mirror of
https://gitee.com/mateos/mateclaw.git
synced 2026-09-14 03:33:43 +08:00
683 lines
31 KiB
Java
683 lines
31 KiB
Java
package vip.mate.agent;
|
||
import lombok.extern.slf4j.Slf4j;
|
||
import org.springframework.ai.chat.client.ChatClient;
|
||
import org.springframework.ai.chat.messages.AssistantMessage;
|
||
import org.springframework.ai.chat.messages.Message;
|
||
import org.springframework.ai.chat.messages.SystemMessage;
|
||
import org.springframework.ai.chat.messages.UserMessage;
|
||
import org.springframework.ai.content.Media;
|
||
import org.springframework.core.io.FileSystemResource;
|
||
import org.springframework.util.MimeType;
|
||
import reactor.core.publisher.Flux;
|
||
import vip.mate.approval.ApprovalPlaceholderUtil;
|
||
import vip.mate.llm.service.ModelCapabilityService;
|
||
import vip.mate.workspace.conversation.ConversationService;
|
||
import vip.mate.workspace.conversation.model.MessageContentPart;
|
||
import vip.mate.workspace.conversation.model.MessageEntity;
|
||
|
||
import java.nio.file.Files;
|
||
import java.nio.file.Path;
|
||
import java.nio.file.Paths;
|
||
import java.util.ArrayList;
|
||
import java.util.EnumSet;
|
||
import java.util.List;
|
||
import java.util.Set;
|
||
import java.util.concurrent.atomic.AtomicReference;
|
||
|
||
/**
|
||
* Agent 抽象基类
|
||
* 定义所有 Agent 的基础行为与状态管理
|
||
*
|
||
* @author MateClaw Team
|
||
*/
|
||
@Slf4j
|
||
public abstract class BaseAgent {
|
||
|
||
protected final ChatClient chatClient;
|
||
protected final ConversationService conversationService;
|
||
protected final AtomicReference<AgentState> state = new AtomicReference<>(AgentState.IDLE);
|
||
|
||
/** Agent 唯一标识 */
|
||
protected String agentId;
|
||
|
||
/** Agent 名称 */
|
||
protected String agentName;
|
||
|
||
/** 系统提示词 */
|
||
protected String systemPrompt;
|
||
|
||
/**
|
||
* Max ReAct iterations (one reasoning + action + observation step counts as one).
|
||
* Default 100, hard ceiling 100 (enforced in AgentGraphBuilder so per-agent DB
|
||
* overrides cannot exceed it).
|
||
*/
|
||
public static final int MAX_ITERATIONS_HARD_CEILING = 100;
|
||
protected int maxIterations = 100;
|
||
|
||
/** 工作区活动目录（限制文件工具访问范围，为空不限制） */
|
||
protected String workspaceBasePath;
|
||
|
||
/** 模型名称 */
|
||
protected String modelName;
|
||
|
||
/**
|
||
* Modalities the chat model can natively consume (resolved at agent build time
|
||
* by {@link vip.mate.llm.service.ModelCapabilityService}). Empty set = unknown model,
|
||
* fall back to text-only behavior. See issue #44.
|
||
*/
|
||
protected Set<ModelCapabilityService.Modality> modelCapabilities = EnumSet.noneOf(ModelCapabilityService.Modality.class);
|
||
|
||
/** 采样温度 */
|
||
protected Double temperature;
|
||
|
||
/** 最大输出 token */
|
||
protected Integer maxTokens;
|
||
|
||
/** 最大输入 token（上下文窗口） */
|
||
protected Integer maxInputTokens;
|
||
|
||
/** Top P */
|
||
protected Double topP;
|
||
|
||
/** 当前运行时是否启用工具调用 */
|
||
protected boolean toolCallingEnabled = true;
|
||
|
||
/** 构建时使用的 provider ID（运行时快照） */
|
||
protected String runtimeProviderId;
|
||
|
||
|
||
protected BaseAgent(ChatClient chatClient, ConversationService conversationService) {
|
||
this.chatClient = chatClient;
|
||
this.conversationService = conversationService;
|
||
}
|
||
|
||
/**
|
||
* 同步对话接口
|
||
*
|
||
* @param userMessage 用户消息
|
||
* @param conversationId 会话ID
|
||
* @return 助手回复
|
||
*/
|
||
public abstract String chat(String userMessage, String conversationId);
|
||
|
||
/**
|
||
* 流式对话接口（SSE）
|
||
*
|
||
* @param userMessage 用户消息
|
||
* @param conversationId 会话ID
|
||
* @return 流式文本 Flux
|
||
*/
|
||
public abstract Flux<String> chatStream(String userMessage, String conversationId);
|
||
|
||
/**
|
||
* 执行复杂任务（Plan-and-Execute 模式）
|
||
*
|
||
* @param goal 任务目标
|
||
* @param conversationId 会话ID
|
||
* @return 执行结果摘要
|
||
*/
|
||
public abstract String execute(String goal, String conversationId);
|
||
|
||
/**
|
||
* 带工具重放的对话接口（审批通过后调用）
|
||
* <p>
|
||
* 默认实现退化为普通 chat，子类可覆盖注入 forced_tool_call。
|
||
*
|
||
* @param userMessage 用户消息
|
||
* @param conversationId 会话 ID
|
||
* @param toolCallPayload 要重放的工具调用 JSON
|
||
* @return 助手回复
|
||
*/
|
||
public String chatWithReplay(String userMessage, String conversationId, String toolCallPayload) {
|
||
return chat(userMessage, conversationId);
|
||
}
|
||
|
||
/**
|
||
* 带工具重放的流式对话接口（Web 端审批通过后调用）
|
||
*/
|
||
public Flux<AgentService.StreamDelta> chatWithReplayStream(String userMessage, String conversationId,
|
||
String toolCallPayload) {
|
||
return chatWithReplayStream(userMessage, conversationId, toolCallPayload, "");
|
||
}
|
||
|
||
public Flux<AgentService.StreamDelta> chatWithReplayStream(String userMessage, String conversationId,
|
||
String toolCallPayload, String requesterId) {
|
||
if (this instanceof StructuredStreamCapable capable) {
|
||
return capable.chatStructuredStream(userMessage, conversationId, requesterId);
|
||
}
|
||
return chatStream(userMessage, conversationId)
|
||
.map(chunk -> new AgentService.StreamDelta(chunk, null));
|
||
}
|
||
|
||
/**
|
||
* 获取当前 Agent 状态
|
||
*/
|
||
public AgentState getState() {
|
||
return state.get();
|
||
}
|
||
|
||
/**
|
||
* 设置 Agent 状态
|
||
*/
|
||
protected void setState(AgentState newState) {
|
||
AgentState old = state.getAndSet(newState);
|
||
log.debug("[{}] Agent state: {} -> {}", agentName, old, newState);
|
||
}
|
||
|
||
/**
|
||
* 判断 Agent 是否空闲
|
||
*/
|
||
public boolean isIdle() {
|
||
return AgentState.IDLE.equals(state.get());
|
||
}
|
||
|
||
public String getAgentId() { return agentId; }
|
||
public String getAgentName() { return agentName; }
|
||
public String getSystemPrompt() { return systemPrompt; }
|
||
|
||
protected ChatClient.ChatClientRequestSpec createConversationRequest(String userMessage, String conversationId) {
|
||
ChatClient.ChatClientRequestSpec request = chatClient.prompt()
|
||
.system(systemPrompt != null ? systemPrompt : "你是一个有帮助的AI助手。");
|
||
|
||
List<Message> historyMessages = buildConversationHistory(conversationId, userMessage);
|
||
if (!historyMessages.isEmpty()) {
|
||
request = request.messages(historyMessages);
|
||
}
|
||
return request.user(userMessage);
|
||
}
|
||
|
||
protected List<Message> buildConversationHistory(String conversationId, String currentUserMessage) {
|
||
// ===== 两阶段加载：短对话全量，长对话分页（递进式） =====
|
||
long totalCount = conversationService.countMessages(conversationId);
|
||
if (totalCount <= 0) {
|
||
return List.of();
|
||
}
|
||
|
||
int windowSize = getEffectiveWindowSize();
|
||
List<MessageEntity> history;
|
||
|
||
if (totalCount <= windowSize) {
|
||
// 短对话：全量加载（与旧逻辑一致）
|
||
history = conversationService.listMessages(conversationId);
|
||
} else {
|
||
// 长对话：只加载最近 windowSize 条
|
||
history = conversationService.listRecentMessages(conversationId, windowSize);
|
||
log.info("[{}] Progressive load: {} of {} messages (window={})",
|
||
agentName, history.size(), totalCount, windowSize);
|
||
}
|
||
|
||
// ===== 识别持久化的压缩摘要：从摘要位置开始，跳过更早消息 =====
|
||
for (int i = 0; i < history.size(); i++) {
|
||
MessageEntity msg = history.get(i);
|
||
if ("system".equals(msg.getRole()) && isCompressionSummary(msg)) {
|
||
history = new ArrayList<>(history.subList(i, history.size()));
|
||
log.info("[{}] Found compression summary, loading from index {} ({} messages)",
|
||
agentName, i, history.size());
|
||
break;
|
||
}
|
||
}
|
||
|
||
// ===== 转换为 Spring AI Message 对象 =====
|
||
int limit = history.size();
|
||
if (limit > 0) {
|
||
MessageEntity last = history.get(limit - 1);
|
||
if ("user".equals(last.getRole()) && currentUserMessage.equals(last.getContent())) {
|
||
limit -= 1;
|
||
}
|
||
}
|
||
|
||
if (limit <= 0) {
|
||
return List.of();
|
||
}
|
||
|
||
List<Message> messages = new ArrayList<>(limit);
|
||
for (int i = 0; i < limit; i += 1) {
|
||
Message springMessage = sanitizeForLlm(history.get(i));
|
||
if (springMessage != null) {
|
||
messages.add(springMessage);
|
||
}
|
||
}
|
||
|
||
// Tail guard — orphan-user strip (issue #47).
|
||
//
|
||
// Invariant: every caller of buildConversationHistory appends the
|
||
// current user message AFTER this history (BaseAgent.buildClient
|
||
// via .user(), StateGraphReActAgent / StateGraphPlanExecuteAgent
|
||
// via messages.add(buildCurrentUserMessage)). So the final prompt is
|
||
// [system, ...history, current_user]
|
||
// and is *always* terminated by a user message. That means a trailing
|
||
// assistant in history is FINE for every provider we support — it
|
||
// produces the correct [..., user, assistant, current_user] alternation
|
||
// (OpenAI, Anthropic, DeepSeek regular/thinking, Gemini, Qwen, …).
|
||
//
|
||
// The actual hazard is the opposite: a trailing USER in history.
|
||
// That happens when the immediately-prior turn's assistant message
|
||
// was dropped by Stage 1 (approval placeholder) or Stage 1.5 (errored
|
||
// turn / "[错误] " row), or never persisted at all (turn interrupted
|
||
// before doOnComplete saved the assistant). In that case the history
|
||
// ends with an orphan unanswered user, and appending the current user
|
||
// produces TWO consecutive user messages. Most providers concatenate
|
||
// those and answer both — leaking the orphan question's answer
|
||
// alongside the current answer. (This was the symptom reported in
|
||
// issue #47, originally caused by a tail guard that stripped trailing
|
||
// ASSISTANT messages instead of trailing USER ones — a direction-
|
||
// reversed version of this loop.)
|
||
//
|
||
// Stripping orphan users is safe: the user re-asked or asked a new
|
||
// question; the orphan turn produced no answer the model can build
|
||
// on. We lose a small amount of conversational context in exchange
|
||
// for clean alternation across every provider.
|
||
while (!messages.isEmpty() && messages.get(messages.size() - 1) instanceof UserMessage) {
|
||
messages.remove(messages.size() - 1);
|
||
}
|
||
return messages;
|
||
}
|
||
|
||
/**
|
||
* History sanitization entry point. Encapsulates *all* steps applied to a
|
||
* persisted message before it reaches an LLM prompt. Returns {@code null}
|
||
* to drop the message, or a Spring AI {@link Message} (possibly with
|
||
* rewritten content) to keep it.
|
||
*
|
||
* <p>Design philosophy (OpenClaw-inspired): keep the conversion + every
|
||
* sanitization stage centralized here so future steps (RFC-052 §9 PII
|
||
* field-level redaction, RFC-049 thinking-block replay strategy, image
|
||
* compression for vision models, etc.) plug in as additional inline
|
||
* stages with clear ordering rather than scattering across the loop.
|
||
*
|
||
* <p>Current stages (in order):
|
||
* <ol>
|
||
* <li><b>Drop approval placeholders</b> — assistant messages whose
|
||
* content is a "[等待审批]" stub from the approval flow are removed
|
||
* entirely so they don't pollute the LLM context.</li>
|
||
* <li><b>Render content</b> — convert {@code MessageEntity} to a string
|
||
* via {@link ConversationService#renderMessageContent}.</li>
|
||
* <li><b>Direct-tool scrub (RFC-052)</b> — assistant messages produced
|
||
* by a returnDirect tool path get their content replaced with a
|
||
* tool-named placeholder; the original DB content is unchanged.</li>
|
||
* <li><b>Type dispatch</b> — wrap into {@code AssistantMessage},
|
||
* {@code SystemMessage}, or {@code UserMessage} (with multimodal
|
||
* Media for image/video parts).</li>
|
||
* </ol>
|
||
*/
|
||
private Message sanitizeForLlm(MessageEntity entity) {
|
||
if (entity == null) {
|
||
return null;
|
||
}
|
||
|
||
// Stage 0: drop cron-run header rows (system role + "📋 " prefix)
|
||
// inserted by CronJobLifecycleService.startRun. These are UI dividers
|
||
// for the unified tasks_<wsId> view and the IM channel-session
|
||
// mirror — they carry no semantic context for the LLM. Without this
|
||
// skip, every subsequent IM turn would feed the model unsolicited
|
||
// SystemMessage rows like "📋 每日新闻 · 定时触发 · 2026-04-30T10:55"
|
||
// and bloat the prompt with scheduler metadata.
|
||
if ("system".equals(entity.getRole())
|
||
&& entity.getContent() != null
|
||
&& entity.getContent().startsWith("📋 ")) {
|
||
return null;
|
||
}
|
||
|
||
// Stage 1: drop approval-placeholder assistant messages
|
||
if ("assistant".equals(entity.getRole()) && isApprovalPlaceholder(entity.getContent())) {
|
||
log.debug("[{}] Filtering approval placeholder from history: msgId={}",
|
||
agentName, entity.getId());
|
||
return null;
|
||
}
|
||
|
||
// Stage 1.5: drop typed-error assistant messages. These are persisted
|
||
// by ChatController.doOnComplete with status='error' (or carry the
|
||
// "[错误] " prefix injected by NodeStreamingChatHelper for legacy
|
||
// rows). Re-sending them as multi-turn context drives a self-replicating
|
||
// failure loop:
|
||
// - DeepSeek thinking mode → 400 "reasoning_content must be passed back"
|
||
// (we never captured a real reasoning_content for the failed turn)
|
||
// - Anthropic Claude → 400 "does not support assistant message prefill"
|
||
// (the trailing-user-dedup at the call site can leave an assistant
|
||
// tail when the prior turn errored)
|
||
// Both providers' 400 then re-persist a fresh "[错误] " row, repeat.
|
||
if ("assistant".equals(entity.getRole())
|
||
&& ("error".equals(entity.getStatus())
|
||
|| (entity.getContent() != null && entity.getContent().startsWith("[错误] ")))) {
|
||
log.debug("[{}] Filtering error assistant message from history: msgId={} status={}",
|
||
agentName, entity.getId(), entity.getStatus());
|
||
return null;
|
||
}
|
||
|
||
// Delegate stages 2-4 to toSpringMessage; the stage 3 scrub is applied
|
||
// there so the rendered content is replaced before the typed Message
|
||
// wrapper is constructed.
|
||
return toSpringMessage(entity);
|
||
}
|
||
|
||
/**
|
||
* 判断消息是否为持久化的压缩摘要。
|
||
*/
|
||
private boolean isCompressionSummary(MessageEntity msg) {
|
||
return msg.getMetadata() != null && msg.getMetadata().contains("compression_summary");
|
||
}
|
||
|
||
/**
|
||
* 动态计算窗口大小：基于模型上下文长度估算能容纳多少条消息。
|
||
* 保守估算：每条消息平均 200 token，预留 30% 给系统提示词和当前消息。
|
||
*/
|
||
private int getEffectiveWindowSize() {
|
||
int contextTokens = maxInputTokens != null && maxInputTokens > 0
|
||
? maxInputTokens : 128000;
|
||
int window = (int) (contextTokens * 0.7) / 200;
|
||
return Math.max(20, Math.min(window, 500));
|
||
}
|
||
|
||
/**
|
||
* 判断是否为审批占位消息（委托给共享工具类）
|
||
*/
|
||
static boolean isApprovalPlaceholder(String content) {
|
||
return ApprovalPlaceholderUtil.isApprovalPlaceholder(content);
|
||
}
|
||
|
||
/**
|
||
* RFC-052: regex matching {@code "directToolNames":["a","b",...]} in the
|
||
* metadata JSON and capturing every tool name in group(1) iterations. The
|
||
* {@code \\s*} guards keep us robust to pretty-printed JSON.
|
||
*
|
||
* <p>Design note (OpenClaw-inspired): rather than a one-shot "is this a
|
||
* direct turn?" boolean we extract the actual tool names and weave them
|
||
* into the placeholder, so the next LLM turn can reason about *which* tool
|
||
* answered (e.g. "the user just asked their salary; you used
|
||
* query_employee_salary; if they ask follow-up questions, call it again").
|
||
* This preserves conversational continuity that a generic placeholder
|
||
* destroys.
|
||
*/
|
||
private static final java.util.regex.Pattern DIRECT_TOOL_NAMES_ARRAY =
|
||
java.util.regex.Pattern.compile(
|
||
"\"directToolNames\"\\s*:\\s*\\[(\\s*\"[^\"]*\"\\s*(?:,\\s*\"[^\"]*\"\\s*)*)\\]");
|
||
private static final java.util.regex.Pattern DIRECT_TOOL_NAMES_INNER =
|
||
java.util.regex.Pattern.compile("\"([^\"]+)\"");
|
||
|
||
/**
|
||
* RFC-052: returns the list of returnDirect tool names recorded in the
|
||
* persisted assistant message's metadata. Empty list means this is NOT a
|
||
* direct-tool message and the content is safe for the LLM.
|
||
*
|
||
* <p>Allocates only when a non-empty {@code directToolNames} array is
|
||
* actually present (the common case — normal assistant turns — exits at
|
||
* the first {@code contains} check with zero allocations).
|
||
*/
|
||
static List<String> directToolNamesIn(MessageEntity msg) {
|
||
if (msg == null) return List.of();
|
||
String metadata = msg.getMetadata();
|
||
if (metadata == null || metadata.isEmpty()) return List.of();
|
||
if (!metadata.contains("\"directToolNames\"")) return List.of();
|
||
java.util.regex.Matcher arrayMatcher = DIRECT_TOOL_NAMES_ARRAY.matcher(metadata);
|
||
if (!arrayMatcher.find()) return List.of();
|
||
String inner = arrayMatcher.group(1);
|
||
java.util.regex.Matcher nameMatcher = DIRECT_TOOL_NAMES_INNER.matcher(inner);
|
||
List<String> names = new ArrayList<>(2);
|
||
while (nameMatcher.find()) {
|
||
names.add(nameMatcher.group(1));
|
||
}
|
||
return names;
|
||
}
|
||
|
||
/**
|
||
* Convenience wrapper preserved for callers that only need the boolean.
|
||
* Keeps the original test surface stable.
|
||
*/
|
||
static boolean isDirectToolMessage(MessageEntity msg) {
|
||
return !directToolNamesIn(msg).isEmpty();
|
||
}
|
||
|
||
/**
|
||
* RFC-052: build the placeholder text used to replace a direct-tool
|
||
* assistant message in next-turn prompts. Includes the originating tool
|
||
* names so the model retains conversational structure (it knows *why*
|
||
* the content is redacted and *which* tool would re-fetch it). The
|
||
* original message stays unchanged in {@code mate_message.content}.
|
||
*
|
||
* <p>Worded as a neutral status line, not as a faux assistant utterance —
|
||
* the model treats it as a system-level note, not as previous output to
|
||
* be continued.
|
||
*/
|
||
static String directToolHistoryPlaceholder(List<String> toolNames) {
|
||
if (toolNames == null || toolNames.isEmpty()) {
|
||
return "[Previous answer was tool data returned directly to the user. " +
|
||
"Content withheld from model context per tool policy.]";
|
||
}
|
||
String joined = toolNames.size() == 1
|
||
? "'" + toolNames.get(0) + "'"
|
||
: toolNames.stream()
|
||
.map(n -> "'" + n + "'")
|
||
.reduce((a, b) -> a + ", " + b)
|
||
.orElse("");
|
||
return "[Previous turn used direct-return tool(s) " + joined + " to deliver " +
|
||
"data straight to the user. Content withheld from model context per tool " +
|
||
"policy. If the user asks a follow-up that requires that data, call the " +
|
||
"tool again.]";
|
||
}
|
||
|
||
private Message toSpringMessage(MessageEntity message) {
|
||
if (message == null) {
|
||
return null;
|
||
}
|
||
String renderedContent = conversationService.renderMessageContent(message);
|
||
if (renderedContent == null || renderedContent.isBlank()) {
|
||
return null;
|
||
}
|
||
// RFC-052: scrub direct-tool content from any subsequent LLM prompt.
|
||
// The DB content stays unchanged; only the in-memory Message handed to
|
||
// the model gets replaced. This is MateClaw's persistence-aware analog
|
||
// of joyagent-jdgenie's Memory.clearToolContext (purely in-memory) and
|
||
// OpenClaw's stripToolResultDetails (structural strip per replay).
|
||
//
|
||
// Unlike a generic "withheld" placeholder, we name the originating
|
||
// tool(s) so the model retains the dialog structure: it knows what
|
||
// kind of data was withheld and which tool would fetch it again. This
|
||
// preserves multi-turn coherence without leaking the payload itself.
|
||
if ("assistant".equals(message.getRole())) {
|
||
List<String> directNames = directToolNamesIn(message);
|
||
if (!directNames.isEmpty()) {
|
||
log.debug("[{}] Scrubbing direct-tool content from history msgId={} tools={} (RFC-052)",
|
||
agentName, message.getId(), directNames);
|
||
renderedContent = directToolHistoryPlaceholder(directNames);
|
||
}
|
||
}
|
||
return switch (message.getRole()) {
|
||
case "assistant" -> new AssistantMessage(renderedContent);
|
||
case "system" -> new SystemMessage(renderedContent);
|
||
// History user messages: text only. Re-injecting Media on every replay
|
||
// accumulates attachments across turns — many providers cap at 1 video
|
||
// per request (e.g. Zhipu GLM-5V returns code 1210). The current turn
|
||
// gets Media via buildCurrentUserMessage, which is the only path that
|
||
// should send raw bytes to the model.
|
||
case "user" -> buildUserMessage(message, renderedContent, false);
|
||
default -> null;
|
||
};
|
||
}
|
||
|
||
private static final long MAX_VIDEO_SIZE_BYTES = 20 * 1024 * 1024; // 20MB
|
||
|
||
/**
|
||
* 判断当前模型是否支持视频输入。
|
||
* 由 {@link ModelCapabilityService} 在 agent 构建时解析并注入到
|
||
* {@link #modelCapabilities}，per-model 粒度（区分如 glm-4v vs glm-4v-plus）。
|
||
*/
|
||
private boolean modelSupportsVideo() {
|
||
return modelCapabilities.contains(ModelCapabilityService.Modality.VIDEO);
|
||
}
|
||
|
||
private boolean modelSupportsVision() {
|
||
return modelCapabilities.contains(ModelCapabilityService.Modality.VISION);
|
||
}
|
||
|
||
/**
|
||
* 构建 UserMessage，支持 multimodal：如果消息包含图片/视频附件，直接注入 Spring AI Media 对象，
|
||
* 让模型在 prompt 中直接看到媒体内容，不需要再调 MCP read_media_file 工具。
|
||
*/
|
||
protected UserMessage buildUserMessage(MessageEntity message, String renderedContent) {
|
||
return buildUserMessage(message, renderedContent, true);
|
||
}
|
||
|
||
/**
|
||
* @param injectMedia when {@code false} (history replay), skip the Media-loading
|
||
* branch entirely and return text-only — providers like Zhipu
|
||
* GLM-5V cap at 1 video per request, so re-injecting historical
|
||
* attachments on every turn breaks the call.
|
||
*/
|
||
protected UserMessage buildUserMessage(MessageEntity message, String renderedContent, boolean injectMedia) {
|
||
if (!injectMedia) {
|
||
return new UserMessage(renderedContent == null ? "" : renderedContent);
|
||
}
|
||
List<MessageContentPart> parts = conversationService.parseMessageParts(message);
|
||
List<Media> mediaList = new ArrayList<>();
|
||
// Reasons for attachments that the model cannot consume — surfaced to the agent
|
||
// via the user message text so it does not hallucinate a tool call to read them.
|
||
// See issue #44.
|
||
List<String> skippedAttachments = new ArrayList<>();
|
||
boolean videoSupported = modelSupportsVideo();
|
||
boolean visionSupported = modelSupportsVision();
|
||
|
||
for (MessageContentPart part : parts) {
|
||
if (part == null) continue;
|
||
String partType = part.getType();
|
||
String contentType = part.getContentType();
|
||
// image 类型的 part 可能没有精确 contentType，补全为 image/jpeg
|
||
if ("image".equals(partType) && (contentType == null || "image/*".equals(contentType))) {
|
||
contentType = "image/jpeg";
|
||
}
|
||
if (contentType == null) continue;
|
||
|
||
boolean isImage = ("image".equals(partType) || "file".equals(partType)) && contentType.startsWith("image/");
|
||
boolean isVideo = ("video".equals(partType) || "file".equals(partType)) && contentType.startsWith("video/");
|
||
|
||
if (!isImage && !isVideo) continue;
|
||
|
||
// SVG 是 XML 文本，不是光栅图片，LLM multimodal API 不支持
|
||
if (isImage && contentType.contains("svg")) {
|
||
log.debug("[{}] Skipping SVG attachment (not supported by multimodal API): {}",
|
||
agentName, part.getFileName());
|
||
skippedAttachments.add(part.getFileName() + "(SVG 格式，多模态 API 不支持)");
|
||
continue;
|
||
}
|
||
|
||
// 图片仅在模型支持视觉时注入；纯文本模型（如 GLM-5-Turbo / DeepSeek-V3）会被跳过
|
||
if (isImage && !visionSupported) {
|
||
log.debug("[{}] Skipping image attachment (model '{}' does not support vision): {}",
|
||
agentName, modelName, part.getFileName());
|
||
skippedAttachments.add(part.getFileName() + "(当前模型 " + modelName + " 不支持图片输入)");
|
||
continue;
|
||
}
|
||
|
||
// 视频仅在模型支持时注入，否则跳过（避免发送给非视觉模型导致 400 错误）
|
||
if (isVideo && !videoSupported) {
|
||
log.debug("[{}] Skipping video attachment (model '{}' does not support video): {}",
|
||
agentName, modelName, part.getFileName());
|
||
skippedAttachments.add(part.getFileName() + "(当前模型 " + modelName + " 不支持视频输入)");
|
||
continue;
|
||
}
|
||
|
||
// 视频文件大小保护
|
||
if (isVideo && part.getFileSize() != null && part.getFileSize() > MAX_VIDEO_SIZE_BYTES) {
|
||
log.warn("[{}] Skipping oversized video attachment ({}MB > 20MB): {}",
|
||
agentName, part.getFileSize() / (1024 * 1024), part.getFileName());
|
||
skippedAttachments.add(part.getFileName() + "(视频超过 20MB 大小限制)");
|
||
continue;
|
||
}
|
||
|
||
// 解析媒体文件路径：先尝试 path，再尝试 mediaId（IM 渠道下载后存在 mediaId 中），再拼接工作目录
|
||
Path mediaPath = resolveImagePath(part.getPath());
|
||
if (mediaPath == null && part.getMediaId() != null) {
|
||
mediaPath = resolveImagePath(part.getMediaId());
|
||
}
|
||
if (mediaPath == null) {
|
||
log.warn("[{}] {} file not found for attachment: {}, path: {}, mediaId: {}",
|
||
agentName, isVideo ? "Video" : "Image", part.getFileName(), part.getPath(), part.getMediaId());
|
||
skippedAttachments.add(part.getFileName() + "(文件未找到)");
|
||
continue;
|
||
}
|
||
try {
|
||
MimeType mimeType = MimeType.valueOf(contentType);
|
||
Media media = new Media(mimeType, new FileSystemResource(mediaPath));
|
||
mediaList.add(media);
|
||
log.debug("[{}] Injected {} into prompt: {} ({})",
|
||
agentName, isVideo ? "video" : "image", part.getFileName(), mediaPath);
|
||
} catch (Exception e) {
|
||
log.warn("[{}] Failed to create Media for {} {}: {}",
|
||
agentName, isVideo ? "video" : "image", part.getFileName(), e.getMessage());
|
||
skippedAttachments.add(part.getFileName() + "(媒体加载失败)");
|
||
}
|
||
}
|
||
|
||
String finalText = renderedContent;
|
||
if (!skippedAttachments.isEmpty()) {
|
||
finalText = (renderedContent == null ? "" : renderedContent)
|
||
+ "\n\n[系统提示] 以下附件未能传入当前模型：" + String.join("、", skippedAttachments)
|
||
+ "。\n请用对话语言清晰、友好地告诉用户：当前模型无法处理这类附件，建议切换到具备相应能力的多模态模型（图片需视觉模型，视频需视频理解模型）后重新上传。"
|
||
+ "不要调用任何工具（包括 ffmpeg、浏览器、文件读取等）尝试解析这些附件。";
|
||
}
|
||
|
||
if (mediaList.isEmpty()) {
|
||
return new UserMessage(finalText);
|
||
}
|
||
return UserMessage.builder()
|
||
.text(finalText)
|
||
.media(mediaList)
|
||
.build();
|
||
}
|
||
|
||
/**
|
||
* 解析图片文件的绝对路径。
|
||
* <p>
|
||
* 上传文件存储在 data/chat-uploads/ 下，是相对于 Spring Boot 工作目录的路径。
|
||
* MCP 工具的工作目录可能不同，所以这里直接解析为绝对路径。
|
||
*/
|
||
/**
|
||
* 构建当前用户消息的 UserMessage（含 multimodal 图片注入）。
|
||
* <p>
|
||
* 从 DB 读取最后一条 user 消息的 contentParts，提取图片附件并注入 Media。
|
||
* 不依赖文本相等匹配（避免重复文本误绑定到错误轮次），而是直接取最后一条 user 消息，
|
||
* 因为 buildInitialState 在 saveMessage 之后调用，最后一条 user 消息就是当前消息。
|
||
*
|
||
* @param conversationId 会话 ID
|
||
* @param userMessageText 用户消息文本（作为 fallback 内容）
|
||
* @return 带图片 Media 的 UserMessage（如果有图片附件），否则纯文本 UserMessage
|
||
*/
|
||
protected UserMessage buildCurrentUserMessage(String conversationId, String userMessageText) {
|
||
try {
|
||
List<MessageEntity> history = conversationService.listMessages(conversationId);
|
||
// 倒序取最后一条 user 消息（buildInitialState 在 saveMessage 后调用，所以最后一条就是当前消息）
|
||
for (int i = history.size() - 1; i >= 0; i--) {
|
||
MessageEntity msg = history.get(i);
|
||
if ("user".equals(msg.getRole())) {
|
||
// 用 DB 中的实际内容（可能包含 contentParts），不用传入的 text
|
||
String content = conversationService.renderMessageContent(msg);
|
||
return buildUserMessage(msg, content != null && !content.isBlank() ? content : userMessageText);
|
||
}
|
||
}
|
||
} catch (Exception e) {
|
||
log.debug("[{}] Failed to load current user message parts for multimodal: {}",
|
||
agentName, e.getMessage());
|
||
}
|
||
return new UserMessage(userMessageText);
|
||
}
|
||
|
||
protected Path resolveImagePath(String relativePath) {
|
||
if (relativePath == null || relativePath.isBlank()) {
|
||
return null;
|
||
}
|
||
// 1. 如果已经是绝对路径且存在，直接用
|
||
Path path = Paths.get(relativePath);
|
||
if (path.isAbsolute() && Files.exists(path)) {
|
||
return path;
|
||
}
|
||
// 2. 相对于 Spring Boot 工作目录解析
|
||
Path resolved = Paths.get(System.getProperty("user.dir")).resolve(relativePath);
|
||
if (Files.exists(resolved)) {
|
||
return resolved;
|
||
}
|
||
// 3. 都找不到
|
||
log.debug("[{}] Image path not found: tried {} and {}", agentName, path, resolved);
|
||
return null;
|
||
}
|
||
|
||
}
|