resource_limit (intrinsic to the frozen input: frame, input or output cap,
SIGXCPU, SIGABRT, SIGSYS) and verdict_refused close the round at the first
occurrence; timeout, unavailable (now also an external SIGKILL) and
lease_expired keep the 2/5/10/20 s retries up to five attempts. pgTAP 021: 89;
sandbox case k closes at attempt 1.
Max frame (document and base 2 MiB, drawings 6 MiB) passes through the full
sandbox; refusals just beyond give caps verdicts; RLIMIT_AS and the argument
filters still stop the verify child; seccomp traced on the max input; the two
tracked misclassifications asserted as current behaviour.
Third audit (BLOCK 1): counting terminal rows while queued let begin spam
pin the ceiling with declared sizes. The ceiling now sums storage.objects
sizes of the bucket (whatever the row's state) plus the declared size of
pending_upload rows whose quarantine object has not landed: a claim dies
with its row, a present object counts whatever its row says, no full scan of
the assets history. confirm_asset_upload treats an object without a size as
upload_missing. pgTAP (plan 138): rejected rows without bytes weigh nothing,
pending claims fill, superseded claims vanish, present bytes fill, unsized
object refused. Worker: a single-label cleartext host must start with an
ASCII letter (0X… hex forms refused, test added); STORAGE.md names the
native-run search-domain residual.
Re-audit findings: the ceiling now also counts terminal rows whose keys still
wait in the purge queue (contract wording kept honest, residual named);
confirm_asset_upload refuses an object larger than the declared size;
can_upload_avatar_object joins profiles.active; pgTAP covers the null
sha256 guard on a live lease, the three service_role revocations, the
queued-purge accounting and the deactivated-account upload (plan 137).
Worker: cleartext http only for localhost, a dotted-quad loopback address or
a non-numeric single-label name (numeric and hex hosts refused, tests added).
Security audit (PASS, 2 MEDIUM + LOW): the global ceiling summed every
non-deleted row, so rejected uploads ratcheted it toward storage_full with no
disk used → in-flight rows by declared size, clean rows by byte_size, nothing
else (pgTAP: rejected rows do not count, in-flight rows do); confirm refuses a
deactivated account like begin; complete_avatar_job refuses a null sha256;
begin / confirm / remove granted to authenticated only. Worker: SUPABASE_URL
over http:// accepted only for loopback or a single-label Docker service
name; ureq proxy(None); the job's keys must embed its owner and asset ids;
backend_failure / claim_failed / sweep_failure lines carry the closed kind and
the HTTP status, never the library's error text. e2e: uuid shape checked
before the psql query. Contract revision 2 (oracle fixes).
packages/supabase/tests/e2e/avatar-flow.e2e.ts (pnpm e2e:avatar through
scripts/avatar-e2e.sh: worker container on the Supabase network, 5 s sweep,
no quarantine grace; a fresh account per run): clean upload and owner-only
download (Bob, anon, quarantine key, direct final-key write refused),
replacement purged after the grace (queue clock shifted through psql), late
bytes through a still-valid signed upload URL swept, hostile and corrupted
uploads rejected with closed codes, removal with the purge queue drained.
Worker: the Storage API answers a missing object with HTTP 400 and
statusCode 404; normalised to 404 so a purge of a never-written key is acked.
docker-compose.yml: service worker under the worker profile (read-only,
cap_drop ALL, no-new-privileges, 512m, 1 cpu, 64 pids, .env.worker, external
Supabase network). docs/development/STORAGE.md: the implementation contract.
Keyed parent on ureq 3 (rustls + ring, bundled roots), no async runtime:
config from the environment (secret never printed, JWT expiry logged), JSON
line logs with a field allowlist (child stderr escaped and truncated), strict
key validation before any URL, Backend trait over PostgREST RPCs and Storage
objects with closed error kinds, one job per claim with the STORAGE.md §3.4
mapping (content codes, processing_error on signals / timeout / overflow,
Broken stops the loop with exit 20, a refused commit deletes nothing), sweeper
(expire pending, purge queue, quarantine leftovers), per-stage watchdog
(exit 21), WORKER_ONCE for tests, run --check extended. 11 loop tests on a
fake backend and decoder; sandbox tests and seccomp derivation unchanged
(the child still makes the same 7 syscalls); image 11.6 MB.
check-all.sh gains a warn-only cargo deny check advisories stage (no nightly
job runs yet) and the ts-rs drift gate uses git status --porcelain so a new
untracked type file fails too (same in ci.yml); rust-env.sh appends an
absolute ~/.cargo/bin only; wasm-build.sh removes the literal
packages/core-wasm/pkg; the worker passes the unused prctl arguments as
c_ulong and its SAFETY comments no longer claim the kernel checks them.
The [gated <date>] markers read as an approval stamp: replaced by plain
SAFETY comments giving, for each of the seven unsafe sites, the FFI contract
and where the surrounding code establishes it (buffer length and stride
checks, non-zero size and non-null output before from_raw_parts, single
WebPFree, prctl integer arguments with the return code checked, the two
negative hooks compiled with test-hooks only). The stride is now computed
with checked_mul instead of width * 4.