chore: restructure llm plugin to be more coherent & fix various bugs in llm ui

Plugins/ZShell/Llm/llmclient.hpp

@@ -1,18 +1,19 @@
 #pragma once
 
-#include "segment.hpp"
+#include "chatstore.hpp"
 #include "tool.hpp"
+#include "turn.hpp"
 
 #include <QByteArray>
 #include <QJsonArray>
 #include <QNetworkAccessManager>
 #include <QObject>
-#include <QPointer>
 #include <QString>
 #include <QStringList>
 #include <QTimer>
 
 #include <functional>
+#include <memory>
 
 class QJsonObject;
 class QNetworkReply;
@@ -21,80 +22,75 @@ namespace ZShell::llm {
 
 class ChatGeneration;
 class ChatSession;
-class LlmSegment;
-class LlmTool;
 
 class LlmClient : public QObject {
 	Q_OBJECT
+	QML_ELEMENT
+	QML_SINGLETON
+
+	Q_PROPERTY(bool busy READ busy NOTIFY busyChanged)
+	Q_PROPERTY(QString model READ model NOTIFY modelChanged)
+	Q_PROPERTY(
+		QStringList availableModels READ availableModels NOTIFY
+			availableModelsChanged)
+	Q_PROPERTY(int contextSize READ contextSize NOTIFY contextSizeChanged)
+	Q_PROPERTY(
+		bool toolsEnabled READ toolsEnabled WRITE setToolsEnabled NOTIFY
+			toolsEnabledChanged)
+	Q_PROPERTY(
+		bool sandboxEnabled READ sandboxEnabled WRITE setSandboxEnabled NOTIFY
+			sandboxEnabledChanged)
+	Q_PROPERTY(QString serverError READ serverError NOTIFY serverErrorChanged)
+	Q_PROPERTY(ZShell::llm::ChatStore* chats READ chats CONSTANT)
 
 	public:
 	explicit LlmClient(QObject* parent = nullptr);
 	~LlmClient() override;
 
-	[[nodiscard]] QString endpoint() const { return m_endpoint; }
+	[[nodiscard]] ChatStore* chats() const { return m_store; }
+
+	[[nodiscard]] bool busy() const { return m_busy; }
 	[[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; }
-	[[nodiscard]] ToolRegistry* tools() const { return m_tools; }
-	[[nodiscard]] QString serverError() const { return m_serverError; }
-	void setEndpoint(const QString& value);
-	void setModel(const QString& value);
-	void setTemperature(double value);
-	void setContextSize(int size);
-	void probeContextSize(std::function<void()> done = {});
-	void refreshAllStates();
-
-	[[nodiscard]] bool busy() const { return m_busy; }
+	[[nodiscard]] int contextSize() const { return m_tools->contextSize(); }
 	[[nodiscard]] bool toolsEnabled() const { return m_tools->enabled(); }
-	void setToolsEnabled(bool value) { m_tools->setEnabled(value); }
-	[[nodiscard]] bool sandboxEnabled() const { return m_sandboxEnabled; }
+	[[nodiscard]] bool sandboxEnabled() const {
+		return m_tools->sandboxEnabled();
+	}
+	[[nodiscard]] QString serverError() const { return m_serverError; }
+	[[nodiscard]] ChatSession* streamingSession() const {
+		return m_turn ? m_turn->session() : nullptr;
+	}
+
+	void setToolsEnabled(bool value);
 	void setSandboxEnabled(bool value);
-	[[nodiscard]] QString activeWorkspace() const;
-	[[nodiscard]] QString streamingChatId() const { return m_streamingChatId; }
-	[[nodiscard]] ChatSession* streamingSession() const { return m_active; }
+
+	Q_INVOKABLE void selectModel(const QString& id);
+	Q_INVOKABLE void stop();
+	Q_INVOKABLE void refreshModels();
+	Q_INVOKABLE void refreshFromServer();
 
 	void startGeneration(ChatSession* session, ChatGeneration* target);
-	void stop();
-	void endStream();
-	void clearOnFinish(ChatSession* session);
-	void sessionRemoved(ChatSession* session);
-
-	void approveTools();
-	void denyTools();
-
-	void refreshModels();
+	void approveTools(ZShell::llm::ChatGeneration* generation);
+	void denyTools(ZShell::llm::ChatGeneration* generation);
 	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 toolsEnabledChanged();
-	void streamingChatIdChanged();
-	void errorOccurred(const QString& message);
+	void sandboxEnabledChanged();
 	void serverErrorChanged();
-	void titleSuggested(ZShell::llm::ChatSession* session, const QString& title);
-	void iconSuggested(ZShell::llm::ChatSession* session, const QString& icon);
+	void generationFailed(
+		ZShell::llm::ChatSession* session, const QString& message);
 
 	private:
-	struct ToolCallBuilder {
-		QString id;
-		QString name;
-		QString arguments;
-		QPointer<LlmSegment> segment;
-		bool seen = false;
-	};
-
 	void sendRound();
-	QJsonArray buildContextMessages(
-		ChatSession* session, int stopBeforeRow, int extraReserveTokens) const;
-	void appendSandboxPreamble(QJsonArray& messages);
 	void applyToolCallDelta(const QJsonObject& call);
 	void roundFinished();
 	void executeAllCalls();
@@ -102,13 +98,17 @@ class LlmClient : public QObject {
 	void setApprovalPending(bool value);
 	void finishPendingCalls(const QString& resultText);
 	void finishTurn();
+	void endStream();
 	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 setServerError(const QString& value);
+	void setModel(const QString& value);
+	void setContextSize(int size);
+	void probeContextSize(std::function<void()> done = {});
+	void refreshAllStates();
 	void shortRequest(
 		const QString& tag,
 		const QString& systemPrompt,
@@ -119,42 +119,22 @@ class LlmClient : public QObject {
 	static QString jsonErrorMessage(const QByteArray& body);
 	static QString serverErrorMessage(
 		const QByteArray& body, const QString& fallback);
+	[[nodiscard]] ToolContext toolContext() const;
 
 	QNetworkAccessManager m_manager;
+	ChatStore* m_store = nullptr;
 	ToolRegistry* m_tools = nullptr;
+	std::unique_ptr<Turn> m_turn;
 	QNetworkReply* m_reply = nullptr;
 	QByteArray m_buffer;
 	QTimer m_refreshTimer;
-	QPointer<ChatSession> m_active;
-	QPointer<ChatGeneration> m_streaming;
-	QPointer<ChatSession> m_pendingClear;
-	bool m_busy = false;
-	QString m_streamingChatId;
 	QString m_endpoint;
 	QString m_model;
 	QStringList m_availableModels;
 	QString m_serverError;
 	double m_temperature = 0.7;
-	int m_contextSize = 0;
-
-	QJsonArray m_transcript;
-	QList<ToolCallBuilder> m_callBuilders;
-	struct ToolCallResult {
-		QString content;
-		bool success = false;
-	};
-	QList<ToolCallResult> m_callResults;
-	QString m_finishReason;
-	int m_round = 0;
-	qsizetype m_contentMark = 0;
-	qsizetype m_reasoningMark = 0;
-	bool m_sandboxEnabled = true;
-	bool m_roundDone = false;
-	bool m_streamDone = false;
-	bool m_toolPhase = false;
-	bool m_approvalPending = false;
-	bool m_contextRetryUsed = false;
-	int m_pendingCalls = 0;
+	bool m_busy = false;
+	bool m_configEnabled = true;
 	static constexpr int kMaxToolRounds = 12;
 };