fix: truncate context logic

Plugins/ZShell/Llm/llmclient.hpp

@@ -10,6 +10,7 @@
 #include <QPointer>
 #include <QString>
 #include <QStringList>
+#include <QTimer>
 
 #include <functional>
 
@@ -42,7 +43,8 @@ class LlmClient : public QObject {
 	void setModel(const QString& value);
 	void setTemperature(double value);
 	void setContextSize(int size);
-	void probeContextSize();
+	void probeContextSize(std::function<void()> done = {});
+	Q_INVOKABLE void refreshFromServer();
 
 	[[nodiscard]] bool busy() const { return m_busy; }
 	[[nodiscard]] bool toolsEnabled() const { return m_tools->enabled(); }
@@ -56,10 +58,6 @@ class LlmClient : public QObject {
 	void clearOnFinish(ChatSession* session);
 	void sessionRemoved(ChatSession* session);
 
-	// Tool calls are executed only after the user approves the batch the
-	// model requested. The per-generation state (toolApprovalPending,
-	// pendingToolCalls) is mirrored onto ChatGeneration; these are the
-	// entry points QML reaches through it.
 	void approveTools();
 	void denyTools();
 
@@ -118,9 +116,7 @@ class LlmClient : public QObject {
 	ToolRegistry* m_tools = nullptr;
 	QNetworkReply* m_reply = nullptr;
 	QByteArray m_buffer;
-	// QPointer: these objects are owned by the ChatStore tree, which Qt
-	// destroys *before* this client on shutdown (children die in creation
-	// order). The pointers must self-null instead of dangling.
+	QTimer m_refreshTimer;
 	QPointer<ChatSession> m_active;
 	QPointer<ChatGeneration> m_streaming;
 	QPointer<ChatSession> m_pendingClear;
@@ -146,6 +142,7 @@ class LlmClient : public QObject {
 	bool m_roundDone = false;
 	bool m_toolPhase = false;
 	bool m_approvalPending = false;
+	bool m_contextRetryUsed = false;
 	int m_pendingCalls = 0;
 	static constexpr int kMaxToolRounds = 12;
 };