fix(compression): 修复上下文压缩边界

说明：
- 以 compressed_until_message_id 解析压缩边界，并保留旧 marker fallback
- 让手动和自动压缩统一写入边界并发送 typed 压缩事件
- 将前端压缩 loading 按会话隔离，并改用后端上下文用量

操作：
- 新增 get_context_usage Tauri command，输入框 token 环优先使用服务端统计

Closes #103

Author: Licoy

src-tauri/src/commands/conversations.rs

@@ -374,4 +374,11 @@ mod tests {
         assert_eq!(trimmed.len(), 1);
         assert_eq!(trimmed[0].role, "user");
     }
+
+    #[test]
+    fn deepseek_v4_flash_budget_does_not_auto_compress_below_threshold() {
+        let history = vec![text_message("user", &"token ".repeat(100_000))];
+
+        assert!(!should_auto_compress(&[], &history, Some(1_000_000)));
+    }
 }

src-tauri/src/context_manager.rs

@@ -269,6 +269,7 @@ pub fn run() {
         commands::conversations::send_system_message,
         commands::conversations::compress_context,
         commands::conversations::get_compression_summary,
+        commands::conversations::get_context_usage,
         commands::conversations::delete_compression,
         commands::conversations::regenerate_conversation_title,
         // conversation categories

src-tauri/src/lib.rs

@@ -91,7 +91,7 @@ import { formatChatTime } from './chatTime';
 import { invoke } from '@/lib/invoke';
 import { registerHighlight } from 'stream-markdown';
 import { useResolvedAvatarSrc } from '@/hooks/useResolvedAvatarSrc';
-import type { Message, Attachment, ConversationStats } from '@/types';
+import type { Message, Attachment, ConversationStats, ConversationSummary } from '@/types';
 
 // ── markstream-react custom thinking component ──────────────────────────
 
@@ -1999,7 +1999,7 @@ export function ChatView() {
   const loadingOlder = useConversationStore((s) => s.loadingOlder);
   const hasOlderMessages = useConversationStore((s) => s.hasOlderMessages);
   const streaming = useConversationStore((s) => s.streaming);
-  const compressing = useConversationStore((s) => s.compressing);
+  const compressingConversationId = useConversationStore((s) => s.compressingConversationId);
   const streamingMessageId = useConversationStore((s) => s.streamingMessageId);
   const streamActivityByMessageId = useConversationStore((s) => s.streamActivityByMessageId);
   const multiModelParentId = useConversationStore((s) => s.multiModelParentId);
@@ -2022,6 +2022,7 @@ export function ChatView() {
   const deleteCompression = useConversationStore((s) => s.deleteCompression);
   const [summaryModalOpen, setSummaryModalOpen] = useState(false);
   const [summaryModalText, setSummaryModalText] = useState('');
+  const [summaryModalSummary, setSummaryModalSummary] = useState<ConversationSummary | null>(null);
   const [previewPayload, setPreviewPayload] = useState<CodeBlockPreviewPayload | null>(null);
   const [previewModalOpen, setPreviewModalOpen] = useState(false);
   const [mermaidPreviewSvg, setMermaidPreviewSvg] = useState<string | null>(null);
@@ -2072,7 +2073,16 @@ export function ChatView() {
   }, []);
 
   const activeConversation = conversations.find((c) => c.id === activeConversationId);
+  const compressing = compressingConversationId === activeConversationId;
   const isTitleGenerating = activeConversationId != null && titleGeneratingConversationId === activeConversationId;
+  const summaryBoundaryLabel = useMemo(() => {
+    const boundaryId = summaryModalSummary?.compressed_until_message_id;
+    if (!boundaryId) return null;
+    const boundaryIndex = messages.findIndex((message) => message.id === boundaryId);
+    if (boundaryIndex < 0) return boundaryId.slice(0, 8);
+    const boundaryMessage = messages[boundaryIndex];
+    return `#${boundaryIndex + 1} - ${formatChatTime(boundaryMessage.created_at)}`;
+  }, [messages, summaryModalSummary?.compressed_until_message_id]);
 
   const renderConvIconForChat = useCallback((size: number, modelId?: string | null) => {
     if (!activeConversation) return <Avatar icon={<Bot size={16} />} style={{ background: token.colorPrimary }} size={size} />;
@@ -3759,6 +3769,7 @@ export function ChatView() {
                 if (!convId) return;
                 const summary = await getCompressionSummary(convId);
                 setSummaryModalText(summary?.summary_text ?? t('chat.noSummary'));
+                setSummaryModalSummary(summary ?? null);
                 setSummaryModalOpen(true);
               }}
             >
@@ -4113,11 +4124,27 @@ export function ChatView() {
       <Modal
         title={t('chat.compressionSummary')}
         open={summaryModalOpen}
-        onCancel={() => setSummaryModalOpen(false)}
+        onCancel={() => {
+          setSummaryModalOpen(false);
+          setSummaryModalSummary(null);
+        }}
         footer={null}
         width={640}
       >
         <div style={{ maxHeight: 480, overflow: 'auto', padding: '8px 0' }}>
+          {summaryModalSummary && (
+            <div style={{ display: 'flex', flexWrap: 'wrap', gap: 6, marginBottom: 12 }}>
+              {summaryModalSummary.model_used && (
+                <Tag>{t('chat.summaryModel', '模型')}: {summaryModalSummary.model_used}</Tag>
+              )}
+              {summaryModalSummary.token_count != null && (
+                <Tag>{t('chat.summaryTokens', '摘要 tokens')}: {summaryModalSummary.token_count.toLocaleString()}</Tag>
+              )}
+              {summaryBoundaryLabel && (
+                <Tag>{t('chat.summaryBoundary', '压缩到')}: {summaryBoundaryLabel}</Tag>
+              )}
+            </div>
+          )}
           <NodeRenderer
             content={summaryModalText}
             isDark={isDarkMode}

src/components/chat/ChatView.tsx

@@ -115,9 +115,10 @@ export function InputArea() {
 
   const { message: messageApi, modal } = App.useApp();
   const streaming = useConversationStore((s) => s.streaming);
-  const compressing = useConversationStore((s) => s.compressing);
+  const compressingConversationId = useConversationStore((s) => s.compressingConversationId);
   const cancelCurrentStream = useConversationStore((s) => s.cancelCurrentStream);
   const activeConversationId = useConversationStore((s) => s.activeConversationId);
+  const compressing = compressingConversationId === activeConversationId;
   const sendMessage = useConversationStore((s) => s.sendMessage);
   const sendAgentMessage = useConversationStore((s) => s.sendAgentMessage);
   const createConversation = useConversationStore((s) => s.createConversation);
@@ -694,23 +695,41 @@ export function InputArea() {
   );
 
   // Context token usage calculation
-  const getCompressionSummary = useConversationStore((s) => s.getCompressionSummary);
-  const [summaryTokenCount, setSummaryTokenCount] = useState<number>(0);
+  const getContextUsage = useConversationStore((s) => s.getContextUsage);
+  const [serverContextUsage, setServerContextUsage] = useState<{
+    usedTokens: number;
+    maxTokens: number;
+    percent: number;
+  } | null>(null);
 
   useEffect(() => {
-    if (!activeConversationId || !activeConversation?.context_compression) {
-      setSummaryTokenCount(0);
+    if (!activeConversationId) {
+      setServerContextUsage(null);
       return;
     }
-    getCompressionSummary(activeConversationId).then((s) => {
-      setSummaryTokenCount(s?.token_count ?? 0);
+    let cancelled = false;
+    getContextUsage(activeConversationId).then((usage) => {
+      if (cancelled) return;
+      if (!usage?.max_tokens) {
+        setServerContextUsage(null);
+        return;
+      }
+      setServerContextUsage({
+        usedTokens: usage.used_tokens,
+        maxTokens: usage.max_tokens,
+        percent: Math.min(Math.round((usage.used_tokens / usage.max_tokens) * 100), 100),
+      });
     });
-  }, [activeConversationId, activeConversation?.context_compression, getCompressionSummary, messages]);
+    return () => {
+      cancelled = true;
+    };
+  }, [activeConversationId, getContextUsage, messages]);
 
-  // TODO: Token estimation only considers loaded messages. When hasOlderMessages is true
-  // and no context-clear marker is found, the token estimate will be lower than actual.
-  // A proper fix would require the backend to return total token counts.
+  // Fallback only: local estimation sees loaded messages, so it can undercount
+  // paginated conversations when the backend usage query is unavailable.
   const contextTokenUsage = useMemo(() => {
+    if (serverContextUsage) return serverContextUsage;
+
     const maxTokens = currentModel?.max_tokens;
     if (!maxTokens) return null;
 
@@ -730,12 +749,9 @@ export function InputArea() {
       usedTokens += estimateTokens(activeConversation.system_prompt) + 4;
     }
 
-    // Add summary tokens
-    usedTokens += summaryTokenCount;
-
     const percent = Math.min(Math.round((usedTokens / maxTokens) * 100), 100);
     return { usedTokens, maxTokens, percent };
-  }, [messages, currentModel?.max_tokens, activeConversation?.system_prompt, summaryTokenCount]);
+  }, [messages, currentModel?.max_tokens, activeConversation?.system_prompt, serverContextUsage]);
 
   const { hasRealtimeVoice, hasReasoning, hasVision } = React.useMemo(() => ({
     hasRealtimeVoice: activeConversation
@@ -1689,7 +1705,12 @@ export function InputArea() {
                   </span>
                 }
               >
-                <svg width={size} height={size} style={{ display: 'block', cursor: 'pointer' }}>
+                <svg
+                  aria-label={t('chat.contextTokenUsage', '上下文 tokens')}
+                  width={size}
+                  height={size}
+                  style={{ display: 'block', cursor: 'pointer' }}
+                >
                   <circle cx={r + stroke} cy={r + stroke} r={r} fill="none" stroke={token.colorBorderSecondary} strokeWidth={stroke} />
                   <circle
                     cx={r + stroke} cy={r + stroke} r={r}

src/components/chat/InputArea.tsx

@@ -1,5 +1,5 @@
 import { App } from 'antd';
-import { fireEvent, render, screen } from '@testing-library/react';
+import { fireEvent, render, screen, waitFor } from '@testing-library/react';
 import userEvent from '@testing-library/user-event';
 import { beforeEach, describe, expect, it, vi } from 'vitest';
 import type { AppSettings } from '@/types';
@@ -19,11 +19,13 @@ const toggleMemoryNamespace = vi.fn();
 const setThinkingBudget = vi.fn();
 const setThinkingLevel = vi.fn();
 const insertContextClear = vi.fn();
+const getContextUsage = vi.fn();
 const setActivePage = vi.fn();
 const setSettingsSection = vi.fn();
 
 const conversationState = {
   streaming: false,
+  compressingConversationId: null as string | null,
   activeConversationId: 'conv-1',
   sendMessage,
   createConversation,
@@ -51,6 +53,7 @@ const conversationState = {
   setThinkingBudget,
   setThinkingLevel,
   insertContextClear,
+  getContextUsage,
 };
 
 const providerState = {
@@ -178,6 +181,8 @@ describe('InputArea', () => {
     conversationState.conversations[0].model_id = 'model-1';
     conversationState.thinkingBudget = null;
     conversationState.thinkingLevel = null;
+    conversationState.compressingConversationId = null;
+    getContextUsage.mockResolvedValue(null);
     settingsState.settings.document_attachment_reading_enabled = false;
   });
 
@@ -296,6 +301,28 @@ describe('InputArea', () => {
     expect(screen.queryByText('XHigh')).not.toBeInTheDocument();
   });
 
+  it('shows backend context usage instead of a loaded-message estimate', async () => {
+    getContextUsage.mockResolvedValueOnce({
+      used_tokens: 720000,
+      max_tokens: 1000000,
+      threshold_tokens: 700000,
+      has_summary: true,
+      compressed_until_message_id: 'msg-1',
+      messages_after_boundary: 3,
+    });
+
+    render(
+      <App>
+        <InputArea />
+      </App>,
+    );
+
+    await waitFor(() => expect(getContextUsage).toHaveBeenCalledWith('conv-1'));
+    await userEvent.hover(screen.getByLabelText('上下文 tokens'));
+
+    expect(await screen.findByText('720,000 / 1,000,000 tokens (72%)')).toBeInTheDocument();
+  });
+
   it('shows document attachment controls for non-vision models when document reading is enabled', () => {
     settingsState.settings.document_attachment_reading_enabled = true;
 

src/components/chat/__tests__/InputArea.test.tsx

@@ -79,7 +79,7 @@ function deferred<T>() {
 }
 
 async function flushPromises() {
-  for (let index = 0; index < 8; index += 1) {
+  for (let index = 0; index < 16; index += 1) {
     await Promise.resolve();
   }
 }
@@ -1743,6 +1743,111 @@ describe('conversationStore pagination', () => {
     expect(useConversationStore.getState().enabledMemoryNamespaceIds).toEqual([]);
   });
 
+  it('tracks context compression loading by conversation id', async () => {
+    const summary = {
+      id: 'summary-1',
+      conversation_id: 'conv-1',
+      summary_text: 'compressed context',
+      compressed_until_message_id: 'msg-1',
+      token_count: 12,
+      model_used: 'summary-model',
+      created_at: 1,
+      updated_at: 1,
+    };
+    const compression = deferred<typeof summary>();
+    invokeMock.mockImplementation((cmd: string, args: Record<string, unknown>) => {
+      if (cmd === 'compress_context') {
+        expect(args.conversationId).toBe('conv-1');
+        return compression.promise;
+      }
+      if (cmd === 'list_messages_page') {
+        return Promise.resolve(makePage([makeMessage(1, 'conv-1')], false));
+      }
+      throw new Error(`unexpected command: ${cmd}`);
+    });
+    const { useConversationStore } = await import('../conversationStore');
+
+    useConversationStore.setState({
+      activeConversationId: 'conv-1',
+      conversations: [
+        makeConversation('conv-1'),
+        makeConversation('conv-2'),
+      ] as never[],
+      messages: [makeMessage(1, 'conv-1')],
+    });
+
+    const pending = useConversationStore.getState().compressContext();
+    await flushPromises();
+
+    expect(useConversationStore.getState().compressingConversationId).toBe('conv-1');
+    useConversationStore.setState({ activeConversationId: 'conv-2' });
+    expect(useConversationStore.getState().compressingConversationId).toBe('conv-1');
+
+    compression.resolve(summary);
+    await pending;
+
+    expect(useConversationStore.getState().compressingConversationId).toBeNull();
+  });
+
+  it('applies compression events only to the matching active conversation', async () => {
+    const listeners = new Map<string, (event: { payload: unknown }) => void>();
+    listenMock.mockImplementation(async (eventName: string, handler: (event: { payload: unknown }) => void) => {
+      listeners.set(eventName, handler);
+      return () => {};
+    });
+    const { useConversationStore } = await import('../conversationStore');
+    const marker = {
+      ...makeMessage(50, 'conv-1'),
+      role: 'system',
+      content: '<!-- context-compressed -->',
+    };
+    const otherMarker = {
+      ...makeMessage(60, 'conv-2'),
+      role: 'system',
+      content: '<!-- context-compressed -->',
+    };
+    const summary = {
+      id: 'summary-1',
+      conversation_id: 'conv-1',
+      summary_text: 'compressed context',
+      compressed_until_message_id: 'msg-1',
+      token_count: 12,
+      model_used: 'summary-model',
+      created_at: 1,
+      updated_at: 1,
+    };
+
+    useConversationStore.setState({
+      activeConversationId: 'conv-1',
+      conversations: [
+        makeConversation('conv-1'),
+        makeConversation('conv-2'),
+      ] as never[],
+      messages: [makeMessage(1, 'conv-1')],
+    });
+
+    await useConversationStore.getState().startStreamListening();
+
+    listeners.get('conversation:compressed')?.({
+      payload: {
+        conversation_id: 'conv-2',
+        marker_message: otherMarker,
+        summary: { ...summary, conversation_id: 'conv-2' },
+      },
+    });
+    expect(useConversationStore.getState().messages.map((message) => message.id)).toEqual(['msg-1']);
+
+    listeners.get('conversation:compressed')?.({
+      payload: {
+        conversation_id: 'conv-1',
+        marker_message: marker,
+        summary,
+      },
+    });
+
+    expect(useConversationStore.getState().messages.map((message) => message.id)).toEqual(['msg-1', 'msg-50']);
+  });
+
   it('persists reasoning level changes separately from legacy thinking budget', async () => {
     invokeMock.mockResolvedValueOnce(makeConversation('conv-1', { thinking_budget: 4096, thinking_level: 'high' }));
     const { useConversationStore } = await import('../conversationStore');