feat: 打通智能体图片媒体运行链路

- 增加稳定媒体引用解析与模型图片能力校验

- 修复图片 Data URI 与模型完整文本读取

- 补充媒体解析和消息兼容测试
This commit is contained in:
2026-07-17 19:45:37 +08:00
parent 66da0c9039
commit f057900f7a
19 changed files with 629 additions and 12 deletions

View File

@@ -17,6 +17,7 @@ import com.easyagents.agent.runtime.memory.AgentMemoryPolicy;
import com.easyagents.agent.runtime.memory.AgentMemorySnapshot;
import com.easyagents.agent.runtime.message.AgentMessage;
import com.easyagents.agent.runtime.message.AgentMessageRole;
import com.easyagents.agent.runtime.message.AgentMediaBlock;
import com.easyagents.agent.runtime.model.AgentModelSpec;
import com.easyagents.agent.runtime.persistence.session.memory.InMemoryAgentSessionStore;
import com.easyagents.agent.runtime.skill.AgentSkillBoxSpec;
@@ -78,6 +79,20 @@ public class AgentScopeStatefulRuntimeTest {
runtime.stream(message);
}
@Test(expected = AgentRuntimeException.class)
public void shouldRejectImageWhenModelCapabilityIsDisabled() {
AgentScopeReActRuntime runtime = fakeRuntime();
AgentInitRequest request = initRequest();
runtime.init(request);
AgentMessage message = new AgentMessage();
message.setRole(AgentMessageRole.USER);
AgentMediaBlock image = new AgentMediaBlock("image");
image.setReference("draft:upload-1");
message.setContentBlocks(List.of(image));
runtime.stream(message);
}
@Test(expected = AgentRuntimeException.class)
public void shouldRejectDuplicateInit() {
AgentScopeReActRuntime runtime = fakeRuntime();

View File

@@ -0,0 +1,117 @@
package com.easyagents.agent.runtime.event.interceptor;
import com.easyagents.agent.runtime.agentscope.AgentScopeMessageAdapter;
import com.easyagents.agent.runtime.media.AgentMediaResource;
import com.easyagents.agent.runtime.message.AgentMediaBlock;
import io.agentscope.core.ReActAgent;
import io.agentscope.core.hook.PreReasoningEvent;
import io.agentscope.core.message.Base64Source;
import io.agentscope.core.message.ImageBlock;
import io.agentscope.core.message.Msg;
import io.agentscope.core.message.MsgRole;
import io.agentscope.core.message.URLSource;
import io.agentscope.core.model.ChatResponse;
import io.agentscope.core.model.GenerateOptions;
import io.agentscope.core.model.Model;
import io.agentscope.core.model.ToolSchema;
import io.agentscope.core.tool.Toolkit;
import org.junit.Assert;
import org.junit.Test;
import reactor.core.publisher.Flux;
import java.nio.charset.StandardCharsets;
import java.util.List;
import java.util.concurrent.atomic.AtomicReference;
/**
* 测试模型调用边界的媒体引用解析。
*/
public class MediaReferenceInterceptorTest {
/**
* 验证私有引用仅在模型调用边界转换为 Base64原始记忆消息保持稳定引用。
*/
@Test
public void shouldResolveReferenceToBase64WithoutMutatingOriginalMessage() {
AgentScopeMessageAdapter adapter = new AgentScopeMessageAdapter();
AgentMediaBlock mediaBlock = new AgentMediaBlock("image");
mediaBlock.setReference("draft:upload-1");
mediaBlock.setMimeType("image/png");
Msg original = Msg.builder()
.id("message-1")
.name("user")
.role(MsgRole.USER)
.content(adapter.toContentBlock(mediaBlock))
.build();
AtomicReference<String> resolvedReference = new AtomicReference<>();
MediaReferenceInterceptor interceptor = new MediaReferenceInterceptor(reference -> {
resolvedReference.set(reference);
return new AgentMediaResource("image/png", "png-data".getBytes(StandardCharsets.UTF_8));
});
ReActAgent agent = ReActAgent.builder()
.name("media-test-agent")
.sysPrompt("system")
.model(new EmptyModel())
.toolkit(new Toolkit())
.build();
PreReasoningEvent event = new PreReasoningEvent(agent, "reasoning-1", null, List.of(original));
interceptor.intercept(event).block();
ImageBlock originalImage = original.getFirstContentBlock(ImageBlock.class);
ImageBlock resolvedImage = event.getInputMessages().get(0).getFirstContentBlock(ImageBlock.class);
Assert.assertTrue(originalImage.getSource() instanceof URLSource);
Assert.assertTrue(((URLSource) originalImage.getSource()).getUrl()
.startsWith(AgentScopeMessageAdapter.MEDIA_REFERENCE_SCHEME));
Assert.assertTrue(resolvedImage.getSource() instanceof Base64Source);
Assert.assertEquals("draft:upload-1", resolvedReference.get());
Assert.assertEquals("cG5nLWRhdGE=", ((Base64Source) resolvedImage.getSource()).getData());
}
/**
* 验证消息适配器序列化后仍能恢复不透明媒体引用。
*/
@Test
public void shouldRoundTripStableReferenceThroughMessageAdapter() {
AgentScopeMessageAdapter adapter = new AgentScopeMessageAdapter();
AgentMediaBlock mediaBlock = new AgentMediaBlock("image");
mediaBlock.setReference("formal:message-1:0:image/png");
AgentMediaBlock restored = (AgentMediaBlock) adapter.toAgentBlock(adapter.toContentBlock(mediaBlock));
Assert.assertEquals("formal:message-1:0:image/png", restored.getReference());
Assert.assertNull(restored.getData());
Assert.assertNull(restored.getUrl());
}
/**
* 仅用于构建真实 AgentScope 事件的空模型。
*/
private static final class EmptyModel implements Model {
/**
* 返回空响应流,当前测试不会实际调用模型。
*
* @param messages 输入消息
* @param toolSchemas 工具定义
* @param options 生成参数
* @return 空响应流
*/
@Override
public Flux<ChatResponse> stream(List<Msg> messages,
List<ToolSchema> toolSchemas,
GenerateOptions options) {
return Flux.empty();
}
/**
* 返回测试模型名称。
*
* @return 测试模型名称
*/
@Override
public String getModelName() {
return "media-test-model";
}
}
}