A raw driver `code` (`SQLITE_ERROR`) still reaches client-facing payloads through the batch verbs' `code` limb — the sibling field #8333 did not cover

[os-zhuang]

Filed by the dev on #8333 while landing the P6–P13 producer sweep (PR #8436). Unassigned; nobody is on it. Not a defect in that PR — a different field, with a different rule, deliberately left out of its scope rather than ridden in.

The measurement

#8333 fixed the error string on the batch verbs' payloads: a caught sentence is quoted back only when the error declared a 4xx status. Several of those same collectors carry a sibling limb built straight off the caught object:

...(e?.code ? { code: e.code } : {})

publishPackageDrafts' failed[] and revertCommit' failed[] both have it. That limb is ungated, and drivers do carry code: better-sqlite3 raises SQLITE_ERROR, node:sqlite ERR_SQLITE_ERROR, postgres a five-character SQLSTATE.

Reproduced end to end on the #8333 branch (so, with the message fix already in place) — an insert on sys_metadata failing with Object.assign(new Error('SQLITE_ERROR: no such table: sys_metadata'), { code: 'SQLITE_ERROR', errno: 1 }), driven through publishPackageDrafts:

[
  {
    "type": "view",
    "name": "acct_view",
    "error": "publish failed",
    "code": "SQLITE_ERROR"
  }
]

The message is withheld exactly as #8333 intended, and the driver's own dialect still ships beside it.

Why this is a different rule, not an oversight in #8333

carryCatalogedErrorCode's docblock states the distinction, and it is load-bearing: a message is free text, so the catalog does not bound it and status-alone is the safe test; a code writes ApiErrorSchema.code, a closed union a driver's dialect must never enter. The two limbs therefore need different predicates, and #8333's positive list is the wrong tool for this one — the right test is membership in StandardErrorCode ∪ ERROR_CODE_LEDGER, which the file already has a helper for.

⇒ The likely fix is to route these limbs through the same catalog gate carryCatalogedErrorCode applies at the re-wrap exits, rather than widening declaresClientRefusal.

Scope note

Same sink as #8333 — failed[] rides on response data, so no HTTP boundary's message withhold reaches it. Consumers that switch on code are the reason it cannot simply be dropped: BATCH_ABORTED and the declared 4xx codes on those same arrays are read by callers, so the fix must filter, not delete.

Sites read from source (packages/metadata-protocol/src/protocol.ts, the failed.push collectors in publishPackageDrafts and revertCommit); only the publishPackageDrafts one was reproduced.

Related

#8333 (the error string, PR #8436) · #8136 (option C, the measured cluster) · #8086 (the ruling recording option C)

---

Generated by Claude Code

[os-zhuang]

Claiming this.

• Session: session_012WMpuAfA2KSdDjGF6tm1bH
• Branch: claude/issue-8441-batch-verb-driver-code
• Declared region (3-concurrent-editor cap on packages/metadata-protocol/src/protocol.ts): the failed[] construction sites in publishPackageDrafts and revertCommit, plus carryCatalogedErrorCode/its catalog helper. Not the field write door (held by meta-field-write-inert: an accepted field PUT never reaches the object — a runtime-created field is stored valid=true and is absent from fields forever #7893), not the error limb ([finding] metadata-protocol's batch verbs still put caught error text on client-facing payloads — the 8 producers option C did not reach #8333, landed as f58b1a88), not seed-loader.ts (seed-loader interpolates driver text into errors[].message, which rides on seedApplied of a 200 publish response #8442).

Plan: verify the premise on current main first, then route the code limb through the same catalog-membership gate carryCatalogedErrorCode applies at the re-wrap exits — catalogued codes pass through unchanged, non-catalogued driver dialect replaced. #8333's positive control (Studio flow CEL draft answering with code + structured issues) must stay green.

---

Generated by Claude Code

[os-zhuang]

{
  "issue": 8441,
  "status": "done",
  "branch": "claude/issue-8441-batch-verb-driver-code",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/8505",
  "premise_still_valid": true,
  "summary": "Premise reproduced on current main at both named sites, with the message fix already in place: publishPackageDrafts answered { error: 'publish failed', code: 'SQLITE_ERROR' } and revertCommit { error: 'revert failed', code: 'SQLITE_ERROR' }. Added clientFacingFailureCode(err) beside clientFacingFailureText and routed both failed[] limbs through it: the rule is CATALOG MEMBERSHIP (ErrorCode = StandardErrorCode union ERROR_CODE_LEDGER, the predicate carryCatalogedErrorCode and toRowApiError already share), not #8333's 4xx question. A catalogued code passes through byte for byte; an uncatalogued one becomes the code its declared status maps to, else INTERNAL_ERROR; an error with no code still yields no key. No error code minted. PM assumptions both held: the fix is confined to the two named sites plus the new helper (protocol.ts only), and seed/sys_capability paths were not touched. The two sibling collectors that build the identical limb, discardPackageDrafts and deletePackage, were MEASURED clean (no code at all under the same driver fault) because deleteMetaItem's re-wrap exits already gate code, so they got no filter and the measurement is pinned instead. head SHA 03db0654d2bb0530114cf166eeaac60370d9b770.",
  "tests": "New pin: packages/metadata-protocol/src/protocol.batch-verb-driver-code.test.ts, 12 cases in six sections — 12 passed. PER-SITE PREMISE on main: publishPackageDrafts failed[] = [{\"error\":\"publish failed\",\"code\":\"SQLITE_ERROR\"}]; revertCommit failed[] = [{\"error\":\"revert failed\",\"code\":\"SQLITE_ERROR\"}]; discardPackageDrafts and deletePackage carried NO code. AFTER: both named sites answer code INTERNAL_ERROR, nothing leaked. PASSES THROUGH: BATCH_ABORTED on collateral rows, NOT_OVERRIDABLE, VERSION_NOT_FOUND, INVALID_METADATA, and ERR_DATASOURCE_UNAVAILABLE on a 503 whose sentence is withheld (the discriminator that proves the two limbs answer different questions). REPLACED: SQLITE_ERROR -> INTERNAL_ERROR; SQLSTATE 42501 on a declared 403 -> PERMISSION_DENIED with the authored sentence kept whole. POSITIVE CONTROL GREEN: the broken-CEL flow draft still reports flows[0].nodes[1].config.approvers[0].value, code INVALID_METADATA, and issues[0].rule approval-expression-invalid. REVERSE VERIFICATION, both directions predicted first: (a) protocol.ts reverted to pre-fix — predicted 4 red / 8 green, MEASURED 4 red / 8 green, every prediction held including the operator-half case predicted red for its payload assertion; (b) the over-broad drop-the-limb variant — predicted 8 red / 4 green, MEASURED 9 red / 3 green, ONE MISS: the operator-half case went red because it asserts expectCataloged before the log — the same miss #8333 recorded for its own P7 log case — recorded in the file header rather than retro-fitted. ANTI-VACUITY: #8333's fake throws a bare Error with no code property, which is why its payload scan could not see this; the harness here attaches a real driver code, and section 1 fails against pre-fix code. SUITES: @objectstack/metadata-protocol 82 files / 1208 tests passed; downstream consumers (prefix filter, i.e. dependents) @objectstack/objectql 197 / 3539 and @objectstack/rest 112 / 1843 passed. GATES (lint.yml parse union re-derived with dispatch-gates.mjs against actual changed paths, all run post-merge): check:nul-bytes, check:error-code-casing, check:engine-double-contract, check:cross-package-test-inputs, check:durability-log-level, check:filter-alias-parity, check:changeset-gate-self-tests, check:objectui-changeset, check:query-options-erasure, check:type-check-coverage, check-adr-0087-registration, check-changeset-no-major, check-empty-changeset — all PASS. check:error-code-casing was NOT in the prompt's list and caught a real defect mid-run (a lowercase pseudo-code literal in my fixture, ADR-0112 D1); the fixture now uses the postgres SQLSTATE 42501. check:objectui-pin-fresh is RED and pre-existing: .objectui-sha untouched here, main already carries 2018 changesets, so it fires identically on main. Changeset added. CI status at report time: in_progress (reporting at draft-PR time per the 2026-08-10 ruling).",
  "open_questions": [],
  "out_of_scope_findings": [
    "filed as #8502: deleteManyData's row errors carry raw driver text in errors[].message — measured {\"code\":\"INTERNAL_ERROR\",\"message\":\"SQLITE_ERROR: no such table: leave_request\"}; toRowApiError gates the code half but not the message half, and the sink is a data batch row rather than a metadata verb, so neither #8136 nor #8333 enumerated it. Searched first (keyword + file-path over open issues); #3566 is closed and a different producer, #8442 is seed-loader. Filed unassigned and unlabeled for PM triage."
  ]
}

---

Generated by Claude Code