Release Notes: v0.3.0¶

Release Date: 2026-02-22

Overview¶

OmniAgent v0.3.0 introduces a powerful skill system compatible with OpenClaw/ClawHub and layered sandboxing for secure tool execution. This release significantly expands the agent's extensibility while maintaining security through WASM and Docker isolation.

Highlights¶

Skill System - Load and use skills from the OpenClaw/ClawHub ecosystem
WASM Sandboxing - Lightweight isolation using wazero runtime
Docker Sandboxing - OS-level container isolation for CLI tools

New Features¶

Skill System¶

OmniAgent now supports skills compatible with the OpenClaw SKILL.md format, enabling you to extend your agent with domain-specific capabilities.

Key capabilities:

Parse SKILL.md files with YAML frontmatter and metadata
Discover skills from multiple directories with deduplication
Check requirements (binaries, environment variables) with install hints
Inject skill instructions into the system prompt

CLI commands:

omniagent skills list      # List all discovered skills
omniagent skills info NAME # Show skill details and requirements
omniagent skills check     # Validate all skill requirements

Configuration:

skills:
  enabled: true
  paths:
    - ~/.omniagent/skills
    - /opt/skills
  disabled:
    - experimental-skill
  max_injected: 20

WASM Sandbox¶

A lightweight sandbox using wazero provides capability-based isolation for tool execution.

Capabilities:

Capability	Description
`fs_read`	Read files from allowed paths
`fs_write`	Write files to allowed paths
`net_http`	Make HTTP requests to allowed hosts
`exec_run`	Execute allowed commands

Security features:

Memory limits (default 16MB, max 4GB)
Timeout enforcement via context
Path validation with symlink resolution
Command allowlist enforcement

Docker Sandbox¶

For stronger OS-level isolation, tools can run inside Docker containers with security hardening.

Default security settings:

NetworkMode: none - No network access
CapDrop: ALL - Drop all Linux capabilities
ReadonlyRootfs: true - Read-only root filesystem
SecurityOpt: no-new-privileges - Prevent privilege escalation
Memory and CPU limits

Volume mounts allow controlled filesystem access:

Mounts: []sandbox.DockerMount{
    {HostPath: "/data/input", ContainerPath: "/input", ReadOnly: true},
    {HostPath: "/data/output", ContainerPath: "/output", ReadOnly: false},
}

Bug Fixes¶

Config tests now clear environment variables to prevent interference from user's shell environment

Dependencies¶

Package	Version	Purpose
`github.com/tetratelabs/wazero`	v1.11.0	WASM runtime
`github.com/moby/moby/client`	v0.2.2	Docker SDK
`github.com/moby/moby/api`	v1.53.0	Docker API types

Upgrade Guide¶

From v0.2.0¶

This release is backwards compatible. No configuration changes are required.

To enable skills:

Skills are enabled by default
Place SKILL.md files in ~/.omniagent/skills/ or configure custom paths
Run omniagent skills list to verify discovery

To use sandboxing:

Sandboxing is available programmatically for tool developers. See the sandbox/ package documentation.

Known Issues¶

The testdata/skills/self-improving-agent directory is committed as a git submodule reference
Docker sandbox tests are skipped when Docker is not available

Full Changelog¶

See CHANGELOG.md for the complete list of changes.