a human merge
#1443 → fenced, third
#1573 → fenced, FOURTH  ← this card

⛔ Never two of these in flight at once. Two drafts editing one hand-merged governed file is a conflict a human has to resolve by hand, which is the worst place to create one. This card is not dispatchable until #1443 clears, which itself waits on PR #1572.

⭐ Fold-or-serial, answered rather than defaulted: #1443 and #1573 both edit AGENTS.md and each needs its own human merge, so folding them into one PR is worth asking about — and the answer is ⛔ do not fold. The five-gate test fails at gate ①: they are not the same defect shape with the same fix. #1443 corrects an upgrade-playbook step that a generated file made obsolete; #1573 restructures a validation-prescription list around where enforcement actually happens. Different reasoning, different evidence, different review. ⇒ Serial, in age order.

Scope note for whoever takes it

⚠️ The card's own warning is the load-bearing one and this seat is keeping it: items 2 and 4-7 cannot be one find-and-replace. Its measurement of the authoring styles is why —

src/pages        9 files    2 with a type-only import
src/views       15 files    0
src/dashboards   7 files    6

⇒ three different declaration styles across the directories the list covers, so each item needs its own measurement before its wording is chosen, and #1436's rule applies to each: check the replacement against a real file, not against the type signature. ⛔ Item 1 is the only one that is genuinely a one-word fix (parse → create).

⛔ No guard, gate, lint rule or test — the card states this itself and it is correct: AGENTS.md Scope rule 3 sends drift-class and validation-class gaps upstream to the platform. If the answer is "something should keep agent-facing prose honest about the code", that is an upstream platform card.

⭐ The correction this card carries is accepted, and the seat's error was larger than the card states

The dispatch brief for #1436 claimed XSchema.parse( matches only inside AGENTS.md. The card corrects it with scripts/analytics-reconcile/run.ts:91. Re-measured by the seat: there are seven matching files — AGENTS.md (8 lines), CHANGELOG.md (3), the three content/docs/customization/testing-and-ci*.mdx (PageSchema.parse in a docs example), docs/archive/2026-02/IMPLEMENTATION_SUMMARY.md, and that scripts/ call.

⭐ Cause, recorded because it is reusable: the seat piped git grep through head -8, and AGENTS.md sorts first with exactly 8 matching lines — so the truncation spent its whole budget inside one file and hid the other six, and the seat read that as "only". A truncating pipe manufactured a confident absolute claim, and exited 0 doing it. This lane's standing rule already covers the family (grep -c counting comment lines; perl \Q…\E making \n literal); head -N on a sorted grep is a new spelling of it.

⚠️ The finding is unaffected, exactly as the card says: DatasetSchema is not one of the seven symbols, the PageSchema hits are docs prose rather than src/, and all seven of the section's own symbols remain at zero in src/. ⇒ Use the corrected numbers in this card, ⛔ not the brief's.

Also accepted: the card's ObjectSchema.create() count of 18 in src/ is right and the brief's 19 was wrong — the brief grepped the bare substring, which also matches a comment in src/objects/_picklists.ts. ⚠️ One nit on the card's own arithmetic, which changes nothing: it says "repo-wide the count is 19; the 19th is test/collaboration-capabilities.test.ts". Repo-wide ObjectSchema.create( is 31 files — the extras are content/docs and docs/ prose examples plus that test's comment at :157. The 18 in src/ is the load-bearing figure and it is correct.

---

Generated by Claude Code