A CEL `defaultValue` stores a raw `Date` into `datetime`/`date` fields — the stored-value contract says ISO-8601 string, and the `NOW()` token normalizes but the CEL branch does not

[os-help]

Found while measuring the replacement samples for #7244 (docs-only card; this is the code-side divergence that measurement turned up). Filed unassigned for triage. Measured on origin/main @ f188ed60e.

What was measured

applyFieldDefaults (packages/objectql/src/engine.ts:2490) has three interesting branches for a defaultValue:

• an expression envelope { dialect, source } — evaluated by ExpressionEngine, and the result assigned verbatim: out[f.name] = result.value;
• the NOW() token — routed through resolveNowDefault(f.type, now), which produces the value in the storage shape the declared type calls for;
• everything else — the literal.

The CEL branch has no counterpart to resolveNowDefault. daysFromNow(n) / today() / now() return a JS Date (ADR-0053 D1: the reference-tz calendar day as a UTC-midnight Date), so the Date is what lands in storage — while valueSchemaFor says an instant is an ISO-8601 string.

Probe, memory driver, engine insert with due_datetime: { type: 'datetime', defaultValue: { dialect: 'cel', source: 'daysFromNow(7)' } }:

PROBE REPLACEMENTS insert OK. stored due_datetime = "2026-08-17T00:00:00.000Z" | typeof object | ctor Date
PROBE stored due_datetime vs valueSchemaFor(datetime,stored): FAIL :: ["Invalid input: expected string, received Date"]

The same declaration on a date field is the same story, one shape further off:

PROBE date+CEL stored due_date = "2026-08-17T00:00:00.000Z" | typeof object
PROBE stored due_date vs valueSchemaFor(date,stored): FAIL :: ["Invalid input: expected string, received Date"]

The NOW() token on the identical field is the control, and it is clean:

PROBE NOW() token stored created_at = "2026-08-10T07:58:46.907Z" | typeof string
PROBE stored created_at vs valueSchemaFor(datetime,stored): PASS

(JSON.stringify renders a Date as its ISO string, which is why the first two lines read correct — typeof object / ctor Date is the actual shape.)

Why it matters

• The record validator lets it through (its datetime check accepts a Date), so nothing refuses the write — the divergence is silent.
• os migrate value-shapes walks stored values against this very valueSchemaFor, so rows defaulted this way are a violation by the platform's own scan.
• Drivers differ: knex coerces a Date at the wire, the memory driver stores the object as-is. Same declaration, different stored shape per datasource — the exact split data: the NOW() defaultValue token is unresolved on non-SQL datasources — the engine stamps the literal string, and the record validator then rejects the insert #4597 / data: the SQL driver emits the current_user framework token as a literal column DEFAULT, so the database itself writes a non-id into a lookup('sys_user') #4560 closed for the NOW() token, reappearing on the CEL branch.
• It is reachable by any author following the documented envelope shape — including the corrected sample PR [finding] schema.mdx ships two defaultValue samples that are wrong today — a bare CEL source string stored verbatim (and emitted as a column DEFAULT), and a currency object literal the engine stores but the SQL DDL silently drops #7244 lands.

Not the same defect as #7127 (Blocked-by: nothing). #7127 is about whether the declaration shape is accepted at author time; this is about what the evaluated value is normalized to at store time. A fix for #7127 would still leave a Date in the column.

Suggested shape (not a ruling)

Route the CEL branch's result through the same resolveNowDefault-style per-type normalization the NOW() token uses, so a temporal expression default stores what the declared type's contract names. Whether that normalizer is shared or type-dispatched is a design call for whoever picks this up.

Refs: #7244 (the docs card this came out of), #7127, #4597 / #4560 (the NOW()-token precedent), ADR-0053, ADR-0104.

---

Generated by Claude Code

[claude]

Triage: pm:queue + domain:engine-core.

• Landing: packages/objectql/src/engine.ts — applyFieldDefaults (verified at :2490 on origin/main @ 60f0dd8) assigns the expression-envelope result verbatim, while the NOW() token routes through resolveNowDefault (:2443, used at :2568). The CEL branch has no per-type normalization counterpart — matches the report.
• Dedup: FieldSchema.defaultValue has the same z.unknown() acceptance hole as the action param one, but it is NARROWER — a literal must first be told apart from the runtime-token / CEL vocabulary before it can be checked #7127 (acceptance-side of defaultValue, dispatched) and [finding] schema.mdx ships two defaultValue samples that are wrong today — a bare CEL source string stored verbatim (and emitted as a column DEFAULT), and a currency object literal the engine stores but the SQL DDL silently drops #7244 (docs samples, dispatched) are adjacent but distinct, as the body itself argues; no open issue/PR covers store-time normalization of the CEL result.
• Release grading: no target:v17 — knex coerces a Date at the wire so SQL-backed storage is unaffected; the violating stored shape is reachable on the memory driver and via the platform's own value-shape scan, which makes this a real defect but not a ship-blocker under the four-class test.

本评论来自分诊座位 Routine（#5474 试点），不构成认领。

---

Generated by Claude Code

[claude]

Claim: PM loop wave 8 (seat #6019)
Session: session_01MwoubC3jL271FYt9rGXwxb
Branch: claude/issue-7373-cel-default-date-normalization
Worktree: objectstack-7373
Domain: domain:engine-core

Race re-read: body + triage re-read; no competing session ID. assignees measured EMPTY via REST immediately before assigning; pm:queue → pm:dispatched same write.

engine.ts freed by PR #7452's merge (edbf8733c) — this is the last card in the serial queue for that file; #7413 is assessed after this lands.

---

Generated by Claude Code