Agent 技能

A Skill is a reusable instruction asset. It tells an Agent how to approach a class of work; it does not grant a tool, data access, or an approval bypass.

AI4J's Agent integration follows the same split that makes Skills safe in local and service hosts:

the SDK discovers, catalogs, scopes, and reads SKILL.md content;
the application decides which Skills the current request is allowed to see;
the application continues to own authentication, tenant policy, tools, approvals, and UI.

The Coding Agent has its own workspace integration. See Coding Agent Skills when that is the runtime you are using.

Start with the right integration

For a local developer Agent, use the workspace convenience method:

import io.github.lnyocly.ai4j.agent.Agent;
import io.github.lnyocly.ai4j.agent.Agents;

import java.nio.file.Paths;

Agent agent = Agents.react()
        .modelClient(modelClient)
        .model("gpt-4.1")
        .skills(Paths.get(System.getProperty("user.dir")))
        .build();

skills(workspace) searches only these workspace-owned roots:

<workspace>/.ai4j/skills
<workspace>/.agents/skills

It does not read the process user's home directory. That is the safe default for both repositories and service processes.

For a local developer tool that deliberately needs process-user Skills, opt in explicitly:

Agent agent = Agents.react()
        .modelClient(modelClient)
        .model("gpt-4.1")
        .skillsIncludingUserHome(Paths.get(System.getProperty("user.dir")))
        .build();

不要用于多租户服务

Do not use skillsIncludingUserHome(...) for a tenant service. It can expose the server process's local Skill roots, which are not tenant policy.

Use a host resolver for services

AgentSkillResolver runs once at the start of each Agent run. It is the service-safe extension point for a database, configuration center, or other host-owned Skill source.

import io.github.lnyocly.ai4j.agent.Agent;
import io.github.lnyocly.ai4j.agent.AgentRequest;
import io.github.lnyocly.ai4j.agent.Agents;
import io.github.lnyocly.ai4j.agent.skill.AgentSkillResolver;
import io.github.lnyocly.ai4j.agent.skill.AgentSkillScope;
import io.github.lnyocly.ai4j.skill.SkillDescriptor;

import java.nio.file.Paths;
import java.util.List;

// tenantSkillService, currentTenantId(), currentUserId(), and namesOf(...)
// are application code, not SDK services.
AgentSkillResolver resolver = new AgentSkillResolver() {
    @Override
    public AgentSkillScope resolve(AgentRequest request) {
        List<SkillDescriptor> allowed = tenantSkillService.findEnabledSkills(
                currentTenantId(), currentUserId(), request);

        return AgentSkillScope.builder()
                .workspaceRoot(Paths.get(System.getProperty("user.dir")))
                .providedSkills(allowed)
                .enabledSkillNames(namesOf(allowed))
                .build();
    }
};

Agent agent = Agents.react()
        .modelClient(modelClient)
        .model("gpt-4.1")
        .skillResolver(resolver)
        .build();

The resolver's result is a request-scoped snapshot. Re-resolve it on the next run when an administrator changes a Skill or its grant. Do not cache a tenant's authorization decision globally in the SDK.

For an in-memory Skill from a tenant store, use a stable virtual location and provide the body directly:

SkillDescriptor refundRunbook = SkillDescriptor.builder()
        .name("invoice-refund")
        .description("Handle a verified invoice-refund request.")
        .content("# Invoice refund\n\nVerify the invoice before requesting a refund.")
        .skillFilePath("tenant://acme/invoice-refund/SKILL.md")
        .source("tenant")
        .disableModelInvocation(true)
        .build();

skillFilePath is a stable identifier for the scoped reader. A tenant://... path is not read from the local filesystem.

Automatic and explicit activation

Skills have two activation paths.

Path	When to use it	Runtime behavior
Automatic catalog	The model may choose a Skill for a task.	The stable system prompt gets metadata only. The model reads the relevant body through the scoped Skill reader on demand.
Explicit selection	A user or host chooses a named Skill for this run.	The selected body is a request-scoped user overlay. It is not stored in long-lived memory or appended to the stable system-prompt prefix.

Set disableModelInvocation(true) for a manual-only Skill. It remains available to a host-selected request but is not offered for automatic model invocation.

import io.github.lnyocly.ai4j.agent.AgentRequest;

AgentRequest request = AgentRequest.builder()
        .input("Refund invoice INV-42")
        .selectedSkills(java.util.Collections.singletonList("invoice-refund"))
        .build();

agent.run(request);

selectedSkills 不是授权凭据

Treat selectedSkills as a request from your UI, not proof of authorization. The resolver must return only Skills the authenticated tenant and user may activate. Unknown, disabled, or unauthorized names must not be added to the scope.

Prompt and tool boundary

The automatic catalog is kept small for prompt-cache reuse. The Agent exposes the scoped reader as read_file unless the host already owns that name; in that case it uses read_skill_file and preserves the host tool unchanged.

Reading a Skill does not grant the instructions inside it any extra authority. A Skill that says "call refund_invoice" still needs that tool to be registered and permitted by the host's existing tool, approval, sandbox, and guardrail policies.

What belongs where

Concern	Owner
`SKILL.md` parsing, workspace discovery, descriptors, prompt catalog, scoped reading	AI4J SDK
Per-run `AgentSkillScope` and explicit selection input	Application host
Tenant storage, versioning, publish state, user/role grants, audit records	Application business layer
Slash-command picker or a "my Skills" UI	Application frontend/backend
Tool registration, authorization, approval, sandboxing, and guardrails	Agent host/runtime configuration

This separation is intentional. A reusable SDK cannot infer a tenant's authorization model or safely expose a server process's global Skills to application users.

Do not confuse these concepts

A file-based Skill is a method or instruction resource.
A Tool is an executable capability exposed by the host.
MCP is a protocol for connecting or publishing external capabilities.
A2ASkill describes an agent capability in an A2A AgentCard; it is not a local SKILL.md file. 用 A2ASkillMapper 可以把已授权的 ai4j SkillDescriptor 发布成 AgentCard 上的 A2A skill（见 A2A — Publish ai4j Skills）。

Read Skill、Tool 与 MCP 对比 for the Core distinction and A2A for the AgentCard capability model.

Start with the right integration​

Use a host resolver for services​

Automatic and explicit activation​

Prompt and tool boundary​

What belongs where​

Do not confuse these concepts​

Next pages​