fix: truncate webfetch content to fit context

Plugins/ZShell/Llm/webfetchtool.hpp

@@ -20,7 +20,9 @@ class WebFetchTool : public LlmTool {
 	static constexpr int MaxResponseBytes = 5 * 1024 * 1024;
 	static constexpr int DefaultTimeoutSeconds = 30;
 	static constexpr int MaxTimeoutSeconds = 120;
-	static constexpr int MaxOutputChars = 64 * 1024;
+	// Where the full (untruncated) output of each call is stored so the
+	// model can page through it with the readfile tool.
+	static const QString StoragePath;
 
 	explicit WebFetchTool(QObject* parent = nullptr);
 	~WebFetchTool() override;
@@ -29,6 +31,7 @@ class WebFetchTool : public LlmTool {
 	QString description() const override;
 	QJsonObject parameters() const override;
 	void execute(
+		const QString& toolCallId,
 		const QJsonObject& args,
 		std::function<void(const QJsonObject& result)> done) override;
 	void cancel() override;
@@ -43,10 +46,15 @@ class WebFetchTool : public LlmTool {
 		QByteArray body;
 		bool tooLarge = false;
 		QString format;
+		QString toolCallId;
 		std::function<void(const QJsonObject& result)> done;
 	};
 
 	void completeJob(Job* job, QJsonObject result);
+	// Writes the full output to StoragePath; returns the file path, or an
+	// empty string when saving failed.
+	QString saveToFile(const Job& job, const QString& content,
+					   const QString& mime) const;
 
 	QNetworkAccessManager m_manager;
 	QList<Job*> m_jobs;