Release notes

Discovery on Windows no longer lists devices that have left the network. The desktop app stays responsive with results of 20,000 rows, and its CSV export now makes control characters safe the way the CLI does. MCP clients see progress on long scans and can cancel them, and Windows names Felix Stubner as the desktop app publisher.

Added

• The MCP server reports progress and can be cancelled. Discover, port scans and sweeps send progress notifications to clients that ask for them, so a long sweep no longer looks stuck. A client that cancels a call now stops the scan, instead of it running to the end and holding one of the server's sixteen request slots.

Changed

• The desktop app stays responsive with large results. It drew every row of a result and redrew all of them on every keypress or scroll, so a full 4,096-port scan took about a tenth of a second per arrow key, and a 20,000-row result such as a long packet capture over half a second. It now draws only the rows on screen, so 4,096 rows respond as fast as 100, and 20,000 in about a twentieth of a second. Progress updates from a running scan are also sent ten times a second instead of once per port.

• Windows lists the desktop app's publisher as Felix Stubner. Installed apps showed "netscli", a default taken from the app's internal identifier, while the code signature, winget and the Microsoft Store listing all say Felix Stubner. Upgrading from an earlier version leaves one entry, under the new name.

Fixed

• Discovery on Windows listed devices that had left the network. The Windows device table keeps an entry for a while after its device goes, and discovery reported those entries as devices. On one network that was 4 of 26, none of which answered anything afterwards. Devices found only in the table are now asked again directly, which can add about 2 seconds and only when the table holds such entries, and are listed only if they answer. Devices that ignore ping still answer, so they are still found.

• The desktop app's CSV export defuses control characters, as the CLI's does. A banner containing an escape sequence went into the file as it was, and a cell such as a tab followed by a number skipped the guard that stops a spreadsheet reading it as a formula. Both exports are now tested against the same list of cases.

• The desktop app no longer leaves a program running after looking for the CLI. To offer MCP setup it runs each netscli it finds with --version and gives up after 3 seconds. A program still running at that point was left behind. It is now stopped.

• MCP clients built on the official SDK can connect. Every reply the MCP server sent carried both a result and an error, one of them empty, which the protocol doesn't allow. The official TypeScript SDK drops a reply like that, so a client built on it waited for the answer to its first message and timed out. Replies now carry one or the other, and CI connects to the server with the SDK whenever the code changes.

• Cancelling an MCP call works when the server is busy. When a 17th call arrived with 16 running, the server stopped reading what the client sent until one finished. A cancel, which is how a client frees a slot, went unread, and so did the client closing the connection. The server now keeps reading. Up to 16 more calls wait for a slot, and any past that are refused at once with an error.

• Cancelling or disconnecting stops an MCP packet capture. A capture ran on after its call was cancelled, timed out or lost its client, for up to an hour, and gave its slot back straight away, so cancelling and restarting got past the limit of four captures. It now stops, and gives the slot back once it has. Background captures stop when the client goes too. In a test on Windows, a server whose client left mid-capture exited within 0.3 seconds, where before it was still running after 20. A capture_pcap call with only maxPackets could also run for an hour. It now stops after the 10 seconds the tool advertises.

• mDNS discovery no longer leaves its listener running. A service type without its final dot, such as _http._tcp.local, or a cancelled MCP call left the listener running in the background, its network sockets open and any searches it had started still repeating, until the program exited. It now always stops.

• An MCP result too large to send keeps its file path and counts. A result over 1 MiB that wasn't a plain list was replaced by an error, so a large packet capture lost its file name, packet count and job status along with its packets. Now only its longest list is cut to fit. A background capture also reports its file name as soon as it starts.

• Smaller MCP server fixes.
  • The server answers ping, which clients use to check that a server is alive. It answered with an error.
  • It agrees a protocol version it supports, instead of whatever version the client asked for.
  • Text from scanned hosts in HTTP headers and mDNS records is cut to 256 characters, as banners already were.
  • discover_mdns takes at most 32 service types, and browses each once.
  • An over-long request line no longer has its tail read as the next message.
  • A request that reuses the id of one just answered can always be cancelled. A race could leave it impossible to cancel.

• Smaller scan fixes.
  • A port behind a firewall that rejects connections with an ICMP message reads as filtered, as nmap reports it, instead of as an error.
  • The HTTP probe sends IPv6 targets a valid Host header, with the address in brackets. The old one was malformed, and a strict server can refuse it.
  • A stray data/oui.json in the folder netscli ran from no longer replaces or empties the vendor list.
  • On a network wider than /16, the default /24 is the one around this machine's address on that network. It could come from another interface or fall back to 192.168.1.0/24.
  • Windows discovery re-checks at most 64 devices at once. Checked all at once, a very long device table could drop devices that were there.
  • The example in the netscli-core README compiles.

Security

• MCP packet captures never overwrite a file or follow a symlink. capture_pcap wrote capture.pcap in the server's working directory, often your project, and replaced any file of that name. On Linux and macOS, a capture.pcap symlink in a repository could send the capture through to wherever it pointed, as root if the server ran as root. Captures now create a new file and refuse a name that is taken, symlinks included, and without outputFile each capture gets a new name.

• DNS lookups no longer go to Cloudflare behind your back. When your own DNS servers could not answer a lookup, including when a name simply had no records of the type asked for, NetsCLI asked Cloudflare's public resolver (1.1.1.1) the same question. That sent names you looked up to a third party, unencrypted and without saying so, and the privacy page said the opposite. It affected dns and every command that turns a host name into an address, in the CLI, the terminal UI, the desktop app and the MCP server. Lookups now go only to the DNS servers your computer is set up to use. Some home routers refuse record types other than A and AAAA, and dns now reports that refusal instead of quietly asking someone else. NETSCLI_DNS_FALLBACK no longer does anything. The privacy page now says what 0.3.4 and earlier did.

Port scans now name the software and version a service reports, can check common UDP services, and report closed ports as closed on Windows instead of filtered. The CLI writes CSV and Markdown tables, inspect gives an OS hint, and the desktop app can update itself. Discovery no longer lists the network broadcast address as a device.

Added

• The CLI can write CSV and Markdown tables. --csv and --md work on every command that returns a list (discover, scan, sweep, dns, ping, arp, interfaces, mdns and pcap). The columns are the same field names --json uses, one row per host, port, record or packet, so a result opens straight in a spreadsheet or pastes into an issue without going through jq. Text a scanned host chose is made safe for each. A banner that starts like a spreadsheet formula is defused, the same way the desktop app's CSV export already does it, and one containing Markdown or HTML is escaped.

• Port scans show the software and version where a service states them. An SSH server's identification line, a web server's Server header, an FTP or mail server's greeting and a MySQL or MariaDB server's connection greeting usually name the software, often with its version: OpenSSH 9.6p1, nginx 1.25.3, MariaDB 10.11.6. Redis, Valkey and Memcached say nothing until asked, so each gets the one read-only question that returns its version. Scans report this as product and version in the JSON, and in a Version column in the CLI, the terminal UI and the desktop app. It's far narrower than nmap's -sV: a service that doesn't announce itself, and isn't one of those three, gets no version.

• UDP scanning. netscli scan <host> --udp checks the UDP services most networks run (DNS, NTP, NetBIOS, SSDP and mDNS), each sent the request it expects, or the ports you give with -p. A reply reads as open, with what came back (NTP v4, stratum 2, a UPnP device's server string, a Windows machine's NetBIOS name). A port-unreachable reads as closed, and silence reads as open|filtered, because UDP can't tell a quiet service from a firewall. It needs no administrator rights. The terminal UI takes /scan <host> --udp, the MCP server's scan_ports takes udp: true, and the desktop app's Port Scan has a TCP/UDP switch.

• Inspect gives an OS hint. netscli inspect now says what the host probably runs, with the clues behind it, such as a Windows machine's exact version and build from the start of an SMB connection (no login), the distribution an SSH banner names, an (Ubuntu) or IIS web server header, Windows' RPC and file-sharing ports, an Apple or Raspberry Pi network card, and the ping reply's TTL. It also shows the host's MAC address and vendor when it's on the same network. It's a hint, not nmap's packet fingerprinting, and needs no administrator rights. The same hint is in the terminal UI, the desktop app's Inspect details and the MCP server's inspect_host.

• The desktop app can update itself. It already told you when a newer release was out and linked to the release page. Now, where it can, the notice opens a dialog with the release notes and an Install and restart button. The update is downloaded, checked against NetsCLI's signing key, installed, and the app reopens. Nothing downloads unless you click it, and the existing setting still turns the check off entirely.

This takes effect from the version after this one. An app can only update itself if it was built with the updater, so 0.3.4 is the first that can, and 0.3.5 is the first update it will install.

Changed

• MCP scan replies leave out closed and filtered ports. scan_ports and inspect_host returned one entry for every port scanned, so a 1,024-port scan of a machine with two open ports came to 136,834 bytes of mostly "closed". They now return open, open|filtered and errored ports only, which made the same scan 178 bytes. Pass include_closed: true for every port. Replies are also sent as compact JSON instead of indented, which took a sweep of the local network from 15,399 bytes to 8,746.

• Defaults now match across the CLI, TUI, desktop app and MCP server.
  • The desktop app scanned five ports by default (22, 80, 443, 8080 and 8443) where everything else scans 22, 80 and 443. It now uses the same three, for port scans and inspect alike.
  • The MCP server applied the timeout it advertises to every step, so an MCP inspect or sweep pinged with 500 ms where the others use 1,000, and gave name lookups 500 or 1,000 ms where the others give 1,500. Without a timeout each step now keeps its usual default.
  • The MCP tool list advertised defaults the server did not use. DNS lookups claimed an A record and look up every type, discover and sweep claimed 192.168.1.0/24 and use your own network, and background captures claimed capture.pcap and name the file after the job. The tool list now says what happens.
  • Ping count tops out at 256 everywhere. The desktop app stopped at 50 or 64, and the TUI had no limit at all.
  • Desktop inspect ignored the concurrency setting in Preferences. It now uses it, like scan, discover and sweep.
  • A desktop packet capture lasts 10 seconds by default, as on the CLI and MCP server, instead of 5.

• Discovery reads the Windows device table directly instead of running arp -a and parsing its text. On an idle machine that makes no measurable difference (arp -a took 65 ms), but it removes a program start from every discover and sweep, and under heavy CPU load arp -a took about 4 seconds on the same machine.

Fixed

• Discovery reported the broadcast address as a device. The network's device table on Windows lists x.x.x.255 with the MAC ff:ff:ff:ff:ff:ff, and discovery listed it as a host that ignored ping (and a sweep then scanned it). The network and broadcast addresses, and broadcast and multicast MACs, are no longer reported.
• Closed ports showed as filtered on Windows. Windows waits about two seconds before reporting a refused connection, and the scan stops waiting after half a second, so a port that was plainly closed came back as filtered. The scan now reports it as closed straight away. A scan of 1,024 ports on a LAN machine went from 1,022 filtered to 1,022 closed.
• The macOS app could be refused as broken on Apple Silicon. Its only signature was the one Apple's linker puts on every arm64 program, which claims the app's files are sealed when nothing sealed them. macOS's own checks fail it with "code has no resources but signature indicates they must be present", a broken app rather than one from an unidentified developer, so the Open Anyway route the site describes may never be offered. The whole app is now ad-hoc signed, which seals it, and the release checks the app inside the finished .dmg before shipping it. It is still not notarized, so the first launch still goes through Open Anyway in System Settings → Privacy & Security.
• Redis, DNS, NFS, SOCKS and Prometheus ports were probed as if they spoke TLS. The scanner treated any service name ending in "s" as a TLS variant, a rule meant for imaps and pop3s, so these ports got a TLS handshake instead of having their greeting read. The TLS services are now listed by name.
• Installing the desktop app with Scoop now adds it to the Start menu. The manifest's shortcut pointed at NetsCLI.exe, a file that is in no version of the package. Scoop extracts the MSI rather than running it, which leaves the app at PFiles\NetsCLI\netscli-gui.exe. Scoop reported "Creating shortcut ... failed" and finished the install anyway, so the app was installed with no way to launch it but finding the folder. The manifest now flattens that folder and names the real executable, and the publish job sets both on every release.
• The desktop app itself is signed, not only its installer. Since 0.3.3 the .msi has carried an Authenticode signature, but netscli-gui.exe inside it did not, and that is the file SmartScreen and antivirus look at when the app runs. It is now signed during the build, before it is packed into the installer, and the release checks the finished MSI's contents before shipping it. Measured on the published 0.3.3 installer. The MSI's signature is valid, the app inside it is unsigned.

Windows executables and the desktop installer are signed, so Windows no longer warns that the publisher is unknown. Connecting the MCP server to a client no longer means installing netscli first and writing the configuration by hand: npx netscli serve works, and each release carries a one-click bundle for clients that support them.

Added

• Windows executables and the desktop installer are signed. Both netscli builds and the .msi now carry an Authenticode signature from a Certum certificate, timestamped so it keeps verifying after the certificate expires. Windows has been showing an unknown-publisher warning on every download since the first release, and SmartScreen treats an unsigned installer from a low-reputation domain as something to discourage rather than merely flag.

• Two new ways to connect the MCP server, so it no longer has to be a hand-written config file. Until now the only route was installing netscli, finding where it landed, and writing the JSON yourself.

npx netscli serve now works: netscli is on npm as a small launcher plus one prebuilt binary per platform, and npm fetches only the one that matches your machine. This is the form every MCP client's documentation already uses.

There are also .mcpb bundles on each release, one per platform. A client that supports MCP bundles installs one in a single action. The binary is inside, so nothing else is needed. The install prompt carries a switch for the local-network default described below, which is the first time that choice has been visible to the person making it rather than buried in an environment variable.

Anyone who already has netscli installed should keep pointing their client at it. That copy is the version you chose, there is only one of it, and it is the only one of the three that can capture packets. Libpcap and Npcap cannot be shipped inside an npm package or a bundle.

Documentation

• The MCP page now says what the local-network default is and why. The server has refused targets outside your own networks since 0.3.1, and the only description of that lived in a source comment, so the first anyone heard of it was an error message naming a variable. It is now written down next to the reason. This is the one surface where a scan can be requested by something a model read rather than by the person at the keyboard.

Repairs the Linux packaging. The AUR desktop package installed a file that could not start, and the AppImage aborted before opening a window on any host with a recent Mesa. A desktop window that opens blank now recovers by itself, netscli with no arguments no longer hangs where there is no terminal, and discover names devices from mDNS when reverse DNS comes back empty.

Added

• Discover names hosts from mDNS when reverse DNS cannot. Consumer routers do not serve PTR records for their own DHCP clients, and appliances ignore LLMNR and NetBIOS, so reverse lookup returned nothing for exactly the devices someone opened the app to identify. Those devices announce their names over mDNS constantly, and netscli has shipped an mDNS browser all along as a separate operation that discover never consulted. It does now, filling only the blanks, so nothing that resolves today changes. On one ordinary /24 that named 3 more of 26 hosts, taking 21 named to 24. Results carry a hostname_source of reverse or mdns, since a name from a device's own announcement is a different kind of claim from one in DNS.
• The desktop app says the MCP server exists. Someone who only ever opens the app had no way to learn that netscli ships an MCP server, let alone connect an agent to one. There is now a panel that looks for a netscli binary and gives you the client configuration to paste, including what to do when it cannot find one. The desktop installers do not carry the CLI.

Fixed

• netscli with no arguments no longer hangs when there is no terminal. With no subcommand it opens the TUI, which needs a terminal to draw on and read from. Without one it did not fail, it blocked forever. Raw mode was entered and the runtime then waited on input that could never arrive. So netscli | head, or netscli from a script or a CI job, ran until something killed it. It now prints what --help prints and exits 0. The TUI is unchanged wherever there is a terminal.

• The docs site lost its navigation and its theme switch between 800px and 1152px wide, and the search button sat stranded beside the wordmark. The header links (Features, Install, FAQ, Docs, Changelog, GitHub) and the light/dark control were both hidden across that range, on the understanding that the mobile menu carried them from there down. The button that opens that menu only appears below 800px, so for 352px of width there was nothing to press and no way to reach any of it. The docs sidebar is not a substitute. It lists the pages of the docs and carries five of those six links nowhere.

The search button had a second fault behind it. The width at which the links hide moved from 900px to 1152px and two rules that depended on that number stayed put, so the layout seam ended up on a hidden element, which takes no part in the layout. Search fell back to the left with up to 861px of empty bar beside it.

A new check measures where the header's controls sit at seventeen widths. Neither fault was visible to the existing accessibility, contrast or performance gates, because both are about position rather than markup, colour or speed.

• netscli-gui-bin on the AUR installed a desktop app that could not start. The PKGBUILD did not set options=('!strip'), and strip is in makepkg's default options. An AppImage is the AppImage runtime (an ordinary static ELF) with a squashfs image appended after everything the ELF headers describe, so stripping it rewrote the file from its section table and threw the appended image away. What reached /usr/bin was the 944,632-byte runtime out of a 79 MB download, and running it said only "This doesn't look like a squashfs image". The package built, installed and verified its checksum at every step, because the truncation happened after the checksum was checked. Reported in #377.
• The Linux desktop AppImage no longer aborts on hosts with a newer Mesa. It failed with Could not create default EGL display: EGL_BAD_PARAMETER before any window appeared. The AppImage carried its own copies of nine display-stack libraries (the wayland client stack, libxkbcommon, and part of the xcb/X11 stack) and put them ahead of the host's, so the host's Mesa was made to talk to the wayland client library from the machine the release was built on. Those libraries are now removed from the image after it is built. Reported in #378 against v0.2.6 on Mesa 26.2.2 (v0.3.1 bundled the same nine).
• A desktop window that opens black or blank now recovers on the next launch. WebKitGTK's hardware compositing can fail against a driver that only partly supports it, and it fails silently. The window opens, nothing paints, and there is nothing on stderr to go on. Seen on virtual machines using vmwgfx. The app now marks each launch and clears the mark once the UI has actually drawn a frame, so a launch that never drew one is noticed by the next, which turns hardware compositing off and says why.

This could not be a setting in the app. Every GUI preference lives in the webview's localStorage, and the webview is the part that is not rendering, so someone looking at a blank window cannot reach any of it. Alongside the automatic recovery there are now --disable-gpu-compositing and --gpu-compositing flags, which are remembered across launches. Linux only: the other two platforms use a web engine with neither the fault nor the setting. Also reported in #378.

• Text in the desktop app's result tables can be selected again. Both the table and its wrapper set user-select: none, so a port, MAC address, vendor string or banner could not be dragged over with the mouse, and those values are on screen precisely so they can go somewhere else. The only route out was the detail pane. Selecting rows is a click, not a drag, so nothing was gained by it. Reported in #417.
• The website's release notes lost their paragraph breaks, and the fade over a long entry read navy rather than matching the page. Both on the changelog page.

Changed

• The website and docs got another pass. The install section's two controls line up and its alternatives stopped shouting. The hero badge shows the released version. The interfaces are shown rather than described. The README says only what a README can and its TUI screenshots work again. The comparison with nmap and the other scanners is fairer in both directions. Docs pages carry structured data, and the docs shell picked up a Lighthouse gate and two fixes it found.

Security

• Three open advisories cleared. rustls 0.23.40 → 0.23.45 (RUSTSEC-2026-0285, medium), which was in 0.3.1's lockfile and so is in the binaries that release produced. The other two are the website's build dependencies rather than anything in a release artifact: adm-zip ≤0.6.0 (GHSA-vwc7-r8mq-g2x9 and GHSA-7q85-xj36-vmfc, high) and devalue <5.9.1 (GHSA-9rgm-9g3h-6x36, moderate).

The desktop app is redesigned around a denser diagnostic workspace, with reorderable tabs, right-click tab management and a refreshed icon. Port scans return richer status data on every interface, probe concurrency is configurable everywhere, and the website and docs were rebuilt alongside.

The headline is the desktop app. Everything listed under it is new to anyone upgrading from 0.2.6. The old dashboard-style GUI is gone, and what replaces it is a different application rather than a revision of that one. The rest is additive work in the core, CLI and MCP server, and a long tail of fixes. The recurring theme in that tail is code that reported success while doing nothing.

Added

• The desktop app is a diagnostic workspace. NetsCLI Desktop replaces the earlier dashboard with a native-like shell built around operation tabs. It runs the same netscli-core operations as the CLI, TUI and MCP server, so its results match the rest of the tool. What it does:
  • Operation tabs you can reorder by drag or keyboard, and close individually, to either side, or all at once from a right-click menu.
  • Sortable and filterable result tables, row detail panes, and a preview of the CLI command each run is equivalent to.
  • Save a whole workspace of results to a file and reopen it later. Export CSV or JSON, whole or selection-only. Exported cells are escaped against spreadsheet formula injection, since banners and hostnames are chosen by the scanned host.
  • Light and dark themes, and a settings dialog covering probe concurrency, default and traffic interfaces, IPv4/IPv6 display preference, history and save behaviour, and notification preferences.
  • Full keyboard operation, and a refreshed icon matching the site's brand.
• Clear the ARP table, then discover. A chevron beside Run offers a cache-flushing variant on discover, sweep and the ARP tab, the tools where a stale neighbour entry changes the answer. Clearing needs administrator rights. When it fails the run is suppressed rather than quietly returning the stale entries it was meant to drop.
• Richer port scan results across every interface. Port scans now report additive status and detail fields (open, closed, filtered, error, latency, banners, HTTP metadata, TLS metadata, and raw previews where available) while keeping the older open, port, service, and error fields intact for compatibility.
• User-configurable probe concurrency. The CLI and MCP server already accepted concurrency limits. The desktop app and TUI settings now expose the same control, so probes can be reduced on fragile networks or raised within the core safety cap.

Changed

• netscli scan --json now reports every port, not just the open ones. Filtering to open ports made "all closed", "all filtered" and "every probe errored" the same empty array, so a script could not tell a clean scan from a host that refused every probe. Each entry carries open and status, so callers that want only open ports can filter for them.
• The MCP server now scans only local networks by default. This is the one surface driven by a model rather than by the person at the keyboard, so the instruction to scan a third party can arrive from a web page or a file someone else wrote, and the packets leave from your machine and your IP. RFC1918, loopback, link-local and the carrier-grade NAT range overlay networks use are allowed. Set NETSCLI_MCP_ALLOW_PUBLIC_TARGETS=1 to reach past them.
• Scan results returned to a model are capped. The full probe response (raw) is no longer included and banners are truncated, both being bytes chosen by the scanned host.
• Tool failures are returned as MCP isError results rather than JSON-RPC errors, so a failed scan no longer reads to a client as a broken server.
• CLI and TUI scan output reflects the richer status data. Human output stays concise, but scanned ports can show closed, filtered and error states with latency where available, instead of only emphasising open ports.
• Windows install guidance prefers Winget for the desktop app. Winget's manifest review and installer hash verification make it the recommended Windows path. Direct GitHub Windows installers remain unsigned and may show warnings until code signing is added.
• Linux/macOS install docs clarified. mDNS is the default pure-Rust capability in published builds. Packet capture remains the optional workflow depending on libpcap/Npcap.
• The website and docs were rebuilt. A consistent shell, unified code and table styling, clearer search, and a layout swept across six widths and both themes. The brand accent moved from a teal-green that read blue in small text to one that reads green at any size.

Fixed

• Pinging your own machine no longer reports 100% loss. On Windows, ping 127.0.0.1 (and the machine's own LAN address) timed out while the system ping answered immediately. Raw ICMP sockets need administrator rights, so an ordinary run fell back to TCP probes on ports 80/443/22 and concluded a host was down when nothing answered. A Windows raw socket does not observe traffic to an address the host owns. Windows now sends echoes through the IP Helper API, which needs no privileges and reaches local addresses. Discover and sweep both start from a ping sweep, so both returned nothing for any range covering this host.
• IPv6 hosts can be pinged. ping ::1 reported total loss because IPv6 had no ICMP path at all and fell through to the same TCP probe. Windows now uses Icmp6SendEcho2.
• netscli arp --clear, --add and --delete no longer claim to have changed the table when they have not. On Windows these print "The requested operation requires elevation" and then exit 0, so checking the exit status alone reported success while nothing was touched. An ordinary run printed "ARP entry added for 192.0.2.77", and "ok": true in JSON, having done nothing. All three now share one check and exit non-zero with the reason. --clear also errors on platforms where it was never implemented, instead of reporting a cleared table.
• Discover reports devices the OS already knows about. Results merge probe replies with the neighbour table, so a device that answers ARP but not ICMP is no longer missing. Each host records whether it was found by probe or by neighbour, since a stale neighbour entry can outlive the device.
• Safety limits were enforced in Ops but not in the engines. The scan, sweep, discover and inspect engines are public API re-exported at the crate root, and called directly they applied no subnet, port or concurrency cap, and 0.0.0.0/0 collected 4,294,967,294 addresses into a Vec before sending a packet. Every engine now enforces its own limits. (#198)
• Safety limits that only one caller was applying. SweepEngine::sweep validates its port list instead of trusting the caller and silently returning "no open ports". mDNS browse duration, ping -c and packet captures given a packet count but no duration all gained the core-side ceiling they were documented to have.
• Port 0 was rejected only by the MCP surface. Now rejected everywhere. (#164)
• MCP server handled one request at a time. The read loop awaited each handler before parsing the next line, so a slow scan blocked every other request on the connection, including cancellation. Handlers now run concurrently under a semaphore. (#169)
• Reading the ARP table blocked a runtime worker. On Windows and macOS it shells out to arp and waits on the child process. Three callers invoked it straight from async code. With MCP handlers capped at 16 concurrent, sixteen of these could stall every worker, including the one reading stdin, so no further request could even be parsed. Moved to a blocking thread. (#196)
• Four ways an MCP client could wedge or kill the server: no overall request deadline, permits acquired after spawning rather than before, a single invalid UTF-8 byte on stdin terminating the process, and client disconnect cancelling nothing.
• The MCP server no longer says it returned everything while truncating. A capped result reported the byte count as its item count, so a 40,000-row scan that returned 11,518 rows said returned: 40000, total: 40000 beside truncated: true.
• Packet captures fetched as a background MCP job are bounded like every other result. get_pcap_capture_result was routed before the limits that strip and truncate remote text, so the one result made entirely of bytes off the wire was the one that skipped them.
• The concurrent packet-capture limit could be bypassed by calling the blocking capture tool, which never registered a job.
• discover_network with no arguments failed on a host whose interface carries a /8, because the substituted default exceeded the /16 cap.
• A database written by a newer netscli is refused rather than read. The version check treated a future schema as "already migrated", so an older build queried tables it had never seen.
• netscli trace no longer prints router-supplied hostnames unsanitised. Hop names come from PTR records controlled by whoever runs those routers, and this was the last plain-text output path without the terminal-safety pass every other one had.
• TUI mis-measured wide characters, so CJK and emoji in a remote-supplied hostname or banner pushed box borders out of alignment. (#173)
• Panic paths in the core, and silent corruption in the OUI generator. (#172)
• winget publishes the version number, not the tag. Both package manifests were passed the tag including its v, which the action only strips when the input is left empty. v0.2.2 through v0.2.6 are already in the public catalog that way, so winget show netscli reports a version this project never issued. Upgrades still work (winget normalises a leading v when comparing) and this release fixes what is displayed. The publish job now asserts MAJOR.MINOR.PATCH rather than trusting the strip, because winget-pkgs accepted all five without complaint and a merged manifest is permanent.
• AUR packages are published against a re-hashed asset. Both AUR jobs took the published .sha256 sidecar on trust rather than downloading the asset and hashing it, which is the circular check the release scripts exist to prevent. The other registries already did this correctly.
• The Windows installer verifies Npcap before running it. install.ps1 downloaded the Npcap installer from an overridable URL and launched it elevated with nothing checked. It now verifies the Authenticode signature and signer, and refuses to run an unsigned or unexpected binary.
• install.sh no longer claims success before installing libpcap. A user who asked for capture support could read "Installed successfully" and get a binary that cannot capture.

Website

• Small grey text was unreadable on card surfaces. The docs footer, built-with row and mobile section labels use --sl-color-gray-3, much of it at 12px. It had been raised once to clear 4.5:1 against the page background, but nothing checked it against the slightly darker card surface, where it sat at 4.44:1.
• Muted text, and several status colours, were below the readability bar. Eight colour tokens failed WCAG AA (4.5:1) against surfaces they are actually painted on. The mint accent failed as text on every light surface, down to 3.85:1. All 84 foreground/surface combinations now clear 4.5:1, and a check keeps them there.
• The install guide has the verification steps the landing page promises. It advertised checksums and Sigstore signatures and linked to a page with none of them on it.
• The site claimed packet capture in builds that do not ship it, and advertised a version that was never released. (#194, #208)
• The interface coverage table no longer conflates separate things. Rows that merged setup with doctor, or reading the ARP table with changing it, are split, and dashes that meant "not applicable" say which.
• The FAQ answers two questions people actually search for, picked from Search Console data rather than guesswork. Is there a netscan command, and does this replace nmap and have a terminal UI?

Installed GUI builds now identify themselves correctly, and the Windows title-bar controls work. Also completes the CLI/TUI refactors that make future interface changes easier to review and test.

Fixed

• The GUI's in-app version display was stuck at 0.1.0. A stale APP_VERSION constant in App.tsx powered both the bottom-bar version readout and the About dialog, but it never got bumped alongside package.json, tauri.conf.json, or the workspace Cargo.tomls. Caught by a Winget moderator on microsoft/winget-pkgs#368471: the v0.2.4 build correctly reported 0.2.4 to the Windows registry (Tauri pulls ProductVersion from tauri.conf.json), but users saw 0.1.0 in the GUI itself. Wired APP_VERSION to read package.json at build time via Vite's define so the in-app display auto-syncs every release going forward.
• GUI title-bar buttons (close, minimize, maximize) didn't work on Windows. Tauri 2's deny-by-default permission system requires explicit core:window:allow-close/minimize/maximize/unmaximize/start-dragging grants. The app was missing its capabilities config entirely. Added src-tauri/capabilities/main.json. (#62)

Closes a DNS resolver security advisory. Windows subnet detection is fixed, so discovery and sweep find real LAN hosts again, and package publishing now covers GUI installers.

Security

• hickory-resolver 0.24 → 0.26 closes RUSTSEC-2026-0119: CPU exhaustion during message encoding due to O(n²) name compression in hickory-proto. The DNS lookup tab and any inspect/discover that resolves hostnames are no longer reachable through the vulnerable encoding path. The 0.26 builder pattern (TokioResolver::builder_tokio()) replaces the deprecated TokioAsyncResolver::tokio constructor. See PR #55 for the source migration. The .cargo/audit.toml ignore added in #52 was removed once the bump landed.

Fixed

• GUI discover/sweep returned only a single host on Windows. Root cause. detect_default_ipv4_subnet iterated ipconfig::Adapter::prefixes() and grabbed the first IPv4 entry, but that list contains the host's own /32, broadcast /32, multicast /4, link-local /16, and the network /24. Windows reports the host /32 first, so the "subnet" was a single IP. New helper pick_ipv4_subnet_from_prefixes filters to network-shaped prefixes (length 1..=30, not multicast, not link-local) and truncates host bits, matching the Linux path. 5 unit tests added that run on every CI platform via cfg(any(windows, test)). (#59)
• The GUI dashboard's "Recent Scans" rendered with wrong colors / not as list rows. .history-item is a <button> (for keyboard accessibility) but the CSS didn't reset user-agent button styles. WebView2 on Windows applied Win32 chrome (color: ButtonText, centered text, content-fit width, system button font), breaking the inherit chain for child labels. Explicit reset added. (#59)

Changed

• Dependencies (all transitive, no API surface impact):
  • crossterm 0.27 → 0.28 + tui-textarea 0.4 → 0.7 had to land together, as tui-textarea 0.7 hardcodes crossterm = "0.28". (#58)
  • mdns-sd 0.13 → 0.19, adapting to the new ScopedIp::to_ip_addr() accessor in netscli-core/src/mdns.rs. (#58)
  • clap 4.5.60 → 4.6.1 (#43), pcap 1.3 → 2.4 (#45), clap_mangen 0.2.33 → 0.3.0 (#46), dialoguer 0.11.0 → 0.12.0 (#48), dirs 5.0.1 → 6.0.0 (#51), tokio 1.52.1 → 1.52.2 + clap_complete 4.6.2 → 4.6.3 (#57).
• ratatui 0.29 → 0.30 deferred: tui-textarea has no version yet that supports ratatui 0.30 (latest 0.7 still pins ratatui 0.29). Tracked via @dependabot ignore on the closed PR #49.

Added

• Release pipeline GUI automation. publish.yml extended with 4 parallel jobs that publish the GUI bundles to Homebrew Cask, Scoop extras (netscli-gui.json), Winget (fstubner.netscli.gui), and AUR (netscli-gui-bin) on every tagged release. (#54, #53, #56)

v0.2.3 built the GUI installers but never attached them. Publishing is repaired here, along with the AUR deploy action that was blocking Linux packages.

Fixed

• GUI bundle path in release.yml's GUI matrix was rooted at apps/netscli-gui/src-tauri/target/${TARGET}/release/bundle/. Cargo workspaces actually use the workspace-root target/ directory regardless of which subcrate's directory cargo was invoked from, so Tauri's bundle output lives at target/${TARGET}/release/bundle/. v0.2.3 built the .deb / .dmg / .msi correctly but the collect step found an empty bundle dir and skipped everything, and sigstore-sign then failed trying to sign nothing.
• AUR deploy action (KSXGitHub/github-actions-deploy-aur) was pinned to @v2.7.0 (April 2024), which has a bash: --command: invalid option regression in its container entrypoint. Bumped to @v4.1.3 (current stable, same input shape).

Notes

• CLI release shipped with 44 assets, sigstore-signed, on the v0.2.3 release page.
• Homebrew, Scoop, Winget, and crates.io all updated to 0.2.3.
• AUR is still on the previous version (failed to push).
• 0 GUI installers attached to v0.2.3 release.