7.9 KiB
Pending issues
Auto-chunking rollout is explicitly held
The user requested on 2026-11-09 that auto-chunking not be brought online while the existing RVBox deployment is in use. Implement and verify it only in the repository and isolated fixtures. Do not restart/upgrade live services or replace the published installer for this feature without a later explicit instruction lifting that hold. See the design.
Legacy Windows client upgrade remains recommended
On 2026-11-16 a legacy client entered a persistent reconnect loop because the server echoed every already-compacted tombstone as a terminal-discard action. The redundant UUID list exceeded that client's 32 KiB WebSocket receive limit. The server correction omits those no-op actions only after the usual immutable hash/revision checks and reconciliation decision. Retained terminal commands still receive discard authorization, and contradictory non-terminal server state is still interrupted and incidented. Tombstones and history are preserved.
The isolated v16 hotfix excludes the held auto-chunking implementation. The new regression reproduces the original error using TLS, no subprotocol, a real 32 KiB reader and 1,100 tombstones, then verifies repeat reconciliation after the fix. Store tests cover missing/terminal server rows, hash mismatch, and a tombstone contradicting running server state. Large genuine action lists, inline commands and stdin can still exceed the legacy reader; upgrade old Windows clients to the verified non-chunking installer. This optimization does not change the remaining 1 MiB logical snapshot limit or add pagination.
Recovery on 2026-11-16 was verified live: server version
0.1.0-gui-manual.16-tombstone.1, based on 474afb3 plus only the store
correction; no internal/agentwire package is present in that build source.
The active fixture's Compose override pins the binary under
.test-runs/workstation-disconnect-20261116/; treat it as a live deployment
artifact, not disposable diagnostics. That directory also holds build
provenance, the preserved pre-upgrade data/config and the guarded recovery script.
Do not delete it while mounted. The existing published v16 installer is
unchanged; its executable was checksum-verified on the user's desktop for
manual upgrade from v11. No Windows service restart or command reissue was
needed for server recovery.
Reconciliation must restore a lost command acceptance
A reconnect can leave the server at COMMAND_DISPATCHED while the client has
durably accepted/executed the command. Current ReconcileResumeDelivery only
retargets its session generation. Replayed script-status events can commit, but
the next running event fails with illegal command lifecycle transition COMMAND_DISPATCHED -> COMMAND_RUNNING, repeatedly closing the whole session.
Temporary recovery was verified on 2026-11-09 without changing the v16 binary:
stop the server, preserve its complete data/config, rehearse on an independent
copy, and use Store.RecordCommandAcceptance for the single affected command.
The recovery checked its owner, UUID, revision, generation, event sequence, and
durable completed-script receipt. It restored only acceptance; client replay
then supplied the actual running/terminal events and existing cancellation
result. No command was reissued and no event sequence or terminal result was
invented. This is an incident-specific repair, not a general SQL-edit recipe.
The permanent correction needs an atomic reconciliation update based on the matching client snapshot: restore a missing acceptance only for compatible client lifecycle, immutable hash, revision, and owning session. Keep ordinary lifecycle transition checks strict. Add lost-acceptance/reconnect coverage with already-persisted script status, running/terminal event replay, fencing/hash/ revision rejection, and proof of no duplicate execution. This correction is separate from the held auto-chunking feature and has not been deployed.
RVBox Windows commands do not reliably show GUI prompts
When a GUI uninstaller is launched through rvc run --shell cmd, its dialog may
not appear on the user's desktop. This was observed with IDA Freeware 7.0 and
HandBrake 1.6.1: their registered uninstallers ran through RVBox without a
visible confirmation dialog, and the apps remained installed after the
non-elevated attempts.
The command status selected the affected desktop's active user in session 1, with the
matching user SID. That rules out a simple Session 0 selection error, but does
not establish that the process shares Explorer's visible desktop. The Windows
launcher in the affected build started the shell with STARTF_USESHOWWINDOW
and SW_HIDE, then CREATE_NEW_CONSOLE, with STARTUPINFO.lpDesktop unset.
That hid the shell and did not explicitly select the user's visible desktop.
Microsoft's STARTUPINFO documentation
describes these startup fields as controlling the process's window station,
desktop, and initial window appearance; ShowWindow
uses the supplied startup state for an application's first window.
The repository's GUI test guidance explicitly says that matching session IDs alone do not prove a process shares Explorer's window station and desktop. Together, the hidden startup state and inherited desktop are the strongest code-based explanation for the missing dialogs. No screenshot or direct window-station inspection was available to confirm which condition prevented these specific uninstaller windows from appearing.
The RVBox troubleshooting guide also says GUI work needs the intended interactive session and desktop, and that token/session data alone is insufficient (troubleshooting.md). Treat a successful command lifecycle as evidence that the launched process exited, not that an uninstall completed. Verify the app's executable and uninstall registration afterward.
Implemented correction; native acceptance pending
The corrected launcher explicitly selects winsta0\default for active-user,
active-user-elevated, and active-system, including the command shell and signal
helper. The authenticated launcher reports its observed desktop and verifies
existing GUI access under the selected token before user code is prepared.
Session 0 contexts retain their noninteractive desktop. No desktop ACLs or
user-global terminal preferences are changed.
Shells now use CREATE_NO_WINDOW with SW_SHOWNORMAL, keeping console
suppression independent of GUI startup appearance. Existing output pipes,
suspended launch/authorization, Job containment, and TERM helper remain in use.
Console-event delivery needs native acceptance on this headless-console route.
Portable policy tests and Windows compilation cover the new path. Native tests exercise actual desktop access, real GUI creation through the launcher, and CTRL_BREAK observed by a controlled root PowerShell handler. The native harness stages/checksums the self-contained GUI probe and checks visible-window evidence for active-user, active-user-elevated, and active-system. These tests have not been executed while Helium is offline; compilation is not a native pass.
Keep this issue open until the actual Windows desktop probe and the affected
uninstallers have been retested. The earlier SW_HIDE on the shell was a
plausible contributor, not proof that every GUI descendant inherited it.
Validate a real window and the application's final effect, including elevated
requirements, rather than inferring success from session/SID or process exit.
See the advanced GUI guide for /wait,
the diagnostic script, locked/background windows, Job lifetime and shell-policy
limits. Vendor unattended flags remain an optional workflow, not proof of a
working interactive launch path.