实现聊天流式输出

src/ai/ConversationManager.cpp

@@ -83,6 +83,53 @@ void ConversationManager::sendUserMessage(const QString &message, ResponseCallba
     });
 }
 
+void ConversationManager::sendUserMessageStreaming(const QString &message, StreamCallback streamCallback, ResponseCallback callback)
+{
+    const QString content = message.trimmed();
+    if (content.isEmpty())
+    {
+        if (callback)
+        {
+            callback({false, {}, QStringLiteral("Message is empty."), 0});
+        }
+        return;
+    }
+
+    if (!m_provider)
+    {
+        if (callback)
+        {
+            callback({false, {}, QStringLiteral("AI provider is not ready."), 0});
+        }
+        return;
+    }
+
+    if (isBusy())
+    {
+        if (callback)
+        {
+            callback({false, {}, QStringLiteral("AI request is already running."), 0});
+        }
+        return;
+    }
+
+    const ChatMessage userMessage{QStringLiteral("user"), content};
+    m_provider->sendStreamingChatRequest(
+        buildRequest(userMessage),
+        std::move(streamCallback),
+        [this, userMessage, callback = std::move(callback)](const ChatResponse &response) mutable {
+            if (response.success)
+            {
+                appendExchange(userMessage, {QStringLiteral("assistant"), response.content});
+            }
+
+            if (callback)
+            {
+                callback(response);
+            }
+        });
+}
+
 void ConversationManager::cancel()
 {
     if (m_provider)

src/ai/ConversationManager.h

@@ -11,6 +11,7 @@ class ConversationManager
 {
 public:
     using ResponseCallback = LLMProvider::ResponseCallback;
+    using StreamCallback = LLMProvider::StreamCallback;
 
     ConversationManager();
     ~ConversationManager();
@@ -20,6 +21,7 @@ public:
     QVector<ChatMessage> history() const;
     bool setProvider(std::unique_ptr<LLMProvider> provider);
     void sendUserMessage(const QString &message, ResponseCallback callback);
+    void sendUserMessageStreaming(const QString &message, StreamCallback streamCallback, ResponseCallback callback);
     void cancel();
     void clear();
 

src/ai/LLMProvider.h

@@ -2,16 +2,28 @@
 
 #include "LLMTypes.h"
 
+#include <QtGlobal>
+
 #include <functional>
+#include <utility>
 
 class LLMProvider
 {
 public:
     using ResponseCallback = std::function<void(const ChatResponse &)>;
+    using StreamCallback = std::function<void(const QString &)>;
 
     virtual ~LLMProvider() = default;
 
     virtual bool isBusy() const = 0;
     virtual void sendChatRequest(const ChatRequest &request, ResponseCallback callback) = 0;
+    virtual void sendStreamingChatRequest(
+        const ChatRequest &request,
+        StreamCallback streamCallback,
+        ResponseCallback callback)
+    {
+        Q_UNUSED(streamCallback);
+        sendChatRequest(request, std::move(callback));
+    }
     virtual void cancel() = 0;
 };

src/ai/OpenAICompatibleProvider.cpp

@@ -1,5 +1,7 @@
 #include "OpenAICompatibleProvider.h"
 
+#include "../util/Logger.h"
+
 #include <QJsonArray>
 #include <QJsonDocument>
 #include <QJsonObject>
@@ -8,21 +10,75 @@
 #include <QNetworkRequest>
 #include <QStringList>
 #include <QUrl>
+#include <QUrlQuery>
 
 #include <utility>
 
 namespace
 {
+constexpr int MaxDiagnosticBodyLength = 1000;
+
 QString trimmedResponseBody(const QByteArray &body)
 {
     const QString text = QString::fromUtf8(body).trimmed();
-    constexpr int MaxErrorBodyLength = 1000;
-    if (text.size() <= MaxErrorBodyLength)
+    if (text.size() <= MaxDiagnosticBodyLength)
     {
         return text;
     }
 
-    return text.left(MaxErrorBodyLength) + QStringLiteral("...");
+    return text.left(MaxDiagnosticBodyLength) + QStringLiteral("...");
+}
+
+QString oneLine(QString text)
+{
+    text.replace(QLatin1Char('\r'), QLatin1Char(' '));
+    text.replace(QLatin1Char('\n'), QLatin1Char(' '));
+    text.replace(QLatin1Char('\t'), QLatin1Char(' '));
+    return text.simplified();
+}
+
+bool isSensitiveQueryName(const QString &name)
+{
+    const QString lowerName = name.toLower();
+    return lowerName.contains(QStringLiteral("key"))
+        || lowerName.contains(QStringLiteral("token"))
+        || lowerName.contains(QStringLiteral("secret"))
+        || lowerName.contains(QStringLiteral("password"))
+        || lowerName.contains(QStringLiteral("authorization"))
+        || lowerName.contains(QStringLiteral("signature"));
+}
+
+QString sanitizedUrlString(QUrl url)
+{
+    url.setUserInfo(QString());
+
+    QUrlQuery query(url);
+    if (!query.isEmpty())
+    {
+        QUrlQuery sanitizedQuery;
+        const auto items = query.queryItems(QUrl::FullyDecoded);
+        for (const auto &item : items)
+        {
+            sanitizedQuery.addQueryItem(
+                item.first,
+                isSensitiveQueryName(item.first) ? QStringLiteral("<redacted>") : item.second);
+        }
+        url.setQuery(sanitizedQuery);
+    }
+
+    return url.toString(QUrl::FullyEncoded);
+}
+
+QString diagnosticContext(const AIConfig &config, const QUrl &url)
+{
+    return QStringLiteral("provider=%1 protocol=%2 model=%3 url=%4 timeoutMs=%5 maxTokens=%6 temperature=%7")
+        .arg(config.provider.trimmed().isEmpty() ? QStringLiteral("<empty>") : config.provider.trimmed())
+        .arg(config.protocol.trimmed().isEmpty() ? QStringLiteral("<empty>") : config.protocol.trimmed())
+        .arg(config.model.trimmed().isEmpty() ? QStringLiteral("<empty>") : config.model.trimmed())
+        .arg(sanitizedUrlString(url))
+        .arg(config.timeoutMs)
+        .arg(config.maxTokens)
+        .arg(QString::number(config.temperature, 'f', 2));
 }
 
 QString errorMessageFromBody(const QByteArray &body)
@@ -113,6 +169,23 @@ bool OpenAICompatibleProvider::isBusy() const
 }
 
 void OpenAICompatibleProvider::sendChatRequest(const ChatRequest &request, ResponseCallback callback)
+{
+    sendChatRequestInternal(request, false, nullptr, std::move(callback));
+}
+
+void OpenAICompatibleProvider::sendStreamingChatRequest(
+    const ChatRequest &request,
+    StreamCallback streamCallback,
+    ResponseCallback callback)
+{
+    sendChatRequestInternal(request, true, std::move(streamCallback), std::move(callback));
+}
+
+void OpenAICompatibleProvider::sendChatRequestInternal(
+    const ChatRequest &request,
+    bool stream,
+    StreamCallback streamCallback,
+    ResponseCallback callback)
 {
     if (isBusy())
     {
@@ -151,13 +224,34 @@ void OpenAICompatibleProvider::sendChatRequest(const ChatRequest &request, Respo
     }
 
     m_callback = std::move(callback);
+    m_streamCallback = std::move(streamCallback);
+    m_streaming = stream;
+    m_streamDone = false;
+    m_streamBuffer.clear();
+    m_streamRawBody.clear();
+    m_streamedContent.clear();
 
-    QNetworkRequest networkRequest(requestUrl());
+    const QUrl url = requestUrl();
+    QNetworkRequest networkRequest(url);
     networkRequest.setHeader(QNetworkRequest::ContentTypeHeader, QStringLiteral("application/json"));
     networkRequest.setRawHeader("Authorization", QByteArray("Bearer ") + m_config.apiKey.toUtf8());
 
-    const QJsonDocument document(buildPayload(request));
-    m_currentReply = m_networkManager.post(networkRequest, document.toJson(QJsonDocument::Compact));
+    const QJsonDocument document(buildPayload(request, stream));
+    const QByteArray payload = document.toJson(QJsonDocument::Compact);
+    Logger::info(QStringLiteral("AI request started: %1 stream=%2 messageCount=%3 payloadBytes=%4")
+        .arg(diagnosticContext(m_config, url))
+        .arg(stream ? QStringLiteral("true") : QStringLiteral("false"))
+        .arg(QString::number(request.messages.size()))
+        .arg(QString::number(payload.size())));
+
+    m_currentReply = m_networkManager.post(networkRequest, payload);
+    if (stream)
+    {
+        m_readyReadConnection = QObject::connect(m_currentReply, &QNetworkReply::readyRead, [this]() {
+            handleStreamReadyRead();
+        });
+    }
+
     m_replyFinishedConnection = QObject::connect(m_currentReply, &QNetworkReply::finished, [this]() {
         if (m_currentReply.isNull())
         {
@@ -166,7 +260,37 @@ void OpenAICompatibleProvider::sendChatRequest(const ChatRequest &request, Respo
 
         QNetworkReply *reply = m_currentReply;
         const QByteArray body = reply->readAll();
-        ChatResponse response = parseResponse(reply, body);
+        ChatResponse response;
+        if (m_streaming)
+        {
+            if (!body.isEmpty())
+            {
+                m_streamRawBody.append(body);
+            }
+            if (reply->error() == QNetworkReply::NoError && !body.isEmpty())
+            {
+                m_streamBuffer.append(body);
+                processStreamBuffer();
+                if (m_currentReply.isNull())
+                {
+                    return;
+                }
+            }
+            if (reply->error() == QNetworkReply::NoError && !m_streamBuffer.isEmpty())
+            {
+                m_streamBuffer.append('\n');
+                processStreamBuffer();
+                if (m_currentReply.isNull())
+                {
+                    return;
+                }
+            }
+            response = finishStreamingResponse(reply, body);
+        }
+        else
+        {
+            response = parseResponse(reply, body);
+        }
         clearReply();
 
         if (m_callback)
@@ -184,6 +308,11 @@ void OpenAICompatibleProvider::cancel()
 {
     m_callback = nullptr;
     QPointer<QNetworkReply> reply = m_currentReply;
+    if (!reply.isNull())
+    {
+        Logger::info(QStringLiteral("AI request canceled: %1")
+            .arg(diagnosticContext(m_config, reply->request().url())));
+    }
     clearReply();
 
     if (!reply.isNull())
@@ -192,7 +321,7 @@ void OpenAICompatibleProvider::cancel()
     }
 }
 
-QJsonObject OpenAICompatibleProvider::buildPayload(const ChatRequest &request) const
+QJsonObject OpenAICompatibleProvider::buildPayload(const ChatRequest &request, bool stream) const
 {
     QJsonArray messages;
     for (const ChatMessage &message : request.messages)
@@ -206,7 +335,7 @@ QJsonObject OpenAICompatibleProvider::buildPayload(const ChatRequest &request) c
     QJsonObject payload;
     payload.insert(QStringLiteral("model"), m_config.model);
     payload.insert(QStringLiteral("messages"), messages);
-    payload.insert(QStringLiteral("stream"), false);
+    payload.insert(QStringLiteral("stream"), stream);
     payload.insert(QStringLiteral("temperature"), m_config.temperature);
     payload.insert(QStringLiteral("max_tokens"), m_config.maxTokens);
     return payload;
@@ -230,12 +359,113 @@ QUrl OpenAICompatibleProvider::requestUrl() const
     return QUrl(baseUrl + path);
 }
 
+void OpenAICompatibleProvider::handleStreamReadyRead()
+{
+    if (m_currentReply.isNull())
+    {
+        return;
+    }
+
+    const QByteArray chunk = m_currentReply->readAll();
+    m_streamRawBody.append(chunk);
+    m_streamBuffer.append(chunk);
+    processStreamBuffer();
+}
+
+void OpenAICompatibleProvider::processStreamBuffer()
+{
+    while (true)
+    {
+        const int lineEnd = m_streamBuffer.indexOf('\n');
+        if (lineEnd < 0)
+        {
+            return;
+        }
+
+        QByteArray line = m_streamBuffer.left(lineEnd);
+        m_streamBuffer.remove(0, lineEnd + 1);
+        line = line.trimmed();
+        if (line.isEmpty() || line.startsWith(':'))
+        {
+            continue;
+        }
+
+        if (!line.startsWith("data:"))
+        {
+            continue;
+        }
+
+        const QByteArray payload = line.mid(5).trimmed();
+        if (!handleStreamPayload(payload))
+        {
+            return;
+        }
+    }
+}
+
+bool OpenAICompatibleProvider::handleStreamPayload(const QByteArray &payload)
+{
+    if (payload == "[DONE]")
+    {
+        m_streamDone = true;
+        return true;
+    }
+
+    QJsonParseError parseError;
+    const QJsonDocument document = QJsonDocument::fromJson(payload, &parseError);
+    if (parseError.error != QJsonParseError::NoError || !document.isObject())
+    {
+        Logger::warning(QStringLiteral("AI stream chunk JSON parse failed: %1 parseError=\"%2\" chunkBytes=%3")
+            .arg(diagnosticContext(m_config, requestUrl()))
+            .arg(oneLine(parseError.errorString()))
+            .arg(QString::number(payload.size())));
+        return true;
+    }
+
+    const QJsonObject root = document.object();
+    if (root.contains(QStringLiteral("error")))
+    {
+        const QString bodyError = errorMessageFromBody(payload);
+        finishWithError(bodyError.isEmpty() ? QStringLiteral("AI stream returned an error.") : bodyError);
+        return false;
+    }
+
+    const QJsonArray choices = root.value(QStringLiteral("choices")).toArray();
+    if (choices.isEmpty() || !choices.first().isObject())
+    {
+        return true;
+    }
+
+    const QJsonObject choice = choices.first().toObject();
+    const QJsonObject delta = choice.value(QStringLiteral("delta")).toObject();
+    const QString content = delta.value(QStringLiteral("content")).toString();
+    if (content.isEmpty())
+    {
+        return true;
+    }
+
+    m_streamedContent += content;
+    if (m_streamCallback)
+    {
+        m_streamCallback(content);
+    }
+
+    return true;
+}
+
 ChatResponse OpenAICompatibleProvider::parseResponse(QNetworkReply *reply, const QByteArray &body) const
 {
     const int httpStatus = reply->attribute(QNetworkRequest::HttpStatusCodeAttribute).toInt();
     if (reply->error() != QNetworkReply::NoError)
     {
         const QString bodyError = errorMessageFromBody(body);
+        Logger::warning(QStringLiteral("AI request network error: %1 qtError=%2 errorString=\"%3\" httpStatus=%4 body=\"%5\"")
+            .arg(diagnosticContext(m_config, reply->request().url()))
+            .arg(static_cast<int>(reply->error()))
+            .arg(oneLine(reply->errorString()))
+            .arg(httpStatus)
+            .arg(oneLine(trimmedResponseBody(body))));
+
         if (!bodyError.isEmpty())
         {
             return {false, {}, bodyError, httpStatus};
@@ -248,6 +478,11 @@ ChatResponse OpenAICompatibleProvider::parseResponse(QNetworkReply *reply, const
     const QJsonDocument document = QJsonDocument::fromJson(body, &parseError);
     if (parseError.error != QJsonParseError::NoError || !document.isObject())
     {
+        Logger::warning(QStringLiteral("AI response JSON parse failed: %1 httpStatus=%2 parseError=\"%3\" body=\"%4\"")
+            .arg(diagnosticContext(m_config, reply->request().url()))
+            .arg(httpStatus)
+            .arg(oneLine(parseError.errorString()))
+            .arg(oneLine(trimmedResponseBody(body))));
         return {false, {}, QStringLiteral("AI response is not valid JSON."), httpStatus};
     }
 
@@ -255,6 +490,10 @@ ChatResponse OpenAICompatibleProvider::parseResponse(QNetworkReply *reply, const
     const QJsonArray choices = root.value(QStringLiteral("choices")).toArray();
     if (choices.isEmpty() || !choices.first().isObject())
     {
+        Logger::warning(QStringLiteral("AI response has no choices: %1 httpStatus=%2 body=\"%3\"")
+            .arg(diagnosticContext(m_config, reply->request().url()))
+            .arg(httpStatus)
+            .arg(oneLine(trimmedResponseBody(body))));
         return {false, {}, QStringLiteral("AI response has no choices."), httpStatus};
     }
 
@@ -262,15 +501,76 @@ ChatResponse OpenAICompatibleProvider::parseResponse(QNetworkReply *reply, const
     const QString content = message.value(QStringLiteral("content")).toString();
     if (content.isEmpty())
     {
+        Logger::warning(QStringLiteral("AI response content is empty: %1 httpStatus=%2 body=\"%3\"")
+            .arg(diagnosticContext(m_config, reply->request().url()))
+            .arg(httpStatus)
+            .arg(oneLine(trimmedResponseBody(body))));
         return {false, {}, QStringLiteral("AI response content is empty."), httpStatus};
     }
 
+    Logger::info(QStringLiteral("AI request completed: %1 httpStatus=%2 responseChars=%3 responseBytes=%4")
+        .arg(diagnosticContext(m_config, reply->request().url()))
+        .arg(httpStatus)
+        .arg(QString::number(content.size()))
+        .arg(QString::number(body.size())));
+
     return {true, content, {}, httpStatus};
 }
 
+ChatResponse OpenAICompatibleProvider::finishStreamingResponse(QNetworkReply *reply, const QByteArray &body) const
+{
+    const int httpStatus = reply->attribute(QNetworkRequest::HttpStatusCodeAttribute).toInt();
+    const QByteArray diagnosticBody = m_streamedContent.isEmpty()
+        ? (m_streamRawBody.isEmpty() ? body : m_streamRawBody)
+        : QByteArray();
+    if (reply->error() != QNetworkReply::NoError)
+    {
+        const QString bodyError = errorMessageFromBody(diagnosticBody);
+        Logger::warning(QStringLiteral("AI streaming request network error: %1 qtError=%2 errorString=\"%3\" httpStatus=%4 body=\"%5\" streamedChars=%6")
+            .arg(diagnosticContext(m_config, reply->request().url()))
+            .arg(static_cast<int>(reply->error()))
+            .arg(oneLine(reply->errorString()))
+            .arg(httpStatus)
+            .arg(oneLine(trimmedResponseBody(diagnosticBody)))
+            .arg(QString::number(m_streamedContent.size())));
+
+        if (!bodyError.isEmpty())
+        {
+            return {false, {}, bodyError, httpStatus};
+        }
+
+        return {false, {}, reply->errorString(), httpStatus};
+    }
+
+    if (m_streamedContent.isEmpty())
+    {
+        Logger::warning(QStringLiteral("AI streaming response content is empty: %1 httpStatus=%2 streamDone=%3 residualBytes=%4 body=\"%5\"")
+            .arg(diagnosticContext(m_config, reply->request().url()))
+            .arg(httpStatus)
+            .arg(m_streamDone ? QStringLiteral("true") : QStringLiteral("false"))
+            .arg(QString::number(m_streamBuffer.size()))
+            .arg(oneLine(trimmedResponseBody(diagnosticBody))));
+        return {false, {}, QStringLiteral("AI streaming response content is empty."), httpStatus};
+    }
+
+    Logger::info(QStringLiteral("AI streaming request completed: %1 httpStatus=%2 streamDone=%3 responseChars=%4")
+        .arg(diagnosticContext(m_config, reply->request().url()))
+        .arg(httpStatus)
+        .arg(m_streamDone ? QStringLiteral("true") : QStringLiteral("false"))
+        .arg(QString::number(m_streamedContent.size())));
+
+    return {true, m_streamedContent, {}, httpStatus};
+}
+
 void OpenAICompatibleProvider::finishWithError(const QString &message, int httpStatus)
 {
     QPointer<QNetworkReply> reply = m_currentReply;
+    const QUrl url = reply.isNull() ? requestUrl() : reply->request().url();
+    Logger::warning(QStringLiteral("AI request finished with error: %1 httpStatus=%2 error=\"%3\"")
+        .arg(diagnosticContext(m_config, url))
+        .arg(httpStatus)
+        .arg(oneLine(message)));
+
     clearReply();
 
     if (!reply.isNull())
@@ -294,10 +594,22 @@ void OpenAICompatibleProvider::clearReply()
         QObject::disconnect(m_replyFinishedConnection);
         m_replyFinishedConnection = {};
     }
+    if (m_readyReadConnection)
+    {
+        QObject::disconnect(m_readyReadConnection);
+        m_readyReadConnection = {};
+    }
 
     if (!m_currentReply.isNull())
     {
         m_currentReply->deleteLater();
         m_currentReply.clear();
     }
+
+    m_streamCallback = nullptr;
+    m_streamBuffer.clear();
+    m_streamRawBody.clear();
+    m_streamedContent.clear();
+    m_streaming = false;
+    m_streamDone = false;
 }

src/ai/OpenAICompatibleProvider.h

@@ -3,6 +3,7 @@
 #include "LLMProvider.h"
 #include "../config/AIConfig.h"
 
+#include <QByteArray>
 #include <QJsonObject>
 #include <QMetaObject>
 #include <QNetworkAccessManager>
@@ -20,12 +21,25 @@ public:
 
     bool isBusy() const override;
     void sendChatRequest(const ChatRequest &request, ResponseCallback callback) override;
+    void sendStreamingChatRequest(
+        const ChatRequest &request,
+        StreamCallback streamCallback,
+        ResponseCallback callback) override;
     void cancel() override;
 
 private:
-    QJsonObject buildPayload(const ChatRequest &request) const;
+    QJsonObject buildPayload(const ChatRequest &request, bool stream) const;
     QUrl requestUrl() const;
     ChatResponse parseResponse(QNetworkReply *reply, const QByteArray &body) const;
+    void sendChatRequestInternal(
+        const ChatRequest &request,
+        bool stream,
+        StreamCallback streamCallback,
+        ResponseCallback callback);
+    void handleStreamReadyRead();
+    void processStreamBuffer();
+    bool handleStreamPayload(const QByteArray &payload);
+    ChatResponse finishStreamingResponse(QNetworkReply *reply, const QByteArray &body) const;
     void finishWithError(const QString &message, int httpStatus = 0);
     void clearReply();
 
@@ -33,6 +47,13 @@ private:
     QNetworkAccessManager m_networkManager;
     QPointer<QNetworkReply> m_currentReply;
     QMetaObject::Connection m_replyFinishedConnection;
+    QMetaObject::Connection m_readyReadConnection;
     QTimer m_timeoutTimer;
     ResponseCallback m_callback;
+    StreamCallback m_streamCallback;
+    QByteArray m_streamBuffer;
+    QByteArray m_streamRawBody;
+    QString m_streamedContent;
+    bool m_streaming = false;
+    bool m_streamDone = false;
 };

src/ui/ChatBubble.cpp

@@ -69,14 +69,13 @@ ChatBubble::ChatBubble(QWidget *parent)
     });
 }
 
-void ChatBubble::showMessage(const QString &message, const QPoint &anchorPosition, int durationMs)
+void ChatBubble::showMessage(const QString &message, const QPoint &anchorPosition, int durationMs, bool scrollToBottom)
 {
     m_dismissOnExternalInteraction = false;
     m_anchorPosition = anchorPosition;
     m_autoHideDurationMs = durationMs;
     const QString trimmed = message.trimmed();
     m_textEdit->setPlainText(trimmed);
-    m_textEdit->verticalScrollBar()->setValue(0);
 
     const QSize bubbleSize = preferredBubbleSize(trimmed);
     m_textEdit->setFixedSize(bubbleSize);
@@ -84,10 +83,17 @@ void ChatBubble::showMessage(const QString &message, const QPoint &anchorPositio
     updatePosition();
     show();
 
+    m_textEdit->verticalScrollBar()->setValue(
+        scrollToBottom ? m_textEdit->verticalScrollBar()->maximum() : 0);
+
     if (durationMs > 0)
     {
         m_hideTimer.start(durationMs);
     }
+    else
+    {
+        m_hideTimer.stop();
+    }
 }
 
 void ChatBubble::updateAnchorPosition(const QPoint &anchorPosition)

src/ui/ChatBubble.h

@@ -11,7 +11,11 @@ class ChatBubble : public QWidget
 public:
     explicit ChatBubble(QWidget *parent = nullptr);
 
-    void showMessage(const QString &message, const QPoint &anchorPosition, int durationMs = 10000);
+    void showMessage(
+        const QString &message,
+        const QPoint &anchorPosition,
+        int durationMs = 10000,
+        bool scrollToBottom = false);
     void updateAnchorPosition(const QPoint &anchorPosition);
     void setDismissOnExternalInteraction(bool enabled);
     void resetAutoHideTimer();

src/ui/PetWindow.cpp

@@ -44,6 +44,7 @@ QString previewImagePath()
 
 constexpr int MaxUserMessageLength = 4000;
 constexpr int ChatInputLowerOffsetY = 48;
+constexpr int StreamBubbleUpdateIntervalMs = 80;
 
 bool populateRuntimeApiKey(AIConfig &config, QString *errorMessage)
 {
@@ -163,6 +164,11 @@ PetWindow::PetWindow(QWidget *parent)
         returnToIdleFromBehavior();
     });
 
+    m_streamBubbleUpdateTimer.setSingleShot(true);
+    connect(&m_streamBubbleUpdateTimer, &QTimer::timeout, this, [this]() {
+        flushStreamingBubble(false);
+    });
+
     QPointer<PetWindow> window(this);
     m_chatInputDialog->setSubmitCallback([window](const QString &message) {
         return !window.isNull() && window->submitChatMessage(message);
@@ -231,14 +237,7 @@ void PetWindow::contextMenuEvent(QContextMenuEvent *event)
     topAction->setCheckable(true);
     topAction->setChecked(m_alwaysOnTop);
 
-    QMenu *bubbleTestMenu = menu.addMenu(QStringLiteral("气泡测试"));
-    QAction *shortBubbleAction = bubbleTestMenu->addAction(QStringLiteral("短文本"));
-    QAction *maxBubbleAction = bubbleTestMenu->addAction(QStringLiteral("接近最大尺寸"));
-    QAction *scrollBubbleAction = bubbleTestMenu->addAction(QStringLiteral("超长滚动文本"));
-
     const bool aiRequestRunning = hasActiveAIRequest();
-    QAction *aiTestAction = menu.addAction(QStringLiteral("AI 测试"));
-    aiTestAction->setEnabled(!aiRequestRunning);
     QAction *chatAction = menu.addAction(QStringLiteral("聊天"));
     chatAction->setEnabled(!aiRequestRunning);
     QAction *showConversationAction = menu.addAction(QStringLiteral("显示对话"));
@@ -258,22 +257,6 @@ void PetWindow::contextMenuEvent(QContextMenuEvent *event)
     {
         setAlwaysOnTop(!m_alwaysOnTop);
     }
-    else if (selectedAction == shortBubbleAction)
-    {
-        showBubbleMessage(QStringLiteral("收到，马上处理。"));
-    }
-    else if (selectedAction == maxBubbleAction)
-    {
-        showBubbleMessage(QStringLiteral("这是一段用于测试气泡最大显示区域附近表现的文本。它应该自动换行，并在不出现滚动条的情况下尽量接近最大宽度和高度，方便观察边距、圆角、阴影和整体位置是否自然。"));
-    }
-    else if (selectedAction == scrollBubbleAction)
-    {
-        showBubbleMessage(QStringLiteral("这是一段用于测试超长 AI 回复的文本。第一段内容会让气泡快速达到最大高度，并触发垂直滚动条。用户应该可以通过滚动条查看后续内容，同时气泡本身不能继续无限变高。第二段内容继续补充更多文字，用来确认滚动区域、文本换行、右侧滚动条样式和顶部位置是否正常。第三段内容模拟模型输出较长解释时的情况：文本仍然需要保持清晰可读，不能遮挡整个桌面，也不能因为内容太多导致窗口尺寸失控。第四段内容用于确认自动隐藏计时仍然生效，并且再次打开气泡时滚动条会回到顶部。"));
-    }
-    else if (selectedAction == aiTestAction)
-    {
-        startAITest();
-    }
     else if (selectedAction == chatAction)
     {
         startChat();
@@ -309,65 +292,8 @@ void PetWindow::contextMenuEvent(QContextMenuEvent *event)
     }
 }
 
-void PetWindow::startAITest()
-{
-    if (m_aiTestProvider && m_aiTestProvider->isBusy())
-    {
-        showBubbleMessage(QStringLiteral("AI 测试请求正在进行。"));
-        return;
-    }
-
-    if (m_conversationManager && m_conversationManager->isBusy())
-    {
-        showBubbleMessage(QStringLiteral("AI 回复正在进行。"));
-        return;
-    }
-
-    ConfigManager configManager;
-    AIConfig config = configManager.loadAIConfig();
-    QString errorMessage;
-    if (!prepareRuntimeAIConfig(config, &errorMessage))
-    {
-        playState(QStringLiteral("error"), false);
-        showBubbleMessage(errorMessage);
-        return;
-    }
-
-    ChatRequest request;
-    request.messages.append({QStringLiteral("system"), QStringLiteral("你是桌宠的最小连通性测试助手。")});
-    request.messages.append({QStringLiteral("user"), QStringLiteral("请用一句话回复测试成功。")});
-
-    m_aiTestProvider = std::make_unique<OpenAICompatibleProvider>(config);
-    playState(QStringLiteral("think"), false);
-    showBubbleMessage(QStringLiteral("正在测试 AI 连接..."));
-
-    QPointer<PetWindow> window(this);
-    m_aiTestProvider->sendChatRequest(request, [window](const ChatResponse &response) {
-        if (window.isNull())
-        {
-            return;
-        }
-
-        if (response.success)
-        {
-            window->playState(QStringLiteral("talk"), false);
-            window->showBubbleMessage(response.content);
-            return;
-        }
-
-        window->playState(QStringLiteral("error"), false);
-        window->showBubbleMessage(QStringLiteral("AI 测试失败：") + userVisibleErrorMessage(response));
-    });
-}
-
 void PetWindow::startChat()
 {
-    if (m_aiTestProvider && m_aiTestProvider->isBusy())
-    {
-        showBubbleMessage(QStringLiteral("AI 测试请求正在进行。"));
-        return;
-    }
-
     if (!m_chatInputDialog)
     {
         return;
@@ -378,12 +304,6 @@ void PetWindow::startChat()
 
 bool PetWindow::submitChatMessage(const QString &message)
 {
-    if (m_aiTestProvider && m_aiTestProvider->isBusy())
-    {
-        showBubbleMessage(QStringLiteral("AI 测试请求正在进行。"));
-        return false;
-    }
-
     if (!m_conversationManager || m_conversationManager->isBusy())
     {
         showBubbleMessage(QStringLiteral("AI 回复正在进行。"));
@@ -420,26 +340,39 @@ bool PetWindow::submitChatMessage(const QString &message)
     }
 
     playState(QStringLiteral("think"), false);
-    showBubbleMessage(QStringLiteral("正在思考..."));
+    m_streamingAssistantText.clear();
+    m_streamBubbleUpdateTimer.stop();
+    m_chatBubble->showMessage(QStringLiteral("正在思考..."), bubbleAnchorPosition(), 0);
 
     QPointer<PetWindow> window(this);
-    m_conversationManager->sendUserMessage(message, [window](const ChatResponse &response) {
-        if (window.isNull())
-        {
-            return;
-        }
+    m_conversationManager->sendUserMessageStreaming(
+        trimmedMessage,
+        [window](const QString &delta) {
+            if (!window.isNull())
+            {
+                window->handleChatStreamDelta(delta);
+            }
+        },
+        [window](const ChatResponse &response) {
+            if (window.isNull())
+            {
+                return;
+            }
 
-        if (response.success)
-        {
-            window->playState(QStringLiteral("talk"), false);
-            window->showBubbleMessage(response.content);
-            window->refreshChatHistoryPanel();
-            return;
-        }
+            window->m_streamBubbleUpdateTimer.stop();
+            if (response.success)
+            {
+                window->m_streamingAssistantText = response.content;
+                window->flushStreamingBubble(true);
+                window->playState(QStringLiteral("talk"), false);
+                window->refreshChatHistoryPanel();
+                return;
+            }
 
-        window->playState(QStringLiteral("error"), false);
-        window->showBubbleMessage(QStringLiteral("AI 回复失败：") + userVisibleErrorMessage(response));
-    });
+            window->m_streamingAssistantText.clear();
+            window->playState(QStringLiteral("error"), false);
+            window->showBubbleMessage(QStringLiteral("AI 回复失败：") + userVisibleErrorMessage(response));
+        });
 
     return true;
 }
@@ -452,12 +385,9 @@ void PetWindow::clearConversation()
     }
 
     const bool hadActiveRequest = hasActiveAIRequest();
-    if (m_aiTestProvider && m_aiTestProvider->isBusy())
-    {
-        m_aiTestProvider->cancel();
-    }
-
     m_conversationManager->clear();
+    m_streamBubbleUpdateTimer.stop();
+    m_streamingAssistantText.clear();
     refreshChatHistoryPanel();
     showBubbleMessage(hadActiveRequest
         ? QStringLiteral("已取消 AI 请求，并清空对话。")
@@ -467,33 +397,22 @@ void PetWindow::clearConversation()
 
 void PetWindow::cancelActiveAIRequest()
 {
-    bool canceled = false;
-    if (m_aiTestProvider && m_aiTestProvider->isBusy())
-    {
-        m_aiTestProvider->cancel();
-        canceled = true;
-    }
-
     if (m_conversationManager && m_conversationManager->isBusy())
     {
         m_conversationManager->cancel();
-        canceled = true;
-    }
-
-    if (!canceled)
-    {
-        showBubbleMessage(QStringLiteral("没有正在进行的 AI 请求。"));
+        m_streamBubbleUpdateTimer.stop();
+        m_streamingAssistantText.clear();
+        showBubbleMessage(QStringLiteral("AI 请求已取消。"));
+        playState(QStringLiteral("idle"), false);
         return;
     }
 
-    showBubbleMessage(QStringLiteral("AI 请求已取消。"));
-    playState(QStringLiteral("idle"), false);
+    showBubbleMessage(QStringLiteral("没有正在进行的 AI 请求。"));
 }
 
 bool PetWindow::hasActiveAIRequest() const
 {
-    return (m_aiTestProvider && m_aiTestProvider->isBusy())
-        || (m_conversationManager && m_conversationManager->isBusy());
+    return m_conversationManager && m_conversationManager->isBusy();
 }
 
 void PetWindow::showConversationHistory()
@@ -514,6 +433,39 @@ void PetWindow::refreshChatHistoryPanel()
     m_chatHistoryPanel->setMessages(history);
 }
 
+void PetWindow::handleChatStreamDelta(const QString &delta)
+{
+    if (delta.isEmpty())
+    {
+        return;
+    }
+
+    m_streamingAssistantText += delta;
+    if (!isVisible())
+    {
+        return;
+    }
+
+    if (!m_streamBubbleUpdateTimer.isActive())
+    {
+        m_streamBubbleUpdateTimer.start(StreamBubbleUpdateIntervalMs);
+    }
+}
+
+void PetWindow::flushStreamingBubble(bool finalUpdate)
+{
+    if (!isVisible() || m_streamingAssistantText.trimmed().isEmpty())
+    {
+        return;
+    }
+
+    m_chatBubble->showMessage(
+        m_streamingAssistantText,
+        bubbleAnchorPosition(),
+        finalUpdate ? 10000 : 0,
+        true);
+}
+
 void PetWindow::resetBubbleAutoHideTimer()
 {
     if (m_chatBubble)
@@ -529,6 +481,7 @@ QPoint PetWindow::chatInputAnchorPosition() const
 
 void PetWindow::hideEvent(QHideEvent *event)
 {
+    m_streamBubbleUpdateTimer.stop();
     if (m_chatBubble)
     {
         m_chatBubble->hideBubble();

src/ui/PetWindow.h

@@ -21,7 +21,6 @@ class ChatBubble;
 class ChatHistoryPanel;
 class ChatInputDialog;
 class ConversationManager;
-class LLMProvider;
 class PetView;
 
 class PetWindow : public QWidget
@@ -49,13 +48,14 @@ private:
     void loadInitialImage();
     void buildAnimationClips();
     void addStateTestActions(QMenu *menu);
-    void startAITest();
     void startChat();
     bool submitChatMessage(const QString &message);
     void clearConversation();
     void cancelActiveAIRequest();
     void showConversationHistory();
     void refreshChatHistoryPanel();
+    void handleChatStreamDelta(const QString &delta);
+    void flushStreamingBubble(bool finalUpdate);
     bool hasActiveAIRequest() const;
     void resetBubbleAutoHideTimer();
     QPoint chatInputAnchorPosition() const;
@@ -75,15 +75,16 @@ private:
     std::unique_ptr<ChatHistoryPanel> m_chatHistoryPanel;
     std::unique_ptr<ChatInputDialog> m_chatInputDialog;
     std::unique_ptr<ConversationManager> m_conversationManager;
-    std::unique_ptr<LLMProvider> m_aiTestProvider;
     PetView *m_petView;
     QTimer m_idleBehaviorTimer;
     QTimer m_behaviorReturnTimer;
+    QTimer m_streamBubbleUpdateTimer;
     CharacterPackage m_characterPackage;
     QMap<QString, AnimationClip> m_clips;
     FrameAnimator m_frameAnimator;
     PetStateMachine m_stateMachine;
     QPoint m_dragOffset;
+    QString m_streamingAssistantText;
     bool m_dragging;
     bool m_alwaysOnTop;
     bool m_centerNextFrame;