Alibaba Cloud Linux 4 Agentic Edition (ANOLISA) Product Overview

Cosh

cosh is the default interactive shell of Alibaba Cloud Linux 4 Agentic Edition (ANOLISA), replacing bash as the first entry after system login.

The core design concept of cosh is "native Shell host + Agent enhancement". It directly hosts a real bash or zsh session. Shell commands are executed according to native semantics without prefixes and no need to fall back to an independent bash environment. When troubleshooting or processing larger tasks, the user directly describes the intention in Chinese or English, and the Agent carries the nearest terminal context to take over the execution; the slash command at the beginning of the line is used to control the session itself. The three types of input are automatically diverted at the same prompt, and there is no need to switch environments during the whole process.

On the basis of retaining complete bash/zsh compatibility, cosh adds capabilities such as natural language understanding, Skill calling, MCP tool integration, and multi-level approval control. It further consolidates system operation and maintenance operations into a cross-distribution structured command interface cosh-cli, and opens the Agent runtime into an interfaceless integration interface cosh-core. cosh abstracts complex system-level capabilities into natural language interaction and integrates OS Skills instructions, which lowers the threshold for using the operating system and enables both human users and Agent agents to drive the operating system to complete tasks in a simple way.

cosh mainly provides the following capabilities:

OS Skills

OS Skills is an operating system user manual written by ANOLISA for AI Agent.

Traditional operating system documentation is geared toward human users and relies on natural language descriptions, screenshot examples, and underlying industry consensus. When the Agent reads this type of document, it needs to consume a large amount of Tokens to understand it. OS Skills instructions reorganize operating system knowledge into a structured format that Agent can directly understand and execute - SKILL. Agent no longer needs to "read the document and then operate", but "can do it as soon as it is read".

The OS Skills manual covers two major areas:

After receiving the user's intention, the Agent automatically matches the corresponding Skill and executes it without manually specifying the calling path.

AgentSecCore

AgentSecCore is an operating system-level security kernel for the AI ​​Agent running platform. In the context of AI Agents gradually gaining operating system-level execution capabilities (including file reading and writing, network access, process management, etc.), traditional application security boundaries are no longer applicable. AgentSecCore builds an in-depth defense system for Agents from the OS level to ensure that Agents can run safely in a controlled, auditable, and minimally privileged environment.

AgentSecCore has built a three-layer defense system in depth around the two pillars of "intent security" and "system-level security" - even if the previous layer is breached, subsequent layers can still contain it. The architecture from bottom to top is:

AgentSight

AgentSight is an operating system-level observable component for the AI ​​Agent running platform. It solves the problem of Token consumption during Agent running far exceeding expectations and users lacking perception and traceability methods. It realizes fine-grained data collection and correlation analysis of the entire Agent operation link without intruding business logic.

AgentSight mainly provides the following three capabilities:

• Token consumption analysis: Comprehensive measurement and attribution of Token consumption during Agent operation. Supports flexible query by time period or the last N hours, supports splitting consumption sources by multiple dimensions such as agents, tasks, roles, etc., and the analysis granularity can be accurate to a single LLM call.

• behavioral audit: Full-link recording of Agent’s LLM calls and process execution behaviors. It completely retains metadata such as provider and model version of each call, synchronously captures process command line parameters, and supports filtering and visual summary statistics by time, session and other dimensions.

• Dashboard visualization:Provides a web visual interface that supports direct access through a local browser after remote deployment. It can view token consumption trends in real time, monitor Agent process status, and provide abnormal restart capabilities. It also supports layer-by-layer in-depth viewing of the complete trace link of each session, including user input, model prompt words, inference process, and token consumption distribution at each step.

ws-ckpt

ws-ckpt is ANOLISA's file-level snapshot and rollback tool (AI Agent Workspace Checkpoint) for AI Agent workspace.

AI Agent will make a large number of modifications to the workspace files when performing tasks. Once the operation is made incorrectly or the results are not as expected, users often face difficulties in recovering. The ws-ckpt snapshot mechanism provides lightweight snapshot management capabilities for the workspace. Users can manually create snapshots before key operations, or enable automatic snapshots at the end of each round of Q&A. When rollback is needed, restore to any historical state with one click, making Agent operations reversible and traceable.

The core design concept of ws-ckpt is to "insure the Agent's work". Users can manually create snapshots through natural language or CLI before performing dangerous operations, or enable automatic snapshots at the end of each round of Q&A when performing frequent file modification tasks. The system automatically completes initialization when a snapshot is created for the first time, without additional configuration.

ws-ckpt mainly provides the following capabilities:

Tokenless

Tokenless is the token optimization component of ANOLISA, which reduces the token consumption of the interaction between Agent and LLM from two dimensions: context compression and command filtering.

As the tasks undertaken by Agents become increasingly complex, the expansion of tool definitions, the redundancy of structured responses, and the noise of command output will quickly fill up the context window, which not only increases the cost of reasoning, but also squeezes out the space for communicating effective information. Tokenless

Build an intelligent optimization pipeline between Agent and LLM: at the front end, the description information defined by the tool is automatically streamlined, and low-value fields in the response are identified and filtered; at the middle stage, structured data is compactly encoded to further compress the volume; at the back end, interference content in the command execution output is intelligently filtered. The three work together to significantly reduce token overhead without changing the semantics of Agent behavior.

Tokenless is designed to allow Agents to use fewer Tokens to complete the same task. The entire optimization process is automatically intervened through plug-ins and Hook mechanisms, and is completely transparent to the upstream Agent framework without modifying the business code. All compression effects are quantitatively recorded, providing data support for evaluating optimization benefits.

Tokenless supports seven Agent frameworks: cosh, OpenClaw, Hermes Agent, Claude Code, Codex, Qoder CLI, and Qwen Code. Among them, cosh is integrated by default and takes effect automatically after rpm installation without additional configuration. The other frameworks complete plug-in registration with one click through scripts/install.sh in each adapter directory.

Tokenless mainly provides the following capabilities:

AgentMemory

Agent-Memory is ANOLISA's file-form memory MCP server, which provides AI Agents with persistent, searchable, and sandbox-protected memory space.

Agents face the problem of memory fragmentation when performing cross-session tasks: each new conversation starts from scratch, and previous analysis results, decision records, and user preferences cannot be reused. Agent-Memory uses the file system as the memory model and builds a standardized access pipeline between the Agent and the memory warehouse: on the storage side, the Agent reads and writes memory content like operating files, supporting namespace isolation and path sandboxing; on the retrieval side, it uses BM25 and dense vector mixed retrieval to recall relevant memory fragments according to semantics; on the governance side, it provides git version control and snapshot backup to ensure that the memory is traceable and rollable. The three work together to allow the Agent's memory to cross session boundaries and achieve context continuation.

Agent-Memory is designed to allow Agents to use fewer Tokens to complete cross-session tasks. The entire service exposes 37 standard tools through the MCP protocol, supports any stdio MCP client such as Claude Code, Cursor, and Continue, and is completely transparent to the upstream Agent framework without modifying the business code. The kernel-level security sandbox (openat2) protects the memory warehouse from path traversal or symbolic link escape, and prompt injection detection and escape packaging ensure that the injected LLM memory content is safe and controllable.

Agent-Memory mainly provides the following capabilities: