fix: truncate context logic

Modules/Notifications/Sidebar/Chat/Content/ProcessBlock.qml

@@ -154,11 +154,12 @@ Item {
 					Layout.fillWidth: true
 
 					function reparentWrapper(): void {
-						const newParent = open ? rootParent : toolsBg;
+						const newParent = open ? toolsBg.rootParent : toolsBg;
 						const pos = toolsWrapper.mapToItem(newParent, 0, 0);
 						toolsWrapper.parent = newParent;
 						toolsWrapper.x = pos.x;
 						toolsWrapper.y = pos.y;
+						console.log("X, Y:\n" + toolsWrapper.x, toolsWrapper.y, "\nWidth, Height:\n" + toolsWrapper.width, toolsWrapper.height);
 					}
 
 					BlobGroup {
@@ -187,7 +188,7 @@ Item {
 						id: toolsWrapper
 
 						width: toolsBg.width
-						height: toolList.implicitHeight + toolList.anchors.margins * 2
+						height: toolBox.implicitHeight
 
 						states: State {
 							name: "open"
@@ -265,16 +266,24 @@ Item {
 
 						CustomRect {
 							id: toolBox
-							anchors.fill: parent
+							anchors.left: parent.left
+							anchors.right: parent.right
+							height: Math.min(implicitHeight, parent.height)
+							implicitHeight: toolList.implicitHeight + Tokens.padding.medium * 2
 							radius: Tokens.rounding.small
 
 							StateLayer {
-								onClicked: toolsBg.open = true
+								onClicked: {
+									console.log("X, Y:\n" + toolsWrapper.x, toolsWrapper.y, "\nWidth, Height:\n" + toolsWrapper.width, toolsWrapper.height);
+									toolsBg.open = true;
+								}
 							}
 
 							ColumnLayout {
 								id: toolList
-								anchors.fill: parent
+								anchors.left: parent.left
+								anchors.right: parent.right
+								anchors.verticalCenter: parent.verticalCenter
 								anchors.margins: Tokens.padding.medium
 
 								Repeater {

Modules/Notifications/Sidebar/Content.qml

@@ -77,6 +77,7 @@ Item {
 				anchors.top: parent.top
 				active: Config.llm.enabled
 				width: parent.width
+				property bool tre: true
 
 				sourceComponent: Item {
 					id: chatPage

Plugins/ZShell/Llm/chat.cpp

@@ -40,7 +40,6 @@ Chat::Chat(QObject* parent)
 	});
 
 	connect(m_client, &LlmClient::busyChanged, this, [this]() {
-		// A fresh run supersedes the previous error.
 		if (m_client->busy() && !m_lastError.isEmpty()) {
 			m_lastError.clear();
 			Q_EMIT lastErrorChanged();
@@ -161,6 +160,10 @@ void Chat::refreshModels() {
 	m_client->refreshModels();
 }
 
+void Chat::refreshFromServer() {
+	m_client->refreshFromServer();
+}
+
 void Chat::selectModel(const QString& id) {
 	if (id.isEmpty()) return;
 	m_client->setModel(id);

Plugins/ZShell/Llm/chat.hpp

@@ -14,9 +14,6 @@ namespace ZShell::llm {
 
 class LlmClient;
 
-// QML-facing facade. Persistence lives in ChatStore, network streaming in
-// LlmClient; this class only wires them together and exposes the
-// application-wide state.
 class Chat : public QObject {
 	Q_OBJECT
 	QML_ELEMENT
@@ -55,6 +52,7 @@ class Chat : public QObject {
 	Q_INVOKABLE void stop();
 	Q_INVOKABLE void dismissError();
 	Q_INVOKABLE void refreshModels();
+	Q_INVOKABLE void refreshFromServer();
 	Q_INVOKABLE void selectModel(const QString& id);
 
 	static Chat* create(QQmlEngine*, QJSEngine*);

Plugins/ZShell/Llm/codehighlighter.hpp

@@ -37,10 +37,10 @@ class CodeHighlighter : public QObject {
 
 	private:
 	struct State {
-		bool bad = false; // permanent failure, do not retry
+		bool bad = false;
 		void* lib = nullptr;
-		const void* lang = nullptr; // const TSLanguage*
-		void* query = nullptr;		// TSQuery*
+		const void* lang = nullptr;
+		void* query = nullptr;
 	};
 
 	[[nodiscard]] static const QHash<QString, QString>& aliases();
@@ -59,15 +59,14 @@ class CodeHighlighter : public QObject {
 		const QVariantList& spans) const;
 
 	struct SpanCacheEntry {
-		QString code; // re-compared on lookup; a hash collision can
-					  // never deliver the wrong spans
+		QString code;
 		QVariantList spans;
 	};
 
 	mutable QHash<QString, std::shared_ptr<const State>> m_states;
 	mutable QMutex m_stateMutex;
 	mutable QHash<QString, SpanCacheEntry> m_spanCache;
-	mutable QStringList m_spanCacheOrder; // LRU order, oldest first
+	mutable QStringList m_spanCacheOrder;
 	mutable int m_spanCacheBytes = 0;
 	mutable QMutex m_cacheMutex;
 	static CodeHighlighter* s_instance;

Plugins/ZShell/Llm/filetool.cpp

@@ -53,9 +53,9 @@ QJsonObject FileReadTool::parameters() const {
 	QJsonObject limit;
 	limit[QStringLiteral("type")] = QStringLiteral("integer");
 	limit[QStringLiteral("minimum")] = 1;
-	limit[QStringLiteral("description")] =
-		QStringLiteral("Maximum number of characters to return. Defaults "
-					   "to a value that fits the model's context window.");
+	limit[QStringLiteral("description")] = QStringLiteral(
+		"Maximum number of characters to return. Defaults "
+		"to a value that fits the model's context window.");
 
 	QJsonObject properties;
 	properties[QStringLiteral("path")] = path;
@@ -74,6 +74,7 @@ QJsonObject FileReadTool::parameters() const {
 void FileReadTool::execute(
 	const QString& toolCallId,
 	const QJsonObject& args,
+	int outputBudgetChars,
 	std::function<void(const QJsonObject&)> done) {
 	Q_UNUSED(toolCallId);
 	const QString path = args[QStringLiteral("path")].toString().trimmed();
@@ -82,7 +83,8 @@ void FileReadTool::execute(
 		return;
 	}
 	const int offset = qMax(0, args[QStringLiteral("offset")].toInt(0));
-	const int budget = inlineBudgetChars();
+	int budget = inlineBudgetChars();
+	if (outputBudgetChars > 0) budget = qMin(budget, outputBudgetChars);
 	const int limit =
 		qBound(1, args[QStringLiteral("limit")].toInt(budget), budget);
 
@@ -90,8 +92,9 @@ void FileReadTool::execute(
 	if (!file.open(QIODevice::ReadOnly)) {
 		const QFileInfo info(path);
 		if (info.exists() && info.isDir()) {
-			done(makeError(QStringLiteral("Path is a directory, not a "
-										   "file: %1")
+			done(makeError(QStringLiteral(
+							   "Path is a directory, not a "
+							   "file: %1")
 							   .arg(path)));
 		} else {
 			done(makeError(QStringLiteral("Cannot open file: %1").arg(path)));
@@ -106,8 +109,9 @@ void FileReadTool::execute(
 	}
 	if (offset >= fileSize) {
 		file.close();
-		done(makeError(QStringLiteral("Offset %1 is beyond the end of the "
-									   "file (%2 bytes)")
+		done(makeError(QStringLiteral(
+						   "Offset %1 is beyond the end of the "
+						   "file (%2 bytes)")
 						   .arg(offset)
 						   .arg(fileSize)));
 		return;
@@ -117,15 +121,13 @@ void FileReadTool::execute(
 		done(makeError(QStringLiteral("Cannot seek in file: %1").arg(path)));
 		return;
 	}
-	// UTF-8 uses at most 4 bytes per character, so this many bytes always
-	// cover `limit` characters.
 	const qint64 wanted = qint64(limit) * 4 + 16;
 	const QByteArray chunk = file.read(qMin(wanted, fileSize - offset));
 	file.close();
 
 	if (chunk.contains('\0')) {
-		done(makeError(QStringLiteral("Binary files are not supported: %1")
-						   .arg(path)));
+		done(makeError(
+			QStringLiteral("Binary files are not supported: %1").arg(path)));
 		return;
 	}
 	const QString content = QString::fromUtf8(chunk);

Plugins/ZShell/Llm/filetool.hpp

@@ -4,10 +4,6 @@
 
 namespace ZShell::llm {
 
-// Lets the model read text files from the local filesystem, in
-// byte-offset windows small enough to fit the context. Primarily used to
-// page through the full webfetch output stored under
-// /tmp/zshell-llm/webfetch/.
 class FileReadTool : public LlmTool {
 	Q_OBJECT
 
@@ -20,6 +16,7 @@ class FileReadTool : public LlmTool {
 	void execute(
 		const QString& toolCallId,
 		const QJsonObject& args,
+		int outputBudgetChars,
 		std::function<void(const QJsonObject& result)> done) override;
 };
 

Plugins/ZShell/Llm/generation.cpp

@@ -220,8 +220,6 @@ LlmSegment* ChatGeneration::beginToolCall(
 		LlmSegment::Type::ToolCall, QDateTime::currentMSecsSinceEpoch(), this);
 	segment->setName(name);
 	segment->setToolCallId(toolCallId);
-	// Awaiting user approval until LlmClient::approveTools() promotes it
-	// to Running; denyTools()/endStream() finalize it as an error.
 	segment->setStatus(LlmSegment::Status::Pending);
 	segment->begin();
 	addSegment(segment);

Plugins/ZShell/Llm/llmclient.cpp

@@ -12,15 +12,16 @@
 #include <QJsonObject>
 #include <QNetworkReply>
 #include <QNetworkRequest>
+#include <QRegularExpression>
 #include <QSet>
 #include <QUrl>
 
+#include <memory>
+
 namespace ZShell::llm {
 
 namespace {
 
-// Minimum prompt budget kept even when the transcript alone nearly fills
-// the window, so the model always sees the triggering message.
 constexpr int kMinContextTokens = 512;
 
 qsizetype jsonValueChars(const QJsonValue& value) {
@@ -49,16 +50,11 @@ int estimateMessageTokens(const QJsonObject& message) {
 	return int(jsonValueChars(message) / LlmTool::CharsPerToken) + 4;
 }
 
-// One message model row and the JSON messages it produces. Trimming works
-// in whole units so an assistant turn and its tool results always travel
-// together.
 struct ContextUnit {
 	QList<QJsonObject> messages;
 	int tokens = 0;
 };
 
-// Shrink the free-text fields of a unit until it fits the budget. Used
-// when a single unit (e.g. one huge user message) alone exceeds it.
 void shrinkUnitToBudget(QList<QJsonObject>& unit, int budgetTokens) {
 	static const QStringList kTextKeys = {
 		QStringLiteral("content"), QStringLiteral("reasoning_content")};
@@ -90,6 +86,41 @@ void shrinkUnitToBudget(QList<QJsonObject>& unit, int budgetTokens) {
 	}
 }
 
+int estimateArrayTokens(const QJsonArray& messages) {
+	int total = 0;
+	for (const QJsonValue& value : messages)
+		total += estimateMessageTokens(value.toObject());
+	return total;
+}
+
+constexpr int kToolResultStubChars = 512;
+
+constexpr double kPromptEstimateDiscount = 10.0 / 12.0;
+
+constexpr int kMinToolResultTokens = 128;
+
+int completionReserve(int contextSize) {
+	return qBound(2048, contextSize / 8, 8192);
+}
+
+int promptBudgetTokens(int contextSize) {
+	return int(
+		(contextSize - completionReserve(contextSize)) *
+		kPromptEstimateDiscount);
+}
+
+constexpr int kRefreshIntervalMs = 60'000;
+
+constexpr int kProbeTimeoutMs = 2000;
+
+int contextSizeFromErrorMessage(const QString& message) {
+	static const QRegularExpression re(
+		QLatin1String("available context size \\((\\d+)"));
+	const auto match = re.match(message);
+	if (match.hasMatch()) return match.captured(1).toInt();
+	return 0;
+}
+
 } // namespace
 
 QString LlmClient::completionsPath(
@@ -140,7 +171,12 @@ LlmClient::LlmClient(QObject* parent) : QObject(parent) {
 		&ToolRegistry::enabledChanged,
 		this,
 		&LlmClient::toolsEnabledChanged);
-	probeContextSize();
+
+	m_refreshTimer.setParent(this);
+	m_refreshTimer.setInterval(kRefreshIntervalMs);
+	connect(
+		&m_refreshTimer, &QTimer::timeout, this, &LlmClient::refreshFromServer);
+	m_refreshTimer.start();
 }
 
 LlmClient::~LlmClient() {
@@ -153,8 +189,8 @@ void LlmClient::setEndpoint(const QString& value) {
 	if (m_endpoint == value) return;
 	m_endpoint = value;
 	Q_EMIT endpointChanged();
-	probeContextSize();
-	if (m_model.isEmpty()) refreshModels();
+	m_refreshTimer.start();
+	refreshFromServer();
 }
 
 void LlmClient::setModel(const QString& value) {
@@ -170,6 +206,8 @@ void LlmClient::setTemperature(double value) {
 
 void LlmClient::startGeneration(ChatSession* session, ChatGeneration* target) {
 	if (m_busy || !session || !target) return;
+	refreshModels();
+	m_contextRetryUsed = false;
 	m_active = session;
 	m_streaming = target;
 	target->setStreaming(true);
@@ -202,11 +240,15 @@ void LlmClient::startGeneration(ChatSession* session, ChatGeneration* target) {
 	m_reasoningMark = 0;
 	setApprovalPending(false);
 
-	sendRound();
+	probeContextSize([this, session, target]() {
+		if (m_active != session || m_streaming != target) return;
+		sendRound();
+	});
 }
 
 void LlmClient::sendRound() {
 	if (!m_active || !m_streaming) return;
+	probeContextSize();
 	m_finishReason.clear();
 	m_callBuilders.clear();
 	m_callResults.clear();
@@ -234,8 +276,6 @@ void LlmClient::sendRound() {
 		return;
 	}
 
-	// The transcript (this generation so far) is always sent in full, so
-	// reserve its estimated size out of the context budget.
 	int transcriptTokens = 0;
 	for (const QJsonValue& value : m_transcript)
 		transcriptTokens += estimateMessageTokens(value.toObject());
@@ -244,6 +284,30 @@ void LlmClient::sendRound() {
 	for (const QJsonValue& value : m_transcript)
 		messages.append(value);
 
+	if (m_contextSize > 0) {
+		const int promptLimit = promptBudgetTokens(m_contextSize);
+		while (estimateArrayTokens(messages) > promptLimit) {
+			bool stubbed = false;
+			for (qsizetype i = 0; i < messages.size(); ++i) {
+				QJsonObject message = messages.at(i).toObject();
+				if (message.value("role").toString() != QLatin1String("tool"))
+					continue;
+				const QString content = message.value("content").toString();
+				if (content.size() <= kToolResultStubChars) continue;
+				QString stub = content.left(kToolResultStubChars);
+				stub += QStringLiteral(
+							"\n[truncated to fit the context window; was "
+							"%1 characters]")
+							.arg(content.size());
+				message["content"] = stub;
+				messages[i] = message;
+				stubbed = true;
+				break;
+			}
+			if (!stubbed) break;
+		}
+	}
+
 	QJsonObject body;
 	QJsonObject streamOptions;
 	streamOptions[QStringLiteral("include_usage")] = true;
@@ -283,8 +347,21 @@ void LlmClient::sendRound() {
 			roundFinished();
 		else if (error == QNetworkReply::OperationCanceledError)
 			finishTurn();
-		else
-			fail(serverErrorMessage(responseBody, errorString));
+		else {
+			const QString message =
+				serverErrorMessage(responseBody, errorString);
+			const int reported = contextSizeFromErrorMessage(message);
+			if (reported > 0 && reported != m_contextSize &&
+				!m_contextRetryUsed) {
+				qWarning() << "LlmClient: prompt overflow, server reports"
+						   << reported << "tokens; retrying round";
+				setContextSize(reported);
+				m_contextRetryUsed = true;
+				sendRound();
+				return;
+			}
+			fail(message);
+		}
 	});
 }
 
@@ -349,21 +426,17 @@ QJsonArray LlmClient::buildContextMessages(
 	}
 
 	if (m_contextSize > 0) {
-		// Keep headroom for the model's own reply so the prompt cannot
-		// fill the whole window.
-		const int completionReserve = qBound(1024, m_contextSize / 8, 8192);
-		int budget = m_contextSize - completionReserve - extraReserveTokens;
+		int budget = m_contextSize - completionReserve(m_contextSize) -
+					 extraReserveTokens;
 		budget = qMax(budget, kMinContextTokens);
 
 		int total = 0;
 		for (const ContextUnit& unit : units)
 			total += unit.tokens;
-		// Drop oldest units until the prompt fits the budget.
 		while (units.size() > 1 && total > budget) {
 			total -= units.first().tokens;
 			units.removeFirst();
 		}
-		// A conversation cannot start with an assistant turn.
 		while (
 			units.size() > 1 &&
 			units.first().messages.first()[QStringLiteral("role")].toString() ==
@@ -371,7 +444,6 @@ QJsonArray LlmClient::buildContextMessages(
 			total -= units.first().tokens;
 			units.removeFirst();
 		}
-		// A single oversized unit still gets sent, shrunk to fit.
 		if (units.size() == 1 && units.first().tokens > budget)
 			shrinkUnitToBudget(units.first().messages, budget);
 	}
@@ -417,7 +489,21 @@ void LlmClient::roundFinished() {
 	bool hasCalls = false;
 	for (const auto& builder : m_callBuilders)
 		if (builder.seen) hasCalls = true;
-	if (m_finishReason != QLatin1String("tool_calls") && !hasCalls) {
+	if (!hasCalls) {
+		finishTurn();
+		return;
+	}
+	if (m_finishReason != QLatin1String("tool_calls")) {
+		qWarning() << "LlmClient: stream ended" << m_finishReason
+				   << "with incomplete tool calls, ending turn";
+		for (const auto& builder : m_callBuilders) {
+			if (!builder.seen || !builder.segment) continue;
+			builder.segment->finishTool(
+				QStringLiteral(
+					"Stream truncated before the tool call "
+					"completed"),
+				false);
+		}
 		finishTurn();
 		return;
 	}
@@ -458,8 +544,6 @@ void LlmClient::roundFinished() {
 	m_transcript.append(assistant);
 	m_round++;
 
-	// The whole batch of this round waits for a single user decision;
-	// approveTools() runs all of it, denyTools() ends the turn.
 	setApprovalPending(true);
 }
 
@@ -468,6 +552,18 @@ void LlmClient::executeAllCalls() {
 	m_pendingCalls = 0;
 	m_callResults = QList<ToolCallResult>(m_callBuilders.size());
 
+	int batchSize = 0;
+	for (const auto& call : m_callBuilders)
+		if (call.seen) ++batchSize;
+	int perCallChars = 0;
+	if (batchSize > 0 && m_contextSize > 0) {
+		const int available = promptBudgetTokens(m_contextSize) -
+							  estimateArrayTokens(m_transcript);
+		const int perCallTokens =
+			qMax(available / batchSize, kMinToolResultTokens);
+		perCallChars = perCallTokens * LlmTool::CharsPerToken;
+	}
+
 	for (int i = 0; i < m_callBuilders.size(); ++i) {
 		const auto& call = m_callBuilders.at(i);
 		if (!call.seen) continue;
@@ -498,18 +594,22 @@ void LlmClient::executeAllCalls() {
 		}
 
 		++m_pendingCalls;
-		tool->execute(call.id, args, [this, i, call](const QJsonObject& result) {
-			if (!m_streaming) return;
-			const bool success = result.contains(QStringLiteral("output"));
-			const QString content =
-				success ? result[QStringLiteral("output")].toString()
-						: QStringLiteral("Error: ") +
-							  result[QStringLiteral("error")].toString();
-			m_callResults[i] = {content, success};
-			if (LlmSegment* segment = call.segment)
-				segment->finishTool(content, success);
-			if (--m_pendingCalls == 0) flushCallResults();
-		});
+		tool->execute(
+			call.id,
+			args,
+			perCallChars,
+			[this, i, call](const QJsonObject& result) {
+				if (!m_streaming) return;
+				const bool success = result.contains(QStringLiteral("output"));
+				const QString content =
+					success ? result[QStringLiteral("output")].toString()
+							: QStringLiteral("Error: ") +
+								  result[QStringLiteral("error")].toString();
+				m_callResults[i] = {content, success};
+				if (LlmSegment* segment = call.segment)
+					segment->finishTool(content, success);
+				if (--m_pendingCalls == 0) flushCallResults();
+			});
 	}
 	if (m_pendingCalls == 0) flushCallResults();
 }
@@ -533,8 +633,6 @@ void LlmClient::flushCallResults() {
 void LlmClient::setApprovalPending(bool value) {
 	if (m_approvalPending == value) return;
 	m_approvalPending = value;
-	// Mirror onto the streaming generation; QML reads and answers from
-	// there, so only the affected ProcessBlock reacts.
 	if (ChatGeneration* streaming = m_streaming)
 		streaming->setApprovalPending(value);
 }
@@ -548,8 +646,6 @@ void LlmClient::approveTools() {
 		if (segment->type() != LlmSegment::Type::ToolCall ||
 			segment->status() != LlmSegment::Status::Pending)
 			continue;
-		// Restart the segment timer so elapsed time covers execution, not
-		// the wait for approval.
 		segment->setStatus(LlmSegment::Status::Running);
 		segment->begin();
 	}
@@ -592,7 +688,10 @@ void LlmClient::stop() {
 		finishTurn();
 		return;
 	}
-	if (m_reply) m_reply->abort();
+	if (m_reply)
+		m_reply->abort();
+	else
+		endStream();
 }
 
 void LlmClient::endStream() {
@@ -604,8 +703,6 @@ void LlmClient::endStream() {
 	generation->setStreaming(false);
 	m_approvalPending = false;
 	generation->setApprovalPending(false);
-	// Finalize tool calls that never received an approval (turn cancelled,
-	// round limit reached, ...).
 	for (auto* segment : generation->segments())
 		if (segment->type() == LlmSegment::Type::ToolCall &&
 			segment->status() == LlmSegment::Status::Pending)
@@ -633,6 +730,7 @@ void LlmClient::finishTurn() {
 		m_pendingClear.clear();
 	}
 	if (session) session->persist();
+	refreshFromServer();
 }
 
 void LlmClient::clearOnFinish(ChatSession* session) {
@@ -707,6 +805,12 @@ void LlmClient::handleLine(const QByteArray& line) {
 	}
 }
 
+void LlmClient::refreshFromServer() {
+	if (m_endpoint.trimmed().isEmpty()) return;
+	probeContextSize();
+	refreshModels();
+}
+
 void LlmClient::refreshModels() {
 	const QUrl url =
 		QUrl::fromUserInput(completionsPath(m_endpoint, "/models"));
@@ -742,6 +846,9 @@ void LlmClient::refreshModels() {
 		if (m_model.isEmpty()) {
 			m_model = models.first();
 			Q_EMIT modelChanged();
+		} else if (!models.contains(m_model)) {
+			m_model = models.first();
+			Q_EMIT modelChanged();
 		}
 	});
 }
@@ -753,15 +860,22 @@ void LlmClient::setContextSize(int size) {
 	Q_EMIT contextSizeChanged();
 }
 
-void LlmClient::probeContextSize() {
+void LlmClient::probeContextSize(std::function<void()> done) {
 	QString base = m_endpoint.trimmed();
 	while (base.endsWith('/'))
 		base.chop(1);
 	const QUrl url = QUrl::fromUserInput(base + "/props");
-	if (!url.isValid() || url.host().isEmpty()) return;
+	if (!url.isValid() || url.host().isEmpty()) {
+		if (done) done();
+		return;
+	}
 
 	auto* reply = m_manager.get(QNetworkRequest(url));
-	connect(reply, &QNetworkReply::finished, this, [this, reply]() {
+	auto settled = std::make_shared<bool>(false);
+	const auto finish = [this, reply, settled, done]() {
+		if (*settled) return;
+		*settled = true;
+
 		const QNetworkReply::NetworkError error = reply->error();
 		const QByteArray data = reply->readAll();
 		reply->deleteLater();
@@ -786,7 +900,15 @@ void LlmClient::probeContextSize() {
 							   .toInt(0);
 			}
 		}
-		setContextSize(size > 0 ? size : 4096);
+		if (size > 0)
+			setContextSize(size);
+		else if (m_contextSize <= 0)
+			setContextSize(4096);
+		if (done) done();
+	};
+	connect(reply, &QNetworkReply::finished, this, finish);
+	QTimer::singleShot(kProbeTimeoutMs, this, [reply, settled]() {
+		if (!*settled) reply->abort();
 	});
 }
 

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;
 };

Plugins/ZShell/Llm/segment.hpp

@@ -29,8 +29,6 @@ class LlmSegment : public QObject {
 	enum class Type : int { Reasoning = 0, ToolCall, Content };
 	Q_ENUM(Type)
 
-	// Pending is appended last: statuses are persisted as integers in the
-	// chat database, so existing values must not move.
 	enum class Status : int { None = 0, Running, Success, Error, Pending };
 	Q_ENUM(Status)
 

Plugins/ZShell/Llm/tool.cpp

@@ -13,8 +13,6 @@ void LlmTool::setContextSize(int value) {
 
 int LlmTool::inlineBudgetChars() const {
 	if (m_contextSize <= 0) return DefaultInlineChars;
-	// About a quarter of the context window, clamped so tiny contexts
-	// still get a usable result and huge ones stay sane.
 	const int tokens =
 		qBound(512, m_contextSize / 4, DefaultInlineChars / CharsPerToken);
 	return tokens * CharsPerToken;

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:

Plugins/ZShell/Llm/webfetchtool.cpp

@@ -112,10 +112,12 @@ void WebFetchTool::completeJob(Job* job, QJsonObject result) {
 void WebFetchTool::execute(
 	const QString& toolCallId,
 	const QJsonObject& args,
+	int outputBudgetChars,
 	std::function<void(const QJsonObject&)> done) {
 	auto* job = new Job;
 	job->done = std::move(done);
 	job->toolCallId = toolCallId;
+	job->budgetChars = outputBudgetChars;
 	m_jobs.append(job);
 
 	auto fail = [this, job](const QString& message) {
@@ -248,23 +250,24 @@ void WebFetchTool::execute(
 		const QString savedPath = saveToFile(*job, content, mime);
 
 		QString output = content;
-		const int budget = inlineBudgetChars();
+		int budget = inlineBudgetChars();
+		if (job->budgetChars > 0) budget = qMin(budget, job->budgetChars);
 		if (content.size() > budget) {
 			output = content.left(budget);
-			const QString note = QStringLiteral(
-				"\n\n[... truncated: showing %1 of %2 characters")
-								.arg(budget)
-								.arg(content.size());
+			const QString note =
+				QStringLiteral(
+					"\n\n[... truncated: showing %1 of %2 characters")
+					.arg(budget)
+					.arg(content.size());
 			if (!savedPath.isEmpty())
-				output += note +
-						  QStringLiteral(
-							  ". The full content is saved to %1; use "
-							  "the readfile tool to read the rest.")
-							  .arg(savedPath);
+				output += note + QStringLiteral(
+									 ". The full content is saved to %1; use "
+									 "the readfile tool to read the rest.")
+									 .arg(savedPath);
 			else
 				output += note + QStringLiteral(
-								   ". The remaining content is not "
-								   "available.");
+									 ". The remaining content is not "
+									 "available.");
 		}
 		completeJob(job, makeOutput(output));
 	});
@@ -273,14 +276,13 @@ void WebFetchTool::execute(
 QString WebFetchTool::saveToFile(
 	const Job& job, const QString& content, const QString& mime) const {
 	QDir dir(StoragePath);
-	if (!dir.exists() && !dir.mkpath(QStringLiteral(".")))
-		return QString();
+	if (!dir.exists() && !dir.mkpath(QStringLiteral("."))) return QString();
 
 	QString fileName;
 	fileName.reserve(job.toolCallId.size());
 	for (const QChar& c : job.toolCallId)
 		fileName += (c.isLetterOrNumber() || c == QLatin1Char('-') ||
-					  c == QLatin1Char('_'))
+					 c == QLatin1Char('_'))
 						? c
 						: QLatin1Char('_');
 	if (fileName.isEmpty()) fileName = QStringLiteral("fetch");
@@ -293,8 +295,7 @@ QString WebFetchTool::saveToFile(
 	else if (mime.contains(QLatin1String("xml")))
 		extension = QStringLiteral("xml");
 
-	const QString path =
-		dir.filePath(fileName + QLatin1Char('.') + extension);
+	const QString path = dir.filePath(fileName + QLatin1Char('.') + extension);
 	QFile file(path);
 	if (!file.open(QIODevice::WriteOnly | QIODevice::Truncate))
 		return QString();

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;