Changelog

All notable changes. The format follows Keep a Changelog; versions follow Semantic Versioning.

1.4.2 — 2026-11-10

Added

• shell (admins only): opens an interactive window on the host shell, like vm shell does for a VM. It always targets the host, ignores the tab's current system, needs no sudo, and refuses normal users (also when RCONSOLE_ALLOW_USER_HOST is on). The launch is authorized again when the terminal starts.
• Photos and videos in the 📁 file browser. Clicking an image (png jpg jpeg gif webp bmp ico avif svg) shows it; clicking a video (mp4 m4v webm ogv mov) plays it in a player. Previous / next (buttons, ←/→, swipe) move through the folder's media, ⬇ downloads, Esc closes. New /api/media endpoint: login required, same path confinement as the rest of the file browser, whitelisted types only (others get 415), Range requests so videos seek, and a sandboxing Content-Security-Policy so an SVG can never run scripts. Unplayable codecs get a message and a download button.
• tests/test_media_shell.py.

Changed

• help lists shell; COMMANDS.md, FEATURES.md, README.md and structure.md describe the new command and the viewer.

Security

• /api/media serves only a fixed list of image and video types with Content-Security-Policy: sandbox and no script source, so even an SVG opened directly cannot run code with the console's session.

1.4.1 — 2026-11-08

Added

• Arch Linux as a system (vm install arch, alias archlinux; also in vm install all and vm list). It is a rolling release, so there is exactly one release (arch). The root filesystem comes from the linuxcontainers mirrors (amd64 and arm64) and is verified against the published SHA-256, like Debian.
• Everything the other systems have works on Arch: vm shell/run/use/status/doctor/reset/remove, per-user /root, neofetch/fetch (with an Arch logo and the pacman package count), vm setup <profile> for tools, build, python, node, rust, db, all, docker and udocker (Arch package names; the real Docker client comes from Docker's static build because Arch has no client-only package), vm update (a full pacman -Syu with the usual lock, snapshot, verify and rollback) and the dependency backup/restore (pacman -Qqen is recorded and reinstalled after a restore).
• pacman works under proot out of the box: the space check, the download user and the sandbox are switched off (a fake root has no real mounts or users), signature checking stays on, systemd/udev hooks that cannot run in a VM are turned into no-ops, a mirror is written if none is active, and the first pacman -S creates the signing keys by itself (/usr/local/bin/pacman wrapper, the real one stays at /usr/bin/pacman).
• tests/test_arch.py.

Fixed

• Arch: pacman-key failed with /dev/fd/63: No such file or directory on hosts where a VM has no PID namespace of its own (bash process substitution needs /dev/fd -> /proc/self/fd, which the restricted /proc lacks). The host now creates the signing keys itself, with a fixed command, at vm install arch and on vm update arch (which also repairs a half-made keyring). Inside the VM, pacman says to run vm update arch instead of printing errors.

Changed

• vm list wraps its heading lines to the screen width (the longer Arch line overflowed a 34-column phone).

Removed

• A Bengali-script alias for Kali (code, docs, tests and wiki). kali, kali-linux and kalilinux still work.

Notes

• Arch is not tested in a real VM yet: the rootfs download is not reachable from the build sandbox. Everything else is covered by tests with faked mirrors.

1.4.0 — 2026-11-06

Added

• Admin dashboard at /dashboard (admins only). One page to watch the whole instance, live (refreshes every 5 s):
• Overview: CPU, memory, disk and network (container-aware: cgroup limits, so it shows what your plan allows), load, process count, users online, running terminals / jobs / services, uptime, alerts (CPU or memory pegged, disk nearly full, a service stuck in crash backoff, a failed system, backups unconfigured or failing, the factory ark password still in use) and one hour of history as graphs (5 / 15 / 60 min).
• Processes: every service (restart, stop, start, remove, logs), every job (output, kill), every open terminal (close) and the busiest host processes with CPU %, memory and which user / job / terminal they belong to.
• Users: who is online and when they were last seen, per-user terminals, jobs, services, apps and systems; create users, promote and demote admins, reset passwords, stop everything a user runs, delete users — the web counterpart of sudo muser, auser, nauser, cpass and ruser (they share the same code).
• Systems: every user's Linux systems with state and per-system process list (stop a process), and disk usage by area (each system, backups, logs ...), measured in the background.
• Apps: exposed ports with a live "is anything listening?" check, and hosted apps.
• Activity: an audit trail of logins (also failed and locked-out ones), logouts, and every account or process change made from the dashboard or with sudo commands. Kept in memory and in logs/audit.jsonl (rotated at 1 MB).
• Health: server facts, backups (with a "Back up now" button), keep-alive and update status.
• Logout button next to the theme and system buttons (in the ⚙ menu on phones). It does exactly what the logout command does (POST /api/logout): ends the login, keeps your terminals and data.
• Dashboard link in the ⚙ menu, rendered only for admins. The ⚙ button now also shows on wide screens for admins.
• RCONSOLE_METRICS_INTERVAL (default 5): seconds between dashboard samples.
• tests/test_dashboard.py (access control, every section, every action, audit trail, logout) and jsdom checks tests/js/dashboard_check.js plus new menu / logout checks in tests/js/ui_check.js.

Design

• Dashboard: cards with a status-coloured top edge and a trend line under each meter, graphs with dashed gridlines and a hover / touch readout of the nearest sample, status pills, alert banners with icons, a pulsing live indicator, a pill-style tab bar, focus rings and reduced-motion support. Colours come from the console theme variables, so all 17 themes work.
• Dashboard on phones: a bottom tab bar with icons, two-column meter cards, every table turns into one labelled card per row, dialogs open as bottom sheets, and tap targets are at least 36 px.
• Logout button with a power icon that turns red on hover; in the ⚙ menu it is a full-width outlined button. The Dashboard link in the menu is now a highlighted row with an icon, a one-line description and an arrow.

Changed

• sudo auser, nauser, muser, cpass and ruser now also write to the audit trail. ruser shares its implementation with the dashboard (cmd_users.remove_user).
• Non-admins get a plain 404 for /dashboard and 403 for its API; the page link is never sent to them.
• The dashboard refuses to demote, delete or stop yourself (use sudo nauser / ruser); the last admin was already protected.

1.3.11 — 2026-11-05

Added

• Tests for the journalctl, systemd-hwdb and udevadm shims (JournalctlShimTests, HwdbShimTests).

Fixed

• Setting up systemd printed /proc/ is not mounted and Failed to update catalog in Debian, Ubuntu and Kali VMs. systemd's install script runs journalctl --update-catalog, which has nothing to do in a VM (no journal, no real /proc mount). Debian-family VMs now get a journalctl shim in /usr/local/bin that skips just that call; every other journalctl command runs the real binary. Existing VMs get it on their next launch.
• Setting up udev printed Failed to write database /usr/lib/udev/hwdb.bin: Function not implemented. The hardware database cannot be written under proot and a VM has no hardware to describe. Debian-family VMs now get systemd-hwdb and udevadm shims that skip only update / hwdb --update; every other call runs the real tool.

1.3.10 — 2026-11-05

Added

• Tests for the ln shim, the /proc whitelist and the games PATH (KaliLoginNoiseTests).

Changed

• VMs start in /home (new tabs, the interactive window, vm run, services without --cwd, and the fallback when a directory is missing). The host shell is unchanged (/home/rconsole), and $HOME and cd ~ inside a VM are still /root. /home is created in VMs that lack it.

Fixed

• Kali VM printed errors on start and after every command (ln: failed to create symbolic link '/dev/fd': Permission denied, same for /dev/stdin, /dev/stdout, /dev/stderr, and bash: /proc/sys/kernel/random/uuid: No such file or directory, same for boot_id). Kali's login scripts try to re-create the standard /dev links (the host's /dev is read-only here and already has them) and read two kernel files that the restricted /proc did not expose. VMs now get a small ln shim in /usr/local/bin that skips those four links when they already exist, and the restricted /proc now includes sys/kernel/random/uuid and boot_id (harmless, random values). Existing VMs get the shim on their next launch.
• Programs in /usr/games were "command not found" on Debian, Ubuntu and Kali (apt install sl then sl). Debian's /etc/profile resets root's PATH without /usr/games and /usr/local/games; Alpine installs to /usr/bin, which is why it was unaffected. Both directories are now always on PATH (line mode, login shells and scripts).

1.3.9 — 2026-11-05

Fixed

• RCONSOLE_BACKUP_TOKEN and RCONSOLE_BACKUP_PASSPHRASE reached host child processes. The docs said server secrets never reach user commands, hosted apps or services, but only four secrets were stripped. Both are now removed from the environment of every command, terminal, job, service and hosted app on the host shell (VMs never inherited the server environment). Backups are unaffected: the backup engine reads them itself.
• Documentation contradicted the code and itself. README still described the shared PID namespace removed in 1.3.8; DEPLOY said the free Render blueprint has a 5 GB disk (only render.paid.yaml does); "never sleeps", "as many VMs as you like", "services survive redeploys" and "open terminals reconnect" promised more than the implementation does; the architecture diagram said Alpine only; RCONSOLE_VM_ROOTFS_URL was documented as an Alpine-only default; the README listed Python 3.12 as passing though 3.14 is required; the shutdown backup was documented as 25 s (it waits about 22 s); Koyeb was listed as a free option for new users; the settings tables omitted RCONSOLE_VM_PIDNS, RCONSOLE_RESTORE_DEPS, RCONSOLE_UPDATE_API, RCONSOLE_LOG_LEVEL, RCONSOLE_DOMAIN, RCONSOLE_VERSION and the Docker-shim variables, and gave wrong defaults for RCONSOLE_TRUST_PROXY, RCONSOLE_COOKIE_SECURE and RCONSOLE_USER_VM_PROC.

Added

• Documentation website (GitHub Pages, https://salehin-07.github.io/rconsole-ark/) built from the repository's own Markdown by scripts/build_docs.py: quick start, guides, command reference, search, dark mode, mobile layout. New pages: backups and restore, updating and rollback, configuration, security model, architecture, compatibility.
• SECURITY.md (how to report a vulnerability).
• One authoritative environment-variable list (scripts/envvars.py); the tables in README, FEATURES and the site are generated from it, and tests fail when the code, the list or any document disagree.
• Documentation tests for cross-document facts: version, Python version, Render blueprints, backup limits, release asset names, per-user limits, secret stripping, broken links.

Notes

• No restart-sensitive behaviour changed apart from the secret stripping. Host-shell scripts that read RCONSOLE_BACKUP_TOKEN or RCONSOLE_BACKUP_PASSPHRASE from the environment must now get them another way.
• Enable the site once: repository Settings → Pages → Source: GitHub Actions.

1.3.8 — 2026-11-04

Fixed

• VMs shared one PID space. Every VM of every user ran in the server's PID namespace: an admin's VM listed everyone's processes (the server, other users' VMs) and kill (even kill -9 -1) reached them all. Each VM now gets its own PID namespace (unshare --user --pid, new pidns.py, init in pidns_init.py): inside a VM ps, top, kill and /proc show only that VM's processes, with PIDs starting at 1, and kill -9 -1 only reaches that VM. All commands of one VM (terminals, jobs, services, hosted apps) share its namespace, so a service started in one terminal is visible from another. The namespace's init reaps zombies and exits when the server dies, so nothing is orphaned.
• On a host that forbids user namespaces rconsole still works, without isolation: it says so in vm doctor (pid ns: line), and admins' VMs no longer get the host's whole /proc (everyone gets only the harmless system files, as normal users already did). Use vm ps / vm kill there. RCONSOLE_VM_PIDNS=off restores the old behaviour, on is the same as auto but documents intent.
• A restore could start services on a half-built VM. After a redeploy a VM was marked ready as soon as its system was reinstalled, so a service could start before its vm setup profiles were back and crash-loop. A VM is now locked in a restoring state (set before the supervisor's first pass) until its system, setups and packages are back; its services wait, then start by themselves.

Added

• Dependency manifest per VM (deps.json, part of every backup, new vmdeps.py): before each backup rconsole records what is installed in each VM — manually installed system packages (apk world / apt-mark showmanual), pip packages of the system Python, global npm packages, every Python virtualenv under /root with its packages, and every Node project that had a node_modules (with its lockfile type). node_modules and virtualenvs are not in backups, so this list is what brings them back. It is refreshed only when something changed (directory mtimes), never while a VM is being restored, and a partial restore never makes it forget what could not be reinstalled.
• Full restore: after a redeploy each VM is rebuilt, its profiles re-applied, then everything in its manifest reinstalled (apk add/apt-get install, pip install -r, npm install -g, venvs recreated, npm ci / yarn / pnpm install in projects). Failures are per item, logged, and shown by vm deps. RCONSOLE_RESTORE_DEPS=0 skips it.
• vm ps [name], vm kill <pid> [name] [--9], vm deps [name] (capture / restore run it now).
• vm doctor and vm status show the PID-namespace mode and the restoring state.
• Tests: PID isolation between VMs and users, kill -1 containment, the no-namespace fallback, manifest parsing/merge/restore script, and "services wait for a full restore".

Notes

• The Docker image now installs util-linux explicitly (unshare, nsenter). Docker's default seccomp profile blocks user namespaces: on plain Docker add --security-opt seccomp=unconfined (or a profile allowing unshare/setns) to get isolation; vm doctor tells you which mode you are in. Not verified on Render.
• Not captured: files outside /root (/etc, /opt, /usr/local), custom apt repositories, and host-shell packages (admin host mode). Keep projects under /root.

1.3.7 — 2026-11-03

Added

• RCONSOLE_HOST_HOME (default /home/rconsole): where the host shell starts.
• tests/test_cwd.py: default directories, ending a terminal keeps the tab, and a real host shell's directory carried back on exit.

Changed

• Default directories: the host shell now starts in /home/rconsole, a VM in /root. In the Docker image /home/rconsole is linked to the persistent workspace (<data>/workspace) by the entrypoint, so what you keep there is still part of every backup; where /home/rconsole does not exist or is not writable (a plain install), the host falls back to <data>/workspace as before. RCONSOLE_HOST_HOME overrides the path. This needs a redeploy / container restart (the entrypoint runs at start).
• File manager: the always-visible ☑ button next to the new file / folder buttons and the All button in the selection bar are gone. A select-all checkbox now appears at the start of the selection bar once at least one item is selected (empty, or ticked when everything is selected; clicking it selects all / none), like other file managers. Ctrl/Cmd+A still works.
• The login logo is smaller (96 px, was 132 px).
• The app is called Rconsole in the PWA manifest (installed app name), the home-screen title, the page titles and the README.

Fixed

• The prompt in a VM's interactive window no longer shows the server's long host id and a bare #. It reads <system>@root:<dir># (ubuntu@root:~#, alpine@root:/etc#), coloured in bash. It comes from a managed ~/.rconsole_prompt that .bashrc / .profile load (existing VMs get it on their next launch; the old per-VM PS1= line is retired).
• Returning from an interactive window no longer resets the directory. Ending a terminal session used to delete the tab on the server, so the next command ran in a new tab in the default directory (and the tab's history was lost). The tab is kept now. On the host the directory the shell ended in is carried back too (cd inside bash, then exit); inside a VM the console returns to the directory the window was opened from.

1.3.6 — 2026-11-02

Added

• Select, compress, extract and Properties in the 📁 file manager, working like a desktop file manager:
• Selecting. Click the blank space of a row to select it (the name still opens the file or folder); Ctrl/Cmd+click toggles, Shift+click selects a range, dragging a rectangle selects every row it touches (Ctrl/Shift+drag adds; it auto-scrolls near the edges), double-click opens. On phones and tablets a long-press selects a row and dragging without lifting sweeps over more rows; after that a tap toggles rows and tapping a name selects instead of opening, so nothing opens by accident. A plain tap on blank space does nothing while nothing is selected, so scrolling never selects. Select all (the ☑ button, or Ctrl/Cmd+A), Esc clears the selection (a second Esc closes the browser), and the Delete key deletes it.
• A selection bar with the item count and total size and: All, Compress, Extract (only when every selected item is an archive), Properties (only when exactly one item is selected), Delete (one confirmation for the whole selection) and ✕.
• Compress to zip, tar, tar.gz, tar.bz2, tar.xz, tar.zst (Python 3.14+), 7z, RAR and the one-file gz, bz2, xz, zst. Only formats that work on the server are offered; one-file formats are greyed out unless exactly one plain file is selected; the default name follows the selection; an existing archive is never overwritten (name (2).zip). Symlinks are stored as links, never followed.
• Extract zip (also jar, war, ear, apk, whl, egg), tar, tar.gz/bz2/xz, gz, bz2, xz, zst and, through bsdtar, 7z, RAR (RAR3 and RAR5), CAB, ISO, CPIO, DEB, RPM, LZH, XAR. Extract to a new folder (an archive with exactly one top folder becomes that folder, otherwise a folder named after the archive) or Extract here (never overwrites; skipped entries are counted).
• Properties: type, location, size (folders counted without following links), size on disk, modified / accessed / changed, permissions (-rw-r--r-- and octal), owner and group, link target, MIME type and, for zip and tar archives, the number of entries and their unpacked size.
• Compress and extract run as background jobs with a progress bar and a Cancel button (temp files are removed); one job per user at a time. New endpoints: GET /api/props, POST /api/archive, GET /api/archive/<id>, POST /api/archive/<id>/cancel, and the delete_many action of /api/fs.
• RCONSOLE_MAX_EXTRACT_MB (default 2048): the most an archive may expand to when extracted (zip-bomb guard).
• fileops.py (the engine, standard library only) and static/files-core.js (the selection logic, DOM-free), with tests: tests/test_fileops.py (46), tests/test_fileman_api.py, tests/js/files_check.js (15) and a larger tests/js/ui_check.js that drives the real handlers (click, ctrl/shift, rectangle, long-press, dialogs).
• The Docker image now installs libarchive-tools (bsdtar), which gives 7z creation and 7z / RAR / CAB / ISO extraction. This is a Dockerfile change, so it needs a redeploy; the self-updater does not install packages. Without bsdtar those formats are simply not offered (zip, tar and gz/bz2/xz always work).

Changed

• New logo everywhere: favicon, home-screen / PWA icons (including an Android maskable icon), the iOS touch icon, the login page (it showed ASCII art) and the README. static/icon.svg is gone.
• The editor's font-size buttons moved to the top bar next to Save and Close and are now just − and + (the size is in their tooltip; it is still remembered and still 16 px minimum on iOS).
• In a file row, the name now only takes the room it needs; the free space to its right is the click target for selecting.

Safety

• Extraction cannot escape the folder it is given. Every member is created relative to a directory file descriptor with O_NOFOLLOW, so the classic "symlink, then a file below the symlink" archive writes nothing outside. .., absolute and Windows drive-letter names are refused; devices, FIFOs and hard links are skipped; setuid bits are dropped; files are never overwritten; the expanded size, the entry count (200,000) and the free disk space are all checked while writing (zip and gzip bombs stop early). All of it is tested with crafted archives.
• Confined (VM) users only get bsdtar for the external formats, because it refuses to write through symlinks by default; 7-Zip, unrar, unar and rar are used only for the host admin. 7z / RAR creation is refused for trees that contain symlinks, since those tools may follow them out of the folder.
• Archives are written to a hidden temp file and renamed into place, so an interrupted job never leaves a half-written archive.

Fixed

• test_download_and_upload_are_confined_to_the_vm sat inside the if __name__ == "__main__": block of tests/test_app.py, so it never ran under test discovery. It now does (it needs proot, so it is skipped where proot is missing, exactly like its neighbours).

Known limits

• RAR files cannot be created without WinRAR's proprietary rar tool, so RAR only appears in the Compress list for an admin who installed rar. Extraction works (via bsdtar).
• Password-protected archives are not supported (the job says so).
• There is no right-click menu, copy / cut / paste or drag-and-drop of files between folders yet.
• Folder sizes in Properties stop counting after 3 seconds or 500,000 entries and say so.
• Zstandard (tar.zst, .zst) needs Python 3.14 (the Docker image and CI have it).
• Touch behaviour was tested with simulated touch events (jsdom), not on a real phone. The long-press drag relies on the browser letting the page cancel scrolling once the long-press fired, which Chrome and Safari allow.

1.3.5 — 2026-11-01

Added

• A much smarter file editor (edit, and files opened from the 📁 browser). It now behaves like a small code editor instead of a plain text box:
• Auto-indent: Enter keeps the current indentation and adds one level after { [ (, after : (Python, YAML, Makefile), after then / do / else (shell) and after an opening HTML/XML tag. Enter between a pair ({|}, (|), <div>|</div>) opens an indented block with the closer on its own line. Python dedents after return / pass / break / continue / raise. Markdown lists continue (-, *, 1., - [ ]) and an empty item ends the list; /** … */ comments continue with *.
• Tabs or spaces, detected per file: the file's own indentation is detected when it opens (tabs, 2 or 4 spaces; otherwise a language default: tabs for Go and Makefiles, 2 spaces for JS/TS/JSON/YAML/HTML/CSS and friends, 4 for the rest). Tab inserts up to the next tab stop, Shift+Tab outdents, Tab/Shift+Tab on a selection (de)indents every line, and Backspace inside the indentation removes a whole level. The status bar shows the current style; tap it to cycle 2 spaces → 4 spaces → tabs.
• Auto-closing brackets and quotes: ( [ { " ' and the backtick are closed as you type, typing the closer steps over it, Backspace between a pair removes both, and typing one around a selection wraps it (*, _, ~ too in Markdown). Quotes are not paired after a word character (so don't works), Python f"…" / r'…' prefixes and """docstrings""" are understood, apostrophes are not paired in prose/Rust/Lisp-style languages, a closer typed on a blank line steps back one level, and <div> in HTML/XML adds </div> (not for void tags such as <br>).
• Editor shortcuts: Ctrl/Cmd+/ toggle comment (~60 languages, block comments for HTML/CSS/Markdown), Ctrl/Cmd+] / [ indent / outdent, Alt+↑/↓ move line(s), Alt+Shift+↑/↓ copy line(s), Ctrl/Cmd+Shift+K delete line, Ctrl/Cmd+Enter (+Shift) insert line below (above), Ctrl/Cmd+L select line, Ctrl/Cmd+J join lines, smart Home (first non-blank ↔ column 0), Ctrl/Cmd+G go to line.
• Find and replace (Ctrl/Cmd+F, Ctrl/Cmd+H, the Find button, F3): match case, whole word, regular expressions (with $1 in the replacement), next/previous with wrap-around, a "3 of 12" counter, Replace and Replace All. Esc closes the find bar first.
• Font size: A− / A+ in the editor's status bar (9–32 px, 16–32 px on iPhone/iPad where smaller text would make iOS zoom the page). The choice is remembered in the browser.
• Status bar: line and column (tap to go to a line), selection size, line count, indentation style. The current line is highlighted, and closing the tab with unsaved changes now asks first.
• Symbol bar on touch devices: Tab, outdent, undo, redo, Find, comment toggle and the symbols that are hard to reach on a phone keyboard (brackets, quotes, ; : = / \ | & _ - # $ % and more). They go through the same auto-close logic as typing.
• static/editor-core.js: the editor's text logic as a small DOM-free module, with its own tests (tests/js/editor_check.js, 76 checks) plus end-to-end checks of the real handlers in tests/js/ui_check.js.

Changed

• Edits made by the editor go through the browser's own editing command, so Ctrl+Z / Ctrl+Y keep working after auto-indent, auto-close, comment toggles and line moves (the old Tab handler overwrote the text and wiped the undo history).
• Opening a file puts the cursor at the top-left (it used to be at the end while the view showed the top).
• Editor text, line numbers and highlighting share one font-size setting instead of three fixed values.

Known limits

• Typing a closer always steps over an identical closer that is already next to the cursor (it does not track which ones it inserted itself), there is no multi-cursor and no word wrap, and tab characters are drawn 4 wide.
• Auto-close and auto-indent hook the browser's beforeinput event. Browsers do not allow cancelling it during IME composition, so on keyboards that compose words (some Android predictive keyboards) a few of the smart behaviours may not trigger; plain typing, Tab, the symbol bar and all shortcuts are unaffected.
• A few browsers reserve some shortcuts (e.g. Ctrl+L, Ctrl+J, Ctrl+Shift+K); the same actions are reachable with the other keys or, on touch devices, the symbol bar.

1.3.4 — 2026-11-01

Added

• After an Alpine release walk the VM now records which release it is (/etc/alpine-release, /etc/os-release), so vm status, neofetch and cat /etc/os-release show the new release.
• An extra apk upgrade pass and a "nothing left to upgrade" check at the end of the walk.

Fixed

• vm update on Alpine rolled a successful release walk back. All steps ran and the packages were upgraded (3.13 → 3.24 in a real VM), but the final check read /etc/alpine-release, which in an Alpine rootfs belongs to no package, so apk never changes it and it still said 3.13.12: "expected Alpine 3.24 but found 3.13.12", and the safety net restored the old system. The check is now based on facts that are really true afterwards: (1) every repository line points at the target release, (2) apk upgrade has nothing left to do (one extra pass is tried first; leftovers fail the update and roll it back), and only then (3) the VM is told which release it is: /etc/alpine-release and /etc/os-release (so vm status, neofetch and cat /etc/os-release show the new release) are written inside the VM, following symlinks there. A final "programs still start" check closes the update.
• The test double for apk now behaves like the real one (it no longer rewrites /etc/alpine-release itself), which is why the old tests could not see this. New tests: repositories left on another release, packages left behind, the identify script run for real against a regular file and a symlinked os-release, and the real-proot walk using a fake apk that never touches the release file.

1.3.3 — 2026-11-01

Added

• vm update --stay: update the packages only and keep the current release (on any system).
• vm install alpine:<major.minor.patch>: exact Alpine releases such as alpine:3.23.5 (see Fixed).

Fixed

• vm install alpine:3.23.5 / alpine:3.24.1 failed. An exact Alpine point release was turned into a branch called v3.23.5, which does not exist. Exact releases now resolve to their branch's folder and are verified against the published .sha256. A release that really does not exist says so and suggests alpine:3.23.
• Wrong Alpine numbers now get a useful answer. Alpine releases are 3.x; typing alpine:1.24.1 or 1.22 used to say "try alpine:1.24" or a raw "not found: ". It now says there is no Alpine 1.x, shows the valid forms (alpine:3.24, alpine:3.23.5, alpine, alpine:edge) and suggests the likely intent (3.24.1).
• vm install alpine 3.23.5 (a space instead of a colon) now means Alpine 3.23.5 instead of "unknown system '3.23.5'".

Changed

• vm update now moves an Alpine VM to the latest Alpine release (e.g. 3.12 → 3.24.3), not just to the newest packages of its current release. It walks one release at a time (each release ships the signing keys of the next), inside the same lock + snapshot + verify + rollback protection: every step is checked (package database, programs still start) and the final release is confirmed. Any failing step, timeout or shutdown restores the original system, repositories included. The VM's metadata (release, label) follows; a VM installed as plain alpine keeps following "latest".
• Edge systems and customised repository lists (e.g. @testing) are never rewritten: they get packages only and the message says why. If the latest release cannot be looked up (no network) the update is refused rather than guessed.
• The start message now says exactly what will happen ("from Alpine 3.12.12 to 3.24.3: 12 release steps"), and for Debian, Ubuntu and Kali that the release is kept.
• Tests: the planner, the hop script, the whole walk, a failing/never-starting/wrong-version step, --stay, customised repositories, and (real proot) a complete walk and a rollback with a fake apk.

Not included

• Release upgrades for Debian, Ubuntu and Kali are not supported in place (Debian needs full-upgrade and source rewrites, Ubuntu needs do-release-upgrade, Kali is rolling). Install the newer release next to the old one: vm install debian:<codename>, vm install ubuntu:26.04.

1.3.2 — 2026-11-01

Added

• vm update [name]: update a VM to the latest packages of its release, safely. The new module vmupdate.py: 1. locks the VM (nothing can start in it: commands, terminals, files, services, vm run/shell/setup/doctor and vm remove/reset/install all answer "being updated") and stops what ran in it; 2. health-checks the package database (a VM that is already broken is not touched); 3. refreshes the package lists and counts what is new (nothing new: says so, changes nothing); 4. takes a snapshot of the system files: hard links for the big files (almost no extra disk), real copies of /etc and the dpkg/apk databases that packages edit in place; needs ~200 MB free and refuses otherwise; 5. upgrades with no questions and existing config files kept (apt-get upgrade, full-upgrade on Kali, apk upgrade; it never removes packages); 6. verifies: programs still start and the package database is consistent; 7. on success drops the snapshot; on any problem (failed command, 1 h timeout, the server shutting down, even a crash) puts the snapshot back, so the VM is exactly as before. A crash is rolled back by itself at the next start, and the VM stays locked until then. /root is never touched and the update runs without the user's own environment variables.
• Who and what: admins must be in the host shell (vm mode host). A normal user with a single system has everything they run stopped meanwhile (it all lives in it; services restart by themselves afterwards) and may omit the name. With several systems only the updating one is locked and stopped, the others keep working, and the name is required (a picker offers them). Running terminals/jobs ask for --yes first. vm status shows progress and the last result, vm update <name> --log the last run's output.
• Services whose VM is installing or updating now wait quietly (no restart storm) and start when it is usable.
• tests/test_vmupdate.py (fake package manager on real files: snapshot, rollback, lock, rules, recovery, shutdown) and tests/test_vmupdate_e2e.py (a real proot VM with a fake apt-get: success, failure, broken database, timeout, lock, environment).

Changed

• Server shutdown (reboot, self-update, SIGTERM) interrupts a running VM update and rolls it back before the final backup. Renaming or deleting an account, and vm remove/reset, wait while its VM is updating.
• Jobs and terminals remember which VM they run in (needed to stop exactly the right ones).

Not included

• vm update keeps the release (trixie stays trixie). Moving to a newer release is a fresh install next to it, e.g. vm install ubuntu:26.04.

1.3.1 — 2026-10-31

Added

• serve.py, a launcher with a gunicorn fallback. uv run python serve.py starts gunicorn (gunicorn -c gunicorn.conf.py app:app) where it works, and otherwise serves the same app with a built-in threaded server, so testing on a Windows device no longer fails with a missing gunicorn. The fallback is one process, serves the terminal websockets, and on Ctrl+C / Ctrl+Break / SIGTERM runs the same orderly shutdown (pause the supervisor, stop children, final backup). RCONSOLE_SERVER=gunicorn|builtin forces a server and RCONSOLE_HOST sets the built-in server's bind address.
• tests/test_serve.py: server choice, the gunicorn command, and the fallback (requests answered, shutdown run).

Changed

• gunicorn is no longer installed on Windows (sys_platform != 'win32' in pyproject.toml, uv.lock, requirements.txt), because it cannot run there.
• README, FEATURES.md, DEPLOY.md and structure.md describe serve.py. Docker, systemd and Render are unchanged: they still run gunicorn directly.

1.3.0 — 2026-10-31

Added

• rconsole update: checks GitHub for a newer release and updates the whole program to it, safely. --check only reports (everyone), --yes skips the question, --rollback returns to the previous version, --force also allows a git checkout or reinstalling the same / a rolled-back version. Admins install; no sudo prefix needed.
• The console is locked while it runs. Every request except /healthz, static files and the status endpoint gets 503 (header X-Rconsole-Updating): no command can run, no terminal, file operation, login or hosted app can be used, in any tab. Open tabs show a progress screen and reload when it is over (/api/update/status).
• Safe by construction: the release's SHA-256 is checked against its SHA256SUMS (no checksum, no update); the archive is unpacked strictly (no links, .., absolute paths; size limits); the new tree must have the right version, compile and import in a separate process with a throw-away data directory before anything is stopped or replaced. Failure up to that point changes nothing and stops nothing.
• Data stays intact: only program files are replaced (per file, atomically), after a rollback copy of each is saved; data/, .venv, .git, .env are never touched, and only files a previous release installed are ever removed. Services, jobs and terminals are stopped and a final backup is taken first; then rconsole restarts in place and services start again.
• Changed Python dependencies are installed into <data>/update/site-<version>, never into the system.
• Rollback and boot safety: rconsole update --rollback, python updater.py rollback (works with a dead server), and an automatic rollback when a new version keeps failing to start (it must stay up ~25 s to count as healthy).
• Survives container re-creation: the verified release is kept in <data>/update/releases, is part of the encrypted backups, and is re-applied at boot (docker-entrypoint.sh → updater.py sync, or after a restore on a blank disk) when the image was rebuilt from older code.
• RCONSOLE_UPDATE_REPO (use your fork) and RCONSOLE_UPDATE_TOKEN (private repo / rate limit; never passed to user commands) settings.
• FEATURES.md: documents every feature, area by area.
• rconsole version: running version, install directory, last update result, rollback availability.

Changed

• README, COMMANDS.md, DEPLOY.md, FREE_HOSTING.md and structure.md describe the update; a test now keeps the README version, FEATURES.md and the command reference in step with the code.

1.2.4 — 2026-10-30

Added

• docker run now works, without a daemon. A docker command (a shim over udocker) is installed in the image and by vm setup docker / vm setup udocker. It translates run (-e -v -w -u --entrypoint --name --rm --env-file), pull, images, ps, rm, rmi, inspect, search, login; a run container is removed afterwards unless named. On the host shell it uses proot (P1) and runs anything, including static images such as hello-world. Inside a VM it uses fakechroot (F1), because proot cannot run inside proot: normal programs work, static binaries cannot, and it says so. -p, --restart and similar are ignored with a warning; -d, exec, logs, build, compose and the like explain that they need a daemon. DOCKER_HOST plus a real client forwards everything to that engine.
• vm setup udocker: the same as vm setup docker without the real Docker client.

Changed

• vm setup docker now installs udocker and the docker/drun commands as well as the real client (which moved to /usr/local/lib/docker-real when fetched as a static binary, so it never shadows the shim).
• drun is now just docker run, always local.
• On Alpine, vm setup docker/udocker also installs GNU findutils (busybox find makes udocker print "while modifying attributes of image layer" for every layer) and the shim warns if it finds only busybox find.

1.2.3 — 2026-10-30

Added

• drun: run Docker images without a Docker daemon, in the host shell (drun hello-world). A Render service and every rconsole VM is an unprivileged container, so a daemon can never run there; drun wraps udocker (proot mode) instead. It handles images that ship no /etc or /root (hello-world), reuses one container per image, and refuses inside a VM (proot cannot nest). Suits command-line images, not -d, ports, compose or desktops. Installed in the Docker image (udocker==1.3.17).

Changed

• The vm setup docker closing note now points at drun as the no-engine option.

1.2.2 — 2026-10-30

(There is no 1.2.1: its tag was withdrawn before release and cannot be reused; everything from it is here.)

Added

• File browser: download. Every row has a ⬇ button. A file downloads as it is; a folder downloads as a .zip that streams while it is built (constant memory). Symlinks inside a folder are skipped and opened with O_NOFOLLOW, so a link can never pull in a file from outside the folder or the VM.
• File browser: upload. ⬆ Upload files and ⬆ Upload folder (keeps the folder's structure). One request per file with live progress, three at a time; existing files are replaced only after you confirm once. Names are validated (no .., absolute or empty parts), writes are atomic, and a symlink at the target is replaced rather than written through. Limit per file: RCONSOLE_MAX_UPLOAD_MB (default 64).
• vm setup docker — Docker CLI, compose and buildx, client only on every system (never the engine: no dockerd, containerd, runc or iptables, which cannot run under proot and would waste ~300 MB). Where a release has no docker-cli package, Docker's static client and the compose plugin are fetched instead. The script says why there is no daemon and shows how to use DOCKER_HOST with a remote engine.
• vm setup kali — Kali VMs only (refused anywhere else): the command-line Kali toolset (kali-linux-headless, no desktop) plus the basics it needs, with a smaller fallback (kali-tools-top10).

Removed

• sudo serve, sudo gunicorn and sudo sprocess. service add … --expose PORT:NAME (keeps it running) and expose (publishes something you started) cover them; service logs <name> replaces sprocess. The unused _static_host.py and _wsgi_host.py helpers went with them. kprocess stays.

Changed

• vm setup lists docker and kali; vm setup all still means tools, build, python, node, rust and db; it does not include docker or kali.

1.2.0 — 2026-10-21

rconsole now has a clear purpose: free hosting for bots, websites and background workers, for one person or a small team — on free tiers such as Render's, from a phone. This release adds what that needs. See FREE_HOSTING.md.

Added

• service — a process supervisor for bots, workers and servers: named commands that start at boot, restart after crashes (exponential backoff 2 s → 1 min), run on the host or inside a VM, keep logs, and can be exposed on the web (--expose PORT:NAME). Definitions persist and are backed up.
• Encrypted backups and automatic restore (backup): snapshots of users, secrets, settings, the workspace, services, exposures and every VM's /root go — encrypted (Fernet, key from RCONSOLE_BACKUP_PASSPHRASE) — to a private git repo or a folder, every 15 min when something changed and once at shutdown. A fresh instance restores itself at boot (before secrets are generated), rebuilds your VMs and re-applies your vm setup toolchains, and services start again. The git branch is replaced on each push, so it never grows.
• Built-in keep-alive (keepalive): pings the site's own public URL so free instances never go to sleep; also pings extra URLs. On by default on Render.
• reboot — restart rconsole in place (final backup, stop everything, reload) without redeploying; --hard to exit the process.
• Installable app: web manifest and icon, so the console can be added to a phone's home screen.
• RCONSOLE_MAX_USERS (default 10) keeps it a small-team tool; RCONSOLE_DEFAULT_SYSTEM picks the default VM system (e.g. alpine on 512 MB instances); RCONSOLE_PUBLIC_URL.
• FREE_HOSTING.md and a free-tier render.yaml.

Changed

• Host-mode commands now start in a persistent workspace (<data>/workspace), which is part of every backup, instead of the application directory.
• render.yaml targets Render's free plan; the disk-backed variant is render.paid.yaml.
• Dependabot updates are grouped into one pull request per ecosystem.

Fixed

• Slow shutdown/reload: an open browser connection held gunicorn's graceful shutdown for the full 30 s timeout (and the final backup only ran after that wait). Upstream keep-alive is now off, and shutdown work starts the moment SIGTERM arrives. Restarts now take well under a second.
• Orphaned processes: stopping or reloading the server left background jobs, services and hosted apps running, and an old worker's supervisor could restart them while it was exiting ("address already in use" for the new one). Shutdown now pauses the supervisor first, stops everything it started, then takes the final backup.

Security

• Backups are encrypted by default and refuse to run without a passphrase; tokens, passphrases and URL credentials are scrubbed from every message; restore only writes inside the expected folders and never follows symlinks or paths outside them.

1.1.0 — 2026-10-20

The first release of rconsole as a complete, self-hostable Linux console: private Linux VMs for every user, a hardened security model, a phone-friendly UI, one-command publishing of your own servers, full documentation and ready-made deployment files.

Added

• Several virtual machines per user (vm): Alpine, Debian (default), Ubuntu and Kali — any number at once, each with its own root filesystem and /root. Runs under proot, so it needs no host privileges and works on locked-down hosts such as Render. vm list, vm install <name>…|all, vm shell, vm run, vm use, vm mode, vm reset, vm remove, vm doctor, and @vm:<name> one-off prefixes.
• System picker and target dropdown: commands that need a system show an inline dropdown (including installing another); the tab bar has a permanent dropdown, and the prompt shows where you are.
• vm setup <profile>: one command to install Python, pip, a compiler and the headers that psycopg2, mysqlclient, Pillow, lxml and cryptography need (python), plus build, node, rust, db, tools, all.
• expose <port> <name> (and unexpose): publish any server you started — on the host or in a VM — at https://<domain>/<name>/. Private by default, --public for admins, --rewrite for apps unaware of the prefix.
• neofetch / fetch / rfetch that follow your target (a VM's own system, or rconsole's) and fit phone screens.
• Phone layout: the theme button and system dropdown move into a ⚙ settings menu on small screens.
• ✕ close button on every interactive window.
• version command and /healthz reports the version.
• COMMANDS.md: every command documented (a test keeps it complete); DEPLOY.md for any platform; MIT license.
• Deployment files: Dockerfile, docker-compose.yml + Caddy, nginx, Fly.io, Kubernetes, systemd, Render blueprint, GitHub Actions CI and a release workflow that publishes archives and a container image.
• Native commands and a native AI agent (sudo agent) now work inside the VM you are in.

Changed

• Debian is the default system for vm install.
• Plain commands in a VM always have ~/.local/bin, ~/.cargo/bin, ~/.bun/bin, ~/go/bin on PATH, and read ~/.profile exports, so tools installed with curl … | sh (uv, rustup, bun, …) work outside the interactive shell.
• pip install works in a VM without --break-system-packages; installs never prompt (DEBIAN_FRONTEND=noninteractive).
• Configuration is entirely environment-driven (config.py); secrets are generated once and persisted.

Fixed

• Inline system picker did nothing in a real browser: a screen-wide click handler re-focused the command box, closing the native dropdown the moment it opened.
• Interactive shells had no controlling terminal: bash: cannot set terminal process group … no job control, and Ctrl+C did nothing in bash-based VMs (Ubuntu, Debian, Kali).
• Ubuntu 26.04: coreutils: unknown program 'prooted-…' — old proot builds cannot run its Rust coreutils. rconsole now uses a pinned proot 5.5.0 and the install self-test explains the problem instead of leaving a broken VM.
• tzdata's "Please select the geographic area" prompt; apt's "does not have a stable CLI interface" warning.
• Kali: login script failing on /dev/net/tun (each VM gets a private writable /dev/net) and its banner.
• Pipes and redirects were quoted away (echo hi | tr … printed literally); native echo/wc/… shadowed pipelines.
• A corrupt user file silently recreated the default ark/ark admin; it is now quarantined.
• Hosted apps lost multiple Set-Cookie headers (Django's csrftoken + sessionid).

Security

• One authorization choke point (sandbox.py) for every command and file access, including the terminal websocket. Non-admins can never reach the host; file APIs are confined to the user's VM (no ../symlink escapes).
• PBKDF2 hashes with a persisted random pepper (old hashes upgrade on login); login throttling; the factory password must be changed; minimum password length.
• Cross-site request protection (Origin / Sec-Fetch-Site), secure cookie flags, a distinct session cookie name, security headers, server secrets stripped from user processes and hosted apps.
• The VM is a compatibility layer, not a hard security boundary — see the README's security model.

0.1.0

Initial version: browser console, real PTY terminals, hosted apps, users and the AI agent.

Edit this page on GitHub