#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