z-bar-qt/Plugins/ZShell/Llm/llmclient.cpp

#include "llmclient.hpp"

#include "attachment.hpp"
#include "filetool.hpp"
#include "generation.hpp"
#include "message.hpp"
#include "messagemodel.hpp"
#include "prompts.hpp"
#include "session.hpp"
#include "shelltool.hpp"
#include "webfetchtool.hpp"
#include "workspace.hpp"

#include "config/llm.hpp"
#include "config/rootnodes.hpp"

#include <QDebug>
#include <QJsonArray>
#include <QJsonDocument>
#include <QJsonObject>
#include <QNetworkReply>
#include <QNetworkRequest>
#include <QPointer>
#include <QRegularExpression>
#include <QSet>
#include <QTimer>
#include <QUrl>

namespace ZShell::llm {

namespace {

constexpr int kRefreshIntervalMs = 60'000;
constexpr int kProbeTimeoutMs = 2000;

int contextSizeFromErrorMessage(const QString& message) {
	static const QRegularExpression re(
		QLatin1String("available context size \\((\\d+)"));
	const auto match = re.match(message);
	if (match.hasMatch()) return match.captured(1).toInt();
	return 0;
}

} // namespace

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::jsonErrorMessage(const QByteArray& body) {
	const QJsonDocument doc = QJsonDocument::fromJson(body);
	if (!doc.isObject()) return QString();
	const QJsonValue errorValue = doc.object().value("error");
	if (errorValue.isObject())
		return errorValue.toObject().value("message").toString();
	if (errorValue.isString()) return errorValue.toString();
	return QString();
}

QString LlmClient::serverErrorMessage(
	const QByteArray& body, const QString& fallback) {
	const QString message = jsonErrorMessage(body);
	if (!message.isEmpty()) return message;
	return fallback.isEmpty() ? QStringLiteral("Request to LLM server failed")
							  : fallback;
}

void LlmClient::setServerError(const QString& value) {
	if (m_serverError == value) return;
	m_serverError = value;
	Q_EMIT serverErrorChanged();
}

void LlmClient::setBusy(bool value) {
	if (m_busy == value) return;
	m_busy = value;
	Q_EMIT busyChanged();
}

void LlmClient::setModel(const QString& value) {
	if (m_model == value) return;
	m_model = value;
	Q_EMIT modelChanged();
	if (m_model.isEmpty()) refreshModels();
}

void LlmClient::setContextSize(int size) {
	if (size <= 0 || m_tools->contextSize() == size) return;
	m_tools->setContextSize(size);
	Q_EMIT contextSizeChanged();
}

void LlmClient::setToolsEnabled(bool value) {
	if (m_tools->enabled() == value) return;
	m_tools->setEnabled(value);
	Q_EMIT toolsEnabledChanged();
	if (value) refreshAllStates();
}

void LlmClient::setSandboxEnabled(bool value) {
	if (m_tools->sandboxEnabled() == value) return;
	m_tools->setSandboxEnabled(value);
	Q_EMIT sandboxEnabledChanged();
}

LlmClient::LlmClient(QObject* parent) : QObject(parent) {
	m_tools = new ToolRegistry(this);
	m_store = new ChatStore(this, this);

	m_tools->registerTool(new WebFetchTool(m_tools, m_tools));
	m_tools->registerTool(new FileReadTool(m_tools, m_tools));
	m_tools->registerTool(new ShellTool(m_tools, m_tools));

	auto* llm = config::ConfigSingleton::instance()->llm();
	if (llm) {
		m_endpoint = llm->endpoint();
		m_model = llm->model();
		m_temperature = llm->temperature();
		m_configEnabled = llm->enabled();
		m_tools->setEnabled(llm->tools());
		m_tools->setSandboxEnabled(llm->sandbox());

		connect(llm, &config::Llm::endpointChanged, this, [this, llm]() {
			m_endpoint = llm->endpoint();
			refreshAllStates();
		});
		connect(llm, &config::Llm::modelChanged, this, [this, llm]() {
			setModel(llm->model());
		});
		connect(llm, &config::Llm::temperatureChanged, this, [this, llm]() {
			m_temperature = llm->temperature();
		});
		connect(llm, &config::Llm::toolsChanged, this, [this, llm]() {
			setToolsEnabled(llm->tools());
		});
		connect(llm, &config::Llm::sandboxChanged, this, [this, llm]() {
			setSandboxEnabled(llm->sandbox());
		});
		connect(llm, &config::Llm::enabledChanged, this, [this, llm]() {
			m_configEnabled = llm->enabled();
			if (m_configEnabled) {
				m_refreshTimer.setInterval(kRefreshIntervalMs);
				m_refreshTimer.start();
				refreshAllStates();
			} else {
				m_refreshTimer.stop();
			}
		});
	}

	connect(
		m_store, &ChatStore::sessionRemoved, this, [this](ChatSession* session) {
			if (m_turn && m_turn->session() == session) {
				stop();
				endStream();
			}
		});

	m_refreshTimer.setInterval(kRefreshIntervalMs);
	connect(
		&m_refreshTimer, &QTimer::timeout, this, &LlmClient::refreshAllStates);
	if (m_configEnabled) {
		m_refreshTimer.start();
		refreshAllStates();
	}
}

LlmClient::~LlmClient() {
	if (m_reply) {
		QNetworkReply* reply = m_reply;
		m_reply = nullptr;
		reply->abort();
	}
	m_tools->cancelAll();
}

void LlmClient::selectModel(const QString& id) {
	if (id.isEmpty()) return;
	if (auto* config = config::ConfigSingleton::instance())
		config->llm()->set_model(id);
	setModel(id);
}

ToolContext LlmClient::toolContext() const {
	ToolContext context;
	if (m_turn) {
		context.workspace = m_turn->session()->workspace()->path();
		context.outputBudgetChars = m_turn->outputBudgetChars();
	}
	return context;
}

void LlmClient::startGeneration(ChatSession* session, ChatGeneration* target) {
	if (m_busy || !session || !target || !m_configEnabled) return;

	m_turn = std::make_unique<Turn>(session, target, *m_tools);
	if (m_turn->targetRow() < 0) {
		fail(QStringLiteral(
			"Internal error: generation target is not in the session"));
		return;
	}

	target->setStreaming(true);
	setBusy(true);

	probeContextSize([this, session, target]() {
		if (!m_turn || m_turn->session() != session ||
			m_turn->generation() != target)
			return;
		refreshModels();
		sendRound();
	});
}

void LlmClient::sendRound() {
	if (!m_turn) return;
	probeContextSize();

	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;
	}

	Turn* turn = m_turn.get();
	QJsonArray messages = turn->contextMessages();

	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;
	if (m_tools->enabled()) {
		const QJsonArray toolSpecs = m_tools->specifications(toolContext());
		if (!toolSpecs.isEmpty()) body[QStringLiteral("tools")] = toolSpecs;
	}

	QNetworkRequest request(url);
	request.setHeader(QNetworkRequest::ContentTypeHeader, "application/json");
	request.setRawHeader("Accept", "text/event-stream");

	QNetworkReply* reply =
		m_manager.post(request, QJsonDocument(body).toJson());
	if (!reply) {
		fail(QStringLiteral("Failed to start request to LLM server"));
		return;
	}
	m_reply = reply;

	connect(reply, &QNetworkReply::readyRead, this, [this, reply, turn]() {
		if (m_reply != reply || m_turn.get() != turn) return;
		m_buffer.append(reply->readAll());
		drainBuffer();
	});
	connect(reply, &QNetworkReply::finished, this, [this, reply, turn]() {
		if (m_reply != reply) {
			reply->deleteLater();
			return;
		}
		m_reply = nullptr;

		const QNetworkReply::NetworkError error = reply->error();
		const QString errorString = reply->errorString();
		const QByteArray responseBody = reply->readAll();
		m_buffer.append(responseBody);
		reply->deleteLater();

		if (m_turn.get() != turn) return;

		drainBuffer();

		if (error == QNetworkReply::NoError) {
			if (turn->streamDone) {
				roundFinished();
				return;
			}
			const QByteArray tail = m_buffer;
			m_buffer.clear();
			const QString serverMessage = jsonErrorMessage(tail.simplified());
			fail(
				serverMessage.isEmpty()
					? QStringLiteral(
						  "LLM server closed the stream before "
						  "completing the response")
					: serverMessage);
			return;
		}
		if (error == QNetworkReply::OperationCanceledError) {
			finishTurn();
			return;
		}
		const QString message = serverErrorMessage(responseBody, errorString);
		const int reported = contextSizeFromErrorMessage(message);
		if (reported > 0 && reported != m_tools->contextSize() &&
			!turn->contextRetryUsed) {
			qWarning() << "LlmClient: prompt overflow, server reports"
					   << reported << "tokens; retrying round";
			setContextSize(reported);
			turn->contextRetryUsed = true;
			sendRound();
			return;
		}
		fail(message);
	});
}

void LlmClient::applyToolCallDelta(const QJsonObject& call) {
	Turn* turn = m_turn.get();
	if (!turn) return;
	const int index = call[QStringLiteral("index")].toInt(-1);
	if (index < 0) return;
	while (turn->calls().size() <= index)
		turn->calls().append(Turn::Call{});
	Turn::Call& builder = turn->calls()[index];
	builder.seen = true;
	const QString id = call[QStringLiteral("id")].toString();
	if (!id.isEmpty()) builder.id = id;
	const QJsonObject function = call[QStringLiteral("function")].toObject();
	const QString name = function[QStringLiteral("name")].toString();
	if (!name.isEmpty()) builder.name = name;
	const QString arguments = function[QStringLiteral("arguments")].toString();
	if (!arguments.isEmpty()) builder.arguments += arguments;

	if (!builder.segment) {
		turn->generation()->closeOpenSegments();
		builder.segment =
			turn->generation()->beginToolCall(builder.name, builder.id);
	}
	builder.segment->setName(builder.name);
	builder.segment->setToolCallId(builder.id);
	if (!arguments.isEmpty()) builder.segment->appendArguments(arguments);
}

void LlmClient::roundFinished() {
	Turn* turn = m_turn.get();
	if (!turn || turn->roundDone) return;
	turn->roundDone = true;

	if (!turn->roundHasCalls()) {
		finishTurn();
		return;
	}
	if (turn->finishReason != QLatin1String("tool_calls")) {
		qWarning() << "LlmClient: stream ended" << turn->finishReason
				   << "with incomplete tool calls, ending turn";
		for (const auto& builder : turn->calls()) {
			if (!builder.seen || !builder.segment) continue;
			builder.segment->finish(
				QStringLiteral(
					"Stream truncated before the tool call completed"),
				false);
		}
		finishTurn();
		return;
	}
	if (turn->round >= kMaxToolRounds) {
		qWarning() << "LlmClient: tool round limit reached, ending turn";
		finishTurn();
		return;
	}

	turn->generation()->closeOpenSegments();
	turn->appendToTranscript(turn->roundAssistantMessage());
	++turn->round;

	setApprovalPending(true);
}

void LlmClient::executeAllCalls() {
	Turn* turn = m_turn.get();
	if (!turn) return;
	turn->toolPhase = true;
	turn->pendingCalls = 0;
	turn->results() = QList<ToolResult>(turn->calls().size());

	for (int i = 0; i < turn->calls().size(); ++i) {
		const Turn::Call& call = turn->calls().at(i);
		if (!call.seen) continue;

		LlmTool* tool = m_tools->tool(call.name);
		QJsonObject args;
		QString errorText;
		if (!tool) {
			errorText =
				QStringLiteral("Error: unknown tool '%1'").arg(call.name);
		} else if (!call.arguments.isEmpty()) {
			const QJsonDocument doc =
				QJsonDocument::fromJson(call.arguments.toUtf8());
			if (!doc.isObject()) {
				errorText = QStringLiteral(
								"Error: tool arguments are not valid JSON: %1")
								.arg(call.arguments);
			} else {
				args = doc.object();
			}
		}
		if (!errorText.isEmpty()) {
			turn->results()[i] = {false, errorText};
			if (call.segment) call.segment->finish(errorText, false);
			continue;
		}

		ToolContext context = toolContext();
		context.toolCallId = call.id;

		++turn->pendingCalls;
		tool->execute(context, args, [this, turn, i](const ToolResult& result) {
			if (m_turn.get() != turn) return;
			turn->results()[i] = result;
			if (turn->calls().at(i).segment)
				turn->calls().at(i).segment->finish(result.text, result.ok);
			if (--turn->pendingCalls == 0) flushCallResults();
		});
	}
	if (turn->pendingCalls == 0) flushCallResults();
}

void LlmClient::flushCallResults() {
	Turn* turn = m_turn.get();
	if (!turn || !turn->toolPhase) return;
	turn->toolPhase = false;

	for (const QJsonValue& value : turn->toolMessages())
		turn->appendToTranscript(value.toObject());
	sendRound();
}

void LlmClient::setApprovalPending(bool value) {
	Turn* turn = m_turn.get();
	if (!turn) return;
	if (turn->approvalPending == value) return;
	turn->approvalPending = value;
	turn->generation()->setApprovalPending(value);
}

void LlmClient::approveTools(ChatGeneration* generation) {
	Turn* turn = m_turn.get();
	if (!turn || !turn->approvalPending || turn->generation() != generation)
		return;
	setApprovalPending(false);
	for (auto* call : generation->toolCalls()) {
		if (call->status() != LlmToolCall::Status::Pending) continue;
		call->setStatus(LlmToolCall::Status::Running);
		call->begin();
	}
	executeAllCalls();
}

void LlmClient::denyTools(ChatGeneration* generation) {
	Turn* turn = m_turn.get();
	if (!turn || !turn->approvalPending || turn->generation() != generation)
		return;
	setApprovalPending(false);
	for (auto* call : generation->toolCalls())
		if (call->status() == LlmToolCall::Status::Pending)
			call->finish("Denied by user", false);
	finishTurn();
}

void LlmClient::finishPendingCalls(const QString& resultText) {
	Turn* turn = m_turn.get();
	if (!turn) return;
	for (auto* call : turn->generation()->toolCalls())
		if (call->status() == LlmToolCall::Status::Pending)
			call->finish(resultText, false);
}

void LlmClient::stop() {
	if (!m_busy) return;
	Turn* turn = m_turn.get();
	if (!turn) return;

	if (turn->approvalPending) {
		setApprovalPending(false);
		finishPendingCalls(QStringLiteral("Cancelled"));
		finishTurn();
		return;
	}
	if (turn->toolPhase) {
		m_tools->cancelAll();
		for (auto* call : turn->generation()->toolCalls())
			if (call->running())
				call->finish(QStringLiteral("Cancelled"), false);
		finishTurn();
		return;
	}

	if (m_reply) {
		QNetworkReply* reply = m_reply;
		m_reply = nullptr;
		reply->abort();
	}
	endStream();
}

void LlmClient::endStream() {
	Turn* turn = m_turn.get();
	if (!turn) return;
	ChatSession* session = turn->session();
	ChatGeneration* generation = turn->generation();

	if (m_reply) {
		QNetworkReply* reply = m_reply;
		m_reply = nullptr;
		reply->abort();
	}

	generation->setStreaming(false);
	setApprovalPending(false);
	for (auto* call : generation->toolCalls())
		if (call->status() == LlmToolCall::Status::Pending)
			call->finish(QStringLiteral("Cancelled"), false);

	if (generation->content().isEmpty() && generation->reasoning().isEmpty() &&
		generation->toolCallCount() == 0 && !generation->error()) {
		ChatMessage* message = generation->message();
		if (message) {
			if (message->generationCount() <= 1)
				session->removeMessage(message);
			else
				message->removeGeneration(generation);
		}
	}

	m_turn.reset();
	setBusy(false);
}

void LlmClient::finishTurn() {
	ChatSession* session = m_turn ? m_turn->session() : nullptr;
	endStream();
	if (session) {
		m_store->applyPendingClear(session);
		session->persist();
	}
	if (m_configEnabled) refreshAllStates();
}

void LlmClient::fail(const QString& message) {
	qWarning() << "LlmClient:" << message;
	ChatSession* session = m_turn ? m_turn->session() : nullptr;
	if (Turn* turn = m_turn.get()) turn->generation()->setError(message);
	finishTurn();
	if (session && !message.isEmpty())
		Q_EMIT generationFailed(session, 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) {
	Turn* turn = m_turn.get();
	if (!turn || line.isEmpty() || !line.startsWith("data:")) return;

	const QByteArray data = line.mid(5).trimmed();
	if (data == "[DONE]") {
		turn->streamDone = true;
		roundFinished();
		return;
	}

	const QJsonDocument doc = QJsonDocument::fromJson(data);
	if (!doc.isObject()) return;
	const QJsonObject obj = doc.object();
	updateTokenUsage(obj);

	if (obj.contains("error")) {
		const QJsonValue errorValue = obj["error"];
		QString message;
		if (errorValue.isObject())
			message = errorValue.toObject()["message"].toString();
		else if (errorValue.isString())
			message = errorValue.toString();
		fail(
			message.isEmpty() ? QStringLiteral("LLM server returned an error")
							  : message);
		return;
	}

	for (const QJsonValue& choiceValue : obj["choices"].toArray()) {
		if (!m_turn) break;
		const QJsonObject choice = choiceValue.toObject();
		const QJsonObject delta = choice["delta"].toObject();

		const QString finishReason = choice["finish_reason"].toString();
		if (!finishReason.isEmpty()) turn->finishReason = finishReason;

		turn->generation()->appendContent(delta["content"].toString());
		QString reasoning = delta["reasoning_content"].toString();
		if (reasoning.isEmpty()) reasoning = delta["reasoning"].toString();
		turn->generation()->appendReasoning(reasoning);

		for (const QJsonValue& callValue : delta["tool_calls"].toArray()) {
			if (!m_turn) break;
			applyToolCallDelta(callValue.toObject());
		}
	}
}

void LlmClient::refreshAllStates() {
	if (!m_configEnabled) return;
	if (m_endpoint.trimmed().isEmpty()) {
		setServerError(QStringLiteral("LLM endpoint is not configured"));
		return;
	}
	probeContextSize([this]() { refreshModels(); });
}

void LlmClient::refreshModels() {
	if (!m_configEnabled) return;

	const QUrl url =
		QUrl::fromUserInput(completionsPath(m_endpoint, "/models"));
	if (!url.isValid() || url.host().isEmpty()) {
		setServerError(QStringLiteral("LLM endpoint is not configured"));
		return;
	}

	auto* reply = m_manager.get(QNetworkRequest(url));
	connect(reply, &QNetworkReply::finished, this, [this, reply]() {
		const QNetworkReply::NetworkError error = reply->error();
		const QString errorString = reply->errorString();
		const QByteArray data = reply->readAll();
		reply->deleteLater();

		if (error != QNetworkReply::NoError) {
			const QString serverMessage = jsonErrorMessage(data);
			setServerError(
				serverMessage.isEmpty()
					? QStringLiteral("Cannot reach LLM server: %1")
						  .arg(errorString)
					: serverMessage);
			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()) {
			setServerError(QStringLiteral(
				"LLM server is running but has no models loaded"));
			return;
		}
		setServerError(QString());

		m_availableModels = models;
		Q_EMIT availableModelsChanged();

		if (m_model.isEmpty() || !models.contains(m_model)) {
			if (auto* config = config::ConfigSingleton::instance())
				config->llm()->set_model(models.first());
			setModel(models.first());
		}
	});
}

void LlmClient::probeContextSize(std::function<void()> done) {
	if (!m_configEnabled) {
		if (done) done();
		return;
	}

	QString base = m_endpoint.trimmed();
	while (base.endsWith('/'))
		base.chop(1);
	const QUrl url = QUrl::fromUserInput(base + "/props");
	if (!url.isValid() || url.host().isEmpty()) {
		setServerError(
			QStringLiteral("Invalid LLM endpoint: %1").arg(m_endpoint));
		if (done) done();
		return;
	}

	auto* reply = m_manager.get(QNetworkRequest(url));
	auto settled = std::make_shared<bool>(false);
	const auto finish = [this, reply, settled, done]() {
		if (*settled) return;
		*settled = true;

		const QNetworkReply::NetworkError error = reply->error();
		const QString errorString = reply->errorString();
		const QByteArray data = reply->readAll();
		reply->deleteLater();

		int size = 0;
		if (error == QNetworkReply::NoError) {
			setServerError(QString());
			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);
			}
		} else {
			const QString serverMessage = jsonErrorMessage(data);
			setServerError(
				serverMessage.isEmpty()
					? QStringLiteral("Cannot reach LLM server: %1")
						  .arg(errorString)
					: serverMessage);
		}
		if (size > 0)
			setContextSize(size);
		else if (m_tools->contextSize() <= 0)
			setContextSize(4096);
		if (done) done();
	};
	connect(reply, &QNetworkReply::finished, this, finish);
	QTimer::singleShot(kProbeTimeoutMs, this, [reply, settled]() {
		if (!*settled) reply->abort();
	});
}

void LlmClient::updateTokenUsage(const QJsonObject& data) {
	if (!m_turn || m_tools->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_turn->session()->setUsageTokens(static_cast<int>(used));
}

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();

			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();
			onResult(result);
		});
}

void LlmClient::requestTitle(ChatSession* session, const QString& userText) {
	shortRequest(
		QStringLiteral("title"),
		prompts::titleSystemPrompt(),
		userText,
		[this, session = QPointer<ChatSession>(session)](QString title) {
			if (!session) {
				qWarning() << "LlmClient: title request: session gone";
				return;
			}
			title = prompts::cleanTitle(title);
			if (title.length() < 2) {
				qWarning() << "LlmClient: title rejected (too short)" << title;
				return;
			}
			qInfo() << "LlmClient: suggesting title" << title;
			session->setTitle(title);
		});
}

void LlmClient::requestIcon(ChatSession* session, const QString& userText) {
	const QStringList icons = prompts::iconNames();
	shortRequest(
		QStringLiteral("icon"),
		prompts::iconSystemPrompt(),
		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;
			session->setIcon(name);
		});
}

} // namespace ZShell::llm