fix: truncate context logic

Plugins/ZShell/Llm/webfetchtool.hpp

@@ -20,8 +20,6 @@ class WebFetchTool : public LlmTool {
 	static constexpr int MaxResponseBytes = 5 * 1024 * 1024;
 	static constexpr int DefaultTimeoutSeconds = 30;
 	static constexpr int MaxTimeoutSeconds = 120;
-	// 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);
@@ -33,6 +31,7 @@ class WebFetchTool : public LlmTool {
 	void execute(
 		const QString& toolCallId,
 		const QJsonObject& args,
+		int outputBudgetChars,
 		std::function<void(const QJsonObject& result)> done) override;
 	void cancel() override;
 
@@ -47,14 +46,13 @@ class WebFetchTool : public LlmTool {
 		bool tooLarge = false;
 		QString format;
 		QString toolCallId;
+		int budgetChars = 0;
 		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;
+	QString saveToFile(
+		const Job& job, const QString& content, const QString& mime) const;
 
 	QNetworkAccessManager m_manager;
 	QList<Job*> m_jobs;