QA run · priority:P0 (23/23) · 7fa9873f · 2026-11-02 · 21 PASS / 2 PARTIAL / 0 FAIL / 0 BLOCKED / 0 NOT-RUN #28927

domain:devx#1 · os-warren · session_01JQwxbZpHPpovxXPremPLtv
Part of #27377 (wave 4, W4-C1: the standing priority:P0 smoke for the 18.3.0 release candidate)

This run re-judges every P0 item at the wave-4 subject 7fa9873f. It compares each item against the wave-3 P0 record #28545 (at eeec88d1) and the wave-3 retest #28656 (at 6e11a929).

Result: all 23 runnable items have a verdict: 21 PASS, 2 PARTIAL, 0 FAIL, 0 BLOCKED, 0 NOT-RUN.

Against #28545 (20 PASS / 2 PARTIAL / 1 FAIL), one verdict moved:

• approvals.account-app-entry went FAIL → PASS. A non-admin pending approver now opens the request page from Account › Inbox › Approvals, sees the decision panel, and Approve → Confirm answers POST /api/v1/actions/sys_approval_request/approval_approve 200 with the request moving 0/2 → 1/2. Reject, Request info and Reply also land for the non-admin, and the ?request= deep link cold-loads that request's page. This re-judges Approvals: a non-admin approver who opens an approval request from the Approvals inbox sees "You don't have access to Approval Request records" instead of the decision panel (regression from #27713) #28546 and approvals: a non-admin approver's decision from the request page answers 404 RECORD_NOT_FOUND — the action door loads sys_approval_request in the caller's data scope #28657 (PR fix(runtime): a non-admin approver's decision from the request page reaches the approvals routes instead of answering 404 #28708) at this subject.
• The two PARTIALs are the same standing blocked(dependency) clauses as at QA run · priority:P0 (23/23) · eeec88d1 · 2026-11-01 · 20 PASS / 2 PARTIAL / 1 FAIL / 0 BLOCKED / 0 NOT-RUN #28545: access-security.rls-both-sides acceptance[5] (knownGaps[0], qa(checklist): park the blocked partial-coverage gaps of the 17.7 pre-release runs (#21720, #21721, #21782, #21784) — each gets a fixture recipe, a pin, a re-pointed clause or a recorded knownGap #22017 row 11) and access-security.crud-permission-matrix acceptance[5] (security: no door can bind a permission set to the guest anchor — ADR-0138 D3 item 5 has no enforcer (PUT /meta/position/guest answers 403 NOT_OVERRIDABLE; guest suggestions are refused) #23267).

Other findings:

• The off-clause probes produced five low-severity product cards (close-out comment).
• One security finding was reported privately to the PM; its detail is withheld here. Off-clause rows score nothing (RUNNER rule 8), so it moves no item verdict.
• Three wave-3 probe cards were re-checked and hold at this subject: Metadata API: GET /api/v1/meta/app?id= with an empty value ignores the filter and answers the caller's full app list #28548, Datasources: POST /api/v1/datasources accepts and silently drops a top-level password or connectionString that os compile refuses, answering 201 #28549, MCP: aggregate_records silently drops an aggregation entry's filter, so a conditional count answers the unconditional total (REST honours it) #28550.

Environment fingerprint

• Framework: 7fa9873f38f7a0158e3ce17338dcc0f446b113e3 (main, 2026-11-02 02:44Z), a detached checkout including ui/. git status was clean before and after the run.
• Build (cold tree, in this checkout):
  • pnpm install --frozen-lockfile, exit 0 (24 s);
  • pnpm build --concurrency=2 under scripts/pm/os-verify-lock.sh, 117/117 tasks, exit 0, 03:34:39Z–03:47:26Z (12m47s);
  • pnpm console:build --concurrency=2, exit 0: ✓ @objectstack/console dist ready (82304 KB) from ui/apps/console, with the tracked sdui.manifest.json;
  • node scripts/check-dev-prereqs.mjs green (106 artifacts).
• Stale console bundle, ruled out: the dispatcher boot and every browser runner's boot served /_console/ entry assets/index-DqtqClJh.js, the same file as packages/console/dist/assets; R5 and R6 compared the bytes (sha256 prefix 44a80d196d2562bc), identical.
• Toolchain: Node 22.23.4, pnpm 10.31.0. Playwright 1.63.0 resolves chromium build 1243, which /opt/pw-browsers holds, so no executablePath override was needed.
• App: examples/app-showcase.
  • One dispatcher os dev compile boot (port 43001) wrote dist/objectstack.json at 03:47:53Z; every later tree boot used --no-compile --no-watch.
  • Items that compile, install a package, write the external datasource or touch a v17/18.0 DB booted from scratch copies (tar, node_modules symlinked): verify --rls, the datasource matrix, Studio first-run, the MCP os start legs, the dev-boot contract, the ceremony and the boot-refusal legs.
  • After the run examples/app-showcase/.objectstack/ holds only data/ (no installed-packages/).
• Boot line: OS_PORT=PORT OS_STORAGE_LOCAL_ROOT=RUN/uploads setsid pnpm -C … exec objectstack dev --ui --seed-admin --no-compile --no-watch -p PORT -d file:RUN/ITEM.db, with --fresh / --log-level info where an item asks for them.
  • Seed: --seed-admin plus the showcase seed, Seeds: com.example.showcase 132 rows. /api/v1/health reports version 18.2.0, the unreleased line.
• Ports: 43001 (dispatcher), 43101–43103 (R1), 43201–43205 (R2), 43301–43304 (R3), 43401–43413 (R4), 43501–43504 (R5), 43601–43602 (R6), 43701–43702 + 43719 PostgreSQL (R7). Each item ran on its own file DB under its runner's scratch dir. Every server was stopped by process group, and every port was empty at the end.
• Walled leg: os dev with OS_TENANCY_POSTURE=isolated, OS_PLATFORM_OWNER_EMAIL=admin@objectos.ai, OS_AUTH_MEMBERSHIP_POLICY=invite-only. Banner Tenancy: isolated, Organizations mounted, 0 DEGRADED lines; /auth/config multiOrgEnabled true, degradedTenancy false.
• v17-shaped DB: from npm create-objectstack@17.7.0 (--skip-install --skip-skills, then npm install --legacy-peer-deps, cli/runtime/spec 17.7.0), a blank scaffold booted once by its own os dev: 69 tables, 61 rows. 18.0 pre-release DB: the cli-18-prerelease-db restore (born under @objectstack/cli@18.0.0-next.5), recipe step 3 re-dating the token. Only copies were used, with a checksum or row diff around every step.
• Scratch PostgreSQL 16: 127.0.0.1 port 43719, scram-sha-256, data dir under /tmp per the item's knownGaps[2].
• Timestamps (2026-11-02): prep 03:33Z–03:49Z; runners 03:49Z–04:21Z.

Execution

Parallel, per SKILL §2: seven runner subagents (R1–R7), each at effort high on the default tier (dispatch-gates: no path-derived mandate), each with its own port range and file DBs.

• API and CLI items fanned out (R1–R4).
• Browser and mixed items ran on three runners (R5, R6, R7), one browser each.
• Vitest pins ran under the shared heavy-verify lock at --maxWorkers=1.

The dispatcher checked each runner's verdicts against the evidence it returned: pin outputs, HTTP traces, boot logs and response bodies. No P0 fail was recorded, so no rule-7 verifier was owed.

Scope: selector priority:P0, 23 runnable items

checklist-select.mjs reported 23 runnable items and 1 blocked item, hidden: access-security.no-active-org-session-semantics. No item is planned.

Items whose revision moved since #28545: crud-permission-matrix 11→12, owd-sharing-matrix 8→9, walled 7→8, mcp-stdio 8→9, query-contract-matrix 8→9, packaged-action-disabled-dispatch 10→11, account-app-entry 7→8, ceremony 7→8, datasource matrix 7→8, nav-surfaces-render 8→10, builtin-apps-nav-render 11→12, crud-roundtrip 16→17, first-run-loop 8→10. Each was scored at its current revision, and each runner read the new history entry.

Per-clause verdicts

Layer note (RUNNER rule 6). A clause scored from a pin names the pin's layer:

• dogfood pins are in-process HTTP over a real engine and sqlite;
• objectql pins are the engine over a mock or recording driver;
• plugin-unit and dispatcher-domain pins over stubs only corroborate.

Every clause whose oracle is a public door's answer was also driven on a live boot.

Reproduction rules for fails

None: this run recorded no fail.

Pins and the Playwright smoke

• Playwright smoke: examples/app-showcase/e2e/showcase-smoke.spec.ts gave 50 passed, 0 failed (2.1 min) against a live boot on port 43503, from a scratch copy (webServer reused the boot).
• objectstack verify --rls and bare verify, separate invocations on a scratch copy: per-persona 21 PROVEN (21 consistent, 0 HOLES) · 3 NOT PROVEN; total all personas: 38 PROVEN (38 consistent, 0 HOLES) · 226 NOT PROVEN [unit: one object × persona probe]; 10 of 10 positions; bare tail 15 verified, 0 gaps, 0 FAILED, 1 needs-fixture, 8 skipped.
• Dogfood pins (in-process HTTP, --maxWorkers=1, under the lock), all green:
  • R1: private-owd, rls-fixture, owner-anchor-and-bulk-writes, static-readonly, readonly-when-parent, crud-persona-matrix, permission-zoo, guest-grants-channel: 8 files 121/121.
  • R2: private-owd 8, public-read-owd 4, invoice-line-contributor-pair 5, crud-persona-matrix 49, controlled-by-parent 4, invoice-cbp 4, anonymous-deny-surfaces 75, wall-environment-metadata-refusal 23, wall-catalog-refusal 17, wall-package-doors-refusal 23: 212/212 (twice).
  • R3: showcase-expand-crud-gate, packaged-activation-ledger-reach, packaged-action-disabled-dispatch-live: 29/29.
• Engine and unit pins, all green:
  • objectql: engine.test + filter-array-lowering 196; readonly strip 55.
  • rest: request-schema-gate + data-empty-projection-count-door 21; meta-migrate-stored-walled-refusal 6.
  • runtime: action-activation-dispatch + data-empty-projection-count 26; packages-walled 32, automation-walled 12, meta-walled-catalog 19; identity-objects-default-organization pin 4; boot-guard 22 + kernel 10.
  • service-automation flow-activation-ledger 49 (corroboration; the flow item carries no automated).
  • plugin-auth ensure-default-organization 57; plugin-security boot refusal 3.
  • cli ceremony: 8 files 179 (+ unified-db-resolution.pin 7, corroborating).
  • mcp plugin.test.ts 17: plugin unit over a mocked ctx, corroborates only; every ai.mcp-stdio-fail-closed clause was scored on live os start stdio legs.
  • approvals: a non-admin approver's decision from the request page answers 404 RECORD_NOT_FOUND — the action door loads sys_approval_request in the caller's data scope #28657 fix pins, corroboration only: runtime api-action-forward 68; verify action-door-api-forward-denied-subject 3 (HTTP, booted stack).
• Not run, scored by hand instead: ui/e2e/live/record-history-display.spec.ts and the Studio F1 live spec need the console dev server on fixed ports (3000/5180) or write into the shared tree.

Derived item verdicts and fixture gaps

• approvals.account-app-entry: PASS (was FAIL at QA run · priority:P0 (23/23) · eeec88d1 · 2026-11-01 · 20 PASS / 2 PARTIAL / 1 FAIL / 0 BLOCKED / 0 NOT-RUN #28545). All 9 clauses pass; the Approvals: a non-admin approver who opens an approval request from the Approvals inbox sees "You don't have access to Approval Request records" instead of the decision panel (regression from #27713) #28546 and approvals: a non-admin approver's decision from the request page answers 404 RECORD_NOT_FOUND — the action door loads sys_approval_request in the caller's data scope #28657 fixes hold at this subject.
• access-security.rls-both-sides: PARTIAL. acceptance[5] blocked{by: dependency, ref: knownGaps[0] / qa(checklist): park the blocked partial-coverage gaps of the 17.7 pre-release runs (#21720, #21721, #21782, #21784) — each gets a fixture recipe, a pin, a re-pointed clause or a recorded knownGap #22017 row 11}, the D11 spot-check. Unchanged from QA run · priority:P0 (23/23) · eeec88d1 · 2026-11-01 · 20 PASS / 2 PARTIAL / 1 FAIL / 0 BLOCKED / 0 NOT-RUN #28545; the public half holds.
• access-security.crud-permission-matrix: PARTIAL. acceptance[5] blocked{by: dependency, ref: security: no door can bind a permission set to the guest anchor — ADR-0138 D3 item 5 has no enforcer (PUT /meta/position/guest answers 403 NOT_OVERRIDABLE; guest suggestions are refused) #23267}: no door binds a set to the guest anchor. Unchanged from QA run · priority:P0 (23/23) · eeec88d1 · 2026-11-01 · 20 PASS / 2 PARTIAL / 1 FAIL / 0 BLOCKED / 0 NOT-RUN #28545.
• Coverage notes on passing items:
  • platform-core.nav-surfaces-render's ErrorBoundary fallback is still unexercised (knownGap).
  • The marketplace catalog content stays blocked by egress (knownGap).
  • The v17 fixture is a blank-scaffold DB, not a showcase DB.
  • The walled grant was a recorded fixture write outside HTTP, the item's own documented recipe (both HTTP grant doors refuse it, as fixtures.requires[2] says).

Retest of earlier QA cards met in passing

Observed on this run's boots as off-clause controls; they are not clause verdicts.

• Metadata API: GET /api/v1/meta/app?id= with an empty value ignores the filter and answers the caller's full app list #28548: meta/app?id= with an empty value now answers 400 "Query parameter 'id' names no app".
• Datasources: POST /api/v1/datasources accepts and silently drops a top-level password or connectionString that os compile refuses, answering 201 #28549: the runtime datasource door now refuses a top-level password at create and at PATCH (400).
• MCP: aggregate_records silently drops an aggregation entry's filter, so a conditional count answers the unconditional total (REST honours it) #28550: MCP aggregate_records now honours an aggregation's filter ({paid:5, all:13}, equal to REST).
• The QA run · priority:P0 (23/23) · eeec88d1 · 2026-11-01 · 20 PASS / 2 PARTIAL / 1 FAIL / 0 BLOCKED / 0 NOT-RUN #28545 probe "$skip without $top gives total = matches minus offset" is no longer seen: ?$skip=5 returns 9 rows with total 14.

Off-clause probes (RUNNER rule 8)

About 48 runner-minutes in total, after the clauses had verdicts, on the same boots and as the same personas: R1 5, R2 8, R3 5, R4 7, R5 8, R6 9, R7 6.

Doors:

• data CRUD, list, query, by-id, bulk create/update/delete, clone, export, search, analytics, batch;
• sys_user_permission_set, sys_business_unit_member, sys_user_position, sys_metadata, sys_secret;
• meta/app (list, ?id= variants, by name), /api/settings/:ns, /meta writes on the walled boot;
• sign-in, sign-out, get-session, set-active, organization create;
• actions and activation, MCP run_action / query_records / get_record / aggregate_records over HTTP and stdio; automation toggle, trigger, runs and _status;
• approvals request, actions and decisions through the action door and the approvals routes;
• the datasource admin, test, remote-tables and meta doors; packages, metadata draft PUT and publish-drafts;
• os start stdio admission, the ceremony arms, the os serve boot guard, os dev flag parsing; console bad routes.

Personas: admin, walled operator, walled org owner and a non-holder, plain member, contributor, client_liaison, auditor, ops, field_ops_delegate, a non-admin approver and a non-participant, demo personas auditor.demo and phone.demo, API keys (admin, contributor, liaison, revoked), anonymous, and a bogus or empty bearer.

Boundary classes:

• empty, whitespace, unknown, encoded and trailing-slash ids and values; repeated and bracketed (id[]) parameters;
• forged owner, id, credentialsRef and overrides; self-escalation of grants and positions;
• foreign rows through the bulk doors; FLS-hidden predicates in where, orderBy and aggregation filters;
• the opposite posture (walled against single) and the opposite credential (API key against session);
• an empty or null secret; a missing recordId; a missing flag value;
• a wrong or missing plan file, no backup flag, an already-certified DB, a planted-shape DB.

Findings, each in its own card (listed in the close-out comment):

• integration-system.datasource-credential-refusal-matrix · observed beside the clauses: PATCH /api/v1/datasources/:name {"secret":""} replaces the stored credential with an empty secret, and the status reason then misreports the cause.
• platform-core.builtin-apps-nav-render · observed beside the clauses: meta/app?id[]=setup (bracketed array spelling) ignores the filter and serves the caller's unfiltered app list.
• cli.dev-boot-contract · observed beside the clauses: a flag missing its value (os dev -d) exits 2 without the INVOCATION ERROR banner.
• approvals.account-app-entry · observed beside the clauses: an api action called with no recordId answers 404 ENDPOINT_NOT_FOUND instead of naming the missing record id.
• access-security.write-path-guards · observed beside the clauses: the clone and createMany doors report a readonly strip in the body's droppedFields but send no X-ObjectStack-Dropped-Fields header.

Security findings:

• approvals.account-app-entry · observed beside the clauses · security finding — detail withheld, reported to PM

Recorded, not filed:

• showcase_portfolio_snapshot answers a plain member invoices: 12 over REST and MCP: action bodies run elevated by design (buildActionExecutionContext), so this is the specimen app disclosing an aggregate, not a platform hole.
• GET /automation/_status lists every flow's name and binding state to a plain member, who can read /meta anyway.
• The enabled refusal for null / 0 on the toggle door says "not a string".
• A member gets 404 UNKNOWN_NAMESPACE for an unregistered settings namespace and 403 for a registered one; the names are public in the manifests.
• PUT /meta/object/X?mode=draft&package= (empty) saves an env-wide draft; an unknown package is refused 422.
• An unprefixed object draft is accepted at draft time and refused at publish, which blocks the package's later publishes until it is removed.
• A clone re-applies a current_user default to a field that was null on the source.
• The [sharing-rule] share_new_inquiries_with_field_ops expands to NO recipients warning prints on every boot (as at QA run · priority:P0 (23/23) · eeec88d1 · 2026-11-01 · 20 PASS / 2 PARTIAL / 1 FAIL / 0 BLOCKED / 0 NOT-RUN #28545).
• An unknown view in the URL silently renders the default view (as at QA run · priority:P0 (23/23) · eeec88d1 · 2026-11-01 · 20 PASS / 2 PARTIAL / 1 FAIL / 0 BLOCKED / 0 NOT-RUN #28545).

Time

• Environment prep: 16 min wall (03:33Z–03:49Z: protocol reading, install, full build 12m47s, console build, compile boot and bundle check), plus about 46 runner-minutes of per-runner prep (protocol reading, scratch copies, persona minting, the walled boot, the v17 and pre-release DBs, scratch Postgres).
• Testing: 32 min wall (03:49Z–04:21Z), or about 71 runner-minutes of clause work plus 48 of probes.
• Ratio: about 1 : 2 by wall clock (16 to 32 min), and about 1 : 1.9 by agent-minutes (62 to 119).

Checklist-accuracy findings from this run are reported to #27377 for the devx lane; they are not repeated here.

Generated by Claude Code