nolo-cli 0.1.39 → 0.1.41

package/chat/dialog/actions/updateDialogSummaryAction.ts

@@ -0,0 +1,430 @@
+// 文件路径: packages/chat/dialog/actions/updateDialogSummaryAction.ts
+
+import type { RootState } from "../../../app/store";
+import { runLlm } from "../../../ai/agent/agentSlice";
+import {
+    buildBuiltinSummaryContent,
+    BUILTIN_SUMMARY_LLM_CONFIG,
+} from "./builtinDialogLlm";
+import { patch, selectById } from "../../../database/dbSlice";
+// import { selectAllMsgs, selectMsgsByDialogId } from "../../messages/messageSlice"; // CIRCULAR
+import { serializeMessageContent } from "../../messages/messageContent";
+import { DialogConfig, Agent } from "../../../app/types";
+import { selectContextRetention } from "../../../app/settings/settingSlice";
+import { getModelContextWindow } from "../../../ai/llm/getModelContextWindow";
+import { estimateTokenCount } from "../../../ai/context/tokenUtils";
+import { extractCustomId } from "../../../core/prefix";
+import type { Message } from "../../messages/types";
+import {
+    ConversationLoad,
+    planContextUsage,
+} from "../../../ai/context/retention";
+
+// --- 常量 ---
+const MIN_COMPRESS_COUNT = 5; // 至少压缩5条以上才有意义
+const MIN_PROACTIVE_MESSAGE_COUNT = 6;
+const MIN_PROACTIVE_TOKENS = 1800;
+const PROACTIVE_SUMMARY_MAX_INPUT_TOKENS = 12000;
+const ACTIVE_SUMMARY_TAIL_KEEP_COUNT = 2;
+
+
+// --- 辅助函数 ---
+
+const getMessageTokenCount = (msg: any): number => {
+    if (msg.usage?.completion_tokens) {
+        return msg.usage.completion_tokens;
+    }
+    const content = serializeMessageContent(msg.content) || "";
+    return estimateTokenCount(content);
+};
+
+const getMessagesForDialogFromState = (
+    state: RootState,
+    dialogId: string
+): Message[] => {
+    const msgsState = state.message.dialogStateById[dialogId]?.msgs;
+    if (!msgsState || !msgsState.ids) return [];
+
+    return (msgsState.ids as string[])
+        .map(id => msgsState.entities[id])
+        .filter(Boolean) as Message[];
+};
+
+const formatMessagesForSummary = (msgs: Message[]): string =>
+    msgs
+        .map(m => {
+            const content = serializeMessageContent(m.content) || "[非文本内容]";
+            return `${m.role}: ${content}`;
+        })
+        .join("\n");
+
+const findMessageIndexById = (msgs: Message[], id?: string): number =>
+    id ? msgs.findIndex(m => m.id === id) : -1;
+
+const hasOpenEndedToolCall = (msg: Message | undefined): boolean =>
+    !!msg && Array.isArray((msg as any).tool_calls) && (msg as any).tool_calls.length > 0;
+
+const truncateMessagesForSummary = (msgs: Message[]): Message[] => {
+    let total = 0;
+    let start = msgs.length;
+
+    for (let i = msgs.length - 1; i >= 0; i--) {
+        const t = getMessageTokenCount(msgs[i]);
+        if (total + t > PROACTIVE_SUMMARY_MAX_INPUT_TOKENS) break;
+        total += t;
+        start = i;
+    }
+
+    return msgs.slice(start);
+};
+
+const buildPreviousSummaryWithProactive = (
+    dialogConfig: DialogConfig,
+    shouldConsumeProactive: boolean
+): string => {
+    const base = dialogConfig.summary || "";
+    const proactive = shouldConsumeProactive ? dialogConfig.proactiveSummary || "" : "";
+
+    if (!base && !proactive) return "";
+    if (!proactive) return base;
+    if (!base) return `【主动工作摘要】\n${proactive}`;
+
+    return `${base}\n\n【主动工作摘要】\n${proactive}`;
+};
+
+const isActiveSummaryWorthDoing = (
+    pendingTokens: number,
+    contextWindow: number
+): boolean => {
+    const minTokens = Math.min(
+        40_000,
+        Math.max(10_000, Math.floor(contextWindow * 0.05))
+    );
+    return pendingTokens >= minTokens;
+};
+
+// --- Action ---
+const classifyConversationLoad = (msgs: Message[]): ConversationLoad => {
+    const N = 20;
+    if (!Array.isArray(msgs) || msgs.length === 0) return "light";
+
+    const tail = msgs.slice(-N);
+    const tokenSamples = tail.map(getMessageTokenCount);
+    if (tokenSamples.length === 0) return "light";
+
+    const sum = tokenSamples.reduce((acc, v) => acc + v, 0);
+    const avg = sum / tokenSamples.length;
+    const sorted = [...tokenSamples].sort((a, b) => a - b);
+    const p95 = sorted[Math.floor((sorted.length - 1) * 0.95)];
+
+    if (p95 < 200 && avg < 120) {
+        return "light";
+    }
+
+    if (p95 > 2000 || avg > 1200) {
+        return "heavy";
+    }
+
+    return "medium";
+};
+
+
+const summarizingDialogs = new Set<string>();
+
+export const updateDialogSummaryAction = async (
+    args: {
+        dialogKey: string;
+        preFetchedMessages?: Message[];
+        force?: boolean;
+        reason?: "task_completed" | "context_budget";
+    },
+    thunkApi: any
+) => {
+    const { dialogKey, preFetchedMessages, force = false } = args;
+
+    // 0. 并发锁：避免同一个 Dialog 同时进行多个摘要任务
+    if (summarizingDialogs.has(dialogKey)) return;
+    summarizingDialogs.add(dialogKey);
+
+    const { dispatch, getState } = thunkApi;
+
+    try {
+        const state = getState() as RootState;
+        const dialogId = extractCustomId(dialogKey);
+
+        const dialogConfig = selectById(state, dialogKey) as DialogConfig;
+        if (!dialogConfig) return;
+
+        // 1. 获取配置
+        const retention = selectContextRetention(state);
+
+        let contextWindow = 128000;
+        if (dialogConfig.cybots && dialogConfig.cybots.length > 0) {
+            const agentId = dialogConfig.cybots[0];
+            const agent = selectById(state, agentId) as Agent;
+            if (agent?.model) {
+                contextWindow = getModelContextWindow(agent.model);
+            }
+        }
+
+        // 2. 获取消息
+        const allMsgs =
+            preFetchedMessages || getMessagesForDialogFromState(state, dialogId);
+
+
+
+
+        // 3. 找到最后一次压缩的位置
+        const lastSummarizedId = dialogConfig.summarizedBeforeId;
+        let startIndex = 0;
+        if (lastSummarizedId) {
+            const found = allMsgs.findIndex(m => m.id === lastSummarizedId);
+            if (found !== -1) {
+                startIndex = found + 1;
+            }
+        }
+
+        // 待处理的消息（尚未被压缩进 summary）
+        const pendingMsgs = allMsgs.slice(startIndex);
+        if (pendingMsgs.length === 0) return;
+
+        // 4. 计算当前总开销 = 已有 Summary + 待处理消息
+        const summaryTokens = estimateTokenCount(dialogConfig.summary || "");
+        const pendingTokens = pendingMsgs.reduce((sum, msg) => sum + getMessageTokenCount(msg), 0);
+        const totalUsed = summaryTokens + pendingTokens;
+
+        // 5. 基于模型窗口 / slider / 近期负载，规划历史预算
+        const adjustedSummaryTokens = Math.max(summaryTokens, 1000);
+        const recentLoad = classifyConversationLoad(pendingMsgs);
+
+        const { historyBudget, rawMessageBudget } = planContextUsage({
+            contextWindow,
+            retentionSlider: retention,
+            summaryTokens: adjustedSummaryTokens,
+            recentLoad,
+        });
+
+        const shouldRunActiveSummary =
+            force &&
+            !hasOpenEndedToolCall(pendingMsgs[pendingMsgs.length - 1]) &&
+            isActiveSummaryWorthDoing(pendingTokens, contextWindow);
+
+        // 历史 + 待处理总开销未达到预算，且没有明确的主动归档信号，不触发压缩
+        if (totalUsed < historyBudget && !shouldRunActiveSummary) {
+            return;
+        }
+
+        // 6. 需要压缩：决定把哪些消息压进去
+        // 目标：压缩后，summary + 保留的原始消息 ≈ historyBudget，
+        //       且尽量保留最近的若干条消息。
+
+        let tokensToKeep = 0;
+        let keepCount = 0;
+
+        // 从后往前数，保留最近的消息直到填满 rawMessageBudget
+        for (let i = pendingMsgs.length - 1; i >= 0; i--) {
+            const t = getMessageTokenCount(pendingMsgs[i]);
+            if (tokensToKeep + t > rawMessageBudget) break;
+            tokensToKeep += t;
+            keepCount++;
+        }
+
+        // 主动归档时保留最后两条原文，避免刚给用户的结论立刻被折叠进 summary。
+        let compressCount =
+            shouldRunActiveSummary && totalUsed < historyBudget
+                ? Math.max(0, pendingMsgs.length - ACTIVE_SUMMARY_TAIL_KEEP_COUNT)
+                : pendingMsgs.length - keepCount;
+
+        // Guard: Prevent breaking tool chains or compressing open-ended tool calls
+        // OpenAI requires: Assistant(tool_calls) -> Tool(output) must be contiguous.
+
+        // 1. Ensure we do not cut immediately before a 'tool' message.
+        // If pendingMsgs[compressCount] (the first kept message) is 'tool', 
+        // it means we are summarizing its parent 'assistant'. This is illegal.
+        while (compressCount > 0 && compressCount < pendingMsgs.length && pendingMsgs[compressCount].role === 'tool') {
+            compressCount--;
+        }
+
+        // 2. Ensure the last compressed message is not an assistant with tool_calls that needs a future tool output
+        // (This is implicitly covered by #1 if tool output exists, but if tool output hasn't arrived yet, 
+        // we must check the last message itself).
+        // Actually, if tool output hasn't arrived, 'updateDialogSummaryAction' is usually called after 'messageStreamEnd'.
+        // If the last message is Assistant(tool_calls), the NEXT message (not yet in list) will be Tool.
+        // So we should NOT compress the last message if it has tool_calls.
+        if (compressCount > 0) {
+            const lastCompressed = pendingMsgs[compressCount - 1];
+            const hasToolCalls = Array.isArray((lastCompressed as any).tool_calls) && (lastCompressed as any).tool_calls.length > 0;
+            if (hasToolCalls) {
+                // If we are chopping off the end of the conversation, and it ends with a tool call,
+                // we must preserve it until tool output arrives.
+                // Actually, just to be safe, never compress an active tool call.
+                compressCount--;
+            }
+        }
+
+
+        // 2. 可压缩的消息太少，不值得压缩
+        if (compressCount < MIN_COMPRESS_COUNT) {
+            return;
+        }
+
+        const msgsToCompress = pendingMsgs.slice(0, compressCount);
+
+
+        // 最后一条被压缩的消息
+        const newSummarizedBeforeId = msgsToCompress[msgsToCompress.length - 1].id;
+
+        // 7. 提取引用 Keys (pageKey/dialogKey) 并保存
+        // 这一步非常重要：因为这些消息即将从 context 中消失，必须把 key 留下来
+        // 我们复用 addReferenceKeysAction 的逻辑，或者通过 content 提取
+        // 这里直接提取然后 patch dialog
+        const extractedKeys = new Set(dialogConfig.referenceKeys || []);
+        msgsToCompress.forEach(msg => {
+            // 简单遍历 content
+            if (Array.isArray(msg.content)) {
+                msg.content.forEach((p: any) => {
+                    if (p?.pageKey) extractedKeys.add(p.pageKey);
+                    if (p?.dialogKey) extractedKeys.add(p.dialogKey);
+                });
+            } else if (msg.content && typeof msg.content === 'object') {
+                const c: any = msg.content;
+                if (c.pageKey) extractedKeys.add(c.pageKey);
+                if (c.dialogKey) extractedKeys.add(c.dialogKey);
+            }
+        });
+
+        // 8. 生成新 Summary (结构化记忆升级版)
+        // 目标：将线性摘要升级为 "事实 (Facts) + 剧情 (Narrative)" 双轨记忆。
+        // 这能显著提升 Agent 的 "智商"，防止关键配置在长对话中被遗忘。
+
+        const proactiveBeforeIndex = findMessageIndexById(
+            allMsgs,
+            dialogConfig.proactiveSummaryBeforeId
+        );
+        const shouldConsumeProactive =
+            !!dialogConfig.proactiveSummary &&
+            proactiveBeforeIndex !== -1 &&
+            proactiveBeforeIndex < startIndex + compressCount;
+
+        const previousSummary = buildPreviousSummaryWithProactive(
+            dialogConfig,
+            shouldConsumeProactive
+        );
+        const messagesText = formatMessagesForSummary(msgsToCompress);
+
+        const promptContent = buildBuiltinSummaryContent(
+            previousSummary,
+            messagesText
+        );
+
+        try {
+            // 调用内置 Summary LLM
+            const newSummary = await dispatch(
+                runLlm({
+                    llmConfig: BUILTIN_SUMMARY_LLM_CONFIG,
+                    content: promptContent,
+                    billingDialogKey: dialogKey,
+                })
+            ).unwrap();
+
+            if (newSummary && typeof newSummary === "string" && newSummary.trim()) {
+                const currentCount = dialogConfig.compressionCount || 0;
+
+                await dispatch(
+                    patch({
+                        dbKey: dialogKey,
+                        changes: {
+                            summary: newSummary.trim(),
+                            summarizedBeforeId: newSummarizedBeforeId,
+                            referenceKeys: Array.from(extractedKeys),
+                            compressionCount: currentCount + 1,
+                            summaryPending: false, // Explicitly clear pending flag
+                            ...(shouldConsumeProactive
+                                ? {
+                                    proactiveSummary: "",
+                                    proactiveSummaryBeforeId: undefined,
+                                }
+                                : {}),
+                        },
+                    })
+                ).unwrap();
+
+                console.log(`[ContextCompression] Compressed ${compressCount} messages. New summary len: ${newSummary.length}`);
+            }
+        } catch (err) {
+            console.error("[ContextCompression] Failed:", err);
+        }
+    } finally {
+        summarizingDialogs.delete(dialogKey);
+    }
+};
+
+export const updateDialogProactiveSummaryAction = async (
+    args: { dialogKey: string; preFetchedMessages?: Message[] },
+    thunkApi: any
+) => {
+    const { dialogKey, preFetchedMessages } = args;
+    if (summarizingDialogs.has(dialogKey)) return;
+    if (summarizingDialogs.has(`proactive:${dialogKey}`)) return;
+    summarizingDialogs.add(`proactive:${dialogKey}`);
+
+    const { dispatch, getState } = thunkApi;
+
+    try {
+        const state = getState() as RootState;
+        const dialogId = extractCustomId(dialogKey);
+        const dialogConfig = selectById(state, dialogKey) as DialogConfig;
+        if (!dialogConfig) return;
+
+        const allMsgs = preFetchedMessages || getMessagesForDialogFromState(state, dialogId);
+        if (!allMsgs.length) return;
+
+        const hardSummaryIndex = findMessageIndexById(allMsgs, dialogConfig.summarizedBeforeId);
+        const proactiveIndex = findMessageIndexById(allMsgs, dialogConfig.proactiveSummaryBeforeId);
+        const startIndex = Math.max(hardSummaryIndex, proactiveIndex) + 1;
+        const pendingMsgs = allMsgs.slice(startIndex);
+        if (pendingMsgs.length < MIN_PROACTIVE_MESSAGE_COUNT) return;
+
+        if (hasOpenEndedToolCall(pendingMsgs[pendingMsgs.length - 1])) return;
+
+        const pendingTokens = pendingMsgs.reduce((sum, msg) => sum + getMessageTokenCount(msg), 0);
+        const hasToolOutput = pendingMsgs.some(msg => msg.role === "tool");
+        if (!hasToolOutput && pendingTokens < MIN_PROACTIVE_TOKENS) return;
+
+        const msgsToSummarize = truncateMessagesForSummary(pendingMsgs);
+        if (msgsToSummarize.length < MIN_PROACTIVE_MESSAGE_COUNT && !hasToolOutput) return;
+
+        const newProactiveBeforeId = pendingMsgs[pendingMsgs.length - 1].id;
+        const previousSummary = dialogConfig.proactiveSummary || "";
+        const messagesText = formatMessagesForSummary(msgsToSummarize);
+        const promptContent = buildBuiltinSummaryContent(previousSummary, messagesText);
+
+        try {
+            const newSummary = await dispatch(
+                runLlm({
+                    llmConfig: BUILTIN_SUMMARY_LLM_CONFIG,
+                    content: promptContent,
+                    billingDialogKey: dialogKey,
+                })
+            ).unwrap();
+
+            if (newSummary && typeof newSummary === "string" && newSummary.trim()) {
+                await dispatch(
+                    patch({
+                        dbKey: dialogKey,
+                        changes: {
+                            proactiveSummary: newSummary.trim(),
+                            proactiveSummaryBeforeId: newProactiveBeforeId,
+                        },
+                    })
+                ).unwrap();
+
+                console.log(`[ContextCompression] Proactive summary updated. len: ${newSummary.length}`);
+            }
+        } catch (err) {
+            console.error("[ContextCompression] Proactive summary failed:", err);
+        }
+    } finally {
+        summarizingDialogs.delete(`proactive:${dialogKey}`);
+    }
+};

package/chat/dialog/actions/updateDialogTitleAction.ts

@@ -0,0 +1,151 @@
+// chat/dialogs/updateDialogTitleAction.ts
+
+import { runLlm } from "../../../ai/agent/agentSlice";
+import type { RootState } from "../../../app/store";
+import {
+  updateContentTitle,
+} from "../../../create/space/spaceSlice";
+import { normalizeSpaceId } from "../../../create/space/spaceKeys";
+import { patch, selectById } from "../../../database/dbSlice";
+import { extractCustomId } from "../../../core/prefix";
+import { format } from "date-fns";
+import { selectAllMsgs } from "../../messages/messageSlice";
+import { serializeMessageContent } from "../../messages/messageContent";
+import {
+  BUILTIN_TITLE_LLM_CONFIG,
+} from "./builtinDialogLlm";
+import {
+  shouldUpdateTitle,
+} from "./updateDialogTitlePolicy";
+import {
+  buildDialogFallbackTitleFromMessages,
+  resolveDialogTitle,
+} from "../dialogTitle";
+
+// --- 常量 ---
+const MAX_MESSAGES_FOR_CONTEXT = 20;
+
+const isAssistantToolStub = (msg: any) =>
+  msg?.role === "assistant" &&
+  (msg.content == null ||
+    (typeof msg.content === "string" && msg.content.trim().length === 0) ||
+    (Array.isArray(msg.content) && msg.content.length === 0)) &&
+  Array.isArray(msg?.tool_calls) &&
+  msg.tool_calls.length > 0;
+
+const dedupeById = <T extends { id?: string }>(messages: T[]) => {
+  const seen = new Set<string>();
+  return messages.filter((message) => {
+    const id = typeof message.id === "string" ? message.id : "";
+    if (!id || seen.has(id)) return false;
+    seen.add(id);
+    return true;
+  });
+};
+
+const getMessageContextForTitle = (
+  state: RootState,
+  dialogKey: string,
+  selectAllMessages: typeof selectAllMsgs = selectAllMsgs
+) => {
+  const allMsgs = selectAllMessages(state, extractCustomId(dialogKey));
+
+  const flattened = Array.isArray(allMsgs)
+    ? typeof (allMsgs as any).flat === "function"
+      ? (allMsgs as any).flat()
+      : (allMsgs as any).reduce(
+        (acc: any[], cur: any) => acc.concat(cur),
+        []
+      )
+    : [];
+
+  const normalized = flattened
+    .filter((msg: any) => msg?.role !== "tool" && !isAssistantToolStub(msg))
+    .map((msg: any) => {
+      const textContent = serializeMessageContent(msg?.content, "这里是一个图片");
+      if (!textContent) return null;
+
+      return {
+        ...msg,
+        content: textContent as string, // 之后只会用到 role + content
+      };
+    })
+    .filter(Boolean) as Array<{ role: any; content: string }>;
+
+  const earlyUserTurns = normalized.filter((msg) => msg.role === "user").slice(0, 3);
+  const recentTurns = normalized.slice(-12);
+
+  return dedupeById([...earlyUserTurns, ...recentTurns]).slice(
+    0,
+    MAX_MESSAGES_FOR_CONTEXT
+  );
+};
+
+type UpdateDialogTitleDeps = {
+  runLlmAction?: typeof runLlm;
+  patchAction?: typeof patch;
+  selectDialogById?: typeof selectById;
+  selectAllMessages?: typeof selectAllMsgs;
+  updateSpaceContentTitle?: typeof updateContentTitle;
+};
+
+// --- 异步 Thunk Action ---
+export const updateDialogTitleActionWithDeps = async (
+  args,
+  thunkApi,
+  deps: UpdateDialogTitleDeps = {}
+) => {
+  const {
+    runLlmAction = runLlm,
+    patchAction = patch,
+    selectDialogById = selectById,
+    selectAllMessages = selectAllMsgs,
+    updateSpaceContentTitle = updateContentTitle,
+  } = deps;
+  const { dialogKey } = args;
+  const { dispatch, getState } = thunkApi;
+  const state = getState() as RootState;
+
+  const dialogConfig = selectDialogById(state, dialogKey);
+  if (
+    !dialogConfig ||
+    !shouldUpdateTitle(dialogConfig.createdAt, dialogConfig.updatedAt)
+  ) {
+    return dialogConfig;
+  }
+
+  const messageContext = getMessageContextForTitle(state, dialogKey, selectAllMessages);
+  if (messageContext.length === 0) {
+    return dialogConfig;
+  }
+
+  const content = JSON.stringify(
+    messageContext.map((msg) => ({ role: msg.role, content: msg.content }))
+  );
+
+  const generatedTitle = await dispatch(
+    runLlmAction({
+      llmConfig: BUILTIN_TITLE_LLM_CONFIG,
+      content,
+      billingDialogKey: dialogKey,
+    })
+  ).unwrap();
+
+  const fallbackTitle =
+    buildDialogFallbackTitleFromMessages(messageContext) ||
+    `Conversation on ${format(new Date(), "MMM d")}`;
+  const title = resolveDialogTitle(generatedTitle, fallbackTitle);
+
+  const spaceId =
+    dialogConfig?.spaceId && normalizeSpaceId(dialogConfig.spaceId);
+  if (spaceId) {
+    dispatch(updateSpaceContentTitle({ spaceId, contentKey: dialogKey, title }));
+  }
+
+  return await dispatch(
+    patchAction({ dbKey: dialogKey, changes: { title } })
+  ).unwrap();
+};
+
+export const updateDialogTitleAction = async (args, thunkApi) =>
+  updateDialogTitleActionWithDeps(args, thunkApi);

package/chat/dialog/actions/updateDialogTitlePolicy.ts

@@ -0,0 +1,30 @@
+import { differenceInMinutes } from "date-fns";
+
+export const TITLE_UPDATE_INTERVAL_MINUTES = 30;
+
+const FORCE_UPDATE_FOR_TEST = false;
+
+export const shouldUpdateTitle = (
+  createdAt?: string,
+  lastUpdatedAt?: string,
+  now = new Date()
+): boolean => {
+  if (FORCE_UPDATE_FOR_TEST) return true;
+
+  const lastUpdate = lastUpdatedAt ? new Date(lastUpdatedAt) : null;
+  const creation = createdAt ? new Date(createdAt) : null;
+
+  if (
+    !lastUpdate ||
+    !creation ||
+    Number.isNaN(lastUpdate.getTime()) ||
+    Number.isNaN(creation.getTime())
+  ) {
+    return true;
+  }
+
+  return (
+    differenceInMinutes(now, creation) <= TITLE_UPDATE_INTERVAL_MINUTES ||
+    differenceInMinutes(now, lastUpdate) >= TITLE_UPDATE_INTERVAL_MINUTES
+  );
+};