Commit Graph

12803 Commits

Author SHA1 Message Date
Feng Ruohang 94fbb6df6d fix: never record a resident bucket as a CORS load failure
A bucket that is resident keeps its last loaded metadata through a failed
refresh, exactly like every other bucket configuration, so the load-failure
set only ever holds buckets that were never loaded. This makes the CORS
lookup's precedence explicit: resident document first, fail-closed only for
a bucket that has no loaded document.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PvgysXDmhPBBimCReYtA8q
Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-03 00:45:02 +08:00
Feng Ruohang f4c1286c9d fix: apply replicated Object Lock updates only when newer than the stored state
A replicated CopyObject uses the REPLACE metadata directive, so the map the
handler compared replication timestamps against had already been rebuilt
from the request and filtered of Object Lock keys: the stored retention and
legal-hold timestamps were never seen, every replica update was applied
regardless of order, and the legal-hold timestamp was written under the
retention key. A stale replica could turn a newer legal hold off or shorten
a newer retention.

Capture the stored Object Lock state before the metadata is rebuilt, apply a
replica update only when its source timestamp is newer, put the stored state
back when the update is stale, and keep each timestamp under its own key.
Inherited from upstream; recorded in the advisory ledger.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PvgysXDmhPBBimCReYtA8q
Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-03 00:45:02 +08:00
Feng Ruohang 6e112d1856 build: point Helm defaults at the 20260903 release
Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 23:11:32 +08:00
Feng Ruohang e62dc6e023 Merge branch 'codex/prerelease-cleanup': pre-release cleanup and dependency pins
Fifteen reviewed commits: streaming-trailer trust fix and a single
replication-trust evaluator, Object Lock versioning invariant, resident-only
CORS lookup with fail-closed startup and load-failure states, dead-code
removal, compatibility baseline without exported symbols, upstream minio-go
with Console 43f8447fd and mcli 20260901, advisory ledger SN-2026-006..010,
contributor records, verification scripts with a valid root password, and a
deterministic dynamic-timeout test.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PvgysXDmhPBBimCReYtA8q
Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 21:34:30 +08:00
Feng Ruohang 84e1580a47 fix: clear the load-failure bit when GetConfig reloads a bucket on demand
A successful on-demand load makes the bucket resident, so the failure
recorded by an earlier startup or refresh no longer applies. Clearing it here
keeps the set's invariant exact instead of waiting for the next refresh.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PvgysXDmhPBBimCReYtA8q
Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 21:23:30 +08:00
Feng Ruohang ebac0ca73b test: give the dynamic timeout tests a private random source
TestDynamicTimeoutAdjustExponential and TestDynamicTimeoutAdjustNormal
seeded the global generator and then drew from it while other tests in the
package may use the same generator, so the sample was not the one the seed
promised and the exponential case failed once in a full race run. A private
source makes both tests deterministic.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PvgysXDmhPBBimCReYtA8q
Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 20:04:57 +08:00
Feng Ruohang 3f9c79e919 fix: keep load-failed buckets fail-closed in the CORS lookup
Restore the load-failure set that the cleanup removed. A presigned URL is
authenticated by its signature, so for such requests the bucket's CORS
document is the only origin boundary a browser enforces; a real bucket whose
metadata failed to load must therefore not be answered with the global
policy, and without this bit it is indistinguishable from a name that is not
a bucket. Two helpers own the set's lifecycle; the resident-only lookup and
the removal of the internal-namespace special case stay.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PvgysXDmhPBBimCReYtA8q
Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 19:51:54 +08:00
Feng Ruohang ec2979ca48 fix: keep the CORS lookup fail-closed until bucket metadata is loaded
Restore the startup guard removed by the previous cleanup: while bucket
metadata is still loading, a non-resident name may be a bucket with a
restrictive CORS document, so the request gets no CORS answer instead of the
global policy. After startup a non-resident name still falls back to the
global policy without any metadata I/O; the separate load-failure set stays
removed.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PvgysXDmhPBBimCReYtA8q
Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 18:49:17 +08:00
Feng Ruohang 21646eebd2 fix: derive the Object Lock versioning rule from the parsed configuration
The load-time normalization compared the stored lock document with the
canonical enabled document byte for byte, so a lock configuration that also
carries a default retention rule kept a suspended or prefix-excluded
versioning document. Decide from the parsed configuration instead, after it
is parsed, so every writer that goes through Save, including the site
replication versioning and heal paths, ends with plain Enabled versioning on
a locked bucket. Receiving a lock configuration on a bucket created without
lock now enables versioning as well; the test that asserted the opposite is
updated, and a new test covers a rule-bearing lock document with suspended
and prefix-excluded versioning through Update, Get, and reload.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PvgysXDmhPBBimCReYtA8q
Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 18:49:16 +08:00
Feng Ruohang 5711996231 build: use an eight-character root password in the verification scripts
The rebrand shortened the scripts' root password from minio123 to silo123.
The server requires at least eight characters, so every script that starts
a server with it failed at startup and verify-build.sh then waited forever
on mc ready. No workflow runs these scripts, which is why it went unnoticed.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PvgysXDmhPBBimCReYtA8q
Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 15:53:08 +08:00
Feng Ruohang 0079723d35 deps: return to upstream minio-go and embed Console 43f8447fd
The silo-go fork is retired: silo-pkg v3.13.0, Console v2.3.0, and mcli
20260901 all require upstream github.com/minio/minio-go/v7 again. Drop the
replacement and require the same upstream pre-release they use
(v7.3.1-0.20260828014306-0e78d3f18efe, one commit past the fork's base).

Console moves from e07ef01 (v2.2.1 plus pins) to 43f8447fd, the last commit
of the v2.3.0 line before Console adopted the github.com/pgsty/silo-pkg/v3
module path. It carries the six v2.3.0 security fixes (forwarding-header
trust, outbound TLS verification, credential redaction, WebSocket session and
connection caps), the IAM wildcard and session-identity fixes, and their
regenerated assets, while still consuming silo-pkg through the existing
replacement. mc stays on the last commit before its own path migration;
silo-pkg stays on the last commit that declares the github.com/minio/pkg/v3
path, which differs from v3.13.0 only by that path change. CREDITS follows
the module set.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PvgysXDmhPBBimCReYtA8q
Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 14:51:10 +08:00
Feng Ruohang bc3b35f975 build: track only served routes in the compatibility baseline
The baseline recorded 9,051 exported symbols of the main and internal
packages, which nothing outside this module can import, and 112 request
paths that exist only in test fixtures. Both changed with almost every
functional commit and protected no compatibility promise. The guard now
records routes from non-test files only, ignores untracked files, and drops
the symbol set; the baseline shrinks from 522 KB to 46 KB. CONTRIBUTING
explains when to refresh it.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PvgysXDmhPBBimCReYtA8q
Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 14:08:11 +08:00
Feng Ruohang 00d864ed0e docs: record advisories SN-2026-006 to 010 and refresh contributors
Ledger entries for the zero-byte SSE-C key check, GetObjectAttributes
authentication, replication request trust, user and group status
authorization, and DeleteObjectVersion authorization, plus the Go 1.27
toolchain refresh. The ledger names pgsty/silo, CONTRIBUTORS lists the
per-bucket CORS, ChecksumType, and NoSuchBucket contributors, and the CORS
design record states its merged status without the review logs.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PvgysXDmhPBBimCReYtA8q
Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 14:08:11 +08:00
Feng Ruohang 8f2a30d9af build: bundle mcli 20260901 and describe the component pins
mcli RELEASE.2026-09-01 carries the credential redaction fixes for --debug
output. The go.mod comments now state what the Console and silo-pkg pins are:
the last commits that consume silo-pkg through the github.com/minio/pkg/v3
replacement, since silo-pkg v3.13.0 and Console v2.3.0 moved to the
pgsty/silo-pkg module path.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PvgysXDmhPBBimCReYtA8q
Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 14:06:56 +08:00
Feng Ruohang d9766d7378 chore: drop the wait_pipe lint exclusion and use gomodguard_v2
Assigning the two pipe halves before returning them removes the gofumpt and
gofmt disagreement that needed a permanent formatter exclusion. gomodguard is
deprecated in golangci-lint v2.12; the v2 linter takes the same configuration.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PvgysXDmhPBBimCReYtA8q
Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 14:06:56 +08:00
Feng Ruohang 5594d284fc chore: remove dead code left by earlier fixes
checkSSECCopySourceKey duplicated the source-key authentication that both
CopyObject read paths already perform; the DeleteObjects signature-error
branch became unreachable once signatures are verified once per request; the
discrete PostgreSQL and MySQL notification environment constants have had no
reader since DSNs became mandatory.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PvgysXDmhPBBimCReYtA8q
Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 14:06:56 +08:00
Feng Ruohang 632eb4729a refactor: serve non-resident buckets with the global CORS policy
The pre-authentication CORS lookup stays resident-only, so client-supplied
path segments still cause no metadata I/O and no cache growth. The
fail-closed states for startup, load failures, and the internal namespace
are gone: CORS is a browser response policy rather than an authorization
boundary, and failing closed only denied browser clients CORS headers while
bucket metadata was still loading. A bucket whose stored CORS document does
not parse still gets no CORS headers. This removes the loadFailed bookkeeping
and the unused GetCorsConfig, HasAllowedOrigin, and generic Update path for
CORS; tests use the CORS-specific writer.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PvgysXDmhPBBimCReYtA8q
Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 14:06:56 +08:00
Feng Ruohang 3b5de82f5a fix: keep plain Enabled versioning when Object Lock is enabled on a bucket
Site adoption and ForceCreate preserved a suspended or prefix-excluded
versioning configuration while bootstrapping Object Lock, persisting a state
that PutBucketVersioning itself rejects: objects under an excluded prefix in
a WORM bucket were not versioned and escaped retention. enablePeerBucketVersioning
now takes the lock intent and replaces such configurations with plain Enabled
versioning, and metadata loading ignores prefix exclusions on a locked bucket
as it ignored suspension before. The adoption tests assert the normalized
state and keep the timestamp-preservation checks on valid documents.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PvgysXDmhPBBimCReYtA8q
Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 14:06:56 +08:00
Feng Ruohang 76195f1c68 fix: keep streaming trailers visible after replication headers are stripped
A request that does not earn replication trust continues with a clone whose
internal replication headers are removed. r.Clone copies the Trailer map, but
the streaming body reader created from the original request fills the
original map, so a trailing checksum was never seen by the hash reader and
PutObject and UploadPart with STREAMING-UNSIGNED-PAYLOAD-TRAILER failed with
XAmzContentChecksumMismatch whenever an untrusted X-Minio-Source-* header was
present. Share the trailer map with the clone, as the Snowball path already
does for its per-entry requests, and cover both handlers with a test.

The marker evaluation that was copied into six handlers now lives in
evaluateReplicationTrust so the rule (a declared replica without the
replication permission is rejected; trust needs the exact marker plus the
permission) is defined once.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PvgysXDmhPBBimCReYtA8q
Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 14:06:56 +08:00
Feng Ruohang 6586fbfd0d Merge pull request #103 from pgsty/codex/issue-102-metadata-lock
fix: serialize whole-record bucket metadata updates
2026-09-02 07:30:40 +08:00
Feng Ruohang 53b09f4e25 style: apply gofumpt to migration globals
Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 07:20:53 +08:00
Feng Ruohang 59812446ad fix: surface metadata migration failures after locking
Degrade to the loaded in-memory configuration only when the short migration lock cannot be acquired. Preserve parse, I/O, KMS, and persistence errors once the lock is held.

Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 07:20:53 +08:00
Feng Ruohang 1c2f59604d docs: record metadata-lock rolling-upgrade limits
State that old cors-config.lock and new metadata.lock do not interoperate during a rolling upgrade, while the unchanged on-disk format remains rollback-compatible.

Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 07:20:53 +08:00
Feng Ruohang b3a6d5dbf8 fix: bound bucket metadata migration locking
Use a short dedicated migration lock timeout and fall back to the already loaded legacy or target configuration in memory when persistence is contended. Never fall back to an unlocked metadata save, and keep peer fan-out contexts free of stale lock markers.\n\nRefs: #102

Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 07:20:53 +08:00
Feng Ruohang 11de51ec26 fix: limit imported metadata replication to changed fields
Release metadata.lock with defer and dispatch only the configuration types actually present in the import. Avoid stamping unrelated peer metadata with the import timestamp.\n\nRefs: #102

Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 07:20:53 +08:00
Feng Ruohang 10f3a8590b fix: harden bucket metadata creation edge cases
Preserve existing records only for ForceCreate, reject ghost metadata on genuine creation, keep object-lock versioning invariants, and complete metadata saves after caller cancellation. Expand deterministic coverage for peer bulk, lifecycle delete, ghost creation, and cancellation.\n\nRefs: #102

Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 07:20:53 +08:00
Feng Ruohang 55c22abec1 docs: describe the shared bucket metadata lock
Replace the obsolete CORS-only lock and #77 references with metadata.lock and issue #102. Document lock order, fan-out placement, covered writers, and deliberate follow-ups.

Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 07:20:53 +08:00
Feng Ruohang 32a1b81e4c fix: rebase imports and migrations under metadata.lock
Apply only validated import fields to a fresh locked record, block ForceCreate after real read errors, and route legacy or target-config migration saves through the shared lock. Compute lifecycle deletion state from the locked record.\n\nRefs: #102

Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 07:20:53 +08:00
Feng Ruohang 62d8c649fe fix: preserve metadata during bucket creation and adoption
After storage bucket creation, merge required versioning and lock defaults into the latest on-disk metadata under metadata.lock. Avoid ForceCreate and site-adoption rewrites that replaced existing bucket configuration.\n\nRefs: #102

Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 07:20:53 +08:00
Feng Ruohang 312397739e test: reproduce ForceCreate bucket metadata clobber
Create policy and CORS state, force-create the existing bucket, and require the original Created time and both metadata fields to survive.

Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 07:20:53 +08:00
Feng Ruohang 7ade0c045b fix: serialize bucket metadata updates across config types
Use one per-bucket metadata.lock for ordinary updates, CORS transitions, and legacy bulk replication. Persist and update the local cache while locked, then release before peer metadata reload fan-out.\n\nRefs: #102

Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 07:20:53 +08:00
Feng Ruohang f9f9fa6c90 test: reproduce cross-type bucket metadata lost updates
Pause one whole-record writer at the metadata PutObject boundary and let a different config writer commit from the same stale snapshot. Assert that policy+CORS and tagging+SSE both survive on disk and in the resident cache.

Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 07:20:53 +08:00
Feng Ruohang 0a9c777795 Merge pull request #104 from pgsty/codex/issue-58-delete-version-authz
fix: authorize explicit version deletes with DeleteObjectVersion
2026-09-02 07:20:41 +08:00
Feng Ruohang d2d47a41fb chore: refresh delete authorization compatibility snapshot
Accept the additional DeleteObjects route literal introduced by the issue #58 regression matrix.

Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 07:08:39 +08:00
Feng Ruohang dee2c3a02b fix: preserve multi-delete authentication and audit context
Authenticate DeleteObjects before validating entry count, remove the obsolete per-version auth helper, and pin the marker-only request to the ordinary authorization path.\n\nRefs: #58

Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 07:08:02 +08:00
Feng Ruohang c4140609b7 test: verify least-privilege delete replication
Document that replication targets retain the DeleteObject plus ReplicateDelete contract and extend the existing two-site test with a target user that lacks DeleteObjectVersion. Use mc for explicit version deletion so the gate no longer depends on AWS CLI.

Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 07:08:02 +08:00
Feng Ruohang 75a6734e49 fix: authorize explicit version deletes with DeleteObjectVersion
Select DeleteObject or DeleteObjectVersion from each request's effective version ID for single and multi-delete. Authenticate multi-delete once, bind version conditions to each XML entry, and keep the established DeleteObject plus ReplicateDelete target contract with explicit version denies honored only on trusted replication.\n\nRefs: #58

Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 07:08:02 +08:00
Feng Ruohang f8b598f1d3 test: reproduce explicit-version delete authorization mismatch
Prove that DeleteObjectVersion-only principals cannot delete named UUID or null versions while DeleteObject-only principals can, contrary to the S3 action mapping in issue #58.

Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 07:08:02 +08:00
Feng Ruohang fe9dd90b22 Merge pull request #101 from pgsty/codex/fix-cors-replication-trust
fix: harden CORS and replication request trust
2026-09-02 07:07:49 +08:00
Feng Ruohang 04b097fd9f chore: refresh compatibility and lint baselines
Accept the new CORS test routes, resident getter, and replication header literals in the rebrand guard. Apply gofumpt, context-first helper ordering, and spelling fixes required by CI.

Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 02:31:20 +08:00
Feng Ruohang ab3ae99ca3 fix: preserve Snowball request defaults across workers
Snapshot per-entry requests after applying bucket encryption defaults but before streaming trailers are consumed. Keep authorization failures fatal while retaining Snowball ignore-errors behavior for object-lock failures.

Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 00:20:06 +08:00
Feng Ruohang 5db7be4ee4 fix: validate replication within the rule prefix
Place synthetic permission-check objects under each enabled rule's effective prefix, so least-privilege target policies are validated against the namespace they will actually replicate.

Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 00:03:33 +08:00
Feng Ruohang ff44527a3c fix: isolate Snowball replication trust per entry
Evaluate PutObject and ReplicateObject permissions with immutable per-entry request snapshots during concurrent Snowball extraction. Preserve the first API error without sharing mutable handler state, and cover prefix-scoped trust under the race detector.

Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-02 00:02:24 +08:00
Feng Ruohang c9ad746732 fix: verify replication permissions in validity probes
Evaluate ReplicateObject or ReplicateDelete before returning the no-op validation response, so underprivileged target credentials fail during replication setup instead of at runtime.

Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-01 23:16:09 +08:00
Feng Ruohang f3438b2602 fix: validate CORS state in replication status
Count only valid live CORS states in per-site summaries. Treat baselines and tombstones as absent, and diagnose malformed payloads or missing source timestamps.\n\nRefs: #77

Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-01 23:16:04 +08:00
Feng Ruohang 938603458d fix: harden CORS and replication request trust
Keep pre-authentication CORS lookups resident-only so attacker-controlled path segments cannot trigger metadata I/O or grow the metadata cache. Preserve fail-closed behavior for startup, load failures, invalid metadata, and the internal namespace.

Centralize replication request trust after authentication, distinguish general replication from replica-only privileges, and gate SSE-C ciphertext handling, source metadata, object-lock bypasses, event suppression, delete semantics, and replica status on the appropriate permission. Add least-privilege, multipart, PostPolicy, CORS amplification, and compatibility regressions.

Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-09-01 20:50:08 +08:00
Feng Ruohang d5e763b072 Merge pull request #98 from pgsty/fix/attributes-ssec-replication-authz
fix: authorize SSE-C attribute reads by replication permission
2026-08-30 18:42:10 +08:00
Feng Ruohang 74c97d005d fix: authorize SSE-C attribute reads by replication permission
GetObjectAttributes lets a replication peer read SSE-C attributes without
presenting the customer key. That carve-out was keyed on the
X-Minio-Source-Replication-Request header alone, which any client can set,
so a caller holding only s3:GetObject could read an SSE-C object's ETag,
plaintext size, part list, and checksums without the key -- the response was
byte-identical to one made with the correct key.

Gate the carve-out on s3:ReplicateObject for the target object, mirroring
CopyObjectHandler's existing replication check.

The pre-existing test asserted the carve-out with root credentials, which
hold every action and therefore cannot tell a gated check apart from an
ungated one. Add least-privilege cases that do: a reader without
s3:ReplicateObject is now refused, a caller that holds it keeps the
carve-out, and the ordinary key-bearing path is unchanged.

Signed-off-by: Feng Ruohang <rh@vonng.com>
2026-08-30 15:50:24 +08:00
Feng Ruohang 1be684ba29 Merge pull request #97 from pgsty/codex/pin-prerelease-components
deps: pin reviewed pre-release components
2026-08-29 20:49:58 +08:00
Feng Ruohang 1367364c40 Merge remote-tracking branch 'origin/main' into codex/pin-prerelease-components 2026-08-29 20:33:44 +08:00