Verified-safe deletions from the 2026-07-23 deletion audit, applied now
that the permission-consolidation work (which deferred IsOwnerRole) has
landed:
- Server/service/voice.go: VoiceService was constructed in service.New
and never called by any handler, ws routine, or test.
- permissions.IsOwnerRole: zero callers.
- Server/admin/static/admin-mockup.html: 1299 lines embedded into every
release binary via //go:embed static, referenced by nothing.
- .cache/project-map/*.json: tool cache committed before .gitignore
grew the .cache/ rule.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Closes audit finding A-2026-07-16, two defects in the same rule:
- permissions.HasServerPerm (admin bypass OR all-of bit test) replaces
the hand-rolled copies in api.RequirePermission (whose raw test was
any-of for multi-bit masks) and ModerationService.requireBanPermission.
RequirePermission's doc comment now states the scope contract: role
bitfield only, channel overrides deliberately not consulted.
- PermissionService.getOrPopulate and ChannelService.ListVisibleChannels
no longer substitute an empty override map when
GetAllChannelPermissionsForRole errors. That silently dropped every
channel-level deny — and the permission cache then served the degraded
snapshot for permCacheTTL (30s) across ~25 callers. Both fail closed
now; admins skip the fetch entirely (they bypass channel checks).
- PermissionService.HasChannelPerm delegates to Checker.HasChannelPermBatch
and MessageService.GetAccessibleChannelIDs to VisibleChannelIDs — the
missed fifth D9 site, making that closure true rather than aspirational.
- AuthMiddleware rejects a dangling role_id (GetRoleByID returns nil,
nil) with 401 instead of putting a nil role in the request context.
Locked by failing-first tests: override-fetch-error denies (cached and
uncached paths), admin-outage skip, multi-bit all-of, channel allow
override must not grant a server-wide route, 403 locks on both
RequirePermission routes, dangling-role 401.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
An InvalidateUser/InvalidateChannel/InvalidateAll landing between
getOrPopulate's DB read and its cache store was silently overwritten by
the stale snapshot, serving revoked permissions for up to permCacheTTL
(30s). Guard the cache write with a generation counter bumped by every
invalidation; a populate that lost the race returns its snapshot for the
current request but caches nothing (security scan 2026-07-22, F6).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
hasChannelPerm resolved permissions from the connect-time role snapshot (c.user.RoleID), so a user reassigned to a lower role kept the old role's voice privileges (CONNECT_VOICE and the SPEAK/VIDEO grants in the LiveKit token) until reconnect. Resolve the current role via GetRoleForUser(c.userID), matching the V2 handlers. (Security scan F5)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
invokeCommand drove a shared wazero module (allocate/mem.Write/command_dispatch/mem.Read) with no per-instance lock, so concurrent invocations of the same plugin command raced the module's linear-memory buffer. Add a per-Instance mutex around the guest-call sequence. Confirmed under -race. (Security scan F2)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The per-username brute-force lockout keyed on the raw request username while GetUserByUsername matches COLLATE NOCASE, so case variants (admin/Admin/ADMIN) each got an independent 9-attempt bucket, multiplying allowed guesses per account. Lowercase the username before building the login_user_fail/login_user_lock keys so all casings share one bucket. (Security scan F1)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
FetchTextAssetCached served the unauthenticated, unrate-limited client-update
endpoint. On TTL expiry every concurrent caller missed the cache and issued its
own outbound fetch: a measured 25 requests for 25 callers. Failures were not
cached at all, so an upstream outage produced one outbound request per caller
for as long as it lasted.
Guards the refresh with singleflight so a burst issues one fetch, caches
failures for errorCacheTTL (mirroring the release cache's existing cachedErr
idiom), and evicts expired keys so the map no longer grows by one entry per
release for the process lifetime.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
nhooyr.io/websocket now resolves to github.com/nhooyr/websocket-old and its
README is a one-line deprecation pointing at coder/websocket. Its last three
releases (v1.8.15-17) all shipped on 2024-08-10 as the redirect; the fork has
shipped through 2026-06-15.
The version number decreases (v1.8.17 -> v1.8.15) because both paths tagged in
the same space, but the coder release is ~2 years newer. Import path only; the
9 API symbols used are unchanged.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The closure rationale for audit finding #4 claimed in five places that
nothing in the server calls EventSink.Dispatch. That is disprovable by
grep: ws/hub.go:1034 calls Dispatch on every broadcast message, and
api/router.go:134-139 wires h.pluginSink whenever plugins are enabled.
The call site is pre-existing on main, not introduced by this branch.
Restate the closure on the claim the evidence actually supports:
Dispatch has exactly one caller outside the plugin package's tests
(ws/hub.go, on the hub's broadcast goroutine under seqMu), but its loop
body invokes no guest code and no production code calls Subscribe, so
the subscriber set is always empty and no guest code executes on the
event path. Finding #4 stays closed; the reason changes.
Also warn on Subscribe that adding the first production caller turns
Dispatch's loop live on the hub's hot path, and note in the SECURITY
GATE that the call site already exists so wiring delivery is not a new
integration.
Corrected in: plugin/host_events.go (Dispatch + Subscribe comments),
plugin/audit_closure_test.go, docs/audit-2026-04-07.md (row 4 and the
structural-mitigation paragraph), docs/audit-2026-07-19.md §1 row,
docs/plans/audit-2026-07-19-decisions.md D11.
Comments and docs only — no behaviour change.
Closes audit-2026-04-07 CRITICAL #3. Holding the `commands` capability used
to bind whatever names the guest module returned from `list_commands`, so an
admin enabling a plugin could not know which commands it would claim and a
plugin could widen its own command surface after review.
The manifest is now the authority. `plugin.json` gains a `commands` block
(`[{"name": "hello"}]`) and `RegisterCommand` refuses any undeclared name —
the single choke point both auto-registration and direct registration route
through, so no caller can bypass it. Declared names are validated to the
dispatcher's canonical lowercase form, deduplicated, and capped at 64.
The object shape matches docs/plans/slash-commands.md so the richer
per-command schema can land later without a manifest migration.
Also pins the two neighbouring CRITICALs that verification found already
closed, and adds the storage key cap host_storage.go's doc comment already
promised:
- #2 (storage key isolation): TestStorageKeysIsolatedPerPlugin — the KV
namespace is the caller's Instance.ID with no parameter to override it,
and plugin_kv PRIMARY KEY (plugin_id, key) makes the split structural.
- #4 (event rate limit): TestEventDeliveryHasNoGuestPath — EventSink.Dispatch
invokes no guest code and has no callers, so there is nothing to limit yet;
a SECURITY GATE comment requires the limiter in whatever change wires
delivery.
- #5 mitigation: TestEmptyAllowlistDeniesEveryHost — the shipped empty
http_allowlist must fail closed.
BREAKING CHANGE: a plugin declaring the `commands` capability must now list
its commands in the manifest's `commands` block; undeclared names no longer
bind. Only the in-repo `hello` example is affected and is updated here.
The client held the Klipy key in VITE_KLIPY_API_KEY, which Vite inlines into
the shipped bundle by design — a build variable can never hold a secret. Move
the integration behind the server:
- New authenticated GET /api/v1/gif/search and /api/v1/gif/trending. The key
comes from the new `gif.api_key` config section (koanf,
OWNCORD_GIF_API_KEY) and never leaves the server.
- Default-off: with no key, both endpoints return 503 GIF_DISABLED so clients
can hide the picker instead of showing a broken one. Auth is checked first,
so anonymous callers cannot probe whether a key is configured.
- Outbound call reuses the existing SSRF-guarded dialer (exported as
plugin.GuardedDialContext) rather than a bare http.Get: resolve once,
reject private/loopback/link-local/CGN, dial only vetted IPs. Redirects are
not followed and the response body is size-capped.
- Only id/title/media_formats.{tinygif,gif}.url are forwarded — decoding into
the narrow struct is the allowlist, so an upstream that echoed the key
could not leak it. Upstream errors become a generic 502 and the key is
redacted from anything that reaches the logs.
- Dedicated `gif:` rate-limit bucket (30/min per IP) so debounced search
traffic cannot exhaust the shared bucket used by password/TOTP endpoints.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Port the last three V1 message types to typed V2 handlers, then remove
the V1 registry and fallback path so handleMessage has a single dispatch
generation (audit A-2026-07-09 / backlog item 11). Server-internal only —
the envelope wire format is unchanged, no client/protocol edits.
- chat_command: ChatCommandCmd + constructor (empty/args guards) and a
V2 handler returning an ephemeral Reply plus a channel-routed
PluginBroadcastEvent gated by MessageService.CanPost; PluginDeps reads
the registry live (wired post-construction)
- voice_join/voice_leave: V2 handlers gate parse/rate-limit and hand off
to the still-hub-internal handleVoiceJoin/handleVoiceLeave routines via
new Result.JoinVoice / Result.LeaveVoice appliers (those routines are
also called un-throttled on disconnect and channel switch)
- delete HandlerRegistry.handlers/Register/Dispatch/RegisteredTypes/
IsRegisteredV1/hasV2, the MessageHandler type, and the V1-shadowing
guard; NewHub registers only V2
- tests: per-handler V2 tests + a parity guard asserting every command
constructor has a V2 handler and vice versa (locks the migration shut)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Route the four "must mirror" READ_MESSAGES filters — REST
ListVisibleChannels, ws buildReady, reconnect replay
computeAllowedChannels, and hub RefreshChannelVisibility — through a
single permissions.Checker predicate so a drift can never leak a
private channel (audit A-2026-07-07 / backlog item 3).
- add permissions.Checker.VisibleChannelIDs + ChannelRef (skips dm,
fails closed, admin bypass via HasChannelPermBatch)
- delegate the three batch sites; RefreshChannelVisibility uses
HasChannelPerm instead of its inline EffectivePerms copy
- REST/WS agreement test asserting all three sites yield the identical
non-DM set across admin / member-with-deny / denied-everywhere
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Audit writes stay best-effort — a LogAudit failure must never fail or abort
the request — but a failed write must no longer be silently discarded. Add
db.WriteAudit(auditor, actor, action, targetType, targetID, detail), which
logs a failed write with actor/action/target context (never the detail
string, which may be sensitive) and never propagates the error.
The Auditor interface is satisfied structurally by both *db.DB and the
service-layer Store, so api/admin/ws/service all reach the helper without an
import cycle. Converts all ~26 call sites from `_ = LogAudit(...)` (and the
two backup handlers' inline `if err` blocks) to db.WriteAudit. Pinned by
db/audit_test.go: failure logged and not propagated, success logs nothing,
detail never leaks.
Resolves the repo-wide LogAudit policy question flagged by the D8 note.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Implements the two highest-impact gaps from the client UX spec.
Optimistic send:
- messages.store gains addOptimisticMessage / markSendFailed /
removeOptimistic, and confirmSend now stamps the real id + "sent" on
the ack. addMessage reconciles the broadcast by real id (idempotent,
replay-safe) with a defensive author match, so an echo never
duplicates. Message gains status/correlationId/errorCode.
- ChannelController.performSend renders a pending row immediately and
supports retry / delete-draft (retry preserves attachments).
- MessageList renders pending (dimmed) and failed (reason + Retry /
Delete) rows; the hover action bar is limited to confirmed rows.
- Failures are precise: the server echoes the request id on error
replies (buildErrorMsgWithID), so the dispatcher maps SLOW_MODE /
FORBIDDEN / RATE_LIMITED / BAD_REQUEST to the exact row instead of
dropping the code. An offline send is shown failed, not silently lost.
Composer permission + connection gating:
- The server computes an authoritative per-channel can_send in the ready
payload (channelCanSend mirrors MessageService.checkSendPermission:
READ|SEND, MANAGE_MESSAGES for announcement, admin bypass, channel
overrides). channels.store carries it as Channel.canSend.
- MessageInput gains a disabled-with-reason mode; ChannelController
derives the reason from can_send + channel type + connection status and
disables the composer reactively (announcement read-only, no-permission,
reconnecting) rather than accepting a click and failing. Older servers
that omit can_send default permissive.
Docs: the corresponding "Current gap" callouts in docs/architecture/ux
are updated to reflect the implementation.
Verified: full server suite + client tsc + 3204 unit tests + lint + gofmt.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UA17KPvqGBX3XbXYnMf1rA
Deletes Server/store (SQLiteStore, MemStore, the composed Store
interface) and collapses to a single sqlc-backed db package, executing
the prior audit's P4 "single data layer" direction (finding #6).
SQLiteStore was a pure pass-through to *db.DB, so consumers now depend
on narrow interfaces that *db.DB satisfies directly:
- service.Store (service/datastore.go, renamed from store/store.go)
- ws.EventStore (ws/eventstore.go)
- plugin.PluginStore (plugin/pluginstore.go)
The event- and plugin-KV methods that lived in the store's SQLite
implementation move into the db package (db/event_queries.go,
db/plugin_queries.go), keeping their raw-SQL form.
Tests: the MemStore-based unit tests now run against a real in-memory
SQLite db opened per-test with migrations applied, via package-local
seed helpers. Fault-injection tests embed a real *db.DB and override the
single method under test, preserving error-path coverage. Full server
suite and sqlc-verify are green.
Docs: audit finding #6 and A-2026-07-06 marked resolved; decisions D3
updated; architecture server.md / data-model.md diagrams and prose
updated to the api -> service -> db layering.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UA17KPvqGBX3XbXYnMf1rA
Make 'announcement' a real channel type, resolving the contradiction where
it was documented and offered by the admin API but hard-rejected by the
migration-013 DB triggers.
Model: announcement channels are readable like text channels (same
READ_MESSAGES visibility), but posting is restricted to users with
MANAGE_MESSAGES — no new permission bit, migration, or client permission
plumbing needed.
Server:
- migrations/016: recreate the channel-type triggers to allow
text/voice/announcement/dm.
- service/message.go: checkSendPermission now takes the channel type and
rejects posts to announcement channels from users lacking MANAGE_MESSAGES
(SendMessage + CanPost paths). Added a service test.
- Unread counts: ready-payload builder (ws/serve.go) and
GetChannelUnreadCounts (db) now include announcement channels alongside
text, so they track unread/last-message like text channels.
Client:
- ChannelSidebar renders announcement channels with a megaphone icon
(added to the icon set) instead of the '#' text prefix; they otherwise
behave like text channels (already typed in ChannelType).
Specs + trackers (api.md, protocol.md, schema.md incl. migration 016,
architecture/data-model.md, audit A-2026-07-01, decisions D1) updated.
Verified: go build ./...; go test ./service ./db ./ws ./api ./admin;
sqlc-verify; client tsc + oxlint + prettier clean.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UA17KPvqGBX3XbXYnMf1rA
messages/reactions: CreateMessage, GetMessage (messageFromGen mapper),
EditMessage (EditMessageContent), DeleteMessage (SoftDeleteMessage),
AddReaction, RemoveReaction, GetReactions (GetReactionCounts),
SetMessagePinned, UpdateReadState. Retired the obsolete scanMessage.
Kept raw by design (no clean sqlc mapping): FTS search, cursor-paginated
GetMessages/GetMessagesForAPI/GetPinnedMessages, getReactionsBatch,
GetChannelUnreadCounts, GetLatestMessageID (interface{} MAX result).
D2 status: 97 db.DB methods now delegate to dbgen across every domain;
43 raw d.sqlDB calls remain by design (db.go passthroughs, migrate.go,
variable-length IN(), FTS, multi-statement transactions, PRAGMA/VACUUM).
sqlc is no longer dead code — audit A-2026-07-05 resolved. Full rationale
+ the kept-raw list in docs/plans/sqlc-adoption.md.
Verified: go build ./...; go test ./db ./service ./ws ./auth; sqlc-verify;
gofmt + go vet clean.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UA17KPvqGBX3XbXYnMf1rA
Convert the auth_queries.go user + session reads/writes and
profile_queries.go to the sqlc-generated layer, adding shared
userFromGen/sessionFromGen mappers (db/mappers.go) for the
int64/*string -> int/bool/string domain-model narrowing.
Delegated: GetUserByID, GetUserByUsername, UpdateUserStatus,
UpdateUserTOTPSecret, ResetAllUserStatuses, BanUser, UnbanUser,
ListMembers, CreateSession (EvictOldestSessions + InsertSession),
GetSessionByTokenHash, GetSessionWithBanStatus, DeleteSession,
DeleteOtherSessions, DeleteExpiredSessions, TouchSession,
UpdateUserProfile, UpdateUserPassword, ListUserSessions,
DeleteSessionByID (query changed to :execresult so the RowsAffected
ErrNotFound check is preserved). Behavior and public signatures unchanged.
Verified: go build ./...; go test ./db ./service ./auth; sqlc-verify.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UA17KPvqGBX3XbXYnMf1rA
Wire the sqlc-generated dbgen package into db.DB so it stops being dead
code (audit A-2026-07-05) and becomes the real, CI-verified query layer.
db.DB now holds a *dbgen.Queries (initialized in Open via dbgen.New).
Query method bodies delegate to it; sqlc owns the SQL text and parameter
binding (make sqlc-verify), while db keeps its stable public API and
domain model types so no caller in api/admin/ws/service changes. The
migration is incremental — a method either delegates to d.q.* or still
runs raw SQL — so both layers are correct during the transition.
Converted domains (now load-bearing through sqlc):
- blocks: BlockUser, UnblockUser, IsBlocked, IsEitherBlocked,
ListBlockedUsers (added the query to blocks.sql + regenerated).
Empty ListBlockedUsers now returns []int64{} instead of nil, matching
the MemStore backend — a latent inconsistency fixed, not a regression.
- lockouts: UpsertLockout, LoadActiveLockouts, CleanupExpiredLockouts,
DeleteLockout (RFC3339 time formatting/parsing kept in the wrappers).
- roles: GetRoleByID, ListRoles, GetRoleForUser via a shared roleFromGen
mapper (int64 position/is_default -> int/bool). GetUserWithRole stays
raw for now.
Remaining domains stay on raw SQL and are tracked in
docs/plans/sqlc-adoption.md; store/ event+plugin SQL is intentionally
excluded (that layer is removed in D3). Decisions doc + audit closure
updated (A-2026-07-05 -> in progress).
Verified: go build ./...; go test -race ./db ./service ./auth ./ws (api
green non-race, race run matches CI's -timeout 20m); make sqlc-verify and
protocol-verify pass with the regenerated output committed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UA17KPvqGBX3XbXYnMf1rA
Protocol codegen (decision D4, audit A-2026-07-08):
- Add docs/protocol-schema.json as the real single source of truth for
WS message-type constants, making the long-standing 'generated from'
comment in both constant files true.
- Add Server/scripts/genprotocol, a generator emitting both
Server/ws/message_types.go and Client .../lib/protocolTypes.ts
(constants byte-for-byte value-identical to before; only headers,
ordering alignment, and provenance comments changed).
- Add make protocol-generate / protocol-verify and wire protocol-verify
into CI next to sqlc-verify.
Quick wins (decision D8):
- admin: log LogAudit write failures in the backup handlers instead of
discarding them (prior audit #10).
- api: fix self-contradictory upload Cache-Control to 'private,
no-cache' per remediation plan W3-4; drop the now-unused
fileCacheMaxAgeSeconds constant; update test.
- ws: route the hub settings cache through db.GetSetting instead of
inline SQL.
- ws: fix a latent data race — main.go wires SetEventPersister and
SetEventStore after NewRouter has already started the hub Run loop,
which reads those fields on the broadcast/replay paths. They (and
pluginSink, which one test sets post-Run) are now atomic pointers;
the remaining pre-Run-only setters reject late calls with an error
log instead of racing silently.
Update the audit closure table and decisions doc statuses accordingly.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UA17KPvqGBX3XbXYnMf1rA
Three review findings on the #93 feature:
- RefreshChannelVisibility targeted clients by their connect-time role
snapshot; a user whose role changed mid-session was evaluated against the
stale role. Resolve the current role from the DB per client (fail closed).
- Visibility updates are targeted, unsequenced messages, so a client that
disconnected before an override change and later resumed via replay never
converged (stale sidebar until a fresh connect). Track a visibility-change
sequence watermark and force resumes from at/before it onto the
full-ready path.
- The admin SPA interpolated channel/user names into single-quoted JS
strings inside onclick attributes with HTML-escaping only; a name
containing a quote broke out of the string literal (XSS in the admin
panel, reachable by any user allowed to create channels). Add a jsq()
helper (JS-escape then HTML-escape) and use it for every onclick name
interpolation.
Follow-up to #93.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LwtnpHAoSFr1ZibQgQkNQK
A zero-byte or whitespace-only livekit.yaml (truncated write, touch(1)
placeholder) has no auto-generated marker and was permanently treated as a
user-managed config, wedging LiveKit startup with an empty config file.
Follow-up to #111.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LwtnpHAoSFr1ZibQgQkNQK
The read side of channel visibility was already complete — channel_overrides
rows deny ReadMessages and every list/focus/send/voice path enforces them —
but nothing could write those rows. Add the missing write path and UI:
- db: UpsertChannelOverride / DeleteChannelOverride / ListChannelRoleOverrides
(roles LEFT JOIN overrides so the UI gets everything in one call)
- admin API: GET/PUT/DELETE /admin/api/channels/{id}/permissions[/{roleId}]
with unknown permission bits masked via the new permissions.AllPerms,
audit logging, and immediate permission-cache invalidation
- ws: Hub.RefreshChannelVisibility sends targeted channel_create /
channel_delete to connected clients after an override change, unsubscribes
hidden clients from the channel topic, and clears their focus. Sent outside
the sequenced replay path on purpose: a replayed channel_delete would be
filtered by the post-change allowed-channel set, inverting its audience.
- admin panel: per-channel Access modal (lock icon) with per-role
"Can access" checkboxes; unchecking writes deny = ReadMessages|ConnectVoice
Known limits (follow-ups): users offline during a revoke keep a stale
sidebar entry until their next fresh connect (server still denies access),
and users already in a voice channel are not kicked when it goes private.
Closes#93
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LwtnpHAoSFr1ZibQgQkNQK
Servers reachable via both a LAN IP and a public IP could only serve voice
on one of them: config.yaml accepts a single voice.node_ip and OwnCord
regenerates data/livekit.yaml on every start, discarding manual edits.
LiveKit has no multi-IP list, but it does support advertising internal host
candidates alongside the external mapping.
- New voice.advertise_internal_ip (OWNCORD_VOICE_ADVERTISE_INTERNAL_IP):
emits rtc.advertise_internal_ip: true so LAN clients get a reachable
candidate while remote clients keep using node_ip.
- livekit.yaml escape hatch: if the file exists without the auto-generated
marker header, OwnCord leaves it untouched, giving operators access to
every LiveKit option (ips.includes, interfaces, stun_servers, ...). The
generated header documents how to take ownership.
Closes#111
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LwtnpHAoSFr1ZibQgQkNQK
contextcheck (CI lint) flagged the admin handler calling BanUser without
the request context — the service opened its telemetry span from
context.Background(), detaching the ban from its request trace. Both
moderation entrypoints now take ctx; the span joins the caller's trace.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Register and Unregister travelled on two separate channels, and Run's
select picks randomly when both are ready: a fast connect/disconnect could
process the unregister first (a silent no-op for a not-yet-known client)
and then the register — admitting an already-dead connection as a ghost
client that held presence and swallowed broadcasts until the stale sweep
reaped it minutes later. One tagged event channel preserves each
connection's Register→Unregister submission order, making the inversion
structurally impossible. Found via TestHub_ConcurrentRegisterUnregister
failing the P1 gate under -race on windows-latest (2 ghosts after churn);
that test now settles in milliseconds instead of polling out its deadline.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
admin's fileSHA256 duplicated VerifyChecksum's hashing body. One exported
helper now serves both the TOCTOU snapshot in handleApplyUpdate and
VerifyChecksum itself.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
requireChannelBroadcastAccess went through RequireChannelAccess, whose DM
branch checks only participant membership — a blocked user's plugin
broadcast could reach the person who blocked them — and it issued a raw
GetRoleByID per broadcast, bypassing the permission cache. The gate now
delegates to MessageService.CanPost (extracted over checkSendPermission),
so DM blocks, channel permissions, and future posting policy apply from
exactly one place; fails closed when no service is wired. First brick of
the permission-path unification. MemStore.GetDMRecipient gets an honest
implementation so the block path is testable.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
After validating every resolved IP, the guarded dial connected only to
ips[0] — an allowlisted dual-stack or round-robin host whose first record
was down hard-failed despite reachable vetted alternatives. The dial now
tries each vetted address in order (all records still validated before
any dial: one poisoned private record refuses the whole request).
Also removes the redundant rejectPrivateAddrs pre-resolves (initial
request + redirect hop): the guarded dial is the authoritative check and
every path flows through it, so the pre-resolve only cost an extra DNS
round trip while re-opening the rebinding TOCTOU it was meant to close.
Folds the W3-2-adjacent double-resolve cleanup from the plan.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
With trusted_proxies covering client networks (e.g. 10.0.0.0/8 over LAN
clients), the right-to-left XFF walk skipped every entry, exhausted, and
fell back to the proxy's RemoteAddr — collapsing all clients into one
rate-limit/lockout bucket, so one user's failed logins locked out
everyone. On exhaustion the walk now returns the leftmost valid entry
(furthest-upstream hop), the best distinct per-client key such a config
allows. An untrusted RemoteAddr still never gets its headers honoured.
Also (W3-3): the CIDR list parses once per request instead of once per
XFF candidate, config load warns about invalid CIDR entries at startup
(a silently skipped entry silently un-trusts the proxy), and the sample
config documents that trusted_proxies must list only proxy hops.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
installFromDisk replaced r.plugins/r.byName with a fresh *Instance but
left r.commands keyed to the old pointer and the old module running:
re-installing an enabled plugin blocked its own command re-registration
(RegisterCommand compared ownership by pointer) and kept dispatch routing
into the orphaned module until restart. Reinstall now deactivates the old
instance and clears its bindings, and RegisterCommand compares ownership
by plugin identity (manifest name) — the same plugin re-binds freely, a
different plugin still cannot hijack an owned command.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Revocation failure: no error, audit still written, RevokeFailed set,
password committed. Transient failure: absorbed by exactly one retry.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
UpdateUserPassword commits first; when DeleteOtherSessions then errored the
handler returned 500 and skipped the audit row — telling the user the
change failed while the new password was already live, walking them into
retrying with a dead password and tripping the confirm lockout. The
committed change now always audits and reports success; revocation gets
one bounded compensating retry, and a persistent failure surfaces as a
200 + warning (sessions_revoked count) the client can show.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The empty-prefix middleware shared per-IP buckets with verify-totp,
password change, and the sensitive endpoints, so a client's 30/min
auto-poll could 429 its own user's 2FA or password change. Dedicated
"client_update:" prefix, mirroring "livekit_proxy:".
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Simulates an 8-participant call: two back-to-back full rotations (7 offers
each) must pass the limiter, while same-target spam still trips it.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A rotation is a burst of one offer per peer (join/leave and the periodic
re-key), but the limiter was keyed per sender at 5/sec — in calls with 6+
participants the 6th+ peer's offer was silently rate-limited, that peer
never received the rotated key, and their audio never decrypted again.
Keying per (sender, target) admits any rotation burst regardless of
channel size while still capping repeated offers at a single victim,
which is the abuse the limit exists for (an offer can force the target to
re-key or disconnect).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Service level: BAN_MEMBERS refusal (Forbidden even for nonexistent targets
— no id enumeration), equal-rank and owner-target hierarchy refusals,
authorized ban/unban round-trip, self-ban rejection. Admin API level:
equal-rank owner ban 403s, a lower-positioned ADMINISTRATOR cannot ban the
owner, downward bans still work. All existing NewAdminAPI/NewHandler test
callsites now inject a real ModerationService so the production
authorization runs in every PATCH-user test.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
