fix: truncate context logic

Plugins/ZShell/Llm/tool.hpp

@@ -21,29 +21,21 @@ class LlmTool : public QObject {
 	[[nodiscard]] virtual QString description() const = 0;
 	[[nodiscard]] virtual QJsonObject parameters() const = 0;
 
-	// toolCallId identifies the call within the conversation; tools may use
-	// it to name files they write for later retrieval.
 	virtual void execute(
 		const QString& toolCallId,
 		const QJsonObject& args,
+		int outputBudgetChars,
 		std::function<void(const QJsonObject& result)> done) = 0;
 	virtual void cancel();
 
 	[[nodiscard]] QJsonObject specification() const;
 
-	// The endpoint's context size in tokens (0 = unknown). Set by LlmClient.
 	[[nodiscard]] int contextSize() const { return m_contextSize; }
 	void setContextSize(int value);
 
-	// Rough cap, in characters, for tool output embedded in a tool result
-	// message. Scales with the context size so a single result cannot
-	// consume the whole window.
 	[[nodiscard]] int inlineBudgetChars() const;
 
-	// ~4 characters per token; a deliberately coarse estimate used for
-	// budgeting (never for exact accounting).
 	static constexpr int CharsPerToken = 4;
-	// Inline budget when the context size is unknown.
 	static constexpr int DefaultInlineChars = 64 * 1024;
 
 	protected:
@@ -66,7 +58,6 @@ class ToolRegistry : public QObject {
 	[[nodiscard]] QJsonArray specifications() const;
 	void cancelAll();
 
-	// Propagated to every registered tool, including later ones.
 	void setContextSize(int value);
 
 	Q_SIGNALS: