omnyshell 1.24.0

Secure, Hub-centric remote shell platform written in pure Dart. Clients connect to a Hub by node identity (not host:port); the Hub authenticates, authorizes and brokers encrypted sessions to Nodes ove [...]

1.24.0

Added

• service info <hub|node> and service reinstall <hub|node>. info shows an installed service's parameters (status, scope, install time, restart policy, environment) and the actual command the OS runs it with (the real executable path + arguments, plus the native unit/plist/Task-Scheduler definition). reinstall rebuilds the service for the current executable — refreshing the binary while reusing the stored config — or, when install flags are passed, reinstalls with a fresh config.

Changed

• Service management now goes through dart_service_manager 1.3.0 uniformly on every OS, including Windows (Task Scheduler). The bespoke WindowsTaskService backend moved into the package as a first-class WindowsTaskSchedulerDriver, so the registry is the single source of truth and info/reinstall work everywhere. Windows services installed by an older build are not recorded in the registry; reinstall once (with flags) to adopt them.

Fixed

• TAB path/command completion on Windows nodes (bash, PowerShell, cmd). Completion previously emitted a POSIX sh-only snippet and ran it through the node's default exec shell (always cmd.exe on Windows), so nothing completed. The snippet is now shell-family aware (ShellDialect.completionCommand) and the one-off exec is routed to the interactive session's shell family via a new shellFamily hint on the exec request. Command-position TAB now also completes executables on $PATH on every OS.

1.23.0

Added

• Drive sync preserves the executable (+x) bit (omnydrive 1.8.0). When :drive sync pushes or pulls a file that carries a POSIX execute bit, the destination now marks it executable too. The flag rides the drive RPC write and copy operations, so the node applies chmod +x after a remote write. A chmod-only change (identical content) is detected and synced. No-op on Windows.

1.22.0

Fixed

• .omnyignore now excludes nested files (omnydrive 1.7.0). A slash-less pattern such as *.dill previously matched only files at the mount root, so nested artifacts like bin/backend_server.dill were still uploaded. Patterns without a / now match at any depth, matching gitignore semantics; patterns with a leading or internal / (e.g. /build, a/b) stay anchored to the root.
• run --with now honors .omnyignore. When co-mounting dependencies via --with, each member directory's own .omnyignore (or --ignore-file) is now applied, scoped to that member's subtree under the mounted wrapper. Previously the derived include whitelist suppressed the ignore file entirely, so e.g. omnyshell run <node> "dart analyze ." --with ../dep uploaded dep's ignored build output. The member whitelist is now anchored to the wrapper root so the gitignore depth change does not widen it.

1.21.0

Added

• .omnyignore default-ignore file (omnydrive 1.6.0). A directory mounted with neither --include nor --exclude now consults a gitignore-style .omnyignore at its root and turns those patterns into the drive's default exclude set, baked into the mount so every subsequent sync skips the ignored sub-paths. Lines are trimmed; blank lines and # comments are skipped; !-negation is unsupported (ignored). --ignore-file <name> selects a different file name (defaults to .omnyignore); explicit --include/--exclude override the file entirely. Honored across every directory-mount entry point — omnyshell drive mount, the in-session :drive mount, and the ephemeral mounts of omnyshell run / exec --mount (the resolved filter is computed before the mount-reuse lookup, so reuse still matches). Git mounts reject --ignore-file, as they already do for --include/--exclude.

1.20.0

Added

• Live per-file sync progress (omnydrive 1.5.0). Directory-drive sync now reports each concurrent upload as its bytes stream, so the CLI draws a live multi-bar transfer view — one bar per in-flight file, tagged transferred (↑) vs deduplicated/copied (≡) — over an overall N/M files line, and prints a final report of file counts and raw vs on-wire (compressed) bytes. Pass -v/--verbose to omnyshell drive sync to list every transferred / copied / removed path. The in-session :drive sync command prints a per-path line as each file settles, plus the same summary.
• True socket-paced upload progress over the drive channel: Channel.sendStdin gained an optional onFlushed callback that reports cumulative bytes as the credit window lets each chunk onto the wire; DriveRpcClient.write / ChannelContentSource.writeBytes thread it into omnydrive's new ContentSource.writeBytes(onProgress:), so the bars fill at the real transfer pace rather than jumping on completion.
• SyncOutcome now surfaces the run's transferredPaths, copiedPaths, removedPaths, bytesTransferred and bytesOnWire from omnydrive's enriched SyncMetrics.

Changed

• Bumped the omnydrive dependency to ^1.5.0.

1.19.0

Added

• Directory-drive sync deduplicates identical content (omnydrive 1.4.0). When a pushed file's content already exists on the node — in an existing file or in another file sent the same run — the bytes now cross the wire once and the node copies them into place rather than receiving the payload again. Duplicate build artifacts, vendored files and the like sync for the cost of a single transfer. omnyshell tunnels omnydrive's new ContentSource.copy/supportsCopy over the drive channel via a new copy op: ChannelContentSource issues it and NodeDriveService executes the verified in-place copy against its LocalContentSource, falling back to a byte transfer if the source drifted.

Changed

• Bumped the omnydrive dependency to ^1.4.0.

1.18.1

Fixed

• exec/run no longer stall or truncate on long output. A command that produced more than the channel's 256 KiB send window would stop streaming partway through. Two causes: the client never replenished the node's send window as it consumed output (so the credit drained to zero and the stream stalled permanently), and on process exit the node closed the channel — which discards the credit-gated outbox — before that queued output had transmitted (truncating the tail). The client now grants window credit per consumed chunk in ClientRuntime.executeStreaming, and the node waits for the channel outbox to drain (bounded) before closing the session.

1.18.0

Changed

• omnyshell exec and omnyshell run stream output live. Both commands previously buffered all remote stdout/stderr and printed it only after the command exited, so a long-running command (e.g. exec web-01 "make build") showed nothing until completion. They now forward each output chunk to the terminal as it arrives. A new ClientRuntime.executeStreaming(...) delivers chunks via onStdout/onStderr callbacks and returns the exit code; the existing buffered ClientRuntime.execute(...) is unchanged and now builds on top of it.

1.17.0

Added

• Gzip compression for directory-drive sync (omnydrive 1.3.0). Drive content now transfers gzip-compressed over the DriveMessage channel. omnyshell carries drive content over its own Hub-brokered channel rather than omnydrive's HTTP transport, so the HTTP layer's automatic gzip never applied here; this brings the same saving to the channel by reusing omnydrive's transport-agnostic ContentCompression policy (gzip level 4) at the wire edges. File reads, file writes and the JSON manifest are compressed; payloads below 1 KiB and already- compressed file types (jpeg, png, mp4, zip, pdf, …) are sent verbatim. Each message carries a gz header flag so the receiver inflates only what was compressed. Git drives are unaffected (they move data via the git CLI).

1.16.0

Added

• omnyshell run --with <dir> co-mounts sibling dependencies. When a remote run needs a directory the project references by a relative path (e.g. a build that reads ../dependency-project), the repeatable --with option mounts the nearest common ancestor of --dir and every --with — a wrapper — and sets the remote working directory to wrapper/<--dir>, so the same relative reference resolves on the node. Only the named directories are synced (an --include whitelist built from their wrapper-relative paths is applied automatically); the rest of the wrapper is never pushed or pulled. Mount reuse now also keys on the filter, so changing the --with set creates a fresh mount instead of reusing one with a stale whitelist. This is the safe alternative to .. traversal: every synced path stays within the mounted wrapper root.

1.15.0

Added

• Partial directory mounts via --include/--exclude (omnydrive 1.2.0). omnyshell drive mount and the inner :drive mount gain repeatable --include/--exclude options (gitignore-style globs: *, **, ?, trailing-slash subtree matching) so a mount can sync only part of a local directory. Exclude wins over include; include acts as a whitelist. The filter is rejected when combined with --git (directory mounts only). It is persisted on the mount record and enforced node-side at the serving boundary: the node applies it to every manifest/read/write/delete, so excluded files are never hashed, pushed, pulled, or deleted, and baseline hashes stay consistent across both sides of a sync.

Changed

• Bumped omnydrive 1.1.4 → 1.2.0. The upgrade is backward-compatible for existing mounts: the new PathFilter defaults to empty (whole-tree sync) and the Drive/LocalContentSource/ManifestBuilder filter parameters are optional, so unfiltered mounts behave exactly as before.

1.14.0

Changed

• Faster directory manifests (omnydrive 1.1.4). Bumped omnydrive, whose ManifestBuilder now backs each build with a persisted stat-cache (<root>/.omnydrive/manifest-cache.json), reusing a file's recorded hash when its (size, mtime) are unchanged. An unchanged mount with thousands of files now costs one stat() per file instead of a full read + SHA-256. The produced FileManifest — and therefore the content-addressed SyncRef OmnyShell's drive sync compares — is byte-identical to a full rebuild, so no OmnyShell code change was required. The cache is advisory: a missing, stale, or unwritable cache silently falls back to a full rebuild. The .omnydrive directory is excluded from manifests, so it is never synced.

1.13.0

Changed

• Faster directory drive sync (omnydrive 1.1.3). Bumped omnydrive, whose DirectorySynchronizer now transfers changed files concurrently (up to 8 in flight) instead of one at a time, hiding per-file round-trip latency. This is safe over OmnyShell's single drive channel: request frames are written atomically per message (Channel._sendData) and the node serializes effects in arrival order (NodeDriveService), with replies matched by correlation id. No OmnyShell code change was required.

Added

• omnyshell run reuses a matching mount across repeated calls. Running the same local directory against the same node no longer mints a new ephemeral mount and re-pushes the whole tree each time. If a registered read-write directory mount matches, run (and exec --mount) reuses it and only syncs the changed files before running. Matching: an explicit --mount-path matches on node + local dir + remote path; an ephemeral run (no --mount-path) reuses a previous ephemeral mount for the same node + local dir (its recorded path is reused). The most recently mounted match wins. Pass --fresh to force a brand-new mount. Mount records now persist an ephemeral flag to support this.

1.12.1

Fixed

• omnyshell run (and exec --mount) failed on the default directory. Mounting . / ./ threw Invalid drive id "<node>/.": the absolute form of . ends in /., so the derived mount name became .. DriveManager now normalizes the local path before deriving the mount name and sync root.
• omnyshell run ephemeral mount failed when the node's working directory was not writable (PathAccessException: Creation failed, path = '.omnyshell', errno 13). The default ephemeral mount path was relative, so the node resolved it against its process working directory (often a non-writable system dir). It is now rooted at the node user's home (~/.omnyshell/run/...); the node expands a leading ~ in both the drive mount root and the exec working directory, so the mount lands somewhere writable and the command still runs in the same place. A user-supplied ~/... --mount-path / --cwd is expanded too.

1.12.0

Added

• omnyshell run — edit locally, run remotely, get results back. A new top-level command that mounts a local directory onto the node (pushing the files up), runs a command inside it, then syncs whatever the command produced back down to local. It wraps the existing OmnyDrive mount machinery, so it rides the same authenticated wss session — no extra ports. By default the mount is left registered (re-run, drive sync, or drive unmount later); the remote path is ephemeral unless --mount-path is given, and the command's working directory defaults to the mount path. Flags: --dir (local directory, defaults to the current dir), --mount-path, --mount-name, --initial-sync, --sync-interval (periodic sync-back while running), --cwd, --unmount and --clean-remote (delete the node copy on teardown).
• omnyshell exec gains a working directory and the same mount lifecycle. --cwd <dir> sets the remote command's working directory. --mount <local-dir> (with the same --mount-path / --sync-interval / --unmount / --clean-remote options as run) mounts a directory, runs the command in it, and syncs the results back — run is the convenience form that mounts a directory by default.

1.11.1

Fixed

• Windows service ran a stale version after an upgrade. A dart pub global activate install launches as dart <snapshot>, and Windows locks that pub-cache snapshot while the service runs — so pub global activate <new version> could not rewrite it, and uninstalling then reinstalling the service just re-pinned the same stale snapshot, so the latest omnyshell never took effect. service install now stages a private copy of the runtime under %LOCALAPPDATA%\OmnyShell\bin (or %OMNYSHELL_HOME%\bin) and points the Task Scheduler task there: upgrades can rewrite the pub-cache snapshot freely, and each (re)install refreshes the copy from the currently-installed version. The installer also stops a running task first so its lock is released before the copy is replaced, and uninstall cleans up the staged copy.

1.11.0

Added

• TCP tunnels / port forwarding. Expose an internal TCP port — a connected node's, or your own machine's (--local) — on a public port of the Hub, so external clients reach an otherwise-unreachable service (a NAT'd node's database, a localhost dev server) through hub-host:PUBLIC_PORT. Forwarded bytes ride the existing authenticated, multiplexed wss connection — no extra listener on the node and no new credential. The Hub binds the public listener within an operator-configured range (hub start --tunnel-port-range 20000-20100, fail-closed when unset) and authorizes each open with the same RoleBasedAuthorizer. Ownership is by principal, so a node-exposed tunnel outlives the requesting client and can be listed or closed from a later connection. Use omnyshell tunnel open <node> <port> (or --local <port>), omnyshell tunnel list / close, the in-session :tunnel command, or the Client SDK (openTunnel / listTunnels / closeTunnel). Built on tcp_tunnel's PortRange.
• :tree local command. Prints a sized directory tree of a remote path from inside an interactive session.

1.10.2

Added

• :info now reports the shell protocol family and more. The listing adds the remote shell's command-language family (POSIX (sh/bash), PowerShell, or cmd.exe) — useful on Windows nodes where the node picks the shell — alongside the node display name, global UID, connected Hub URI, operator labels, and the session id/mode.

Fixed

• Windows node service no longer leaves a black console window open. A user-scope service install registered the Task Scheduler task with an interactive logon token, so the cmd.exe action's console window stayed visible on the desktop for as long as the node ran. User-scope tasks now use an S4U principal ("run whether the user is logged on or not"), so the daemon runs in a non-interactive session: no window appears, and it keeps running after logoff. System-scope installs (session 0) are unaffected.

• :tree no longer crashes with a FormatException against a non-UTF-8 node. The command decoded the remote find/stat output (and error text) with a strict UTF-8 decoder, so a Windows node's OEM-codepage error message threw FormatException: Missing extension byte and took down the session. Remote command output in the local commands is now decoded tolerantly (allowMalformed: true), matching the rest of the client. (Note: :tree still relies on POSIX find/stat, so it only produces a tree on nodes that provide them; on a bare Windows shell it now reports the error cleanly instead of crashing.)

1.10.1

Fixed

• Windows connect no longer fails when WSL is present but has no distro. The shell probe picked C:\Windows\System32\bash.exe (the WSL launcher) because it exists on %PATH%; with no distro installed it prints "no distributions installed" and exits 1, closing the session immediately. Each bash candidate is now verified by running bash -c "exit 0" and is only chosen when it exits cleanly — an unusable WSL bash.exe is skipped and the probe falls through to PowerShell. A working WSL bash still wins; the check runs at most once per node process.

1.10.0

Added

• connect now works against Windows nodes. The interactive session protocol was POSIX-only (trap, eval, stty, and a printf/git/id prompt marker), so connecting to a Windows node — which spawned cmd.exe — produced only "command not recognized" errors. The node now selects the best interactive shell on Windows, preferring bash (Git Bash / WSL), then PowerShell (pwsh/powershell), then cmd.exe, and reports the chosen shell family to the client (via SessionOpened). The client speaks a matching ShellDialect for that family:

  • bash reuses the proven POSIX protocol unchanged;
  • PowerShell emits a PowerShell-native prompt marker (cwd, git branch/status counts, and an Administrator → root privilege flag) with the per-line prompt suppressed;
  • cmd.exe is a degraded last resort that tracks the working directory (%CD%) only.

  POSIX nodes (Linux/macOS) and Windows-with-bash are byte-identical to before; Windows exec mode still runs through cmd.exe /c unchanged.

1.9.1

Added

• omnyshell --version (-V). Prints omnyshell <version> and exits, following the usual console convention. The program name and version now also head the CLI --help output and the interactive :help listing.

Fixed

• Windows service install no longer fails with "cannot switch encoding". The Task Scheduler definition that schtasks /Create /XML imports must be a UTF-16 file; it was written as UTF-8, so schtasks rejected it ("the XML is malformed … cannot switch encoding"). The file is now declared and written as UTF-16 (little-endian with a BOM).

1.9.0

Added

• :tree local command — a sized directory tree of a remote path. Prints a tree -lh-style listing of a path on the connected node (default: the current remote directory), annotating every file with its size and every directory with its aggregated total in human-readable form. Supports :tree [path], -L <depth> to limit display depth (0 = unlimited), and -a to include hidden entries (dot-files are skipped by default). Symlinks are shown as non-followed leaves. The tree is built client-side from a single find … -exec stat … run on the node (GNU or BSD stat chosen by node OS), so directory totals stay accurate even when the display depth is truncated.

Fixed

• Installing the Node/Hub as a Windows service no longer fails with error 1053. sc start dart_omnyshell_node failed with "The service did not respond to the start or control request in a timely fashion" because a plain Dart console app cannot perform the in-process Service Control Manager handshake (StartServiceCtrlDispatcher → SetServiceStatus), so the SCM killed it after its timeout. On Windows the service commands now run the daemon through Task Scheduler (schtasks.exe) instead of the SCM: a boot-triggered task (system scope, runs as LocalSystem) or logon-triggered task (user scope) with no execution time limit and restart-on-failure, wrapping the run in cmd.exe to set OMNYSHELL_HOME and capture a log. Linux (systemd) and macOS (launchd) are unchanged.

1.8.3

Fixed

• sessions list now shows the real COMMAND and PATH. The COMMAND column was always - and PATH was frozen at the directory each session was opened in (it never followed cd), for both attached and detached sessions. The default PTY backend runs the interactive shell under a script(1) wrapper, so the session pid is the wrapper's — whose controlling terminal is the node's, not the session's pty, and whose cwd never changes. Process inspection now descends to the real shell beneath the wrapper before reading its foreground command and working directory.

1.8.2

Fixed

• Detach no longer leaves stray escape sequences on the local terminal. A local :detach returned early without restoring the terminal, so DEC private modes a full-screen remote program (vim, claude…) had enabled — mouse tracking, bracketed paste, alt-screen, hidden cursor, SGR — stayed on and the terminal kept emitting special characters (e.g. ESC[<…M mouse reports on every mouse move). Both detach paths now share one reset sequence and clean up the terminal.

1.8.1

Changed

• Upgraded omnydrive to ^1.1.2. Picks up the fix that prevents a mount sync from silently discarding local-only changes. OmnyShell already guards this in DriveManager._autoDirection (read-only mounts only push, read-write mounts only pull when the local copy is unchanged, and a two-sided change surfaces a conflict); added regression tests covering remote→local pulls of new node files, read-only mounts never pulling, and divergent syncs refusing to clobber local work.

1.8.0

Added

• Peek at a session's current screen without attaching. The new omnyshell sessions peek <node> <session-id> prints a session's current screen — the same bytes a resume would paint, captured by the node's alt-screen-aware replay buffer — without connecting to it or delivering any input. Works for both running (attached) and detached sessions you own.
• sessions list now shows the current command and path. Each row gains a COMMAND (the foreground command, or - at the prompt) and PATH (the session's working directory) column, queried best-effort by the node from the OS (Linux/macOS).
• New sessions start in your home directory. A freshly opened session now starts in the node user's home directory (like cd ~) instead of wherever the node process was launched. An explicit per-request working directory still takes precedence.

Changed

• Profile PATH is deduplicated on export. When syncing the node profile, the captured PATH now has empty and duplicate entries removed (first occurrence wins, order preserved) before it is written to profile.yaml.
• Clearer omnyshell node profile sync reporting. When nothing changes it reports Node PATH already up to date.; when it writes a change it now also reminds you to restart the node for the new PATH to take effect. (The restart hint is omitted during node start, which re-loads the profile immediately.)

1.7.0

Added

• Node environment profile (~/.omnyshell/profile.yaml). Sessions run the node's shell non-interactively, so no rc file is sourced and $PATH starts bare. The node now applies an env: map from ~/.omnyshell/profile.yaml (values support ${VAR} expansion) as the baseEnvironment of every shell and exec session. On an interactive node start the node derives PATH from your login shell rc (~/.zshrc, ~/.bash_profile, …) and prompts before writing it when it differs; non-interactive starts leave the profile untouched and print a hint. New flags --profile <path> and --no-profile-sync, plus an omnyshell node profile sync [--yes] subcommand to refresh on demand.
• Remember --insecure-skip-verify at login. Logging in with --insecure-skip-verify now asks whether to persist the setting for that Hub; when stored, later commands reusing the saved session skip TLS verification without re-passing the flag. A non-interactive login defaults to not storing it, and re-running login without the flag (or logout) clears it.

1.6.1

• Dependency updates in pubspec.yaml:
  • cryptography: updated from ^2.7.0 to ^2.9.0
  • omnydrive: updated from ^1.1.0 to ^1.1.1
  • lints: updated from ^6.0.0 to ^6.1.0
  • test: updated from ^1.25.6 to ^1.31.1

1.6.0

Added

• Manage OmnyDrive mounts from inside a session with :drive. The new local command is the in-session counterpart of the top-level omnyshell drive CLI: because the session is already attached to one node, the node is implicit, so paths take no <node>: prefix and every operation is scoped to the connected node. Subcommands mirror the CLI — :drive ls, :drive mount <local-dir> <remote-path> (or --git <url> <remote-path>, with --rw, --no-initial-sync, --name, --branch, --depth), :drive status, :drive sync [--push|--pull], :drive resolve [--accept-local|--accept-origin| --reclone], :drive remount, and :drive unmount [--sync-first] [--no-keep-remote]. A mount-id belonging to a different node is refused, and mounts share the same on-disk registry as the CLI.
• Background :drive watch. :drive watch <mount-id> [--interval S] [--debounce MS] auto-syncs a mount in the background while the shell stays usable, logging each sync above the prompt; :drive unwatch [<mount-id>] stops one or all watchers (teardown also runs automatically when the session ends or the mount is unmounted). Background output repaints around the input line via a new LocalCommandContext.printAbove hook.
• Live drive sync progress. Every drive operation now reports progress as it runs instead of only a final count: sync, mount, resolve, remount and watch, for both the omnyshell drive CLI and the in-session :drive command. The top-level CLI renders an in-place bar ([##########----] 71% 5/7 files src/main.dart); in-session prints a throttled syncing N/M: path line above the prompt. Per-file granularity for directory mounts comes from omnydrive ≥ 1.1.0's per-file ProgressEvents; git push/clone show a coarse pushing… / cloning… phase. Threaded through a new DriveManager onProgress callback and a SyncProgressBar renderer.

1.5.1

Fixed

• No more stray characters on the terminal after a session detaches. Two separate leaks were writing to the local terminal once a session was already gone. (1) The connect client never cancelled its remote stdout/stderr listeners on detach, so bytes still buffered in the channel were flushed afterwards — at an idle prompt the line editor repainted around them, smearing erase/prompt escape sequences onto the detached terminal. The listeners are now guarded against the detached state and cancelled on teardown. (2) When a session was detached from another window mid full-screen program, the terminal reset undid the alternate screen, cursor and color attributes but not mouse reporting, so every later mouse move spewed SGR mouse reports (ESC[<…M) onto the terminal. The detach reset now also disables every mouse-tracking mode (1000/1002/1003 and the 1005/1006/1015 encodings) and bracketed paste (2004).

• Interactive sessions no longer freeze on heavy output. The connect client consumed remote stdout/stderr without ever replenishing the channel's send window, so after a cumulative 256 KiB the node's flow-control credit drained and all further output stalled — the session appeared frozen. This surfaced most often with full-screen TUIs that repaint the whole screen on every scroll (e.g. claude's plan view, vim, less, htop), which exhaust the window within a handful of redraws. The client now grants window credit for every chunk it consumes. The node grants stdin credit symmetrically, so a large paste into a full-screen program can't stall input either.

1.5.0

Added

• Detachable sessions. Leave a node without killing the remote shell and reconnect later. From the interactive prompt, :detach [timeout] parks the session — the PTY, shell and child processes keep running on the node — and prints a short id to resume with (:detach, :detach 30m, :detach 2h, :detach 1d; units s/m/h/d). Manage detached sessions from the CLI: omnyshell sessions list <node>, omnyshell sessions resume <node> <id> (a full id, short handle or unambiguous prefix), and omnyshell sessions kill <node> <id>. Sessions are owned by one authenticated user on one node; the node enforces ownership and never reveals another user's sessions. A dropped client connection auto-detaches by default (preserving the shell) rather than terminating it. Output produced while detached is retained in an in-memory capture and replayed on resume. Detached-session state lives only in node memory — nothing is written to disk and it is lost on node restart by design. The Hub only authenticates, routes and correlates replies; it never persists detached-session metadata. New client APIs: RemoteSession.detach(), ClientRuntime.resumeSession(), listDetachedSessions(), killDetachedSession(); new NodeConfig autoDetachOnDisconnect, autoDetachTimeout, cleanupInterval.

• Detach a running session from another window. Because :detach can't be typed while a full-screen program (vim, top, less, a REPL) owns the terminal, omnyshell sessions detach <node> [session-id] [timeout] detaches a running session from a separate terminal — the attached window drops out of the full-screen app with its terminal restored and prints a resume hint, and the remote shell keeps running. With no session-id it targets your sole active session on that node (errors if several). omnyshell sessions list <node> now shows active sessions too (STATUS attached/detached) so you can find the id. New client APIs: ClientRuntime.detachActiveSession(), listSessions(), and RemoteSession.wasDetached / detachOutcome.

• sessions kill terminates running sessions too. omnyshell sessions kill <node> <id> now resolves both active (attached) and detached sessions, so you can kill a running session from another window; the attached client is disconnected. New ClientRuntime.killSession() (the old killDetachedSession remains as a deprecated alias).

• Resume restores full-screen programs. The node now keeps a continuous, alt-screen-aware capture of recent output for every session (not just while detached), so resuming into nano/vim/htop/less repaints the program's current screen — the frame it had drawn before detaching — instead of a blank terminal. Resume into a full-screen program attaches in passthrough without injecting a prompt marker (which previously typed into the program); the program's existing completion marker restores the prompt when it exits. Restoration is at the detached geometry. SessionOpened gains an altScreen flag (exposed as RemoteSession.resumedInAltScreen). The prompt-completion marker token is derived from the stable session id (reported unchanged across connect and resume), so a resumed client recognizes the marker the running program leaves behind and reliably repaints the prompt after the program exits — with the correct working directory.

1.4.0

Added

• Detachable sessions. Leave a node without killing the remote shell and reconnect later. From the interactive prompt, :detach [timeout] parks the session — the PTY, shell and child processes keep running on the node — and prints a short id to resume with (:detach, :detach 30m, :detach 2h, :detach 1d; units s/m/h/d). Manage detached sessions from the CLI: omnyshell sessions list <node>, omnyshell sessions resume <node> <id> (a full id, short handle or unambiguous prefix), and omnyshell sessions kill <node> <id>. Sessions are owned by one authenticated user on one node; the node enforces ownership and never reveals another user's sessions. A dropped client connection auto-detaches by default (preserving the shell) rather than terminating it. Output produced while detached is retained in an in-memory capture and replayed on resume. Detached-session state lives only in node memory — nothing is written to disk and it is lost on node restart by design. The Hub only authenticates, routes and correlates replies; it never persists detached-session metadata. New client APIs: RemoteSession.detach(), ClientRuntime.resumeSession(), listDetachedSessions(), killDetachedSession(); new NodeConfig autoDetachOnDisconnect, autoDetachTimeout, cleanupInterval.

• Detach a running session from another window. Because :detach can't be typed while a full-screen program (vim, top, less, a REPL) owns the terminal, omnyshell sessions detach <node> [session-id] [timeout] detaches a running session from a separate terminal — the attached window drops out of the full-screen app with its terminal restored and prints a resume hint, and the remote shell keeps running. With no session-id it targets your sole active session on that node (errors if several). omnyshell sessions list <node> now shows active sessions too (STATUS attached/detached) so you can find the id. New client APIs: ClientRuntime.detachActiveSession(), listSessions(), and RemoteSession.wasDetached / detachOutcome.

• sessions kill terminates running sessions too. omnyshell sessions kill <node> <id> now resolves both active (attached) and detached sessions, so you can kill a running session from another window; the attached client is disconnected. New ClientRuntime.killSession() (the old killDetachedSession remains as a deprecated alias).

• Resume restores full-screen programs. The node now keeps a continuous, alt-screen-aware capture of recent output for every session (not just while detached), so resuming into nano/vim/htop/less repaints the program's current screen — the frame it had drawn before detaching — instead of a blank terminal. Resume into a full-screen program attaches in passthrough without injecting a prompt marker (which previously typed into the program); the program's existing completion marker restores the prompt when it exits. Restoration is at the detached geometry. SessionOpened gains an altScreen flag (exposed as RemoteSession.resumedInAltScreen). The prompt-completion marker token is derived from the stable session id (reported unchanged across connect and resume), so a resumed client recognizes the marker the running program leaves behind and reliably repaints the prompt after the program exits — with the correct working directory.

Changed

• Interactive connect now hands the terminal to the remote while a command runs, instead of guessing from the command text. The managed prompt is drawn only when the shell is idle; the moment a command is dispatched the client enters raw passthrough and lets the remote program own the terminal until the CwdMarker completion signal returns. This fixes full-screen programs (vim, nano, less, top) and interactive line-readers (read, y/N confirmations, REPLs) corrupting — or being corrupted by — the local prompt, and replaces the fragile alternate-screen detection plus hardcoded foreground-program list.

• Cooked-mode input now echoes correctly. Because the remote shell runs with stty -echo, the dispatched command is wrapped to re-enable echo just for the program's runtime input (stty echo ; eval '<cmd>' ; <marker> ; stty -echo), so read/cat/y-N prompts show what you type while password prompts stay hidden (those programs disable echo themselves). No node-side change is needed.

Fixed

• Output arriving at the idle prompt no longer tangles with the input line. A backgrounded job (cmd &) printing while you type now erases and repaints the prompt around its output.

1.3.3

• dart_service_manager: ^1.2.2

1.3.2

Added

• --verbose (-v) flag on every omnyshell service subcommand. Drops the service manager's console logger to debug level so the underlying install/lifecycle steps (and the 1.2.0 user-systemd/privilege diagnostics) are printed; without it, info and warnings are shown as before.

• dart_service_manager: ^1.2.1

1.3.1

Changed

• Bumped dart_service_manager to ^1.2.0. omnyshell service install now benefits from the package's new install-time safeguards with no flag changes: on Linux, user-scoped installs auto-configure a persistent per-user systemd environment (enables lingering, resolves the user D-Bus bus and XDG_RUNTIME_DIR) so a --user Hub/Node survives logout/reboot and no longer hits Failed to connect to bus; and install now warns when the requested scope mismatches the current privilege level (e.g. running under sudo with the default user scope, or --system without elevation).

1.3.0

Added

• omnyshell service — install the Hub or Node as a native OS service. A new command group registers the running omnyshell executable with the platform service manager (systemd on Linux, launchd on macOS, the Service Control Manager on Windows) via dart_service_manager, so a Hub or Node starts at boot and is restarted on failure. The flags passed to service install <hub|node> … are the exact hub start / node start flags and are captured into the generated unit/plist; path flags (--cert, --key, --ca, --authorized-keys) are absolutized. Subcommands: install, uninstall, start, stop, restart, status, and reconfigure. Installs to the current user by default; --system installs machine-wide and runs with OMNYSHELL_HOME=/var/lib/omnyshell (override with --data-dir). --dry-run prints the rendered definition without touching the system, and --force replaces an existing service.

1.2.2

Fixed

• Nodes now report their real OmnyShell version. NodeConfig.agentVersion defaults to omnyShellVersion instead of a hardcoded 0.1.0, so a node's advertised Agent: version (shown in :info and node details) reflects the actual build. omnyShellVersion is now the single source of truth and is kept in sync with pubspec.yaml, guarded by a new version-tagged test.

1.2.1

Added

• --insecure-skip-verify flag to bypass TLS verification. Available on all connection commands (login, node start, connect, exec, drive, …), it disables both CA-trust and hostname verification so clients/nodes can reach a Hub presenting a self-signed certificate or a cert whose CN/SAN does not match the connection hostname (common with self-hosted hubs). Opt-in and insecure: it prints a [security] WARNING to stderr whenever active and is intended only for trusted self-signed/dev hubs. The setting is per-invocation and is not persisted into saved sessions.

1.2.0

Added

• omnyshell drive mounts directories and git repos onto remote node paths (OmnyDrive integration). A new command group binds a local directory — or a git repository — to a path on a connected node and synchronizes the two over the same authenticated wss session (no extra ports or credentials). Built on OmnyDrive: content-addressed manifests and explicit conflict detection (never a silent merge).
  • drive mount <local-dir> <node>:<remote-path> (and --git <url> with --branch/--depth) creates a mount; --rw makes the node mirror writable, --name overrides the mount name, --no-initial-sync skips the first push.
  • drive ls, drive status <id> inspect mounts (read local state, no Hub); drive sync <id> (--push/--pull/auto) synchronizes once; drive watch <id> (--interval/--debounce) auto-syncs live on filesystem changes.
  • drive resolve <id> (--accept-local/--accept-origin/--reclone) clears a conflict; drive unmount <id> (--sync-first, --no-keep-remote) tears one down; drive remount <id> re-establishes it after a node/CLI restart.
  • Direction is automatic: read-only mounts push, read-write mounts push/pull/no-op based on which side changed; a two-sided change surfaces a conflict instead of clobbering work.
  • Transport: a new SessionMode.drive carries a framed request/response RPC (NodeDriveService on the node serves a sandboxed content source and git ops; the client runs OmnyDrive's directory synchronizer against a ChannelContentSource). Mount state persists in ~/.omnyshell/mounts.json.
  • Nodes advertise a drive capability and accept mounts by default; the NodeConfig.driveEnabled / driveRoots options gate and path-restrict them.
  • New client APIs: DriveManager, MountStore/MountRecord, ChannelContentSource, DriveRpcClient.

1.1.0

Added

• omnyshell cert gen generates the TLS files a Hub needs. A new cert gen command builds a local CA and a Hub server certificate signed by it, writing ca.crt / ca.key / server.crt / server.key — the set the Hub uses (hub start --cert/--key) and that nodes and clients trust (--ca). Options: --out (output directory, default certs), --host (repeatable, adds SAN entries beyond the default localhost/127.0.0.1), --cn, --days, --ca-days, and --force (overwrite existing certs). It is the built-in equivalent of the tool/gen-dev-certs.sh script; both shell out to openssl, which must be on PATH. The generation logic is also exposed in the API as CertGenerator.

1.0.0

Added

• :download can fetch a remote path as a compressed archive. Pass --gz or --zip for a file, or --tar.gz or --zip for a directory, e.g. :download /var/log --tar.gz or :download /etc/hosts --gz. The archive is built on the node (via gzip/tar/zip in a temp file, removed afterward) so only the compressed bytes are transferred. The local file is named <base>.<ext> by default, or written into a destination directory / explicit path if given. Invalid combinations (e.g. --gz on a directory) and missing remote tools are reported clearly. Plain :download is unchanged.

• :ping accepts a count, e.g. :ping 3 sends three pings in sequence and prints each round-trip plus a min · avg · max summary. :ping with no argument behaves as before (a single ping).

• Prefix-aware history search in connect. When you have typed something at the prompt, Up/Down now walk only the history entries that start with that prefix (the text before the cursor), newest-first — e.g. type git then press Up to cycle just your previous git commands. With an empty line it behaves as before, walking all entries; editing the line recomputes the prefix.

• TAB completion in connect, like a normal ssh shell. Pressing Tab now completes the word under the cursor: command names (first word) are resolved by scanning the node's $PATH, and arguments are completed as file/directory paths (directories get a trailing /). A unique match is inserted (with a trailing space for non-directories); several matches complete the longest common prefix, and pressing Tab again lists them. Candidates are produced by a one-off remote exec run in the session's current directory (portable POSIX sh, no bash-only compgen), so relative paths resolve correctly.

Fixed

• Interactive confirmation prompts now work (e.g. :download's "Proceed? [y/N]"). The line editor previously paused all input while a local command was running, so the answer line could never be read — the prompt appeared to ignore y + Enter. Prompts are now read through the editor directly while a command is in flight; Ctrl-C cancels a prompt (counts as "no"), and non-interactive sessions auto-proceed. Stray keystrokes during a running transfer are still ignored.

• The cwd/git prompt marker no longer interferes with foreground programs. When a command launches an interactive program (editor/pager/monitor such as nano/vim/less/top, or a bare REPL like python/node), the connect client now switches to raw passthrough immediately — detected from the typed command, not just from the alternate-screen output sequence. This fixes the marker being injected into the program's stdin (e.g. on each Enter inside nano) on backends where the alternate screen is never reported to the client, such as the macOS script(1) PTY whose stdout is not a tty. The Ctrl-C prompt resync is likewise suppressed while a foreground program owns the terminal.

• Enter is now recognised inside remote full-screen apps and their prompts (e.g. confirming nano's "File Name to Write" on Ctrl-X). Raw passthrough now forwards the Enter key as a carriage return (\r), matching ssh: Dart's raw mode leaves ICRNL enabled, so the local terminal delivers Enter as \n, which some apps accept in their editor body but ignore at status-bar prompts.

Changed

• No git queries after read-only commands. Commands that cannot change the working directory or git state (e.g. ls, cat, pwd, git status/log/diff) now enqueue a lightweight completion ping instead of the full cwd/git marker: the prompt still repaints in the right place (after the command's output) and keeps its current cwd/branch/status, but the remote no longer runs git rev-parse/git status. Blank input lines repaint the prompt locally with no remote round-trip at all.

Removed

• Disabled the portable_pty/native PTY backend: the PTY exports are removed.

Added

• Default PTY backend now uses the system script(1) utility — no FFI, no native library. Nodes serve interactive connect shells on a real pseudo-terminal allocated by the OS script command, launched as an ordinary child process. The child gets a genuine tty (isatty() true; full-screen apps like vim/htop work) at the client's requested geometry. Selectable via node start --pty-backend script|native|none (default script). This avoids the native portable_pty crash entirely. Trade-off: this backend cannot propagate live resize (SIGWINCH) to the remote terminal — only the initial geometry is honoured; use --pty-backend native if you need live resize. The client still advertises its local TERM/columns/rows when opening the session; when no PTY is available (Windows, or script missing) the node falls back to the pipe-based shell and conveys the initial geometry via TERM/COLUMNS/LINES environment variables.

• Command history keyed by node UID, with change detection. Interactive history is now scoped to the node's deterministic UID rather than its logical id, under ~/.omnyshell/history/<user>@<nodeUid>.history. The last-seen UID for each <user>@<node> connection target is tracked under ~/.omnyshell/node-uids/; when a node reconnects with a changed UID the user is alerted and — interactively — prompted whether to migrate the prior UID's history into the new UID's history (non-interactive sessions migrate automatically). The old history file is always left intact as a backup. Nodes that report no UID fall back to the legacy <user>@<node> key.

• Deterministic global UIDs for nodes and hubs. Each node and hub now derives a stable identifier from its own identity material rather than a random/time seed, so the same machine resolves to the same UID on every start and across hubs. A node UID (nod_…) combines the node's Ed25519 public key (empty for token/keyless nodes) with stable hardware/platform attributes — a per-OS machine id (/etc/machine-id, macOS IOPlatformUUID, Windows MachineGuid), os, arch and hostname. A hub UID (hub_…) combines the TLS certificate's public key (SPKI, so it survives cert renewal when the keypair is reused) with the same hardware/platform attributes. Inputs are length-prefixed (TLV) and SHA-256 hashed under a per-kind domain-separation tag, then rendered as URL-safe base64 — every UID is also a valid node id. The UID is persisted under ~/.omnyshell/{node,hub}.uid, recomputed on every start, and a change is reported loudly (the previous value is retired into the file's history). The node advertises its UID in node.register (surfaced in discovery and :info); the hub advertises its UID in the challenge hello so peers can identify and pin it. Both are printed at startup by the CLI.

• Command history with arrow-key navigation. While connected to a node, the interactive prompt now supports a real line editor: Up/Down walk backward/forward through previously entered commands, and Left/Right, Home/End (also Ctrl-A/Ctrl-E), Backspace, Delete, Ctrl-C (discard line) and Ctrl-D (EOF on an empty line) edit the line. History is persisted per node + user under ~/.omnyshell/history/ (mode 600), so different nodes or principals never share a history; blank lines and consecutive duplicates are skipped and the file is capped at 1000 entries. Both remote shell commands and local :commands are recorded; confirmation-prompt answers are not. The prompt switches stdin to raw mode on a TTY and restores it on exit; piped/non-interactive input falls back to plain line reading with history disabled.

Changed

• connect banners now span the full terminal width. The welcome banner's horizontal rules stretch to the terminal width (no longer capped at 72 columns), and exiting/disconnecting now prints a full-width rule before Session closed, visually separating the finalized session from the local terminal output. The closing line also names where you were connected (Session closed (exit 0) · <node> @ <hub>).

• connect prompt colors refreshed. The working directory is now cyan and the git segment is blue with a red branch name and green status counts (was a blue cwd and a yellow git segment).

• The portable_pty (FFI) PTY backend is temporarily deprecated. PtyShellBackend/PtyShellSession are retained and still opt-in via node start --pty-backend native (they support live resize), but are no longer the default: the underlying native library has a SIGCHLD-handler memory-safety bug that races the Dart VM's child reaper and can intermittently crash the node (EXC_BAD_ACCESS inside portable_pty_open). Reported upstream; once fixed this backend will be promoted back to the default and the deprecation removed.

Fixed

• Ctrl-C interrupts the remote command instead of closing connect. Raw mode clears ICANON but not ISIG, so the terminal raised SIGINT on Ctrl-C and terminated omnyshell before the keystroke ever reached the line editor. Interactive sessions now intercept SIGINT at the process level (so omnyshell stays alive) and relay it to the remote foreground command, discarding the local input line first (or passing straight through to a full-screen app). The remote shell installs a no-op INT trap at session start so it survives the signal — interrupting a running command without killing the (non-interactive) shell, while the command itself still receives the default disposition and stops. Non-interactive runs keep the default behaviour so a scripted session can still be killed with Ctrl-C.

• Full-screen apps (nano/vim/less/top) now work over connect. Two problems are fixed. (1) The cwd-marker printf was sent on its own line right after the command, so the non-interactive remote shell left it in the PTY input buffer where a foreground program read it as typed input (every newline in nano echoed the marker, and a stray pico.save could appear). The command and marker are now sent as one logical line — eval '<cmd>' ; <marker> — so the shell consumes both before executing; the marker runs only after the command/app exits. eval keeps this valid for any command (pipes, trailing &, cd). (2) The client line editor kept buffering keystrokes while a full-screen app was running. The client now watches the output stream for the alternate-screen sequences (ESC[?1049h/ESC[?1049l) and, while the app owns the screen, switches the editor to raw passthrough so keystrokes reach the app verbatim; on exit it restores local line editing and repaints the prompt. (Non-alternate- screen interactive programs such as a bare REPL still have no client-observable signal and remain best-effort.)

• connect no longer shows the remote shell's own prompt over a real PTY. On a PTY the node's shell ran interactively and printed its own PS1/theme prompt and echoed every keystroke — duplicating the client's managed prompt. The script(1) backend now runs the shell non-interactively for shell mode, reading its command stream from /dev/stdin with terminal echo disabled, so the client's prompt is the only one shown. A real PTY is still allocated (full-screen apps and the requested geometry keep working), matching the prompt-free behaviour of the pipe fallback the CwdMarker design assumes.

• Prompt no longer corrupts the cursor on a real PTY. A PTY's ONLCR rewrites the cwd-marker line's trailing \n as \r\n; the stray \r was captured as the marker's last (privilege) field, so a non-root session rendered (⚠ \r) into the prompt and the carriage return jumped the cursor to column 0. The marker parser now drops a trailing \r before splitting fields.

• PTY sessions now terminate correctly on Linux. The portable_pty native library keeps the pty slave fd open for the handle's lifetime, so on Linux the master never reports EOF after the child exits (macOS does), leaving the output stream open forever — interactive sessions appeared to hang and the real-PTY tests timed out on CI. The session now detects child exit explicitly via tryWait() once all readable output has drained, instead of relying solely on master EOF.

0.3.0

Changed

• Local commands now use a : prefix (:help, :info, :exit, …) instead of /. The old / prefix collided with ordinary shell input that legitimately starts with / — most notably absolute binary paths such as /bin/bash, which were intercepted as unknown local commands instead of running. A colon never begins a real shell command, so local commands and remote shell input are no longer ambiguous. Breaking: scripts or muscle-memory using /help, /exit, etc. must switch to :help, :exit.

Added

• :download / :upload file transfer. Inside an interactive session, :download <remotePath> [localDest] and :upload <localPath> [remoteDest] move files and directories between the client and the node. Transfers run over a separate, parallel Hub connection (a dedicated transfer-mode session), so the interactive shell stays responsive. The payload is streamed per file and compressed with GZip level 4 (built-in dart:io codec), resumable by byte offset (re-run to continue a partial copy), and every file's SHA-256 is verified end-to-end — a mismatch drops the file so a re-run fetches it cleanly. Implemented purely over the existing binary channel + credit-window flow control, with no Hub changes (handshake/metadata ride a self-framed record stream on stdin/stdout). A progress bar is shown on a TTY.
  • Destination may be a file or a directory (cp/scp semantics, resolved on the receiving side): an existing directory or a path ending in / means write into it (keeping the source's top-level name); otherwise the destination names the result itself (a single file is written to exactly that path; a directory copied onto a non-existent path becomes the new root); copying a directory onto an existing file is refused.
  • Pre-transfer confirmation spells out the resolved destination, the chosen mode, and the exact target path of each file (tagged new / overwrite / resume) before anything is written.
• Git branch, status, and privilege in the prompt. When the remote working directory is git-managed, the connect prompt now shows the branch and a compact status — user@node:cwd git(branch +S ~M ?U) $ — where +S ~M ?U counts staged/modified/untracked files and appears only when there are changes. A superuser session also shows a (⚠ root) indicator. Both are gathered over the existing per-command $PWD marker (no extra round trip) and ANSI-colorized on a TTY (branch yellow, status counts red, root warning bold red; NO_COLOR honored).
• Welcome banner on connect. Opening an interactive session now prints a rule-separated welcome banner summarizing the connection: the OmnyShell CLI version, the node (id, display name, online status, platform, hostname, agent version), advertised capabilities (shells, features, max sessions), operator labels, the authenticated user and roles, the Hub URL, measured round-trip latency, and the session id/mode. Colorized on a TTY (honors NO_COLOR) and falls back to a plain banner when piped.
• login / logout commands. omnyshell login authenticates to a Hub once (verifying the credentials with a real auth handshake) and saves the session to ~/.omnyshell/credentials.json (file mode 600). Subsequent client commands (connect, exec, nodes list, whoami) then run without credential flags. Sessions are keyed by Hub URL with a remembered default, so multiple Hubs are supported; explicit --principal/--token/--key still take precedence. omnyshell logout removes a saved session (--hub) or all of them (--all). Key-based logins reference the existing Ed25519 seed file by path rather than copying the secret.

Documentation

• Document the login / logout flow and the credential-free command usage in the README, and refresh the badge row (status, tag, commits, PRs, code size).

0.2.0

Added

• Interactive prompt line. omnyshell connect now shows a user@node:cwd $ prompt before each command. The working directory is tracked live via lightweight shell integration (a hidden per-session marker that reports $PWD after each command), and the prompt is ANSI-colorized when stdout is a TTY (honoring NO_COLOR).

0.1.0

Initial release — Stage 1: secure core and a working Client → Hub → Node vertical slice.

Added

• Hub-centric architecture. Clients connect to a Hub by node identity, not by host:port. The Hub discovers nodes, authenticates and authorizes principals, and brokers sessions.
• WebSocket-on-TLS transport. All connections are encrypted; there is no plaintext or raw-TCP mode. Nodes dial the Hub outbound and hold a persistent control connection (NAT-friendly).
• Multiplexed channel protocol. SSH-channel-style multiplexing over a single connection: JSON control messages on WebSocket text frames, binary stream data (stdin/stdout/stderr) behind a compact 10-byte header on binary frames.
• Pluggable authentication. Authenticator contract with two implementations: PublicKeyAuthenticator (Ed25519, authorized_keys-style, replay-resistant nonce challenge) and TokenAuthenticator (bearer).
• Authorization. Authorizer contract with a default role-based implementation enforced by the Hub on every session open.
• Node runtime. Connect → authenticate → register → advertise capabilities → serve sessions, with automatic reconnect and exponential backoff.
• Hub broker. NodeRegistry, SessionRouter (tunnel relay), and a Clock-driven HeartbeatMonitor.
• Client SDK. execute() for one-shot commands and startInteractiveShell() for real-time streaming sessions, plus an extensible local / command system.
• Process shell backend. ProcessShellBackend runs commands via Process.start behind a ShellBackend interface (PTY backend can plug in later).
• CLI. omnyshell hub start, node start, connect, exec, nodes list, and whoami, all built on the public Dart APIs.
• Tests. Unit, integration and end-to-end coverage over real wss loopback connections with a self-signed test certificate.