Repository navigation
Commit b289090
fix(plugin-security)!: an update or delete of a controlled_by_parent detail asks for edit access to its master on every write shape (ADR-0055 §3) (#25042)
Fixes #25012
Clause-②: no (narrowing)
## What this changes
ADR-0055 §3 (`docs/adr/0055-master-detail-controlled-by-parent.md`):
"editing/deleting/creating a detail requires **edit** access to its
master". The `controlled_by_parent` detail write gate
(`assertControlledByParentWrite`, step 2.8 of the security middleware)
decided one write shape the engine routes without asking that question.
It now asks it on every shape.
- **The remaining shape is judged per master.** The gate collects the
masters of the detail rows the write would reach, as the caller's own
read returns them, and runs on each the same chain walk a by-id write
runs. The first refusal is the answer, in the by-id leg's own envelope
(`PERMISSION_DENIED`, 403, the same message), and nothing is written. On
an on-behalf-of write both principals are asked, as on a by-id write.
- **A verdict, never a scope.** No master set is composed into the
write. The read derivation (`computeControlledByParentFilter`),
`packages/objectql/**` and `packages/spec/**` are untouched, and
#24932's ruling B (`6303721609`) on the master-scope mechanism gets no
second spelling here.
- **The chain walk is extracted verbatim** into
`assertControlledByParentChainEditable`, so the insert, by-id and new
legs run one copy. Git shows the walk's lines as unchanged context; one
comment line in it is reworded.
- **It fails closed where the rows cannot be enumerated.** A refused
read of the detail refuses the write with that read's own refusal. Above
the existing 10,000-row read cap, the existing `INVALID_FILTER` refusal
answers.
**Unchanged:** inserts and by-id writes (same path, same walk), system
contexts, the read scope, the platform ownership floor (still applied on
this shape; nothing here vouches it away), and every object that does
not declare `controlled_by_parent`.
**Reach**, stated abstractly under the disclosure ruling on #9387
(`5323310017`): it was re-measured before the change on a real
`ObjectQL` + `SqlDriver` + `SecurityPlugin` + sharing stack, across
every refusal leg, three principal classes and both tenancy postures.
The measurement covers what bounded the rows. The readings are held
privately by the seat.
## Tests
New file:
`packages/plugins/plugin-security/src/controlled-by-parent-write-shape-master-edit.test.ts`,
15 tests on a real stack (`ObjectQL` over better-sqlite3, the real
`SecurityPlugin`, the real sharing service and middleware, contexts from
`resolveAuthzContext` + `assembleExecutionContext`, `isolated` tenancy).
Pinned behaviours:
1. A master the caller may not edit refuses the write. The refusal
equals the by-id door's answer for the same caller and master (code,
status and message, compared to the live by-id answer, not to a
literal). This holds for each refusal leg, for a member, an org admin
and an agent acting for a member, and for both verbs. Nothing is
written.
2. One such master among editable ones refuses the whole write, and
nothing is written under either.
3. A write whose rows all sit under editable masters lands as before.
The rows it writes are unchanged: own rows only for a member and its
agent (the floor), every row for an org admin.
4. A row the caller cannot read is not judged and is not written.
5. A caller whose read of the detail is refused gets that read's own
refusal, and nothing is written.
Ablation, one leg per pinned behaviour. Each leg was mutated through
`scripts/ablation-replace.mjs` (anchor hit 1 to 0, blob changed on disk)
and restored with `git checkout HEAD -- PATH`. Each restore was proven
by blob equality with HEAD (`cdc1192e52`) and an empty `git diff HEAD`.
No build step was needed: the suite imports the subject from `src/`
relatively.
| leg | mutation | observed |
|---|---|---|
| A | the new leg returns without judging | 11 red / 4 green: every
refusal pin red, plus pin 5; pins 3 and 4 stay green |
| B | every such write refused outright | 15 red / 0 green. The first
attempt was a no-op: the replacement contained its anchor and the tool
refused, so nothing ran. Re-anchored and re-run. |
| C | rows enumerated as system instead of through the caller's read | 2
red: pins 4 and 5 |
| D | the read's refusal swallowed as "no rows" | 1 red: pin 5 |
| E | only the first master judged | 1 red: pin 2 |
## Verification (head `2f80e0db95`, base `origin/main` `3b3830ca22`)
- `pnpm --workspace-concurrency=2 --filter
'@objectstack/plugin-security^...' build`: lock `VERDICT command-exit
0`.
- `pnpm --filter @objectstack/plugin-security test`: 200/200 files, 4129
passed, 45 skipped (taken at `44eef53692`; the merge from main touches
none of this package's sources or dependencies).
- `pnpm --filter @objectstack/plugin-security typecheck`: exit 0. This
includes `check:test-typecheck`, which reports 0 files and 0 errors held
in the debt ledger, so the new test compiles.
- `node scripts/pm/dispatch-gates.mjs --commands --repo
objectstack-ai/objectstack` derives 69 commands (identical at
`44eef53692` and `2f80e0db95`). All 69 ran under the lock at
`2f80e0db95`, each exit captured before any pipe: 69 exit 0. `--ran`
reconciliation: 69 derived, 69 run, 0 NOT-MEASURED (a derived zero), 0
UNRUN. At `44eef53692`, before the dogfood closure existed,
`check:dual-build-cjs-loads` and `check:i18n` had exited 3 (PREREQUISITE
NOT MET). They are green here.
- Full dogfood battery (owed by `(narrowing)` and the title's `!`):
dogfood closure built, then `pnpm --filter @objectstack/dogfood test`
under the lock with no shard and no path. `VERDICT command-exit 0`: Test
Files 284 passed and 3 skipped (287); Tests 2361 passed and 18 skipped
(2379); no failures.
- `node scripts/check-nul-bytes.mjs`: OK.
## Acceptance notes
- **Two docblock sentences outside the claimed region now describe the
old behaviour.** I left them untouched under the claim's hard stop (no
other region of `security-plugin.ts`). The seat can approve a
comment-only rider, or name a carrier.
- The third bullet of the `RlsFilterOptions.masterGateCoversThisWrite`
docblock says the gate "returns early and nothing replaces the floor"
when no single id is addressed. Proposed: "does not vouch: the master
gate judges this shape, but the floor stays, because dropping it here
would widen the rows a member writes (a separate decision)".
- Step 2.7's coverage-vouch parenthetical "(`extractSingleId`, or it
returns early)". Proposed: "(`extractSingleId`; with no single id step
2.8 judges every master the write reaches instead)".
- **Writes issued from inside another operation are judged too**, under
the caller's context (a hook's or a cascade's), as step 2.8 judges their
by-id twins. The write doors' readable narrowing excludes such writes,
so for them the judged set is the rows the caller can read.
- **Cost.** Each such write on a `controlled_by_parent` detail adds one
read under the caller's context, one system read of the master
references, and one chain walk per distinct master, each hop about three
reads. The bound is the existing 10,000-row cap.
- **Object-grained `security/explain`** (no record) for update or delete
on a `controlled_by_parent` detail answers `allowed` with no master leg.
It is a record-agnostic question; noted, not filed.
## Open question for the seat
- **Refuse versus narrow, as the long-term answer for this shape.** This
PR refuses: a verdict, with no master scope composed. Narrowing to
editable masters would compose the caller's editable-master set into the
write. That is a master scope, which ruling B decides, and ADR-0139 D1
refuses the engine leaf on write verbs. Refusing now is
forward-compatible: a later move to narrowing would be a widening.
- **Whether the floor should yield on this shape as it does by-id**
(commit 6feac91) is a widening, and it is not taken here.
---
**Seat's append** (`domain:services` seat 1, from the dev's patch-round
report `6307856898` on #25012; the dev writes its PR body once):
## Patch round 1 (head `033fc75a09`)
- **The two docblock sentences named in the Acceptance notes are now
fixed in this PR**, as a comment-only rider. This follows the seat's
ruling on the open question, letter A. `git diff HEAD~1 --
packages/plugins/plugin-security/src/security-plugin.ts` shows 7 lines
in and 3 out, and every changed line is a comment line (0 non-comment
lines). No code token changed.
- The refuse-versus-narrow question is ruled A: keep refusing.
- `origin/main` (`dfe60e3a0c`) was merged first, because it moved files
the suite depends on (`packages/objectql`,
`packages/drivers/driver-sql`, `packages/core`).
- Re-verified on `033fc75a09` under the lock:
- closure build exit 0;
- `pnpm --filter @objectstack/plugin-security test`: 200/200 files, 4129
passed, 45 skipped;
- `typecheck` exit 0;
- full dogfood battery (no shard, no path): `VERDICT command-exit 0`,
Test Files 284 passed and 3 skipped (287), Tests 2361 passed and 18
skipped (2379), no failures.
**Seat's append** (from the dev's patch-round-2 report `6308106727` on
#25012):
## Patch round 2 (head `e71bf1e576`)
- The changeset headline and its ADR-0087 note are scoped to update and
delete, as the reviewer asked; the diff from `033fc75a09` is that one
changeset file, with no code change, so the `033fc75a09` dogfood battery
stands.
---
_Generated by [Claude
Code](https://claude.ai/code/session_011TdZ8GjixNZFLz1xJD9zzK)_
---------
Co-authored-by: Claude <noreply@anthropic.com>1 parent 63085c3 commit b289090
3 files changed
Lines changed: 619 additions & 9 deletions
File tree
- .changeset
- packages/plugins/plugin-security/src
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
| 1 | + | |
| 2 | + | |
| 3 | + | |
| 4 | + | |
| 5 | + | |
| 6 | + | |
| 7 | + | |
| 8 | + | |
| 9 | + | |
| 10 | + | |
| 11 | + | |
| 12 | + | |
| 13 | + | |
| 14 | + | |
| 15 | + | |
| 16 | + | |
| 17 | + | |
| 18 | + | |
| 19 | + | |
| 20 | + | |
0 commit comments