zach/z-bar-qt
1
1
Fork 0
Code Issues 24 Pull Requests 2 Actions Packages Projects 1 Releases Wiki Activity

better chat list view + anchor to bottom. sqlite db for chats + LIFO-ordered

Browse Source
This commit is contained in:
zach
2026-08-23 16:37:11 +02:00
parent 37c6a9986e
commit aae3757daf
36 changed files with 2653 additions and 1459 deletions
Download Patch File Download Diff File Expand all files Collapse all files
Plugins/ZShell/Llm/CMakeLists.txt
+5 -2
View File
@@ -2,9 +2,12 @@ qml_module(ZShell-llm
URI ZShell.Llm
SOURCES
chat.hpp chat.cpp
session.hpp session.cpp
message.hpp message.cpp
chatstore.hpp chatstore.cpp
generation.hpp generation.cpp
llmclient.hpp llmclient.cpp
message.hpp message.cpp
messagemodel.hpp messagemodel.cpp
session.hpp session.cpp
LIBRARIES
Qt::Network
Qt::Sql
Plugins/ZShell/Llm/chat.cpp
+102 -609
View File
@@ -2,111 +2,119 @@
#include "config.hpp"
#include "llm.hpp"
#include "llmclient.hpp"
#include <QDateTime>
#include <QDebug>
#include <QJsonArray>
#include <QJsonDocument>
#include <QJsonObject>
#include <QNetworkReply>
#include <QNetworkRequest>
#include <QSet>
#include <QUrl>
namespace ZShell {
namespace ZShell::llm {
QString Chat::titleFrom(const QString& content) {
const QString flat = content.simplified();
if (flat.isEmpty())
return QString();
if (flat.size() <= 48)
return flat;
return flat.left(47) + QStringLiteral("…");
}
QString Chat::completionsPath(const QString& endpoint, const QString& subpath) {
QString base = endpoint.trimmed();
while (base.endsWith('/'))
base.chop(1);
if (!base.endsWith("/v1"))
base += "/v1";
return base + subpath;
}
QString Chat::serverErrorMessage(
const QByteArray& body, const QString& fallback) {
const QJsonDocument doc = QJsonDocument::fromJson(body);
if (doc.isObject()) {
const QJsonObject obj = doc.object();
if (obj.contains("error")) {
const QJsonValue errorValue = obj["error"];
if (errorValue.isObject()) {
const QString message = errorValue.toObject()["message"].toString();
if (!message.isEmpty())
return message;
} else if (!errorValue.toString().isEmpty()) {
return errorValue.toString();
}
}
}
return fallback.isEmpty() ? QStringLiteral("Request to LLM server failed")
: fallback;
}
void Chat::setBusy(Chat* chat, bool value) {
if (chat->m_busy == value)
return;
chat->m_busy = value;
Q_EMIT chat->busyChanged();
}
void Chat::setStreamingChatId(Chat* chat, const QString& id) {
if (chat->m_streamingChatId == id)
return;
chat->m_streamingChatId = id;
Q_EMIT chat->streamingChatIdChanged();
}
Chat::Chat(QObject* parent) : QObject(parent), m_store(new ChatStore(this)) {
Chat::Chat(QObject* parent)
: QObject(parent), m_store(new ChatStore(this)), m_client(new LlmClient(this)) {
if (!config::Config::instance())
new config::Config();
m_store->setLlmClient(m_client);
const auto* llm = config::Config::instance()->llm();
m_endpoint = llm->endpoint();
m_temperature = llm->temperature();
m_model = llm->model();
m_client->setEndpoint(llm->endpoint());
m_client->setModel(llm->model());
m_client->setTemperature(llm->temperature());
connect(llm, &config::Llm::endpointChanged, this, [this, llm]() {
if (m_endpoint != llm->endpoint()) {
m_endpoint = llm->endpoint();
Q_EMIT endpointChanged();
probeContextSize();
}
if (m_model.isEmpty())
refreshModels();
m_client->setEndpoint(llm->endpoint());
});
connect(llm, &config::Llm::modelChanged, this, [this, llm]() {
if (m_model == llm->model())
return;
m_model = llm->model();
Q_EMIT modelChanged();
if (m_model.isEmpty())
refreshModels();
m_client->setModel(llm->model());
});
connect(llm, &config::Llm::temperatureChanged, this, [this, llm]() {
m_client->setTemperature(llm->temperature());
});
connect(m_client, &LlmClient::busyChanged, this, [this]() {
// A fresh run supersedes the previous error.
if (m_client->busy() && !m_lastError.isEmpty()) {
m_lastError.clear();
Q_EMIT lastErrorChanged();
}
Q_EMIT busyChanged();
});
connect(
llm,
&config::Llm::temperatureChanged,
m_client, &LlmClient::endpointChanged, this, &Chat::endpointChanged);
connect(
m_client, &LlmClient::modelChanged, this, &Chat::modelChanged);
connect(
m_client,
&LlmClient::availableModelsChanged,
this,
[this, llm]() { m_temperature = llm->temperature(); });
if (m_model.isEmpty())
refreshModels();
probeContextSize();
&Chat::availableModelsChanged);
connect(
m_client,
&LlmClient::contextSizeChanged,
this,
&Chat::contextSizeChanged);
connect(
m_client,
&LlmClient::streamingChatIdChanged,
this,
&Chat::streamingChatIdChanged);
connect(
m_client,
&LlmClient::errorOccurred,
this,
[this](const QString& message) {
m_lastError = message;
Q_EMIT lastErrorChanged();
Q_EMIT errorOccurred(message);
});
connect(
m_store,
&ChatStore::sessionRemoved,
this,
[this](ChatSession* session) { m_client->sessionRemoved(session); });
connect(
m_client,
&LlmClient::titleSuggested,
this,
[this](ChatSession* session, const QString& title) {
qInfo() << "Chat: applying generated title" << session->id()
<< title << "(was" << session->title() << ")";
session->setTitle(title);
m_store->saveMeta(session);
});
connect(
m_client,
&LlmClient::iconSuggested,
this,
[this](ChatSession* session, const QString& icon) {
qInfo() << "Chat: applying generated icon" << session->id()
<< icon << "(was" << session->icon() << ")";
session->setIcon(icon);
m_store->saveMeta(session);
});
}
Chat::~Chat() {
if (m_reply)
m_reply->abort();
bool Chat::busy() const {
return m_client->busy();
}
QString Chat::endpoint() const {
return m_client->endpoint();
}
QString Chat::model() const {
return m_client->model();
}
QStringList Chat::availableModels() const {
return m_client->availableModels();
}
int Chat::contextSize() const {
return m_client->contextSize();
}
QString Chat::streamingChatId() const {
return m_client->streamingChatId();
}
Chat* Chat::s_instance = nullptr;
@@ -117,283 +125,8 @@ Chat* Chat::create(QQmlEngine*, QJSEngine*) {
return s_instance;
}
void Chat::send(const QString& chatId, const QString& content) {
const QString text = content.trimmed();
if (text.isEmpty() || m_busy)
return;
ChatSession* session = m_store->sessionById(chatId);
if (!session) {
qWarning() << "Chat: unknown chat id" << chatId;
return;
}
if (!m_lastError.isEmpty()) {
m_lastError.clear();
Q_EMIT lastErrorChanged();
}
session->ensureLoaded();
session->appendMessage(
ChatMessage::Role::User,
text,
QDateTime::currentMSecsSinceEpoch());
while (session->messageCount() > 200)
session->removeMessage(session->messages().first());
if (session->title().isEmpty()) {
session->setTitle(titleFrom(text));
qInfo() << "Chat: new session" << chatId
<< "fallback title" << session->title()
<< "- requesting generated title and icon, model" << m_model
<< "endpoint" << m_endpoint;
requestTitle(chatId, text);
requestIcon(chatId, text);
}
if (m_contextSize > 0 && m_lastTokenCount > m_contextSize * 4 / 5)
trimHistory(session, m_contextSize);
m_active = session;
beginAssistant();
m_store->persist(session);
}
void Chat::beginAssistant() {
m_streaming = m_active->appendMessage(
ChatMessage::Role::Assistant,
QString(),
QDateTime::currentMSecsSinceEpoch());
m_streaming->setStreaming(true);
setBusy(this, true);
setStreamingChatId(this, m_active->id());
const QUrl url =
QUrl::fromUserInput(completionsPath(m_endpoint, "/chat/completions"));
if (!url.isValid() || url.host().isEmpty()) {
fail(QStringLiteral("Invalid LLM endpoint: %1").arg(m_endpoint));
return;
}
QNetworkRequest request(url);
request.setHeader(QNetworkRequest::ContentTypeHeader, "application/json");
request.setRawHeader("Accept", "text/event-stream");
QJsonArray messages;
for (const auto* message : m_active->messages()) {
if (message == m_streaming)
continue;
QJsonObject messageObj;
messageObj[QStringLiteral("role")] =
message->role() == ChatMessage::Role::User
? QStringLiteral("user")
: QStringLiteral("assistant");
messageObj[QStringLiteral("content")] = message->content();
if (!message->reasoning().isEmpty())
messageObj[QStringLiteral("reasoning_content")] = message->reasoning();
messages.append(messageObj);
}
QJsonObject body;
QJsonObject streamOptions;
streamOptions[QStringLiteral("include_usage")] = true;
body[QStringLiteral("stream_options")] = streamOptions;
body[QStringLiteral("messages")] = messages;
body[QStringLiteral("stream")] = true;
body[QStringLiteral("temperature")] = m_temperature;
if (!m_model.isEmpty())
body[QStringLiteral("model")] = m_model;
m_buffer.clear();
m_reply = m_manager.post(request, QJsonDocument(body).toJson());
connect(m_reply, &QNetworkReply::readyRead, this, [this]() {
if (m_reply)
m_buffer.append(m_reply->readAll());
drainBuffer();
});
connect(m_reply, &QNetworkReply::finished, this, [this]() {
QNetworkReply* reply = m_reply;
if (!reply)
return;
m_reply = nullptr;
const QNetworkReply::NetworkError error = reply->error();
const QString errorString = reply->errorString();
const QByteArray responseBody = reply->readAll();
reply->deleteLater();
drainBuffer();
if (!m_streaming)
return;
if (error == QNetworkReply::NoError ||
error == QNetworkReply::OperationCanceledError)
finalize();
else
fail(serverErrorMessage(responseBody, errorString));
});
}
void Chat::shortRequest(
const QString& tag,
const QString& systemPrompt,
const QString& userText,
std::function<void(QString result)> onResult) {
const QUrl url =
QUrl::fromUserInput(completionsPath(m_endpoint, "/chat/completions"));
if (!url.isValid() || url.host().isEmpty())
return;
qInfo() << "Chat:" << tag << "request POST" << url.toString()
<< "model=" << m_model;
QNetworkRequest request(url);
request.setHeader(QNetworkRequest::ContentTypeHeader, "application/json");
request.setRawHeader("Accept", "application/json");
QJsonArray messages;
QJsonObject system;
system[QStringLiteral("role")] = QStringLiteral("system");
system[QStringLiteral("content")] = systemPrompt;
messages.append(system);
QJsonObject user;
user[QStringLiteral("role")] = QStringLiteral("user");
user[QStringLiteral("content")] = userText.simplified().mid(0, 512);
messages.append(user);
QJsonObject body;
if (!m_model.isEmpty())
body[QStringLiteral("model")] = m_model;
body[QStringLiteral("stream")] = false;
body[QStringLiteral("temperature")] = 0.3;
body[QStringLiteral("max_tokens")] = 128;
QJsonObject templateKwargs;
templateKwargs[QStringLiteral("enable_thinking")] = false;
body[QStringLiteral("chat_template_kwargs")] = templateKwargs;
body[QStringLiteral("messages")] = messages;
auto* reply = m_manager.post(request, QJsonDocument(body).toJson());
connect(
reply,
&QNetworkReply::finished,
this,
[this, reply, tag, onResult = std::move(onResult)]() {
const QByteArray data = reply->readAll();
reply->deleteLater();
qInfo() << "Chat:" << tag << "request finished"
<< "error=" << reply->error() << reply->errorString()
<< "http=" << reply->attribute(
QNetworkRequest::HttpStatusCodeAttribute).toInt()
<< "response="
<< QString::fromUtf8(data.left(400)).simplified();
if (reply->error() != QNetworkReply::NoError)
return;
const QJsonDocument doc = QJsonDocument::fromJson(data);
const QJsonArray choices =
doc.object()[QStringLiteral("choices")].toArray();
if (choices.isEmpty())
return;
const QString result =
choices.at(0).toObject()[QStringLiteral("message")].toObject()
[QStringLiteral("content")].toString()
.trimmed();
qInfo() << "Chat:" << tag << "raw result" << result;
onResult(result);
});
}
void Chat::requestTitle(const QString& chatId, const QString& userText) {
shortRequest(
QStringLiteral("title"),
QStringLiteral(
"Write a short, concise title for a chat conversation starting "
"with the user's message below. At most six words, no quotation "
"marks, no trailing punctuation. Reply with the title only."),
userText,
[this, chatId](QString title) {
ChatSession* session = m_store->sessionById(chatId);
if (!session) {
qWarning() << "Chat: title request: session gone" << chatId;
return;
}
const auto isQuote = [](QChar c) {
return c == QLatin1Char('"') || c == QLatin1Char('\'') ||
c == QChar(u'\u201C') || c == QChar(u'\u201D') ||
c == QChar(u'\u2018') || c == QChar(u'\u2019');
};
while (title.size() >= 2 && isQuote(title.at(0)) &&
isQuote(title.at(title.size() - 1)))
title = title.mid(1, title.size() - 2).simplified();
while (!title.isEmpty() &&
(title.endsWith(QLatin1Char('.')) ||
title.endsWith(QLatin1Char('!')) ||
title.endsWith(QLatin1Char('?'))))
title.chop(1);
if (title.size() < 2) {
qWarning() << "Chat: title rejected (too short)" << chatId
<< title;
return;
}
if (title.size() > 48)
title = title.left(47) + QStringLiteral("…");
qInfo() << "Chat: applying generated title" << chatId << title
<< "(was" << session->title() << ")";
session->setTitle(title);
m_store->saveMeta(session);
});
}
void Chat::requestIcon(const QString& chatId, const QString& userText) {
const QStringList icons = {
"chat", "lightbulb", "code",
"description", "article", "school",
"work", "build", "science",
"palette", "music_note", "sports_esports",
"takeout_dining", "flight", "photo_camera",
"psychology_alt", "favorite", "savings",
"gamepad", "auto_awesome",
};
const QString prompt =
QStringLiteral(
"Pick the single icon name from this list that best matches the "
"topic of the user's message below: %1. Reply with only the icon "
"name, exactly as written in the list, and nothing else.")
.arg(icons.join(QStringLiteral(", ")));
shortRequest(
QStringLiteral("icon"),
prompt,
userText,
[this, chatId, icons](QString name) {
ChatSession* session = m_store->sessionById(chatId);
if (!session) {
qWarning() << "Chat: icon request: session gone" << chatId;
return;
}
name = name.simplified().toLower();
while (name.size() >= 2 &&
(name.at(0) == QLatin1Char('"') ||
name.at(0) == QLatin1Char('\'')))
name = name.mid(1).left(name.size() - 2).simplified();
name.replace(QLatin1Char(' '), QLatin1Char('_'));
if (!icons.contains(name)) {
qWarning() << "Chat: icon not in list, using default" << chatId
<< name;
name = QStringLiteral("chat");
}
qInfo() << "Chat: applying generated icon" << chatId << name
<< "(was" << session->icon() << ")";
session->setIcon(name);
m_store->saveMeta(session);
});
}
void Chat::stop() {
if (!m_busy)
return;
if (m_reply)
m_reply->abort();
m_client->stop();
}
void Chat::dismissError() {
@@ -403,256 +136,16 @@ void Chat::dismissError() {
Q_EMIT lastErrorChanged();
}
void Chat::clearConversation(const QString& chatId) {
ChatSession* session = m_store->sessionById(chatId);
if (!session)
return;
if (m_busy && m_active == session) {
m_pendingClearChatId = chatId;
stop();
return;
}
session->ensureLoaded();
session->clearMessages();
m_store->persist(session);
}
void Chat::finalize() {
if (!m_streaming)
return;
ChatSession* session = m_active;
endStream();
if (session && m_pendingClearChatId == session->id()) {
session->clearMessages();
m_pendingClearChatId.clear();
}
if (session)
m_store->persist(session);
}
void Chat::endStream() {
if (!m_streaming)
return;
m_streaming->setStreaming(false);
if (m_streaming->content().isEmpty() && m_streaming->reasoning().isEmpty() && m_active)
m_active->removeMessage(m_streaming);
m_streaming = nullptr;
m_active = nullptr;
setBusy(this, false);
setStreamingChatId(this, QString());
}
void Chat::fail(const QString& message) {
qWarning() << "Chat:" << message;
const bool overflow = message.contains("overflow") ||
message.contains("exceed") ||
m_lastTokenCount > m_contextSize;
const QString shown = overflow
? QStringLiteral(
"%1 (using ~%2 of %3 tokens; the oldest messages were auto-removed "
"so the next one will fit)")
.arg(message, QString::number(m_lastTokenCount),
QString::number(m_contextSize))
: message;
m_lastError = shown;
Q_EMIT lastErrorChanged();
if (m_streaming) {
ChatSession* session = m_active;
endStream();
if (overflow)
trimHistory(session, m_contextSize);
if (session && m_pendingClearChatId == session->id()) {
session->clearMessages();
m_pendingClearChatId.clear();
}
if (session)
m_store->persist(session);
}
Q_EMIT errorOccurred(shown);
}
void Chat::drainBuffer() {
while (true) {
const qsizetype newline = m_buffer.indexOf('\n');
if (newline < 0)
break;
const QByteArray line = m_buffer.left(newline).trimmed();
m_buffer.remove(0, newline + 1);
handleLine(line);
}
}
void Chat::handleLine(const QByteArray& line) {
if (line.isEmpty() || line.startsWith('#') || !line.startsWith("data:"))
return;
const QByteArray data = line.mid(5).trimmed();
if (data == "[DONE]") {
finalize();
return;
}
const QJsonDocument doc = QJsonDocument::fromJson(data);
if (!doc.isObject())
return;
const QJsonObject obj = doc.object();
updateTokenUsage(obj);
if (obj.contains("error")) {
const QJsonObject error = obj["error"].toObject();
const QString message = error["message"].toString();
fail(message.isEmpty() ? QStringLiteral("LLM server returned an error") : message);
return;
}
for (const QJsonValue& choiceValue : obj["choices"].toArray()) {
const QJsonObject delta = choiceValue.toObject()["delta"].toObject();
if (!m_streaming)
continue;
m_streaming->appendContent(delta["content"].toString());
QString reasoning = delta["reasoning_content"].toString();
if (reasoning.isEmpty())
reasoning = delta["reasoning"].toString();
m_streaming->appendReasoning(reasoning);
}
}
void Chat::refreshModels() {
const QUrl url = QUrl::fromUserInput(completionsPath(m_endpoint, "/models"));
if (!url.isValid() || url.host().isEmpty())
return;
auto* reply = m_manager.get(QNetworkRequest(url));
connect(reply, &QNetworkReply::finished, this, [this, reply]() {
const QNetworkReply::NetworkError error = reply->error();
const QByteArray data = reply->readAll();
reply->deleteLater();
if (error != QNetworkReply::NoError) {
qWarning() << "Chat: failed to fetch models:" << error;
return;
}
const QJsonArray dataArr =
QJsonDocument::fromJson(data).object()["data"].toArray();
QStringList models;
QSet<QString> seen;
for (const QJsonValue& value : dataArr) {
const QString id = value.toObject()["id"].toString();
if (!id.isEmpty() && !seen.contains(id)) {
seen.insert(id);
models.append(id);
}
}
if (models.isEmpty())
return;
m_availableModels = models;
Q_EMIT availableModelsChanged();
if (m_model.isEmpty()) {
m_model = models.first();
Q_EMIT modelChanged();
}
});
m_client->refreshModels();
}
void Chat::selectModel(const QString& id) {
if (id.isEmpty())
return;
m_model = id;
Q_EMIT modelChanged();
config::Config::instance()->llm()->set_model(id);
m_client->setModel(id);
if (auto* config = config::Config::instance())
config->llm()->set_model(id);
}
void Chat::setContextSize(int size) {
if (size <= 0 || m_contextSize == size)
return;
m_contextSize = size;
Q_EMIT contextSizeChanged();
}
void Chat::probeContextSize() {
QString base = m_endpoint.trimmed();
while (base.endsWith('/'))
base.chop(1);
const QUrl url = QUrl::fromUserInput(base + "/props");
if (!url.isValid() || url.host().isEmpty())
return;
auto* reply = m_manager.get(QNetworkRequest(url));
connect(
reply,
&QNetworkReply::finished,
this,
[this, reply]() {
const QNetworkReply::NetworkError error = reply->error();
const QByteArray data = reply->readAll();
reply->deleteLater();
int size = 0;
if (error == QNetworkReply::NoError) {
const QJsonDocument doc = QJsonDocument::fromJson(data);
if (doc.isArray()) {
for (const auto& value : doc.array()) {
const QJsonObject slot = value.toObject();
if (slot.contains("n_ctx")) {
size = slot["n_ctx"].toInt(0);
if (size > 0)
break;
}
}
} else if (doc.isObject()) {
const QJsonObject obj = doc.object();
size = obj["n_ctx"].toInt(0);
if (size <= 0)
size = obj["default_generation_settings"].toObject()["n_ctx"].toInt(0);
}
}
setContextSize(size > 0 ? size : 4096);
});
}
void Chat::updateTokenUsage(const QJsonObject& data) {
if (m_contextSize <= 0)
return;
const QJsonObject usage = data["usage"].toObject();
if (usage.isEmpty())
return;
const double used = usage.value("prompt_tokens").toDouble() +
usage.value("completion_tokens").toDouble();
if (used > 0)
m_lastTokenCount = qMin(static_cast<int>(used), m_contextSize * 2);
}
void Chat::trimHistory(ChatSession* session, int contextSize) {
if (!session || contextSize <= 0)
return;
const int budget = contextSize * 4 / 5;
auto estimate = [&](const ChatMessage* message) -> qsizetype {
return (message->content().size() +
message->reasoning().size()) /
4;
};
qsizetype total = 0;
for (const auto* message : session->messages())
total += estimate(message);
while (total > budget && session->messageCount() >= 2) {
ChatMessage* first = session->messages().first();
const qsizetype used = estimate(first);
session->removeMessage(first);
total -= used;
if (session->messageCount() >= 2 &&
session->messages().first()->role() == ChatMessage::Role::Assistant) {
ChatMessage* second = session->messages().first();
const qsizetype usedSecond = estimate(second);
session->removeMessage(second);
total -= usedSecond;
}
}
}
} // namespace ZShell
} // namespace ZShell::llm
Plugins/ZShell/Llm/chat.hpp
+15 -61
View File
@@ -1,23 +1,22 @@
#pragma once
#include <QByteArray>
#include <QList>
#include <QNetworkAccessManager>
#include <QObject>
#include <QString>
#include <QStringList>
#include <QtQml>
#include <functional>
#include "chatstore.hpp"
class QQmlEngine;
class QJSEngine;
class QNetworkReply;
namespace ZShell {
namespace ZShell::llm {
class LlmClient;
// QML-facing facade. Persistence lives in ChatStore, network streaming in
// LlmClient; this class only wires them together and exposes the
// application-wide state.
class Chat : public QObject {
Q_OBJECT
QML_ELEMENT
@@ -29,27 +28,22 @@ class Chat : public QObject {
Q_PROPERTY(QStringList availableModels READ availableModels NOTIFY availableModelsChanged)
Q_PROPERTY(int contextSize READ contextSize NOTIFY contextSizeChanged)
Q_PROPERTY(QString lastError READ lastError NOTIFY lastErrorChanged)
Q_PROPERTY(ChatStore* chats READ chats CONSTANT)
Q_PROPERTY(ZShell::llm::ChatStore* chats READ chats CONSTANT)
Q_PROPERTY(QString streamingChatId READ streamingChatId NOTIFY streamingChatIdChanged)
public:
explicit Chat(QObject* parent = nullptr);
~Chat();
[[nodiscard]] bool busy() const { return m_busy; }
[[nodiscard]] QString endpoint() const { return m_endpoint; }
[[nodiscard]] QString model() const { return m_model; }
[[nodiscard]] QStringList availableModels() const { return m_availableModels; }
[[nodiscard]] int contextSize() const { return m_contextSize; }
[[nodiscard]] int lastTokenCount() const { return m_lastTokenCount; }
[[nodiscard]] bool busy() const;
[[nodiscard]] QString endpoint() const;
[[nodiscard]] QString model() const;
[[nodiscard]] QStringList availableModels() const;
[[nodiscard]] int contextSize() const;
[[nodiscard]] QString lastError() const { return m_lastError; }
[[nodiscard]] ChatStore* chats() const { return m_store; }
[[nodiscard]] QString streamingChatId() const { return m_streamingChatId; }
[[nodiscard]] ChatSession* streamingSession() const { return m_active; }
[[nodiscard]] QString streamingChatId() const;
Q_INVOKABLE void send(const QString& chatId, const QString& content);
Q_INVOKABLE void stop();
Q_INVOKABLE void clearConversation(const QString& chatId);
Q_INVOKABLE void dismissError();
Q_INVOKABLE void refreshModels();
Q_INVOKABLE void selectModel(const QString& id);
@@ -67,51 +61,11 @@ class Chat : public QObject {
void streamingChatIdChanged();
private:
void beginAssistant();
void endStream();
void finalize();
void fail(const QString& message);
void handleLine(const QByteArray& line);
void drainBuffer();
void probeContextSize();
void setContextSize(int size);
void updateTokenUsage(const QJsonObject& data);
static void trimHistory(ChatSession* session, int contextSize);
void shortRequest(
const QString& tag,
const QString& systemPrompt,
const QString& userText,
std::function<void(QString result)> onResult);
void requestTitle(const QString& chatId, const QString& userText);
void requestIcon(const QString& chatId, const QString& userText);
static QString titleFrom(const QString& content);
static QString completionsPath(const QString& endpoint, const QString& subpath);
static QString serverErrorMessage(
const QByteArray& body, const QString& fallback);
static void setBusy(Chat* chat, bool value);
static void setStreamingChatId(Chat* chat, const QString& id);
QNetworkAccessManager m_manager;
ChatStore* m_store = nullptr;
QNetworkReply* m_reply = nullptr;
QByteArray m_buffer;
ChatSession* m_active = nullptr;
ChatMessage* m_streaming = nullptr;
QString m_pendingClearChatId;
bool m_busy = false;
QString m_endpoint;
QString m_model;
QStringList m_availableModels;
LlmClient* m_client = nullptr;
QString m_lastError;
QString m_streamingChatId;
double m_temperature = 0.7;
int m_contextSize = 0;
int m_lastTokenCount = 0;
static Chat* s_instance;
friend class ChatStore;
};
} // namespace ZShell
} // namespace ZShell::llm
Plugins/ZShell/Llm/chatstore.cpp
+117 -62
View File
@@ -1,6 +1,6 @@
#include "chatstore.hpp"
#include "chat.hpp"
#include "llmclient.hpp"
#include "message.hpp"
#include <QDateTime>
@@ -15,7 +15,7 @@
#include <algorithm>
namespace ZShell {
namespace ZShell::llm {
ChatStore::ChatStore(QObject* parent)
: QObject(parent), m_connectionName(QUuid::createUuid().toString()) {
@@ -67,38 +67,42 @@ void ChatStore::openDb() {
")"));
}
{
QSqlQuery info(db);
bool hasIcon = false;
if (info.exec(QStringLiteral("PRAGMA table_info(sessions)")))
while (info.next())
if (info.value(1).toString() == QLatin1String("icon")) {
hasIcon = true;
break;
}
if (!hasIcon) {
QSqlQuery alter(db);
if (!alter.exec(QStringLiteral(
"ALTER TABLE sessions ADD COLUMN icon TEXT NOT NULL "
"DEFAULT ''")))
qWarning() << "ChatStore: failed to add icon column:"
<< alter.lastError().text();
}
QSqlQuery query(db);
query.exec(
QStringLiteral(
"CREATE TABLE IF NOT EXISTS messages (\n"
" id INTEGER PRIMARY KEY AUTOINCREMENT,\n"
" session_id TEXT NOT NULL REFERENCES sessions (id) "
"ON DELETE CASCADE,\n"
" role TEXT NOT NULL,\n"
" timestamp INTEGER NOT NULL\n"
")"));
}
{
QSqlQuery query(db);
query.exec(
QStringLiteral(
"CREATE TABLE IF NOT EXISTS messages (\n"
" session_id TEXT NOT NULL REFERENCES sessions (id) ON DELETE "
"CASCADE,\n"
" role TEXT NOT NULL,\n"
" content TEXT,\n"
" reasoning TEXT,\n"
"CREATE TABLE IF NOT EXISTS generations (\n"
" id INTEGER PRIMARY KEY AUTOINCREMENT,\n"
" message_id INTEGER NOT NULL REFERENCES messages "
"(id) ON DELETE CASCADE,\n"
" content TEXT,\n"
" reasoning TEXT,\n"
" timestamp INTEGER NOT NULL,\n"
" reasoning_elapsed_ms INTEGER NOT NULL DEFAULT 0,\n"
" content_elapsed_ms INTEGER NOT NULL DEFAULT 0\n"
" content_elapsed_ms INTEGER NOT NULL DEFAULT 0,\n"
" is_active INTEGER NOT NULL DEFAULT 1\n"
")"));
}
{
QSqlQuery query(db);
query.exec(QStringLiteral(
"CREATE INDEX IF NOT EXISTS idx_messages_session "
"ON messages (session_id)"));
query.exec(QStringLiteral(
"CREATE INDEX IF NOT EXISTS idx_generations_message "
"ON generations (message_id)"));
}
}
int ChatStore::count() const {
@@ -121,7 +125,7 @@ ChatSession* ChatStore::at(int index) const {
ChatSession* ChatStore::insert(int index) {
const qint64 now = QDateTime::currentMSecsSinceEpoch();
const QString id = QString::number(now);
const QString id = QUuid::createUuid().toString();
{
QSqlQuery query(db());
query.prepare(
@@ -136,8 +140,7 @@ ChatSession* ChatStore::insert(int index) {
}
auto* session = new ChatSession(id, this);
session->setMeta(QString(), now, now, 0);
const int pos =
index >= 0 && index <= m_sessions.size() ? index : 0;
const int pos = index >= 0 && index <= m_sessions.size() ? index : 0;
m_sessions.insert(pos, session);
Q_EMIT countChanged();
Q_EMIT valuesChanged();
@@ -155,12 +158,8 @@ void ChatStore::remove(ChatSession* chat) {
void ChatStore::removeSession(ChatSession* session) {
if (!session || !m_sessions.contains(session))
return;
if (auto* chat = qobject_cast<Chat*>(parent()))
if (chat->m_active == session) {
chat->stop();
chat->endStream();
}
const QList<ChatSession*> before = m_sessions;
Q_EMIT sessionRemoved(session);
{
QSqlQuery query(db());
query.prepare("DELETE FROM sessions WHERE id = :id");
@@ -174,7 +173,7 @@ void ChatStore::removeSession(ChatSession* session) {
void ChatStore::move(int from, int to) {
if (from < 0 || from >= m_sessions.size() || to < 0 ||
to >= m_sessions.size() || from == to)
to >= m_sessions.size() || from == to)
return;
m_sessions.move(from, to);
Q_EMIT valuesChanged();
@@ -193,6 +192,10 @@ ChatSession* ChatStore::sessionById(const QString& id) {
return nullptr;
}
void ChatStore::setLlmClient(LlmClient* client) {
m_llmClient = client;
}
void ChatStore::persist(ChatSession* session) {
if (!session || !m_sessions.contains(session))
return;
@@ -243,30 +246,58 @@ bool ChatStore::saveSession(ChatSession* session) {
ok = query.exec();
}
if (ok) {
QSqlQuery insert(handle);
ok = insert.prepare(
"INSERT INTO messages (session_id, role, content, reasoning, "
"timestamp, reasoning_elapsed_ms, content_elapsed_ms) "
"VALUES (:id, :role, :content, :reasoning, :timestamp, "
":reasoning_elapsed_ms, :content_elapsed_ms)");
for (const auto* message : session->messages()) {
insert.bindValue(":id", session->id());
insert.bindValue(
QSqlQuery messageInsert(handle);
ok = messageInsert.prepare(
"INSERT INTO messages (session_id, role, timestamp) "
"VALUES (:id, :role, :timestamp)");
QSqlQuery generationInsert(handle);
ok = ok && generationInsert.prepare(
"INSERT INTO generations (message_id, content, reasoning, "
"timestamp, reasoning_elapsed_ms, content_elapsed_ms, "
"is_active) VALUES (:mid, :content, :reasoning, :timestamp, "
":reasoning_elapsed_ms, :content_elapsed_ms, :is_active)");
// The model holds messages most recent first; the database keeps
// natural rowid order, so iterate from the oldest row up.
const auto* model = session->messagesModel();
for (int row = model->rowCount() - 1; ok && row >= 0; --row) {
const auto* message = model->at(row);
messageInsert.bindValue(":id", session->id());
messageInsert.bindValue(
":role",
message->role() == ChatMessage::Role::User
? QStringLiteral("user")
: QStringLiteral("assistant"));
insert.bindValue(":content", message->content());
insert.bindValue(":reasoning", message->reasoning());
insert.bindValue(":timestamp", message->timestamp());
insert.bindValue(":reasoning_elapsed_ms", message->reasoningElapsedMs());
insert.bindValue(":content_elapsed_ms", message->contentElapsedMs());
if (!insert.exec()) {
messageInsert.bindValue(":timestamp", message->timestamp());
if (!messageInsert.exec()) {
ok = false;
qWarning() << "ChatStore: saveSession" << id << "insert failed:"
<< insert.lastError().text();
qWarning() << "ChatStore: saveSession" << id
<< "message insert failed:"
<< messageInsert.lastError().text();
break;
}
const int messageId = messageInsert.lastInsertId().toInt();
for (int i = 0; ok && i < message->generationCount(); ++i) {
const auto* generation = message->generation(i);
generationInsert.bindValue(":mid", messageId);
generationInsert.bindValue(":content", generation->content());
generationInsert.bindValue(
":reasoning", generation->reasoning());
generationInsert.bindValue(":timestamp", generation->timestamp());
generationInsert.bindValue(
":reasoning_elapsed_ms", generation->reasoningElapsedMs());
generationInsert.bindValue(
":content_elapsed_ms", generation->contentElapsedMs());
generationInsert.bindValue(
":is_active",
i == message->activeGenerationIndex() ? 1 : 0);
if (!generationInsert.exec()) {
ok = false;
qWarning() << "ChatStore: saveSession" << id
<< "generation insert failed:"
<< generationInsert.lastError().text();
break;
}
}
}
}
if (!ok || !handle.commit()) {
@@ -280,28 +311,52 @@ bool ChatStore::saveSession(ChatSession* session) {
}
void ChatStore::loadMessagesInto(ChatSession* session) {
// Newest first so the model receives rows in display order.
QSqlQuery query(db());
query.prepare(
"SELECT role, content, reasoning, timestamp, reasoning_elapsed_ms, "
"content_elapsed_ms FROM messages WHERE session_id = :id ORDER BY rowid");
"SELECT id, role, timestamp FROM messages WHERE session_id = :id "
"ORDER BY rowid DESC");
query.bindValue(":id", session->id());
if (!query.exec()) {
qWarning() << "ChatStore: failed to load messages for" << session->id()
<< ":" << query.lastError().text();
return;
}
auto* model = session->messagesModel();
QList<ChatMessage*> messages;
while (query.next()) {
auto* message = new ChatMessage(
query.value(0).toString() == QLatin1String("user")
const int messageId = query.value(0).toInt();
auto* message = model->createMessage(
query.value(1).toString() == QLatin1String("user")
? ChatMessage::Role::User
: ChatMessage::Role::Assistant,
query.value(1).toString(),
query.value(3).toLongLong(),
session);
message->setReasoning(query.value(2).toString());
message->setElapsedMs(
query.value(4).toLongLong(), query.value(5).toLongLong());
query.value(2).toLongLong());
QSqlQuery generationQuery(db());
generationQuery.prepare(
"SELECT content, reasoning, timestamp, reasoning_elapsed_ms, "
"content_elapsed_ms, is_active FROM generations "
"WHERE message_id = :mid ORDER BY rowid");
generationQuery.bindValue(":mid", messageId);
int activeIndex = 0;
if (generationQuery.exec()) {
int index = 0;
while (generationQuery.next()) {
message->addGeneration(
generationQuery.value(2).toLongLong(),
generationQuery.value(0).toString(),
generationQuery.value(1).toString(),
generationQuery.value(3).toLongLong(),
generationQuery.value(4).toLongLong());
if (generationQuery.value(5).toInt() != 0)
activeIndex = index;
++index;
}
} else {
qWarning() << "ChatStore: failed to load generations for message"
<< messageId << ":"
<< generationQuery.lastError().text();
}
message->setActiveGeneration(activeIndex);
messages.append(message);
}
session->adoptMessages(messages);
@@ -351,4 +406,4 @@ void ChatStore::notify(const QList<ChatSession*>& before) {
Q_EMIT valuesChanged();
}
} // namespace ZShell
} // namespace ZShell::llm
Plugins/ZShell/Llm/chatstore.hpp
+11 -8
View File
@@ -7,9 +7,9 @@
#include <QString>
#include <QVariantList>
namespace ZShell {
namespace ZShell::llm {
class Chat;
class LlmClient;
class ChatStore : public QObject {
Q_OBJECT
@@ -20,19 +20,22 @@ class ChatStore : public QObject {
public:
explicit ChatStore(QObject* parent = nullptr);
~ChatStore();
~ChatStore() override;
[[nodiscard]] int count() const;
[[nodiscard]] QVariantList values() const;
[[nodiscard]] ChatSession* at(int index) const;
Q_INVOKABLE ChatSession* insert(int index = -1);
Q_INVOKABLE ZShell::llm::ChatSession* insert(int index = -1);
Q_INVOKABLE void remove(int index);
Q_INVOKABLE void remove(ChatSession* chat);
Q_INVOKABLE void remove(ZShell::llm::ChatSession* chat);
Q_INVOKABLE void move(int from, int to);
Q_INVOKABLE void clear();
[[nodiscard]] ChatSession* sessionById(const QString& id);
[[nodiscard]] LlmClient* llmClient() const { return m_llmClient; }
void setLlmClient(LlmClient* client);
void persist(ChatSession* session);
void saveMeta(ChatSession* session);
void loadMessagesInto(ChatSession* session);
@@ -40,6 +43,7 @@ class ChatStore : public QObject {
Q_SIGNALS:
void countChanged();
void valuesChanged();
void sessionRemoved(ZShell::llm::ChatSession* session);
private:
void openDb();
@@ -50,11 +54,10 @@ class ChatStore : public QObject {
void notify(const QList<ChatSession*>& before);
QList<ChatSession*> m_sessions;
LlmClient* m_llmClient = nullptr;
QString m_connectionName;
[[nodiscard]] QSqlDatabase db() const;
friend class Chat;
};
} // namespace ZShell
} // namespace ZShell::llm
Plugins/ZShell/Llm/generation.cpp
+126
View File
@@ -0,0 +1,126 @@
#include "generation.hpp"
#include <QDateTime>
namespace ZShell::llm {
ChatGeneration::ChatGeneration(qint64 timestamp, QObject* parent)
: QObject(parent), m_timestamp(timestamp) {
m_timer.setParent(this);
m_timer.setInterval(500);
m_timer.setTimerType(Qt::CoarseTimer);
connect(&m_timer, &QTimer::timeout, this, [this]() {
if (!reasoningInFlight() && !contentInFlight()) {
m_timer.stop();
return;
}
Q_EMIT elapsedMsChanged();
});
}
qint64 ChatGeneration::reasoningElapsedMs() const {
if (m_reasoningStartedAt <= 0)
return 0;
const qint64 end = m_reasoningEndedAt > 0
? m_reasoningEndedAt
: QDateTime::currentMSecsSinceEpoch();
return end - m_reasoningStartedAt;
}
qint64 ChatGeneration::contentElapsedMs() const {
if (m_contentStartedAt <= 0)
return 0;
const qint64 end = m_contentEndedAt > 0
? m_contentEndedAt
: QDateTime::currentMSecsSinceEpoch();
return end - m_contentStartedAt;
}
void ChatGeneration::updateReasoningActive() {
const bool active = m_streaming && m_content.isEmpty();
if (m_reasoningActive == active)
return;
m_reasoningActive = active;
Q_EMIT reasoningActiveChanged();
}
void ChatGeneration::setContent(const QString& value) {
if (m_content == value)
return;
m_content = value;
Q_EMIT contentChanged();
}
void ChatGeneration::setReasoning(const QString& value) {
if (m_reasoning == value)
return;
m_reasoning = value;
Q_EMIT reasoningChanged();
}
void ChatGeneration::appendContent(const QString& piece) {
if (piece.isEmpty())
return;
if (m_content.isEmpty()) {
const qint64 now = QDateTime::currentMSecsSinceEpoch();
if (reasoningInFlight())
m_reasoningEndedAt = now;
m_contentStartedAt = now;
if (!m_timer.isActive())
m_timer.start();
}
m_content += piece;
Q_EMIT contentChanged();
updateReasoningActive();
Q_EMIT elapsedMsChanged();
}
void ChatGeneration::appendReasoning(const QString& piece) {
if (piece.isEmpty())
return;
if (m_reasoning.isEmpty()) {
if (m_reasoningStartedAt <= 0)
m_reasoningStartedAt = QDateTime::currentMSecsSinceEpoch();
if (!m_timer.isActive())
m_timer.start();
}
m_reasoning += piece;
Q_EMIT reasoningChanged();
updateReasoningActive();
Q_EMIT elapsedMsChanged();
}
void ChatGeneration::setElapsedMs(qint64 reasoningMs, qint64 contentMs) {
if (reasoningMs > 0) {
m_reasoningStartedAt = m_timestamp;
m_reasoningEndedAt = m_timestamp + reasoningMs;
}
if (contentMs > 0) {
m_contentStartedAt = m_timestamp;
m_contentEndedAt = m_timestamp + contentMs;
}
}
void ChatGeneration::setStreaming(bool value) {
if (m_streaming == value)
return;
m_streaming = value;
Q_EMIT streamingChanged();
if (value) {
if (m_reasoningStartedAt <= 0)
m_reasoningStartedAt = QDateTime::currentMSecsSinceEpoch();
if (!m_timer.isActive())
m_timer.start();
} else {
const qint64 now = QDateTime::currentMSecsSinceEpoch();
if (reasoningInFlight())
m_reasoningEndedAt = now;
if (contentInFlight())
m_contentEndedAt = now;
m_timer.stop();
Q_EMIT elapsedMsChanged();
}
updateReasoningActive();
}
} // namespace ZShell::llm
Plugins/ZShell/Llm/generation.hpp
+70
View File
@@ -0,0 +1,70 @@
#pragma once
#include <QObject>
#include <QString>
#include <QTimer>
#include <QtQml>
namespace ZShell::llm {
class ChatGeneration : public QObject {
Q_OBJECT
QML_ELEMENT
QML_UNCREATABLE("Chat generations are managed by ChatMessage")
Q_PROPERTY(QString content READ content WRITE setContent NOTIFY contentChanged)
Q_PROPERTY(QString reasoning READ reasoning NOTIFY reasoningChanged)
Q_PROPERTY(qint64 reasoningElapsedMs READ reasoningElapsedMs NOTIFY elapsedMsChanged)
Q_PROPERTY(qint64 contentElapsedMs READ contentElapsedMs NOTIFY elapsedMsChanged)
Q_PROPERTY(bool streaming READ streaming NOTIFY streamingChanged)
Q_PROPERTY(bool reasoningActive READ reasoningActive NOTIFY reasoningActiveChanged)
Q_PROPERTY(qint64 timestamp READ timestamp CONSTANT)
public:
explicit ChatGeneration(qint64 timestamp, QObject* parent = nullptr);
[[nodiscard]] QString content() const { return m_content; }
[[nodiscard]] QString reasoning() const { return m_reasoning; }
[[nodiscard]] bool reasoningActive() const { return m_reasoningActive; }
[[nodiscard]] qint64 reasoningElapsedMs() const;
[[nodiscard]] qint64 contentElapsedMs() const;
[[nodiscard]] bool streaming() const { return m_streaming; }
[[nodiscard]] qint64 timestamp() const { return m_timestamp; }
void setContent(const QString& value);
void setReasoning(const QString& value);
void appendContent(const QString& piece);
void appendReasoning(const QString& piece);
void setElapsedMs(qint64 reasoningMs, qint64 contentMs);
void setStreaming(bool value);
Q_SIGNALS:
void contentChanged();
void reasoningChanged();
void reasoningActiveChanged();
void elapsedMsChanged();
void streamingChanged();
private:
void updateReasoningActive();
[[nodiscard]] bool reasoningInFlight() const {
return m_reasoningStartedAt > 0 && m_reasoningEndedAt <= 0;
}
[[nodiscard]] bool contentInFlight() const {
return m_contentStartedAt > 0 && m_contentEndedAt <= 0;
}
QTimer m_timer;
QString m_content;
QString m_reasoning;
bool m_reasoningActive = false;
bool m_streaming = false;
qint64 m_timestamp;
qint64 m_reasoningStartedAt = 0;
qint64 m_reasoningEndedAt = 0;
qint64 m_contentStartedAt = 0;
qint64 m_contentEndedAt = 0;
};
} // namespace ZShell::llm
Plugins/ZShell/Llm/llmclient.cpp
+568
View File
@@ -0,0 +1,568 @@
#include "llmclient.hpp"
#include "message.hpp"
#include "messagemodel.hpp"
#include "session.hpp"
#include <QDebug>
#include <QJsonArray>
#include <QJsonDocument>
#include <QJsonObject>
#include <QNetworkReply>
#include <QNetworkRequest>
#include <QSet>
#include <QUrl>
namespace ZShell::llm {
QString LlmClient::completionsPath(
const QString& endpoint, const QString& subpath) {
QString base = endpoint.trimmed();
while (base.endsWith('/'))
base.chop(1);
if (!base.endsWith("/v1"))
base += "/v1";
return base + subpath;
}
QString LlmClient::serverErrorMessage(
const QByteArray& body, const QString& fallback) {
const QJsonDocument doc = QJsonDocument::fromJson(body);
if (doc.isObject()) {
const QJsonObject obj = doc.object();
if (obj.contains("error")) {
const QJsonValue errorValue = obj["error"];
if (errorValue.isObject()) {
const QString message =
errorValue.toObject()["message"].toString();
if (!message.isEmpty())
return message;
} else if (!errorValue.toString().isEmpty()) {
return errorValue.toString();
}
}
}
return fallback.isEmpty()
? QStringLiteral("Request to LLM server failed")
: fallback;
}
void LlmClient::setBusy(bool value) {
if (m_busy == value)
return;
m_busy = value;
Q_EMIT busyChanged();
}
void LlmClient::setStreamingChatId(const QString& id) {
if (m_streamingChatId == id)
return;
m_streamingChatId = id;
Q_EMIT streamingChatIdChanged();
}
LlmClient::LlmClient(QObject* parent) : QObject(parent) {
probeContextSize();
}
LlmClient::~LlmClient() {
if (m_reply)
m_reply->abort();
endStream();
}
void LlmClient::setEndpoint(const QString& value) {
if (m_endpoint == value)
return;
m_endpoint = value;
Q_EMIT endpointChanged();
probeContextSize();
if (m_model.isEmpty())
refreshModels();
}
void LlmClient::setModel(const QString& value) {
if (m_model == value)
return;
m_model = value;
Q_EMIT modelChanged();
if (m_model.isEmpty())
refreshModels();
}
void LlmClient::setTemperature(double value) {
m_temperature = value;
}
void LlmClient::startGeneration(
ChatSession* session, ChatGeneration* target) {
if (m_busy || !session || !target)
return;
m_active = session;
m_streaming = target;
m_streaming->setStreaming(true);
setBusy(true);
setStreamingChatId(session->id());
const QUrl url =
QUrl::fromUserInput(completionsPath(m_endpoint, "/chat/completions"));
if (!url.isValid() || url.host().isEmpty()) {
fail(QStringLiteral("Invalid LLM endpoint: %1").arg(m_endpoint));
return;
}
QNetworkRequest request(url);
request.setHeader(QNetworkRequest::ContentTypeHeader, "application/json");
request.setRawHeader("Accept", "text/event-stream");
const auto* model = session->messagesModel();
const int targetRow =
model->rowOf(qobject_cast<ChatMessage*>(target->parent()));
if (targetRow < 0) {
fail(QStringLiteral("Internal error: generation target is not in the "
"session"));
return;
}
// Context: every message older than the target, oldest first.
QJsonArray messages;
for (int row = model->rowCount() - 1; row > targetRow; --row) {
const auto* message = model->at(row);
const auto* generation = message->activeGeneration();
if (!generation)
continue;
QJsonObject messageObj;
messageObj[QStringLiteral("role")] =
message->role() == ChatMessage::Role::User
? QStringLiteral("user")
: QStringLiteral("assistant");
messageObj[QStringLiteral("content")] = generation->content();
if (!generation->reasoning().isEmpty())
messageObj[QStringLiteral("reasoning_content")] =
generation->reasoning();
messages.append(messageObj);
}
QJsonObject body;
QJsonObject streamOptions;
streamOptions[QStringLiteral("include_usage")] = true;
body[QStringLiteral("stream_options")] = streamOptions;
body[QStringLiteral("messages")] = messages;
body[QStringLiteral("stream")] = true;
body[QStringLiteral("temperature")] = m_temperature;
if (!m_model.isEmpty())
body[QStringLiteral("model")] = m_model;
m_buffer.clear();
m_reply = m_manager.post(request, QJsonDocument(body).toJson());
connect(m_reply, &QNetworkReply::readyRead, this, [this]() {
if (m_reply)
m_buffer.append(m_reply->readAll());
drainBuffer();
});
connect(m_reply, &QNetworkReply::finished, this, [this]() {
QNetworkReply* reply = m_reply;
if (!reply)
return;
m_reply = nullptr;
const QNetworkReply::NetworkError error = reply->error();
const QString errorString = reply->errorString();
// Data not already consumed by readyRead is only reachable here.
const QByteArray responseBody = reply->readAll();
m_buffer.append(responseBody);
reply->deleteLater();
drainBuffer();
if (!m_streaming)
return;
if (error == QNetworkReply::NoError ||
error == QNetworkReply::OperationCanceledError)
finalize();
else
fail(serverErrorMessage(responseBody, errorString));
});
}
void LlmClient::shortRequest(
const QString& tag,
const QString& systemPrompt,
const QString& userText,
std::function<void(QString result)> onResult) {
const QUrl url =
QUrl::fromUserInput(completionsPath(m_endpoint, "/chat/completions"));
if (!url.isValid() || url.host().isEmpty())
return;
qInfo() << "LlmClient:" << tag << "request POST" << url.toString()
<< "model=" << m_model;
QNetworkRequest request(url);
request.setHeader(QNetworkRequest::ContentTypeHeader, "application/json");
request.setRawHeader("Accept", "application/json");
QJsonArray messages;
QJsonObject system;
system[QStringLiteral("role")] = QStringLiteral("system");
system[QStringLiteral("content")] = systemPrompt;
messages.append(system);
QJsonObject user;
user[QStringLiteral("role")] = QStringLiteral("user");
user[QStringLiteral("content")] = userText.simplified().mid(0, 512);
messages.append(user);
QJsonObject body;
if (!m_model.isEmpty())
body[QStringLiteral("model")] = m_model;
body[QStringLiteral("stream")] = false;
body[QStringLiteral("temperature")] = 0.3;
body[QStringLiteral("max_tokens")] = 128;
QJsonObject templateKwargs;
templateKwargs[QStringLiteral("enable_thinking")] = false;
body[QStringLiteral("chat_template_kwargs")] = templateKwargs;
body[QStringLiteral("messages")] = messages;
auto* reply = m_manager.post(request, QJsonDocument(body).toJson());
connect(
reply,
&QNetworkReply::finished,
this,
[this, reply, tag, onResult = std::move(onResult)]() {
const QByteArray data = reply->readAll();
reply->deleteLater();
qInfo() << "LlmClient:" << tag << "request finished"
<< "error=" << reply->error() << reply->errorString()
<< "http=" << reply->attribute(
QNetworkRequest::HttpStatusCodeAttribute)
.toInt()
<< "response="
<< QString::fromUtf8(data.left(400)).simplified();
if (reply->error() != QNetworkReply::NoError)
return;
const QJsonDocument doc = QJsonDocument::fromJson(data);
const QJsonArray choices =
doc.object()[QStringLiteral("choices")].toArray();
if (choices.isEmpty())
return;
const QString result =
choices.at(0).toObject()[QStringLiteral("message")].toObject()
[QStringLiteral("content")].toString()
.trimmed();
qInfo() << "LlmClient:" << tag << "raw result" << result;
onResult(result);
});
}
void LlmClient::requestTitle(ChatSession* session, const QString& userText) {
shortRequest(
QStringLiteral("title"),
QStringLiteral(
"Write a short, concise title for a chat conversation starting "
"with the user's message below. At most six words, no quotation "
"marks, no trailing punctuation. Reply with the title only."),
userText,
[this, session = QPointer<ChatSession>(session)](QString title) {
if (!session) {
qWarning() << "LlmClient: title request: session gone";
return;
}
const auto isQuote = [](QChar c) {
return c == QLatin1Char('"') || c == QLatin1Char('\'') ||
c == QChar(u'\u201C') || c == QChar(u'\u201D') ||
c == QChar(u'\u2018') || c == QChar(u'\u2019');
};
while (title.size() >= 2 && isQuote(title.at(0)) &&
isQuote(title.at(title.size() - 1)))
title = title.mid(1, title.size() - 2).simplified();
while (!title.isEmpty() &&
(title.endsWith(QLatin1Char('.')) ||
title.endsWith(QLatin1Char('!')) ||
title.endsWith(QLatin1Char('?'))))
title.chop(1);
if (title.size() < 2) {
qWarning() << "LlmClient: title rejected (too short)"
<< title;
return;
}
if (title.size() > 48)
title = title.left(47) + QStringLiteral("…");
qInfo() << "LlmClient: suggesting title" << title;
Q_EMIT titleSuggested(session, title);
});
}
void LlmClient::requestIcon(ChatSession* session, const QString& userText) {
const QStringList icons = {
"chat", "lightbulb", "code",
"description", "article", "school",
"work", "build", "science",
"palette", "music_note", "sports_esports",
"takeout_dining", "flight", "photo_camera",
"psychology_alt", "favorite", "savings",
"gamepad", "auto_awesome",
};
const QString prompt =
QStringLiteral(
"Pick the single icon name from this list that best matches the "
"topic of the user's message below: %1. Reply with only the icon "
"name, exactly as written in the list, and nothing else.")
.arg(icons.join(QStringLiteral(", ")));
shortRequest(
QStringLiteral("icon"),
prompt,
userText,
[this, session = QPointer<ChatSession>(session), icons](
QString name) {
if (!session) {
qWarning() << "LlmClient: icon request: session gone";
return;
}
name = name.simplified().toLower();
const auto isQuote = [](QChar c) {
return c == QLatin1Char('"') || c == QLatin1Char('\'');
};
while (name.size() >= 2 && isQuote(name.at(0)) &&
isQuote(name.at(name.size() - 1)))
name = name.mid(1, name.size() - 2).simplified();
name.replace(QLatin1Char(' '), QLatin1Char('_'));
if (!icons.contains(name)) {
qWarning() << "LlmClient: icon not in list, using default"
<< name;
name = QStringLiteral("chat");
}
qInfo() << "LlmClient: suggesting icon" << name;
Q_EMIT iconSuggested(session, name);
});
}
void LlmClient::stop() {
if (!m_busy)
return;
if (m_reply)
m_reply->abort();
}
void LlmClient::endStream() {
if (!m_streaming)
return;
auto* generation = m_streaming;
auto* session = m_active;
m_streaming = nullptr;
m_active = nullptr;
generation->setStreaming(false);
if (generation->content().isEmpty() &&
generation->reasoning().isEmpty()) {
if (auto* message = qobject_cast<ChatMessage*>(generation->parent())) {
if (message->generationCount() <= 1) {
if (session)
session->removeMessage(message);
} else {
message->removeGeneration(generation);
}
}
}
setBusy(false);
setStreamingChatId(QString());
}
void LlmClient::finalize() {
if (!m_streaming)
return;
ChatSession* session = m_active;
endStream();
if (session && m_pendingClear == session) {
session->clearMessages();
m_pendingClear.clear();
}
if (session)
session->persist();
}
void LlmClient::clearOnFinish(ChatSession* session) {
m_pendingClear = session;
}
void LlmClient::sessionRemoved(ChatSession* session) {
if (m_pendingClear == session)
m_pendingClear.clear();
if (m_active == session) {
stop();
endStream();
}
}
void LlmClient::fail(const QString& message) {
qWarning() << "LlmClient:" << message;
if (m_streaming) {
ChatSession* session = m_active;
endStream();
if (session && m_pendingClear == session) {
session->clearMessages();
m_pendingClear.clear();
}
if (session)
session->persist();
}
Q_EMIT errorOccurred(message);
}
void LlmClient::drainBuffer() {
while (true) {
const qsizetype newline = m_buffer.indexOf('\n');
if (newline < 0)
break;
const QByteArray line = m_buffer.left(newline).trimmed();
m_buffer.remove(0, newline + 1);
handleLine(line);
}
}
void LlmClient::handleLine(const QByteArray& line) {
if (!m_streaming || line.isEmpty() || !line.startsWith("data:"))
return;
const QByteArray data = line.mid(5).trimmed();
if (data == "[DONE]") {
finalize();
return;
}
const QJsonDocument doc = QJsonDocument::fromJson(data);
if (!doc.isObject())
return;
const QJsonObject obj = doc.object();
updateTokenUsage(obj);
if (obj.contains("error")) {
const QJsonObject error = obj["error"].toObject();
const QString message = error["message"].toString();
fail(message.isEmpty()
? QStringLiteral("LLM server returned an error")
: message);
return;
}
for (const QJsonValue& choiceValue : obj["choices"].toArray()) {
if (!m_streaming)
continue;
const QJsonObject delta =
choiceValue.toObject()["delta"].toObject();
m_streaming->appendContent(delta["content"].toString());
QString reasoning = delta["reasoning_content"].toString();
if (reasoning.isEmpty())
reasoning = delta["reasoning"].toString();
m_streaming->appendReasoning(reasoning);
}
}
void LlmClient::refreshModels() {
const QUrl url =
QUrl::fromUserInput(completionsPath(m_endpoint, "/models"));
if (!url.isValid() || url.host().isEmpty())
return;
auto* reply = m_manager.get(QNetworkRequest(url));
connect(reply, &QNetworkReply::finished, this, [this, reply]() {
const QNetworkReply::NetworkError error = reply->error();
const QByteArray data = reply->readAll();
reply->deleteLater();
if (error != QNetworkReply::NoError) {
qWarning() << "LlmClient: failed to fetch models:" << error;
return;
}
const QJsonArray dataArr =
QJsonDocument::fromJson(data).object()["data"].toArray();
QStringList models;
QSet<QString> seen;
for (const QJsonValue& value : dataArr) {
const QString id = value.toObject()["id"].toString();
if (!id.isEmpty() && !seen.contains(id)) {
seen.insert(id);
models.append(id);
}
}
if (models.isEmpty())
return;
m_availableModels = models;
Q_EMIT availableModelsChanged();
if (m_model.isEmpty()) {
m_model = models.first();
Q_EMIT modelChanged();
}
});
}
void LlmClient::setContextSize(int size) {
if (size <= 0 || m_contextSize == size)
return;
m_contextSize = size;
Q_EMIT contextSizeChanged();
}
void LlmClient::probeContextSize() {
// llama.cpp-specific endpoint; other servers fall back to 4096.
QString base = m_endpoint.trimmed();
while (base.endsWith('/'))
base.chop(1);
const QUrl url = QUrl::fromUserInput(base + "/props");
if (!url.isValid() || url.host().isEmpty())
return;
auto* reply = m_manager.get(QNetworkRequest(url));
connect(
reply,
&QNetworkReply::finished,
this,
[this, reply]() {
const QNetworkReply::NetworkError error = reply->error();
const QByteArray data = reply->readAll();
reply->deleteLater();
int size = 0;
if (error == QNetworkReply::NoError) {
const QJsonDocument doc = QJsonDocument::fromJson(data);
if (doc.isArray()) {
for (const auto& value : doc.array()) {
const QJsonObject slot = value.toObject();
if (slot.contains("n_ctx")) {
size = slot["n_ctx"].toInt(0);
if (size > 0)
break;
}
}
} else if (doc.isObject()) {
const QJsonObject obj = doc.object();
size = obj["n_ctx"].toInt(0);
if (size <= 0)
size = obj["default_generation_settings"].toObject()
["n_ctx"].toInt(0);
}
}
setContextSize(size > 0 ? size : 4096);
});
}
void LlmClient::updateTokenUsage(const QJsonObject& data) {
if (!m_active || m_contextSize <= 0)
return;
const QJsonObject usage = data["usage"].toObject();
if (usage.isEmpty())
return;
const double used =
usage.value("prompt_tokens").toDouble() +
usage.value("completion_tokens").toDouble();
if (used > 0)
m_active->setLastTokenCount(static_cast<int>(used));
}
} // namespace ZShell::llm
Plugins/ZShell/Llm/llmclient.hpp
+103
View File
@@ -0,0 +1,103 @@
#pragma once
#include <QByteArray>
#include <QNetworkAccessManager>
#include <QObject>
#include <QPointer>
#include <QString>
#include <QStringList>
#include <functional>
class QJsonObject;
class QNetworkReply;
namespace ZShell::llm {
class ChatGeneration;
class ChatSession;
// The only component that talks to the LLM server: owns the network
// manager, the in-flight streaming state and the SSE parsing.
class LlmClient : public QObject {
Q_OBJECT
public:
explicit LlmClient(QObject* parent = nullptr);
~LlmClient() override;
[[nodiscard]] QString endpoint() const { return m_endpoint; }
[[nodiscard]] QString model() const { return m_model; }
[[nodiscard]] double temperature() const { return m_temperature; }
[[nodiscard]] QStringList availableModels() const {
return m_availableModels;
}
[[nodiscard]] int contextSize() const { return m_contextSize; }
void setEndpoint(const QString& value);
void setModel(const QString& value);
void setTemperature(double value);
void setContextSize(int size);
void probeContextSize();
[[nodiscard]] bool busy() const { return m_busy; }
[[nodiscard]] QString streamingChatId() const { return m_streamingChatId; }
[[nodiscard]] ChatSession* streamingSession() const { return m_active; }
// Streams a new assistant reply into `target` (the active generation of
// a message of `session`). The context sent to the model is every
// message of the session that is older than the target message.
void startGeneration(ChatSession* session, ChatGeneration* target);
void stop();
void endStream();
// Clears the session's conversation once the current stream ends.
void clearOnFinish(ChatSession* session);
// A session is about to be destroyed; drop any state pointing at it.
void sessionRemoved(ChatSession* session);
void refreshModels();
void requestTitle(ChatSession* session, const QString& userText);
void requestIcon(ChatSession* session, const QString& userText);
Q_SIGNALS:
void busyChanged();
void endpointChanged();
void modelChanged();
void availableModelsChanged();
void contextSizeChanged();
void streamingChatIdChanged();
void errorOccurred(const QString& message);
void titleSuggested(ZShell::llm::ChatSession* session, const QString& title);
void iconSuggested(ZShell::llm::ChatSession* session, const QString& icon);
private:
void finalize();
void fail(const QString& message);
void handleLine(const QByteArray& line);
void drainBuffer();
void updateTokenUsage(const QJsonObject& data);
void setBusy(bool value);
void setStreamingChatId(const QString& id);
void shortRequest(
const QString& tag,
const QString& systemPrompt,
const QString& userText,
std::function<void(QString result)> onResult);
static QString completionsPath(const QString& endpoint, const QString& subpath);
static QString serverErrorMessage(
const QByteArray& body, const QString& fallback);
QNetworkAccessManager m_manager;
QNetworkReply* m_reply = nullptr;
QByteArray m_buffer;
ChatSession* m_active = nullptr;
ChatGeneration* m_streaming = nullptr;
QPointer<ChatSession> m_pendingClear;
bool m_busy = false;
QString m_streamingChatId;
QString m_endpoint;
QString m_model;
QStringList m_availableModels;
double m_temperature = 0.7;
int m_contextSize = 0;
};
} // namespace ZShell::llm
Plugins/ZShell/Llm/message.cpp
+67 -100
View File
@@ -1,123 +1,90 @@
#include "message.hpp"
#include <QDateTime>
#include "messagemodel.hpp"
#include "session.hpp"
namespace ZShell {
namespace ZShell::llm {
ChatMessage::ChatMessage(
Role role,
const QString& content,
namespace {
ChatSession* sessionOf(const ChatMessage* message) {
if (auto* model = qobject_cast<ChatMessageModel*>(message->parent()))
return model->session();
return nullptr;
}
} // namespace
ChatMessage::ChatMessage(Role role, qint64 timestamp, QObject* parent)
: QObject(parent), m_role(role), m_timestamp(timestamp) {}
ChatGeneration* ChatMessage::addGeneration(
qint64 timestamp,
QObject* parent)
: QObject(parent), m_role(role), m_content(content), m_timestamp(timestamp) {
m_timer.setParent(this);
m_timer.setInterval(500);
m_timer.setTimerType(Qt::CoarseTimer);
connect(&m_timer, &QTimer::timeout, this, [this]() {
if (!reasoningInFlight() && !contentInFlight()) {
m_timer.stop();
return;
}
Q_EMIT elapsedMsChanged();
});
const QString& content,
const QString& reasoning,
qint64 reasoningElapsedMs,
qint64 contentElapsedMs) {
auto* generation = new ChatGeneration(timestamp, this);
generation->setContent(content);
generation->setReasoning(reasoning);
generation->setElapsedMs(reasoningElapsedMs, contentElapsedMs);
m_generations.append(generation);
if (m_active < 0)
m_active = static_cast<int>(m_generations.size() - 1);
Q_EMIT generationsChanged();
return generation;
}
qint64 ChatMessage::reasoningElapsedMs() const {
if (m_reasoningStartedAt <= 0)
return 0;
const qint64 end = m_reasoningEndedAt > 0
? m_reasoningEndedAt
: QDateTime::currentMSecsSinceEpoch();
return end - m_reasoningStartedAt;
ChatGeneration* ChatMessage::appendGeneration(qint64 timestamp) {
auto* generation =
addGeneration(timestamp, QString(), QString(), 0, 0);
setActiveInternal(static_cast<int>(m_generations.size() - 1));
return generation;
}
qint64 ChatMessage::contentElapsedMs() const {
if (m_contentStartedAt <= 0)
return 0;
const qint64 end = m_contentEndedAt > 0
? m_contentEndedAt
: QDateTime::currentMSecsSinceEpoch();
return end - m_contentStartedAt;
}
void ChatMessage::updateReasoningActive() {
const bool active = m_streaming && m_content.isEmpty();
if (m_reasoningActive == active)
void ChatMessage::removeGeneration(ChatGeneration* generation) {
const int index = static_cast<int>(m_generations.indexOf(generation));
if (index < 0)
return;
m_reasoningActive = active;
Q_EMIT reasoningActiveChanged();
const bool wasActive = index == m_active;
m_generations.removeAt(index);
delete generation;
if (m_generations.isEmpty()) {
m_active = -1;
} else if (m_active >= m_generations.size()) {
m_active = static_cast<int>(m_generations.size() - 1);
}
Q_EMIT generationsChanged();
if (wasActive)
Q_EMIT activeGenerationChanged();
}
void ChatMessage::appendContent(const QString& piece) {
if (piece.isEmpty())
void ChatMessage::setActiveInternal(int index) {
if (index < 0 || index >= m_generations.size() || index == m_active)
return;
if (m_content.isEmpty()) {
const qint64 now = QDateTime::currentMSecsSinceEpoch();
if (reasoningInFlight())
m_reasoningEndedAt = now;
m_contentStartedAt = now;
if (!m_timer.isActive())
m_timer.start();
}
m_content += piece;
Q_EMIT contentChanged();
updateReasoningActive();
Q_EMIT elapsedMsChanged();
m_active = index;
Q_EMIT activeGenerationChanged();
}
void ChatMessage::appendReasoning(const QString& piece) {
if (piece.isEmpty())
return;
if (m_reasoning.isEmpty()) {
if (m_reasoningStartedAt <= 0)
m_reasoningStartedAt = QDateTime::currentMSecsSinceEpoch();
if (!m_timer.isActive())
m_timer.start();
}
m_reasoning += piece;
Q_EMIT reasoningChanged();
updateReasoningActive();
Q_EMIT elapsedMsChanged();
void ChatMessage::setActiveGeneration(int index) {
setActiveInternal(index);
}
void ChatMessage::setReasoning(const QString& value) {
if (m_reasoning == value)
return;
m_reasoning = value;
Q_EMIT reasoningChanged();
void ChatMessage::edit(const QString& newContent) {
if (auto* generation = activeGeneration())
generation->setContent(newContent);
if (auto* session = sessionOf(this))
session->persist();
}
void ChatMessage::setElapsedMs(qint64 reasoningMs, qint64 contentMs) {
if (reasoningMs > 0) {
m_reasoningStartedAt = m_timestamp;
m_reasoningEndedAt = m_timestamp + reasoningMs;
}
if (contentMs > 0) {
m_contentStartedAt = m_timestamp;
m_contentEndedAt = m_timestamp + contentMs;
}
void ChatMessage::retry() {
if (auto* session = sessionOf(this))
session->retry(this);
}
void ChatMessage::setStreaming(bool value) {
if (m_streaming == value)
return;
m_streaming = value;
Q_EMIT streamingChanged();
if (value) {
if (m_reasoningStartedAt <= 0)
m_reasoningStartedAt = QDateTime::currentMSecsSinceEpoch();
if (!m_timer.isActive())
m_timer.start();
} else {
const qint64 now = QDateTime::currentMSecsSinceEpoch();
if (reasoningInFlight())
m_reasoningEndedAt = now;
if (contentInFlight())
m_contentEndedAt = now;
m_timer.stop();
Q_EMIT elapsedMsChanged();
}
updateReasoningActive();
void ChatMessage::generate() {
if (auto* session = sessionOf(this))
session->continueFrom(this);
}
} // namespace ZShell
} // namespace ZShell::llm
Plugins/ZShell/Llm/message.hpp
+48 -52
View File
@@ -1,27 +1,29 @@
#pragma once
#include "generation.hpp"
#include <QList>
#include <QObject>
#include <QString>
#include <QTimer>
#include <QtQml>
namespace ZShell {
class Chat;
namespace ZShell::llm {
class ChatMessage : public QObject {
Q_OBJECT
QML_ELEMENT
QML_UNCREATABLE("Chat messages are created by the Chat singleton")
Q_PROPERTY(Role role READ role NOTIFY roleChanged)
Q_PROPERTY(QString content READ content NOTIFY contentChanged)
Q_PROPERTY(QString reasoning READ reasoning NOTIFY reasoningChanged)
Q_PROPERTY(bool reasoningActive READ reasoningActive NOTIFY reasoningActiveChanged)
Q_PROPERTY(qint64 reasoningElapsedMs READ reasoningElapsedMs NOTIFY elapsedMsChanged)
Q_PROPERTY(qint64 contentElapsedMs READ contentElapsedMs NOTIFY elapsedMsChanged)
Q_PROPERTY(bool streaming READ streaming NOTIFY streamingChanged)
Q_PROPERTY(Role role READ role CONSTANT)
Q_PROPERTY(qint64 timestamp READ timestamp CONSTANT)
Q_PROPERTY(int generationCount READ generationCount NOTIFY generationsChanged)
Q_PROPERTY(
QList<ZShell::llm::ChatGeneration*> generations READ generations
NOTIFY generationsChanged)
Q_PROPERTY(
ZShell::llm::ChatGeneration* activeGeneration READ activeGeneration
NOTIFY activeGenerationChanged)
Q_PROPERTY(int activeGenerationIndex READ activeGenerationIndex NOTIFY activeGenerationChanged)
public:
enum class Role : int {
@@ -31,57 +33,51 @@ class ChatMessage : public QObject {
Q_ENUM(Role)
explicit ChatMessage(
Role role,
const QString& content,
qint64 timestamp,
QObject* parent = nullptr);
Role role, qint64 timestamp, QObject* parent = nullptr);
[[nodiscard]] Role role() const { return m_role; }
[[nodiscard]] QString content() const { return m_content; }
[[nodiscard]] QString reasoning() const { return m_reasoning; }
[[nodiscard]] bool reasoningActive() const { return m_reasoningActive; }
[[nodiscard]] qint64 reasoningElapsedMs() const;
[[nodiscard]] qint64 contentElapsedMs() const;
[[nodiscard]] bool streaming() const { return m_streaming; }
[[nodiscard]] qint64 timestamp() const { return m_timestamp; }
[[nodiscard]] int generationCount() const {
return static_cast<int>(m_generations.size());
}
[[nodiscard]] QList<ChatGeneration*> generations() const {
return m_generations;
}
[[nodiscard]] ChatGeneration* activeGeneration() const {
return generation(m_active);
}
[[nodiscard]] int activeGenerationIndex() const { return m_active; }
[[nodiscard]] ChatGeneration* generation(int index) const {
if (index < 0 || index >= m_generations.size())
return nullptr;
return m_generations.at(index);
}
void appendContent(const QString& piece);
void appendReasoning(const QString& piece);
void setReasoning(const QString& value);
void setElapsedMs(qint64 reasoningMs, qint64 contentMs);
void setStreaming(bool value);
Q_INVOKABLE void setActiveGeneration(int index);
Q_INVOKABLE void edit(const QString& newContent);
Q_INVOKABLE void retry();
Q_INVOKABLE void generate();
ChatGeneration* addGeneration(
qint64 timestamp,
const QString& content,
const QString& reasoning,
qint64 reasoningElapsedMs,
qint64 contentElapsedMs);
ChatGeneration* appendGeneration(qint64 timestamp);
void removeGeneration(ChatGeneration* generation);
Q_SIGNALS:
void roleChanged();
void contentChanged();
void reasoningChanged();
void reasoningActiveChanged();
void elapsedMsChanged();
void streamingChanged();
void generationsChanged();
void activeGenerationChanged();
private:
void updateReasoningActive();
void setActiveInternal(int index);
[[nodiscard]] bool reasoningInFlight() const {
return m_reasoningStartedAt > 0 && m_reasoningEndedAt <= 0;
}
[[nodiscard]] bool contentInFlight() const {
return m_contentStartedAt > 0 && m_contentEndedAt <= 0;
}
QTimer m_timer;
Role m_role;
QString m_content;
QString m_reasoning;
bool m_reasoningActive = false;
bool m_streaming = false;
qint64 m_timestamp;
qint64 m_reasoningStartedAt = 0;
qint64 m_reasoningEndedAt = 0;
qint64 m_contentStartedAt = 0;
qint64 m_contentEndedAt = 0;
friend class Chat;
QList<ChatGeneration*> m_generations;
int m_active = -1;
};
} // namespace ZShell
} // namespace ZShell::llm
Plugins/ZShell/Llm/messagemodel.cpp
+107
View File
@@ -0,0 +1,107 @@
#include "messagemodel.hpp"
#include "session.hpp"
namespace ZShell::llm {
ChatMessageModel::ChatMessageModel(ChatSession* session, QObject* parent)
: QAbstractListModel(parent), m_session(session) {}
ChatMessageModel::~ChatMessageModel() = default;
int ChatMessageModel::rowCount(const QModelIndex& parent) const {
if (parent.isValid()) return 0;
return static_cast<int>(m_messages.size());
}
QVariant ChatMessageModel::data(const QModelIndex& index, int role) const {
if (role != Qt::UserRole || !index.isValid() || index.row() < 0 ||
index.row() >= m_messages.size())
return {};
return QVariant::fromValue(m_messages.at(index.row()));
}
QHash<int, QByteArray> ChatMessageModel::roleNames() const {
return {{Qt::UserRole, "modelData"}};
}
ChatMessage* ChatMessageModel::at(int row) const {
if (row < 0 || row >= m_messages.size()) return nullptr;
return m_messages.at(row);
}
int ChatMessageModel::rowOf(const ChatMessage* message) const {
for (int i = 0; i < m_messages.size(); ++i)
if (m_messages.at(i) == message) return i;
return -1;
}
ChatMessage* ChatMessageModel::createMessage(
ChatMessage::Role role, qint64 timestamp) {
return new ChatMessage(role, timestamp, this);
}
ChatMessage* ChatMessageModel::appendNewest(
ChatMessage::Role role, const QString& content, qint64 timestamp) {
auto* message = new ChatMessage(role, timestamp, this);
message->addGeneration(timestamp, content, QString(), 0, 0);
beginInsertRows(QModelIndex(), 0, 0);
m_messages.prepend(message);
endInsertRows();
emit lastMessageChanged();
return message;
}
void ChatMessageModel::removeMessage(ChatMessage* message) {
const int row = rowOf(message);
if (row < 0) return;
const bool wasLastMessage = row == 0;
beginRemoveRows(QModelIndex(), row, row);
m_messages.removeAt(row);
endRemoveRows();
delete message;
if (wasLastMessage) emit lastMessageChanged();
}
void ChatMessageModel::removeRange(int firstRow, int lastRow) {
if (firstRow < 0 || firstRow > lastRow || lastRow >= m_messages.size())
return;
const bool changesLastMessage = firstRow == 0;
beginRemoveRows(QModelIndex(), firstRow, lastRow);
for (int row = lastRow; row >= firstRow; --row)
delete m_messages.takeAt(row);
endRemoveRows();
if (changesLastMessage) emit lastMessageChanged();
}
void ChatMessageModel::clear() {
if (m_messages.isEmpty()) return;
beginRemoveRows(QModelIndex(), 0, static_cast<int>(m_messages.size() - 1));
qDeleteAll(m_messages);
m_messages.clear();
endRemoveRows();
emit lastMessageChanged();
}
void ChatMessageModel::loadMessages(QList<ChatMessage*> messages) {
beginResetModel();
qDeleteAll(m_messages);
m_messages = std::move(messages);
endResetModel();
emit lastMessageChanged();
}
} // namespace ZShell::llm
Plugins/ZShell/Llm/messagemodel.hpp
+68
View File
@@ -0,0 +1,68 @@
#pragma once
#include "message.hpp"
#include <QAbstractListModel>
#include <QHash>
#include <QList>
#include <QtQml>
#include <qtmetamacros.h>
namespace ZShell::llm {
class ChatSession;
// Owns a session's messages, most recent first: row 0 is always the newest
// message. Items are exposed through a role named "modelData" (like
// FileSystemModel), so delegates receive each ChatMessage as `modelData`.
class ChatMessageModel : public QAbstractListModel {
Q_OBJECT
QML_ELEMENT
QML_UNCREATABLE("Chat message models are owned by ChatSession")
Q_PROPERTY(
ZShell::llm::ChatMessage* lastMessage READ lastMessage
NOTIFY lastMessageChanged)
public:
explicit ChatMessageModel(ChatSession* session, QObject* parent = nullptr);
~ChatMessageModel() override;
[[nodiscard]] int rowCount(
const QModelIndex& parent = QModelIndex()) const override;
[[nodiscard]] QVariant data(
const QModelIndex& index, int role = Qt::DisplayRole) const override;
[[nodiscard]] QHash<int, QByteArray> roleNames() const override;
[[nodiscard]] ChatSession* session() const { return m_session; }
// Most recent message first.
[[nodiscard]] QList<ChatMessage*> messages() const { return m_messages; }
[[nodiscard]] ChatMessage* at(int row) const;
[[nodiscard]] int rowOf(const ChatMessage* message) const;
[[nodiscard]] ChatMessage* lastMessage() const {
return m_messages.isEmpty() ? nullptr : m_messages.first();
}
// Creates a message owned by this model without inserting it.
ChatMessage* createMessage(ChatMessage::Role role, qint64 timestamp);
// Appends a new message as the newest one (row 0) with a single
// generation holding `content`.
ChatMessage* appendNewest(
ChatMessage::Role role, const QString& content, qint64 timestamp);
void removeMessage(ChatMessage* message);
void removeRange(int firstRow, int lastRow);
void clear();
// Replaces every row; takes ownership of the given messages, most recent
// first.
void loadMessages(QList<ChatMessage*> messages);
signals:
void lastMessageChanged();
private:
ChatSession* m_session;
QList<ChatMessage*> m_messages;
};
} // namespace ZShell::llm
Plugins/ZShell/Llm/session.cpp
+157 -43
View File
@@ -1,102 +1,216 @@
#include "session.hpp"
#include "chatstore.hpp"
#include "llmclient.hpp"
#include <QDebug>
#include <QDateTime>
#include <QtGlobal>
namespace ZShell {
namespace ZShell::llm {
namespace {
QString titleFrom(const QString& content) {
const QString flat = content.simplified();
if (flat.isEmpty()) return QString();
if (flat.size() <= 48) return flat;
return flat.left(47) + QStringLiteral("…");
}
} // namespace
ChatSession::ChatSession(const QString& id, QObject* parent)
: QObject(parent), m_id(id) {}
: QObject(parent), m_id(id) {
m_model = new ChatMessageModel(this, this);
connect(
m_model,
&QAbstractListModel::rowsInserted,
this,
&ChatSession::onModelRowsChanged);
connect(
m_model,
&QAbstractListModel::rowsRemoved,
this,
&ChatSession::onModelRowsChanged);
connect(
m_model,
&QAbstractListModel::modelReset,
this,
&ChatSession::onModelRowsChanged);
}
void ChatSession::onModelRowsChanged() {
setCount(m_model->rowCount());
}
void ChatSession::setTitle(const QString& value) {
if (m_title == value)
return;
if (m_title == value) return;
m_title = value;
Q_EMIT titleChanged();
persist();
}
void ChatSession::setIcon(const QString& value) {
if (m_icon == value)
return;
if (m_icon == value) return;
m_icon = value;
Q_EMIT iconChanged();
}
void ChatSession::setUpdatedAt(qint64 value) {
if (m_updatedAt == value)
return;
if (m_updatedAt == value) return;
m_updatedAt = value;
Q_EMIT updatedAtChanged();
}
void ChatSession::setPinned(bool value) {
if (m_pinned == value)
return;
if (m_pinned == value) return;
m_pinned = value;
Q_EMIT pinnedChanged();
}
void ChatSession::setCount(int value) {
if (m_messageCount == value)
return;
if (m_messageCount == value) return;
m_messageCount = value;
Q_EMIT messageCountChanged();
}
void ChatSession::setMeta(
const QString& title,
qint64 createdAt,
qint64 updatedAt,
int messageCount) {
const QString& title, qint64 createdAt, qint64 updatedAt, int messageCount) {
m_title = title;
m_createdAt = createdAt;
m_updatedAt = updatedAt;
m_messageCount = messageCount;
}
void ChatSession::setLastTokenCount(int value) {
m_lastTokenCount = value;
}
LlmClient* ChatSession::client() const {
if (auto* store = qobject_cast<ChatStore*>(parent()))
return store->llmClient();
return nullptr;
}
void ChatSession::ensureLoaded() {
if (m_loaded)
return;
if (m_loaded) return;
m_loaded = true;
if (auto* store = qobject_cast<ChatStore*>(parent()))
store->loadMessagesInto(this);
}
void ChatSession::adoptMessages(QList<ChatMessage*> messages) {
qDeleteAll(m_messages);
m_messages = messages;
setCount(m_messages.size());
Q_EMIT messagesChanged();
ChatMessageModel* ChatSession::messagesModel() {
ensureLoaded();
return m_model;
}
ChatMessage* ChatSession::appendMessage(
void ChatSession::adoptMessages(QList<ChatMessage*> messages) {
m_model->loadMessages(std::move(messages));
}
void ChatSession::persist() {
if (auto* store = qobject_cast<ChatStore*>(parent())) store->persist(this);
}
ChatMessage* ChatSession::appendNewest(
ChatMessage::Role role, const QString& content, qint64 timestamp) {
ensureLoaded();
auto* message = new ChatMessage(role, content, timestamp, this);
m_messages.append(message);
setCount(m_messages.size());
Q_EMIT messagesChanged();
return message;
return m_model->appendNewest(role, content, timestamp);
}
void ChatSession::removeMessage(ChatMessage* message) {
if (!message || !m_messages.removeOne(message))
return;
delete message;
setCount(m_messages.size());
Q_EMIT messagesChanged();
m_model->removeMessage(message);
}
void ChatSession::clearMessages() {
ensureLoaded();
if (m_messages.isEmpty())
return;
qDeleteAll(m_messages);
m_messages.clear();
setCount(0);
Q_EMIT messagesChanged();
m_model->clear();
}
} // namespace ZShell
void ChatSession::startGeneration(ChatMessage* target) {
if (auto* generation = target->activeGeneration()) {
if (auto* clientObject = client())
clientObject->startGeneration(this, generation);
}
persist();
}
void ChatSession::sendMessage(const QString& text) {
const QString trimmed = text.trimmed();
if (trimmed.isEmpty()) return;
if (auto* clientObject = client()) {
if (clientObject->busy()) return;
}
const qint64 now = QDateTime::currentMSecsSinceEpoch();
appendNewest(ChatMessage::Role::User, trimmed, now);
while (m_model->rowCount() > 200)
m_model->removeMessage(m_model->at(m_model->rowCount() - 1));
if (m_title.isEmpty()) {
m_title = titleFrom(trimmed);
Q_EMIT titleChanged();
qInfo() << "ChatSession:" << m_id << "new conversation,"
<< "fallback title" << m_title
<< "- requesting generated title and icon";
if (auto* clientObject = client()) {
clientObject->requestTitle(this, trimmed);
clientObject->requestIcon(this, trimmed);
}
}
auto* assistant =
appendNewest(ChatMessage::Role::Assistant, QString(), now);
startGeneration(assistant);
}
void ChatSession::retry(ChatMessage* target) {
if (!target || target->role() != ChatMessage::Role::Assistant) return;
if (auto* clientObject = client()) {
if (clientObject->busy()) return;
}
const int row = m_model->rowOf(target);
if (row < 0) return;
// Drop everything newer than the target, then regenerate from the
// context ending at the user message before it.
m_model->removeRange(0, row - 1);
if (m_model->rowCount() < 2) return;
const qint64 now = QDateTime::currentMSecsSinceEpoch();
target->appendGeneration(now);
startGeneration(target);
}
void ChatSession::continueFrom(ChatMessage* message) {
if (!message) return;
if (auto* clientObject = client()) {
if (clientObject->busy()) return;
}
const int row = m_model->rowOf(message);
if (row < 0) return;
m_model->removeRange(0, row - 1);
const qint64 now = QDateTime::currentMSecsSinceEpoch();
if (message->role() == ChatMessage::Role::Assistant) {
if (m_model->rowCount() < 2) return;
message->appendGeneration(now);
startGeneration(message);
} else {
auto* assistant =
appendNewest(ChatMessage::Role::Assistant, QString(), now);
startGeneration(assistant);
}
}
void ChatSession::clear() {
if (auto* clientObject = client()) {
if (clientObject->busy() && clientObject->streamingSession() == this) {
clientObject->clearOnFinish(this);
clientObject->stop();
return;
}
}
m_model->clear();
persist();
}
} // namespace ZShell::llm
Plugins/ZShell/Llm/session.hpp
+29 -15
View File
@@ -1,14 +1,16 @@
#pragma once
#include "messagemodel.hpp"
#include "message.hpp"
#include <QDateTime>
#include <QList>
#include <QObject>
#include <QString>
#include <QtQml>
namespace ZShell {
namespace ZShell::llm {
class LlmClient;
class ChatSession : public QObject {
Q_OBJECT
@@ -16,13 +18,14 @@ class ChatSession : public QObject {
QML_UNCREATABLE("Chat sessions are managed by Chat.chats")
Q_PROPERTY(QString id READ id CONSTANT)
Q_PROPERTY(QString title READ title NOTIFY titleChanged)
Q_PROPERTY(QString title READ title WRITE setTitle NOTIFY titleChanged)
Q_PROPERTY(QString icon READ icon NOTIFY iconChanged)
Q_PROPERTY(QDateTime createdAt READ createdAt CONSTANT)
Q_PROPERTY(QDateTime updatedAt READ updatedAt NOTIFY updatedAtChanged)
Q_PROPERTY(bool pinned READ pinned NOTIFY pinnedChanged)
Q_PROPERTY(int messageCount READ messageCount NOTIFY messageCountChanged)
Q_PROPERTY(QList<ChatMessage*> messages READ messages NOTIFY messagesChanged)
Q_PROPERTY(
ZShell::llm::ChatMessageModel* messagesModel READ messagesModel CONSTANT)
public:
explicit ChatSession(const QString& id, QObject* parent = nullptr);
@@ -40,10 +43,14 @@ class ChatSession : public QObject {
[[nodiscard]] qint64 updatedAtMs() const { return m_updatedAt; }
[[nodiscard]] bool pinned() const { return m_pinned; }
[[nodiscard]] int messageCount() const { return m_messageCount; }
[[nodiscard]] QList<ChatMessage*> messages() {
ensureLoaded();
return m_messages;
}
// Loads the messages from the store on first access.
[[nodiscard]] ChatMessageModel* messagesModel();
void ensureLoaded();
[[nodiscard]] bool isLoaded() const { return m_loaded; }
[[nodiscard]] LlmClient* client() const;
[[nodiscard]] int lastTokenCount() const { return m_lastTokenCount; }
void setLastTokenCount(int value);
void setTitle(const QString& value);
void setIcon(const QString& value);
@@ -56,23 +63,29 @@ class ChatSession : public QObject {
qint64 updatedAt,
int messageCount);
void ensureLoaded();
[[nodiscard]] bool isLoaded() const { return m_loaded; }
// Replaces the model's rows with `messages` (most recent first).
void adoptMessages(QList<ChatMessage*> messages);
ChatMessage* appendMessage(
ChatMessage::Role role, const QString& content, qint64 timestamp);
void persist();
void removeMessage(ChatMessage* message);
void clearMessages();
Q_INVOKABLE void sendMessage(const QString& text);
Q_INVOKABLE void retry(ZShell::llm::ChatMessage* target);
Q_INVOKABLE void continueFrom(ZShell::llm::ChatMessage* message);
Q_INVOKABLE void clear();
Q_SIGNALS:
void titleChanged();
void iconChanged();
void updatedAtChanged();
void pinnedChanged();
void messageCountChanged();
void messagesChanged();
private:
void onModelRowsChanged();
ChatMessage* appendNewest(
ChatMessage::Role role, const QString& content, qint64 timestamp);
void startGeneration(ChatMessage* target);
void setCount(int value);
QString m_id;
@@ -82,8 +95,9 @@ class ChatSession : public QObject {
qint64 m_updatedAt = 0;
bool m_pinned = false;
int m_messageCount = 0;
QList<ChatMessage*> m_messages;
ChatMessageModel* m_model = nullptr;
bool m_loaded = false;
int m_lastTokenCount = 0;
};
} // namespace ZShell
} // namespace ZShell::llm