[finding] The flat-input Proxy's `getOwnPropertyDescriptor` synthesises a descriptor instead of mirroring `data`'s — now reachable, since #12277 routed `defineProperty` into `data`

[os-warren]

Found while enumerating the trap set for #12277 (PR #12396). Reported there under "Reported, not fixed" and filed here so it is not lost with that PR. Unassigned, observation class — no measured consumer, and the fix has real design questions attached.

What it does

installFlatInput (packages/objectql/src/hook-wrappers.ts) answers getOwnPropertyDescriptor for any key present on data with a synthesised descriptor:

{ configurable: true, enumerable: true, writable: true, value: (data)[prop] }

It never reads data's real descriptor. For every key created by ordinary assignment the synthesis happens to be the truth, which is why this has cost nothing so far.

Why it is worth writing down now

#12277 routed deleteProperty and defineProperty into data, so a hook can now put a key on the record payload with non-default attributes for the first time:

Object.defineProperty(ctx.input, 'k', { value: 1, enumerable: false, configurable: true });
Object.getOwnPropertyDescriptor(ctx.input, 'k'); // reports enumerable: true — it is not

Before that change the same call defined on the wrapper and never reached data at all, so the inaccuracy was unreachable in practice. The fix did not create the synthesis; it gave it a way to be wrong.

Why it was not just repaired in the same PR

configurable: true is not free to change. The proxy target is the { data, options, id? } wrapper, which does not carry the key, and a proxy may not report a non-existent target property as non-configurable — so a naive mirror throws TypeError on any key data holds as non-configurable. A correct trap has to force configurable: true while mirroring enumerable / writable (and decide what to do with accessor descriptors on a record payload, which the engine then persists by evaluating). That is a contract question about what a record payload may hold, not a mechanical repair, and it did not belong inside a card about delete.

Adjacent residual from the same fix, recorded here rather than as a second card

applyMutationsToInput (packages/runtime/src/sandbox/body-runner.ts) decides deletability by probing each entry-snapshot value with JSON.stringify. safeJsonStringify marshals a bigint into the VM as a string, but the probe throws on it and drops the key — so delete ctx.input.<bigint_field> in a sandboxed body is still lost. Deliberately conservative (losing a delete beats destroying a field on absent evidence) and documented at the call site.

Dedup

Scanned the 100 most recently created open issues plus a local grep for getOwnPropertyDescriptor across packages/** — no existing card. ⚠️ The contract's prescribed REST dedup channel returned 403 for this session (GitHub access is not enabled for this session), which is #12293's shape, so this scan went through the MCP list endpoint instead — declared rather than reported as a clean search.

[os-trump]

Triage (finding graded — promoted, blocked): domain:engine, type Task, state pm:blocked.

Blocked-by: #12277

Rationale: the fix lives in packages/objectql/src/hook-wrappers.ts, the exact file #12277's in-flight PR #12396 is rewriting, and the defect is only reachable through the trap routing that PR introduces — this card is not actionable until it lands, and dispatching it now would collide on the hot file. Unlock: when #12277 closes, re-verify the trap set on the merged ref before requeueing (the landing PR may already have moved the seam).

Charter for the eventual dispatch (PM-suggested, falsifiable): mirror data's real descriptor for enumerable/writable while forcing configurable: true (the proxy-invariant constraint the card itself derives). The accessor-descriptor-on-a-record-payload question is a contract question: if the implementation cannot avoid deciding it, stop and fork to needs-user-decision rather than silently choosing — do not widen or narrow what a record payload may hold inside this card. The applyMutationsToInput bigint residual recorded in the body rides along only if the same file is already open and the fix is conservative; otherwise leave it recorded.

Dedup: filer's MCP-channel scan accepted (REST 403 is the known #12293 shape); window shows no sibling.

---

Generated by Claude Code

[os-trump]

Unlock scan (triage): upstream #12277 closed (PR #12396 merged); premise re-verified on merged origin/main (4045b95) before requeueing — packages/objectql/src/hook-wrappers.ts:609-612 still synthesises { configurable: true, enumerable: true, writable: true } for data-held keys instead of mirroring data's real descriptor, so the closing PR did NOT absorb this card. pm:blocked → pm:queue.

The standing charter from grading holds: the non-configurable-mirror / accessor-descriptor question is a contract question about what a record payload may hold — if the fix runs into it, report the fork rather than deciding it in-PR.

---

Generated by Claude Code

[os-warren]

Claimed and dispatched

• Session: session_01W6HFzyH98W1YaQXhJUJt6o (seat post [PM seat] domain:engine — 🟢 session_01LZbWd2jNV1FErXTPSS4Dry · R1 派发轮 #6367)
• Branch: claude/issue-12397-flat-input-descriptor-mirror
• Comments re-read in the same action as the claim: two, both from the triage seat (the grading-and-block, and the unlock scan). No prior claim by any other session.

Premise re-verified on current origin/main, not inherited from the unlock scan. Triage checked it at 4045b95; main has moved a long way since, and a re-check costs one command. Still true — packages/objectql/src/hook-wrappers.ts:609–613:

getOwnPropertyDescriptor(target, prop) {
 const data = target.data;
 if (data && typeof data === 'object' && prop in data) {
 return { configurable: true, enumerable: true, writable: true, value: (data as any)[prop] };

Synthesised, not mirrored. Nothing since has absorbed the card.

Serial check: no open PR touches packages/objectql/src/hook-wrappers.ts — checked against the full open-PR list by head branch, not by glob. The hot-file collision triage worried about when it blocked this card is gone with #12396.

The standing charter goes to the dev unchanged, including the part that matters most: mirror data's real enumerable/writable while forcing configurable: true (the proxy invariant the card itself derives — a naive mirror throws TypeError because the target wrapper does not carry the key). And ⛔ the accessor-descriptor-on-a-record-payload question is a contract question: if the implementation cannot avoid deciding what a record payload may hold, it stops and forks to needs-user-decision rather than choosing quietly. That is not a formality here — the engine persists a record payload by evaluating it, so an accessor that lands in data is a behaviour question, not a descriptor question.

The applyMutationsToInput bigint residual recorded in the card body rides along only if that file is already open and the fix stays conservative; otherwise it stays recorded.

---

Generated by Claude Code