Cero 2.2.3 for Linux

Cero 2.2.3 for Linux

A fix release for 2.2.1. Install it over 2.2.0 or 2.2.1; your settings, workflows and keys carry over.
There is no 2.2.2 release: that number was a local test build, and everything in it is included here.

Downloads

• Arch, Manjaro, EndeavourOS: cero-terminal-bin-2.2.3-1-x86_64.pkg.tar.zst
• Ubuntu 22.04 or later, Debian 12, Mint, Pop!_OS: Cero_2.2.3_amd64.deb
• Fedora 38 or later, openSUSE: Cero-2.2.3-1.x86_64.rpm
• Any other x86_64 distribution: Cero_2.2.3_amd64.AppImage

Checksums: SHA256SUMS.txt.

What is fixed

You saw	Now
A request with several actions ("open folder gitBrains in VS Code ... and also open zen browser") did the first and silently dropped the rest	Each action runs as its own step, with the same folder search and installed-app lookup as when asked alone, and every step is reported.
"Create a Next.js project named x and open it in VS Code" stopped or did nothing, because the project tool waited for answers no one could give	The project is created with Next.js's recommended defaults, then opened; a command that still asks questions opens in a terminal pane for you to answer, and the remaining steps continue when it finishes.
Approval requests appeared sometimes as a centred dialog and sometimes as a small card	Every approval is the same card in the bottom-right corner. It never takes the keyboard, so a prompt you are typing is never interrupted.
The prompt cursor could become invisible after switching tabs, closing Settings or the window losing focus	The cursor is redrawn whenever the prompt gets focus or the window is shown again, and a lost graphics context falls back to the standard renderer.
Right arrow did not accept the grey suggestion; only Tab did	Right at the end of the line accepts it exactly like Tab, once. Turn this off in Settings, Terminal, "Accept suggestion with Right arrow".
Command history (Ctrl+R) kept only the first commands of a session	Every command and > request is saved again. The cause was the invisible markers newer systemd versions add around each command, which hid the shell prompt from Cero.

Changed behaviour

• Approvals are answered with a click on Run or Deny; Enter and Esc no longer answer them, because those keys belong to the prompt you are typing.
• Right arrow now accepts a visible suggestion by default (see above).

Not verified yet: these packages were built on Arch Linux. The fixes are covered by automated tests, and the
multi-action and Next.js requests were run end to end with the agent; the cursor, approval card, Right arrow and
history changes have not yet been checked by hand in the installed app. The deb, rpm and AppImage were built but
not run on their own distributions. Please report problems.

Cero 2.2.1 for Linux

Cero 2.2.1 for Linux

A fix release for 2.2.0. Install it over 2.2.0; your settings, workflows and keys carry over.

Downloads

• Arch, Manjaro, EndeavourOS: cero-terminal-bin-2.2.1-1-x86_64.pkg.tar.zst
• Ubuntu 22.04 or later, Debian 12, Mint, Pop!_OS: Cero_2.2.1_amd64.deb
• Fedora 38 or later, openSUSE: Cero-2.2.1-1.x86_64.rpm
• Any other x86_64 distribution: Cero_2.2.1_amd64.AppImage

Checksums: SHA256SUMS.txt.

What is fixed

You saw	Now
Ctrl+C and the prompt queue could disagree about what was running	Ctrl+C and the queue always keep one consistent state; the queue panel is more compact, and its "not saved across restarts" note is a tooltip on the title.
An approval request could take over a prompt you were writing	An approval request never takes over a prompt being written.
Arrow keys and paste could commit or duplicate suggestion (ghost) text	Arrows and paste can no longer commit or duplicate suggestion text.
Esc handling in Settings was not reliably tested	Esc closes Settings through one tested handler that consumes the key.
The status bar and restart did not always follow the model in use	The status bar and restart restore follow the real active model.
"Open folder X" failed for some ways of naming a folder or app	More ways to name an existing folder or app are understood, checked against a real directory tree.
A saved workflow was not always found again by name, and some "save as workflow" phrasings were ignored	A saved workflow is found again by name, and more save phrasings are understood.
The built-in model and API models could be held to different standards	Both are held to the same decision call; grammar whitespace is bounded.

Docs for Ctrl+C and queue behaviour, ghost-text accept, Esc in Settings, and saving workflows by sentence were updated.

Not verified yet: these packages were built on Arch Linux. Only the Arch package has been installed and run here; the deb, rpm and AppImage were built but not run on their own distributions. Please report problems.

Cero 2.2.0 for Windows

Cero 2.2.0 for Windows

Downloads (built by CI from the release/windows branch)

• the .msi installer for Windows 10 (1803 or later) and Windows 11
• the setup .exe (per user)

The installers are not code-signed, so SmartScreen shows "Windows protected your PC": choose "More info", then "Run anyway". Not run on a Windows PC for this release; the .flow registration and icon are written but unchecked.

What changed (answers to twelve findings from the first Linux test)

#	You saw	Now
1	"...and save this as a workflow" did nothing	Run the task, then save what really ran as a .flow. Every request ends with Saved workflow "x" (N steps) to <path> or Not saved: <reason>.
2	Arrow keys moved the grey suggestion and appended it	Left, Right, Up, Down, Home and End only move the cursor. Tab, or Right at the end of the line, accepts. A setting turns Right-to-accept off.
3	Queued prompts could not be seen	A queue panel and /queue commands: view, remove, reorder, clear.
4	A prompt worked with an API model, not the built-in one	One shared place sets sampling for every provider; one tolerant parser reads every model's answer; after two failed checks the built-in model offers to retry that one request on your API model (it asks first).
5	Esc did not close Settings	It does, and focus returns to the terminal.
6	"Open folder gitBrains in VS Code (inside /padhai_in_linux/Projects/)" made files and opened a second window	The real folder is found (the place you name, this folder, usual project folders, then home), opened in one editor window, and nothing is created unless you say "create it".
7	Wrong folder or app names were not handled	Spelling and case are forgiven, typos ask "Did you mean ...?", two matches ask which, a missing folder or app is reported. Answers are remembered (forget gitbrains clears one). "opn firefox" is read as "open firefox".
8	The chosen provider and model were forgotten	Remembered across restarts; if the provider is not reachable yet it is retried and the reason is shown, never silently replaced by the built-in model.
9	The status bar said "AI: Off" for a working API or local model	Shows AI: local, AI: API or AI: Ollama; "Off" is only for a stopped built-in engine.
10	Ctrl+C did not stop a task	Ctrl+C stops the running request and its command; a second press forces it.
11	The built-in model was right sometimes, random the next time	Decisions are deterministic (temperature 0, one choice, fixed seed, no prompt cache); the prompt is fitted to the context; every action is checked before it runs. Measured numbers: docs/MODEL_RELIABILITY.md.
12	.flow files were saved as .json; no icon; not tied to the app	Only .flow is written (old .json flows are migrated, originals kept as .json.bak); a document icon; registered on macOS, Windows, Linux (deb, rpm, Arch, Flatpak) and by the AppImage on first run; opening one while Cero is running joins that window.

Also: API keys are stored in the macOS Keychain or Windows Credential Manager (a private 0600 file on Linux)
instead of browser storage, after the new copy is read back; "list my workflows"; "what do you remember about X".

Not verified yet: the Linux packages, the AppImage registration, the Windows installer registration and the macOS
Finder icon were built from configuration and unit-tested but not run on those systems.

Teaching Cero. Cero never learns by watching you. /learn teaches the request just above it (only if it worked), /learn <request> -> <command> teaches a pair, /forget removes one, and /learning on|off controls whether it keeps a record of its own failed requests (off by default). See LEARNING.md.

Smaller model option. A 0.5B model (about 470 MB) joins the 1.5B, 3B and 4B choices for slow connections and old machines. Compressing a model file does not shrink it much (measured: 3.4%), so a smaller model is the way to a smaller download. Interrupted downloads resume.

Renamed. Sentinel Terminal is now Cero. On first launch ~/.sentinel moves to ~/.cero, saved settings and stored API keys carry over, and nothing has to be downloaded again. The sentinel command is now cero.

Checksums: SHA256SUMS.txt.

Cero 2.2.0 for macOS

Cero 2.2.0 for macOS

Downloads

• Cero_2.2.0_aarch64.dmg: Apple Silicon (M1 and later). Built and run on macOS 26.

First launch. The app is ad-hoc signed, not notarized. macOS blocks the first launch: open System Settings, then Privacy & Security, and choose "Open Anyway". Drag the new app over an old Sentinel Terminal or Cero in Applications. .flow files open with the copy that is installed.

What changed (answers to twelve findings from the first Linux test)

#	You saw	Now
1	"...and save this as a workflow" did nothing	Run the task, then save what really ran as a .flow. Every request ends with Saved workflow "x" (N steps) to <path> or Not saved: <reason>.
2	Arrow keys moved the grey suggestion and appended it	Left, Right, Up, Down, Home and End only move the cursor. Tab, or Right at the end of the line, accepts. A setting turns Right-to-accept off.
3	Queued prompts could not be seen	A queue panel and /queue commands: view, remove, reorder, clear.
4	A prompt worked with an API model, not the built-in one	One shared place sets sampling for every provider; one tolerant parser reads every model's answer; after two failed checks the built-in model offers to retry that one request on your API model (it asks first).
5	Esc did not close Settings	It does, and focus returns to the terminal.
6	"Open folder gitBrains in VS Code (inside /padhai_in_linux/Projects/)" made files and opened a second window	The real folder is found (the place you name, this folder, usual project folders, then home), opened in one editor window, and nothing is created unless you say "create it".
7	Wrong folder or app names were not handled	Spelling and case are forgiven, typos ask "Did you mean ...?", two matches ask which, a missing folder or app is reported. Answers are remembered (forget gitbrains clears one). "opn firefox" is read as "open firefox".
8	The chosen provider and model were forgotten	Remembered across restarts; if the provider is not reachable yet it is retried and the reason is shown, never silently replaced by the built-in model.
9	The status bar said "AI: Off" for a working API or local model	Shows AI: local, AI: API or AI: Ollama; "Off" is only for a stopped built-in engine.
10	Ctrl+C did not stop a task	Ctrl+C stops the running request and its command; a second press forces it.
11	The built-in model was right sometimes, random the next time	Decisions are deterministic (temperature 0, one choice, fixed seed, no prompt cache); the prompt is fitted to the context; every action is checked before it runs. Measured numbers: docs/MODEL_RELIABILITY.md.
12	.flow files were saved as .json; no icon; not tied to the app	Only .flow is written (old .json flows are migrated, originals kept as .json.bak); a document icon; registered on macOS, Windows, Linux (deb, rpm, Arch, Flatpak) and by the AppImage on first run; opening one while Cero is running joins that window.

Also: API keys are stored in the macOS Keychain or Windows Credential Manager (a private 0600 file on Linux)
instead of browser storage, after the new copy is read back; "list my workflows"; "what do you remember about X".

Not verified yet: the Linux packages, the AppImage registration, the Windows installer registration and the macOS
Finder icon were built from configuration and unit-tested but not run on those systems.

Teaching Cero. Cero never learns by watching you. /learn teaches the request just above it (only if it worked), /learn <request> -> <command> teaches a pair, /forget removes one, and /learning on|off controls whether it keeps a record of its own failed requests (off by default). See LEARNING.md.

Smaller model option. A 0.5B model (about 470 MB) joins the 1.5B, 3B and 4B choices for slow connections and old machines. Compressing a model file does not shrink it much (measured: 3.4%), so a smaller model is the way to a smaller download. Interrupted downloads resume.

Renamed. Sentinel Terminal is now Cero. On first launch ~/.sentinel moves to ~/.cero, saved settings and stored API keys carry over, and nothing has to be downloaded again. The sentinel command is now cero.

Checksums: SHA256SUMS.txt.

Cero 2.2.0 for Linux

Cero 2.2.0 for Linux

Downloads (built by CI from the release/linux branch)

• Arch, Manjaro, EndeavourOS: the .pkg.tar.zst
• Ubuntu 22.04 or later, Debian 12, Mint, Pop!_OS: the .deb
• Fedora 38 or later, openSUSE: the .rpm
• Any other x86_64 distribution: the .AppImage (it offers to register the .flow file type on first run)

These packages were not run on a Linux desktop for this release. The .flow file type, icon and AppImage registration are implemented and unit-tested, not yet checked on real systems. See LINUX_TEST_REPORT.md. Please report problems.

What changed (answers to twelve findings from the first Linux test)

#	You saw	Now
1	"...and save this as a workflow" did nothing	Run the task, then save what really ran as a .flow. Every request ends with Saved workflow "x" (N steps) to <path> or Not saved: <reason>.
2	Arrow keys moved the grey suggestion and appended it	Left, Right, Up, Down, Home and End only move the cursor. Tab, or Right at the end of the line, accepts. A setting turns Right-to-accept off.
3	Queued prompts could not be seen	A queue panel and /queue commands: view, remove, reorder, clear.
4	A prompt worked with an API model, not the built-in one	One shared place sets sampling for every provider; one tolerant parser reads every model's answer; after two failed checks the built-in model offers to retry that one request on your API model (it asks first).
5	Esc did not close Settings	It does, and focus returns to the terminal.
6	"Open folder gitBrains in VS Code (inside /padhai_in_linux/Projects/)" made files and opened a second window	The real folder is found (the place you name, this folder, usual project folders, then home), opened in one editor window, and nothing is created unless you say "create it".
7	Wrong folder or app names were not handled	Spelling and case are forgiven, typos ask "Did you mean ...?", two matches ask which, a missing folder or app is reported. Answers are remembered (forget gitbrains clears one). "opn firefox" is read as "open firefox".
8	The chosen provider and model were forgotten	Remembered across restarts; if the provider is not reachable yet it is retried and the reason is shown, never silently replaced by the built-in model.
9	The status bar said "AI: Off" for a working API or local model	Shows AI: local, AI: API or AI: Ollama; "Off" is only for a stopped built-in engine.
10	Ctrl+C did not stop a task	Ctrl+C stops the running request and its command; a second press forces it.
11	The built-in model was right sometimes, random the next time	Decisions are deterministic (temperature 0, one choice, fixed seed, no prompt cache); the prompt is fitted to the context; every action is checked before it runs. Measured numbers: docs/MODEL_RELIABILITY.md.
12	.flow files were saved as .json; no icon; not tied to the app	Only .flow is written (old .json flows are migrated, originals kept as .json.bak); a document icon; registered on macOS, Windows, Linux (deb, rpm, Arch, Flatpak) and by the AppImage on first run; opening one while Cero is running joins that window.

Also: API keys are stored in the macOS Keychain or Windows Credential Manager (a private 0600 file on Linux)
instead of browser storage, after the new copy is read back; "list my workflows"; "what do you remember about X".

Not verified yet: the Linux packages, the AppImage registration, the Windows installer registration and the macOS
Finder icon were built from configuration and unit-tested but not run on those systems.

Teaching Cero. Cero never learns by watching you. /learn teaches the request just above it (only if it worked), /learn <request> -> <command> teaches a pair, /forget removes one, and /learning on|off controls whether it keeps a record of its own failed requests (off by default). See LEARNING.md.

Smaller model option. A 0.5B model (about 470 MB) joins the 1.5B, 3B and 4B choices for slow connections and old machines. Compressing a model file does not shrink it much (measured: 3.4%), so a smaller model is the way to a smaller download. Interrupted downloads resume.

Renamed. Sentinel Terminal is now Cero. On first launch ~/.sentinel moves to ~/.cero, saved settings and stored API keys carry over, and nothing has to be downloaded again. The sentinel command is now cero.

Checksums: SHA256SUMS.txt.

Sentinel Terminal 2.1.0 for Windows

Cero was called Sentinel Terminal until 2.2.0. This is an older release; the newest builds are on the Releases page (v2.2.0-macos, v2.2.0-linux, v2.2.0-windows). Install the new one over this one: your data and settings carry over.

Sentinel Terminal 2.1.0 for Windows

Downloads

• Sentinel.Terminal_2.1.0_x64_en-US.msi: installer for Windows 10 (1803 or later) and Windows 11.
• Sentinel.Terminal_2.1.0_x64-setup.exe: the same app as a per-user setup program.

The installers are not code-signed, so SmartScreen shows "Windows protected your PC": choose "More info", then "Run anyway".

Open a .flow file and it runs. A .flow file (from a tutorial, a teammate or you) lists what to install, run and open. Sentinel picks the commands for this OS.

• A flow that only opens apps and links (Chrome, YouTube, VS Code) runs without showing the terminal.
• A flow that installs or runs things opens the terminal, lists every command, waits for you to click Run, then types each step so you can see output and answer prompts.

See FLOW_FILES.md and the example flows.

New in 2.1.0

• Make a workflow by asking: "make me a workflow that ..." writes a .flow file from the steps you list, asks where to keep it (Desktop, this folder, Sentinel workflows, or a path) and never replaces an existing file. Steps it does not understand are named and left out, never guessed.
• Close a port by number: "close port 8765" finds what is listening on exactly that port, shows its name and PID, asks, stops it normally and checks that the port is free.
• Wi-Fi is exact: turning Wi-Fi on or off and joining a network use the exact command, shown before it runs. The built-in model is no longer involved (it once invented a password).
• Fixed: a workflow with relative folders stopped at step 3; opening Sentinel on a folder named $(...) could run text from the name; several drivers quoted names unsafely.
• Security: Sentinel no longer asks for your login password (high-risk commands need a click on Run). ~/.sentinel is readable by you only. See SECURITY.md.

What works on Windows in this release

• The terminal (PowerShell), tabs and splits, and talking to other terminals.
• AI requests through the built-in model, Ollama or a cloud key. The built-in engine now downloads and unpacks natively on Windows (Vulkan build when vulkan-1.dll is present, else CPU), with no console windows popping up.
• The agent's shell commands (run in PowerShell), multi-step chains, workflows (including .flow files) and the app functions by request.
• No separate title bar: minimize, maximize and close sit at the right of the tab bar, and the tab bar moves the window.
• Quitting an app by name ("quit notepad") checks the running apps first and asks with the exact name.
• System settings by request, with Windows commands:
  • Wi-Fi and Bluetooth: turn on or off (the same radio switch as the quick settings), status, nearby networks, joining a saved network, paired devices.
  • Brightness (built-in displays), volume and mute, dark mode.
  • Battery, lock screen, and any Settings page ("open display settings").
• Saying only a topic ("bluetooth", "brightness") lists what Sentinel can do for it on Windows ("Did you mean: turn Bluetooth on, ..."). Typing > requests completes them.
• When a switch is blocked (no radio, policy, external monitor without brightness control), Sentinel says why and opens the matching Settings page.

Known gaps on Windows.

• Some system-information tools still use Unix commands.
• The error watcher's checks.
• ROS 2 detection.

These installers are built by CI from the release/windows branch. They were not run on a Windows PC for this release; please report problems.

Checksums: SHA256SUMS.txt.

Sentinel Terminal 2.1.0 for macOS

Cero was called Sentinel Terminal until 2.2.0. This is an older release; the newest builds are on the Releases page (v2.2.0-macos, v2.2.0-linux, v2.2.0-windows). Install the new one over this one: your data and settings carry over.

Sentinel Terminal 2.1.0 for macOS

Downloads

• Sentinel.Terminal_2.1.0_aarch64.dmg: Apple Silicon (M1 and later). Built and tested on macOS 26.
• Sentinel.Terminal_2.1.0_x64.dmg: Intel Macs. Built by CI; not yet tested on Intel hardware.

First launch. The app is ad-hoc signed, not notarized. macOS blocks the first launch: open System Settings, then Privacy & Security, and choose "Open Anyway". On a new machine the app downloads the llama.cpp engine (about 12 MB) and the Qwen2.5-Coder 3B model (about 2 GB) from Settings, AI, Built-in model. Or use Ollama or a cloud key instead.

Updating. Drag the new app over the old one in Applications and choose Replace. macOS opens .flow files with the copy that is installed, so an older Sentinel keeps handling them until it is replaced.

Open a .flow file and it runs. A .flow file (from a tutorial, a teammate or you) lists what to install, run and open. Sentinel picks the commands for this OS.

• A flow that only opens apps and links (Chrome, YouTube, VS Code) runs without showing the terminal.
• A flow that installs or runs things opens the terminal, lists every command, waits for you to click Run, then types each step so you can see output and answer prompts.

See FLOW_FILES.md and the example flows.

New in 2.1.0

• Make a workflow by asking: "make me a workflow that ..." writes a .flow file from the steps you list, asks where to keep it (Desktop, this folder, Sentinel workflows, or a path) and never replaces an existing file. Steps it does not understand are named and left out, never guessed.
• Close a port by number: "close port 8765" finds what is listening on exactly that port, shows its name and PID, asks, stops it normally and checks that the port is free.
• Wi-Fi is exact: turning Wi-Fi on or off and joining a network use the exact command, shown before it runs. The built-in model is no longer involved (it once invented a password).
• Fixed: a workflow with relative folders stopped at step 3; opening Sentinel on a folder named $(...) could run text from the name; several drivers quoted names unsafely.
• Security: Sentinel no longer asks for your login password (high-risk commands need a click on Run). ~/.sentinel is readable by you only. See SECURITY.md.
• System settings by request, with the right commands for each OS: "turn wifi off", "set brightness to 60", "mute", "turn on dark mode", "show paired bluetooth devices", "open sound settings". Saying only a topic ("bluetooth") lists what Sentinel can do for it on this OS. Changes ask first.
• Ask for the app's own functions in plain language: "open settings", "show my command history", "switch to zen mode", "search the terminal for ERROR", "go to tab 2", "rename this tab to api".
• Talk to other terminals: "open a new tab here and run npm run dev", "run ls in tab 2", "stop the server", "what's running in my terminals". The model reads other terminals only when you ask, and only what changed.
• Ready-made, tested commands for tricky requests: CSV totals and conversion, syntax checks, folder sizes, text replacement with backups, safe renames, port owners and git status of a named folder.
• "fix buggy.py so it runs": runs the file, shows the change for approval (original kept as .bak), then re-runs it.
• "why is npm test failing in api?": runs the command there and explains the failure from the code it points at.
• Security: commands that read private keys or credential files always ask first, and keys and tokens are masked before anything reaches a model.
• Quitting an app by name ("quit textedit", "terminate the claude application") checks the running apps first, matches the name in any case and asks with the exact app.
• No separate title bar: the window buttons sit in the tab bar.
• Command history (Ctrl+R) lists commands you actually ran, and never records a password typed into a running program.

Checksums: SHA256SUMS.txt (Intel, from CI) and SHA256SUMS-apple-silicon.txt.

Sentinel Terminal 2.1.0 for Linux

Cero was called Sentinel Terminal until 2.2.0. This is an older release; the newest builds are on the Releases page (v2.2.0-macos, v2.2.0-linux, v2.2.0-windows). Install the new one over this one: your data and settings carry over.

Sentinel Terminal 2.1.0 for Linux

Downloads

File	For
sentinel-terminal-bin-2.1.0-1-x86_64.pkg.tar.zst	Arch Linux, Manjaro, EndeavourOS: sudo pacman -U sentinel-terminal-bin-2.1.0-1-x86_64.pkg.tar.zst (built on Arch; also installs the sentinel and sentinel-shell launchers and a Dolphin "Open in Sentinel" menu)
Sentinel.Terminal_2.1.0_amd64.deb	Ubuntu 22.04 or later, Debian 12, Mint, Pop!_OS: sudo apt install ./Sentinel.Terminal_2.1.0_amd64.deb
Sentinel.Terminal-2.1.0-1.x86_64.rpm	Fedora 38 or later, openSUSE: sudo dnf install ./Sentinel.Terminal-2.1.0-1.x86_64.rpm
Sentinel.Terminal_2.1.0_amd64.AppImage	Any other x86_64 distribution: chmod +x it and run it

Double-clicking a .flow file opens it with Sentinel (the packages register the file type). All packages need WebKitGTK 4.1 (installed automatically by pacman, apt and dnf). The built-in model downloads its engine (the Vulkan build when libvulkan.so.1 is present, else the CPU build) and the 2 GB model from Settings, AI, Built-in model. Ollama and cloud keys work too.

These packages are built by CI from the release/linux branch. They were not run on a Linux desktop for this release; please report problems.

Open a .flow file and it runs. A .flow file (from a tutorial, a teammate or you) lists what to install, run and open. Sentinel picks the commands for this OS.

• A flow that only opens apps and links (Chrome, YouTube, VS Code) runs without showing the terminal.
• A flow that installs or runs things opens the terminal, lists every command, waits for you to click Run, then types each step so you can see output and answer prompts.

See FLOW_FILES.md and the example flows.

New in 2.1.0 (all of it applies on Linux):

• Make a workflow by asking: "make me a workflow that ..." writes a .flow file from the steps you list, asks where to keep it (Desktop, this folder, Sentinel workflows, or a path) and never replaces an existing file. Steps it does not understand are named and left out, never guessed.
• Close a port by number: "close port 8765" finds what is listening on exactly that port, shows its name and PID, asks, stops it normally and checks that the port is free.
• Wi-Fi is exact: turning Wi-Fi on or off and joining a network use the exact command, shown before it runs. The built-in model is no longer involved (it once invented a password).
• Fixed: a workflow with relative folders stopped at step 3; opening Sentinel on a folder named $(...) could run text from the name; several drivers quoted names unsafely.
• Security: Sentinel no longer asks for your login password (high-risk commands need a click on Run). ~/.sentinel is readable by you only. See SECURITY.md.

See also the macOS release notes for the rest; everything there applies on Linux. The window has no separate title bar: minimize, maximize and close sit at the right of the tab bar. ROS 2 pipelines ("run the ros2 talker and listener demo") open each node in its own terminal when ROS 2 is installed.

Checksums: SHA256SUMS.txt.

Sentinel Terminal v2.0.0 — Production Release

Cero was called Sentinel Terminal until 2.2.0. This is an older release; the newest builds are on the Releases page (v2.2.0-macos, v2.2.0-linux, v2.2.0-windows). Install the new one over this one: your data and settings carry over.

🚀 Sentinel Terminal v2.0.0 is officially released!

This major production release introduces the Self-Evolving Reinforcement Learning (SERL) engine, unconditional refusal interception, autonomous self-healing diagnostics, and 100% offline Metal-accelerated inference.

🌟 Key Highlights in v2.0.0

• Decoupled Refusal Interception: Stops canned generic LLM evasions ("as an AI language model I cannot...") and forces active command execution with multi-turn referential context resolution.
• Autonomous Self-Healing Diagnostics: Stderr and non-zero exit code feedback loop with 3-strike autonomous auto-remediation.
• Speculative Shadow-PTY Simulation Engine: Sub-millisecond candidate rollout (=3$) in an ephemeral sandbox with cross-platform syntax translation (Linux vs macOS vs Windows).
• On-Device DPO Preference Engine: Automated capture of chosen/rejected preference pairs for continuous fine-tuning.
• 101 Canonical Tools: Full desktop orchestration across Wi-Fi, Bluetooth, port inspection, DHCP lease renewal, Spotlight search, and IDE launchers.
• Google Colab Training Package: Complete turnkey ChatML SFT/DPO export and notebook for free T4 GPUs.
• Hardware-Accelerated UI: Glassmorphic xterm.js WebGL terminal with real-time newline normalization preventing staircase rendering.
• 983 Automated Tests: 100% passing across 155 test suites.

📥 Installation (macOS)

1. Download Sentinel_Terminal_v2.0.0_macOS.dmg below.
2. Drag Sentinel Terminal.app into your /Applications folder.
3. Open Sentinel Terminal and type > followed by any instruction!

Sentinel v1.0.0 - Production Release

Cero was called Sentinel Terminal until 2.2.0. This is an older release; the newest builds are on the Releases page (v2.2.0-macos, v2.2.0-linux, v2.2.0-windows). Install the new one over this one: your data and settings carry over.

🚀 Sentinel Terminal v1.0.0 is officially live!

This marks the first production build of Sentinel, bridging the gap between traditional shell precision and offline, natural language desktop automation.

✨ What's Included

• Contextual Conversation Memory: Sentinel now natively tracks your last 10 commands so you can follow-up naturally!
• 100% Offline Embedded AI: True zero-trust local inference on Apple Silicon.
• Glassmorphic Multi-pane Workspace: Vertical and horizontal splits with customizable themes.
• 97+ Canonical Native Tools: From filesystem navigation to docker container inspection, Sentinel controls it all.

📥 Installation

1. Download Sentinel_Terminal_v0.1.0_macOS.dmg below.
2. Open it and drag Sentinel into your /Applications/ folder.
3. Enjoy an elevated terminal experience!