Alibaba Cloud Linux 4 Agentic Edition (ANOLISA) uses cosh (cosh-ng) as the first entry after system login. It is an AI-native Linux terminal: bash/zsh works as usual. When encountering complex tasks, directly use natural language to ask the Agent to check or operate. This article introduces cosh's interaction design concepts, basic usage methods (Shell commands, natural language, approval mode, Agent Composer) and a complete slash command and shortcut key reference.
Interactive portal
The interactive shell of Alibaba Cloud Linux 4 Agentic Edition (ANOLISA) defaults to cosh (cosh-ng), which is the first entry after system login. cosh-ng targets different access objects and has different interaction characteristics:
-
For human users: It is an AI native terminal based on bash/zsh. Familiar commands are run directly, and complex tasks are described in natural language.
-
For agents: It is a standard interface for calling system capabilities, integrated through structured JSON (cosh-cli) and JSONL (cosh-core headless mode)
cosh interaction design
The interaction design of cosh follows "Shell first" and has the following main features:
-
Shell command : Familiar commands are executed directly using Shell syntax. Interactive programs (vim, top, ssh), pipes, redirection, task control, and Ctrl+C all work as usual in the foreground terminal.
-
natural language model : Use natural language to describe the troubleshooting task or operation goal, and the Agent will start the analysis. Constraints such as "check only" and "ask before modifying" can be written to control the agent's behavioral boundaries.
The two modes can be freely mixed: Shell input and Agent requests alternate in the same terminal.
The relationship between cosh and traditional shells
In Alibaba Cloud Linux 4 Agentic Edition (ANOLISA), cosh retains full bash/zsh compatibility - it does not replace the shell, but wraps a real front-end bash or zsh. On this basis, capabilities such as natural language understanding, Skills, approval cards, and resumable sessions are added, allowing both human users and agents to invoke system capabilities in a natural way.
Basic usage
After entering Alibaba Cloud Linux 4 Agentic Edition (ANOLISA), you will enter cosh directly. Shell commands can be executed as usual, or tasks can be described in natural language. Pass for first time use /auth Configure the large model backend.
Shell command
Just enter the command directly to execute it without any prefix or mode switch:
$ git status
$ top -bn1 | head -20
$ ls -alnatural language model
Simply describe your intention in Chinese or English, and the system will automatically understand and execute it:
> Check the memory usage of the current system
> Help me install nginx and configure it to start automatically at boot
> Help me deploy OpenClawThe system shell will convert the natural language into the corresponding system operation, and the plan will be displayed and wait for confirmation before execution (depending on the approval mode setting, see "Configuration Management - Approval Mode" for details).
Approval mode
pass /mode approval Set tool approval level:
# View current approval mode
> /mode approval
# Agent only explains and makes suggestions, but does not initiate tool calls.
> /mode approval recommend
# Read-only tools automatically approve, risky operations require confirmation
> /mode approval auto
# All tools are automatically approved (need to enter confirm to confirm the switch)
> /mode approval trust confirm
Safety tips:trust mode will skip all operation confirmations and is recommended to be used only in isolated test environments. Please use recommend or auto mode in production environment.
There is also a separate analysis mode /mode analysis, control the active analysis behavior of failed commands:
-
smart(Default): When the command fails or the system diagnostic output is abnormal, analysis suggestions are displayed after evaluation. -
auto: Only high-confidence failures automatically start Agent analysis -
manual: Turn off proactive analysis and failure insights
Use /agent to reference files and skills
run /agent Open Agent Composer and you can edit one-time requests: the first token can be a Skill, and subsequent @path Reference a file or directory within the current workspace:
/skill:repo-review inspect @Cargo.toml @src
@/etc/nginx/nginx.conf Help me check if there is any problem with this configuration
Things to note:
-
@pathIt must point to a file or directory in the workspace. Absolute paths and parent directory jumps will be rejected. -
Accept up to 16 valid citations per submission
-
Multi-line drafts can also be used
??Enter directly at the beginning,Shift+Enternewline -
Plain text requests without specifying a skill or reference are also valid
cosh command reference
Overview of command categories
|
prefix |
type |
Function |
Example |
|
|
slash command |
cosh control command |
|
|
|
Agent request |
Multi-line draft initiates Agent task |
|
|
(no prefix) |
Shell command |
Execute directly in the foreground Shell |
|
List of slash commands
Session management
|
Order |
illustrate |
usage |
|
|
Manage Agent Sessions |
|
|
|
Restore previous session ( |
|
|
|
New Session ( |
|
|
|
Clear local shell status |
|
Tool and model management
|
Order |
illustrate |
usage |
|
|
Select or update provider certification |
|
|
|
View or modify approval/analysis modes |
|
|
|
List/View/Start/Stop Skills |
|
|
|
Manage MCP server |
|
|
|
Manage expansion packs and settings |
|
|
|
Check Hook status and trust |
|
|
|
Manage local input suggestions |
|
|
|
Open Agent Composer to edit one-time requests |
|
Interface and configuration
|
Order |
illustrate |
usage |
|
|
View/set interface language |
|
|
|
Display model and tool statistics |
|
|
|
Display version, provider, runtime status |
|
Information and help
|
Order |
illustrate |
usage |
|
|
Show help information |
|
|
|
Display version information ( |
|
|
|
Run local health check |
|
|
|
Analyze the latest failed command |
|
|
|
Cancel the current Agent task |
|
|
|
View approval/activity details |
|
|
|
Audit status and export |
|
shortcut key
|
shortcut key |
Function |
illustrate |
|
Shift+Enter |
Insert newline |
Multi-line draft input |
|
Tab |
Complete |
Completion in Agent Composer |
|
Esc |
Cancel |
Cancel a card, panel or draft |
|
Ctrl+C |
interrupt |
Cancel the current command or Agent request |
|
Up arrow key |
Enter history |
Shell input and slash command history |