@librechat/agents 3.3.8 → 3.3.10

package/dist/esm/graphs/Graph.mjs

@@ -1,15 +1,14 @@
-import { resolveLangfuseConfig } from "../langfuseConfig.mjs";
+import { hasToolOutputTracingConfig, resolveLangfuseConfig, resolveToolOutputTracingConfig } from "../langfuseConfig.mjs";
 import { resolveLangfuseRuntimeScope, withLangfuseRuntimeScope } from "../langfuseRuntimeScope.mjs";
 import { createLangfuseHandler, createLangfuseTraceMetadata, disposeLangfuseHandler, isLangfuseCallbackHandler } from "../langfuse.mjs";
-import { shouldTraceToolNodeForLangfuse } from "../langfuseToolOutputTracing.mjs";
-import { appendCallbacks, findCallback } from "../utils/callbacks.mjs";
 import { PREEMPT_BOUNDARY_HOOK_TIMEOUT_MS } from "../common/constants.mjs";
 import { GraphNodeKeys } from "../common/enum.mjs";
 import "../common/index.mjs";
+import { shouldTraceToolNodeForLangfuse } from "../langfuseToolOutputTracing.mjs";
+import { appendCallbacks, findCallback } from "../utils/callbacks.mjs";
 import { calculateMaxToolResultChars } from "../utils/truncation.mjs";
 import { compactToolContent, getToolContentCharLength, serializeToolContentBounded } from "../utils/toolContent.mjs";
 import { apportionTokenCounts } from "../utils/tokens.mjs";
-import { resolveMaxSeals } from "../llm/preempt.mjs";
 import { initializeLangfuseTracing } from "../instrumentation.mjs";
 import { addBedrockTailCacheControl, addTailCacheControl, resolveBedrockPromptCacheTtl, resolvePromptCacheTtl, supportsBedrockToolCache } from "../messages/cache.mjs";
 import { convertMessagesToContent, projectAnthropicArtifactContent, projectArtifactPayload, projectToolStreamContentForProvider } from "../messages/core.mjs";
@@ -27,14 +26,15 @@ import { appendPredecessorHandoffCue, removePredecessorHandoffCue } from "../mes
 import { messagesStateReducer as messagesStateReducer$1 } from "../messages/reducer.mjs";
 import { splitAtRecencyBoundary } from "../messages/recency.mjs";
 import "../messages/index.mjs";
-import { joinKeys, resetIfNotEmpty } from "../utils/graph.mjs";
-import { isAnthropicLike, isGoogleLike, isOpenAILike } from "../utils/llm.mjs";
-import { handleToolCalls } from "../tools/handlers.mjs";
 import { ToolOutputReferenceRegistry, annotateMessagesForLLM } from "../tools/toolOutputReferences.mjs";
+import { partitionAndMarkBedrockToolCache } from "../llm/bedrock/toolCache.mjs";
+import { resolveMaxSeals } from "../llm/preempt.mjs";
+import { handleToolCalls } from "../tools/handlers.mjs";
+import { isAnthropicLike, isGoogleLike, isOpenAILike } from "../utils/llm.mjs";
+import { joinKeys, resetIfNotEmpty } from "../utils/graph.mjs";
 import { HandlerRegistry } from "../events.mjs";
 import { sleep } from "../utils/run.mjs";
 import "../utils/index.mjs";
-import { partitionAndMarkBedrockToolCache } from "../llm/bedrock/toolCache.mjs";
 import { initializeModel } from "../llm/init.mjs";
 import { attemptInvoke, getFallbackErrorContext, getFallbackOverflowCandidates, projectMessagesForProvider, resolveServingModelId, tryFallbackProviders } from "../llm/invoke.mjs";
 import { getBlindRecoveryBudget, planContextOverflowRecovery, translateRecoveryBudget } from "../llm/contextOverflowRecovery.mjs";
@@ -56,12 +56,12 @@ import { createSummarizeNode } from "../summarization/node.mjs";
 import { createSchemaOnlyTools } from "../tools/schema.mjs";
 import { AgentContext } from "../agents/AgentContext.mjs";
 import { createFakeStreamingLLM } from "../llm/fake.mjs";
+import { nanoid } from "nanoid";
 import { AIMessageChunk, HumanMessage, ToolMessage } from "@langchain/core/messages";
 import { Annotation, END, START, StateGraph } from "@langchain/langgraph";
 import { tool } from "@langchain/core/tools";
-import { nanoid } from "nanoid";
-import { ContextOverflowError } from "@langchain/core/errors";
 import { v4 } from "uuid";
+import { ContextOverflowError } from "@langchain/core/errors";
 //#region src/graphs/Graph.ts
 const { AGENT, TOOLS, SUMMARIZE } = GraphNodeKeys;
 const EMPTY_PREEMPT_BOUNDARY = {
@@ -271,6 +271,10 @@ function getDispatchableFinalReasoningContent({ agentContext, responseReasoningC
 	return responseReasoningContent;
 }
 var Graph = class {
+	/** Returns a snapshot of deferred tools discovered by this graph. */
+	getDiscoveredTools(_agentId) {
+		return [];
+	}
 	messageStepHasTextDeltas = /* @__PURE__ */ new Set();
 	messageStepHasToolCalls = /* @__PURE__ */ new Map();
 	messageIdsByStepKey = /* @__PURE__ */ new Map();
@@ -354,6 +358,15 @@ var Graph = class {
 	eagerEventToolUsageCountsByAgentId = /* @__PURE__ */ new Map();
 	eagerEventToolCallChunks = /* @__PURE__ */ new Map();
 	/**
+	* Per-run eager prestart circuit breaker, shared by reference with every
+	* ToolNode this graph compiles. When a prestarted execution's args turn
+	* out to differ from the final request, ToolNode records the tool name
+	* here and the stream handler stops prestarting that tool for the rest of
+	* the run — the retry then executes normally instead of re-diverging in a
+	* loop (LibreChat#14371).
+	*/
+	eagerEventToolSuppressions = /* @__PURE__ */ new Set();
+	/**
 	* Run-scoped execution backend for built-in code tools. Defaults to the
 	* remote Code API sandbox when unset.
 	*/
@@ -401,6 +414,7 @@ var Graph = class {
 		this.eagerEventToolExecutions.clear();
 		this.clearEagerEventToolUsageCounts();
 		this.eagerEventToolCallChunks.clear();
+		this.eagerEventToolSuppressions.clear();
 		this.toolExecution = void 0;
 		this.handlerDispatchedEventCounts.clear();
 		/**
@@ -519,12 +533,24 @@ var StandardGraph = class StandardGraph extends Graph {
 	messages = [];
 	/** Cached run messages preserved before clearHeavyState() so getRunMessages() works after cleanup. */
 	cachedRunMessages;
+	/** Per-agent discovery snapshots preserved before contexts are reset on cleanup. */
+	cachedDiscoveredTools;
 	/** Ids of AI turns the agent node returned THIS run; see isRunProducedMessage. */
 	runProducedAiMessageIds = /* @__PURE__ */ new Set();
 	/** Checkpoint scope whose messages match index-keyed tool snapshots. */
 	originalToolContentCheckpointScope;
 	runId;
 	/**
+	* Identity used to stamp Langfuse runtime scopes and handlers (see
+	* `LangfuseRuntimeContext.runId`). Carries an opaque per-instance
+	* component: public run ids are unrestricted and may repeat across
+	* concurrently executing runs (retries, duplicate submissions,
+	* tenant-local message ids), and equal stamps would let those runs adopt
+	* each other's scopes. One graph instance = one execution's stamp, shared
+	* by the stream handler and every graph-level scope of that execution.
+	*/
+	langfuseScopeRunId;
+	/**
 	* Boundary between historical messages (loaded from conversation state)
 	* and messages produced during the current run.  Set once in the state
 	* reducer when messages first arrive.  Used by `getRunMessages()` and
@@ -596,6 +622,7 @@ var StandardGraph = class StandardGraph extends Graph {
 	constructor({ runId, signal, agents, langfuse, tokenCounter, indexTokenCountMap, calibrationRatio, subagentUsageSink, subagentScope, preemption }) {
 		super();
 		this.runId = runId;
+		this.langfuseScopeRunId = `${runId ?? "graph"}:${nanoid()}`;
 		this.signal = signal;
 		this.langfuse = langfuse;
 		this.subagentUsageSink = subagentUsageSink;
@@ -612,6 +639,7 @@ var StandardGraph = class StandardGraph extends Graph {
 	resetValues(keepContent, checkpointScope) {
 		this.messages = [];
 		this.cachedRunMessages = void 0;
+		this.cachedDiscoveredTools = void 0;
 		this.config = resetIfNotEmpty(this.config, void 0);
 		if (keepContent !== true) {
 			this.contentData = resetIfNotEmpty(this.contentData, []);
@@ -629,6 +657,7 @@ var StandardGraph = class StandardGraph extends Graph {
 		this.eagerEventToolExecutions.clear();
 		this.clearEagerEventToolUsageCounts();
 		this.eagerEventToolCallChunks.clear();
+		this.eagerEventToolSuppressions.clear();
 		this.handlerDispatchedStepIds = resetIfNotEmpty(this.handlerDispatchedStepIds, /* @__PURE__ */ new Set());
 		this.handlerDispatchedEventCounts = resetIfNotEmpty(this.handlerDispatchedEventCounts, /* @__PURE__ */ new Map());
 		this.messageIdsByStepKey = resetIfNotEmpty(this.messageIdsByStepKey, /* @__PURE__ */ new Map());
@@ -646,6 +675,7 @@ var StandardGraph = class StandardGraph extends Graph {
 	}
 	clearHeavyState() {
 		this.cachedRunMessages = this.messages.slice(this.startIndex);
+		this.cachedDiscoveredTools = new Map(Array.from(this.agentContexts, ([agentId, context]) => [agentId, context.getDiscoveredTools()]));
 		super.clearHeavyState();
 		this.messages = [];
 		this.overrideModel = void 0;
@@ -824,6 +854,17 @@ var StandardGraph = class StandardGraph extends Graph {
 		if (this.messages.length === 0 && this.cachedRunMessages != null) return this.cachedRunMessages;
 		return this.messages.slice(this.startIndex);
 	}
+	getDiscoveredTools(agentId) {
+		if (agentId != null) {
+			const current = this.agentContexts.get(agentId)?.getDiscoveredTools() ?? [];
+			if (current.length > 0 || this.cachedDiscoveredTools == null) return current;
+			return [...this.cachedDiscoveredTools.get(agentId) ?? []];
+		}
+		const discoveredTools = /* @__PURE__ */ new Set();
+		for (const context of this.agentContexts.values()) for (const toolName of context.getDiscoveredTools()) discoveredTools.add(toolName);
+		if (discoveredTools.size === 0 && this.cachedDiscoveredTools != null) for (const snapshot of this.cachedDiscoveredTools.values()) for (const toolName of snapshot) discoveredTools.add(toolName);
+		return Array.from(discoveredTools);
+	}
 	/**
 	* True when THIS RUN produced `message` — the provenance the handoff cue
 	* gate needs. Tracked as an id set rather than inferred from `startIndex`
@@ -930,6 +971,7 @@ var StandardGraph = class StandardGraph extends Graph {
 				codeSessionToolNames: this.codeSessionToolNames,
 				eagerEventToolExecutions: this.eagerEventToolExecutions,
 				eagerEventToolUsageCount: this.getEagerEventToolUsageCount(agentContext?.agentId),
+				eagerEventToolSuppressions: this.eagerEventToolSuppressions,
 				toolExecution: this.toolExecution,
 				directToolNames: directToolNames.size > 0 ? directToolNames : void 0,
 				interruptingToolNames: this.interruptingToolNames != null && this.interruptingToolNames.length > 0 ? new Set(this.interruptingToolNames) : void 0,
@@ -1630,7 +1672,12 @@ var StandardGraph = class StandardGraph extends Graph {
 				...config,
 				metadata: {
 					...config.metadata ?? {},
-					...traceMetadata
+					...traceMetadata,
+					/** Canonical agent identity, stamped OUTSIDE trace-metadata
+					*  filtering: `createLangfuseTraceMetadata` drops values over its
+					*  length cap, but scope trust (`isForeignScope`) needs the id
+					*  verbatim regardless of length. */
+					agentId
 				}
 			};
 			initializeLangfuseTracing(langfuse);
@@ -1641,7 +1688,9 @@ var StandardGraph = class StandardGraph extends Graph {
 					sessionId: config.configurable?.thread_id,
 					traceMetadata,
 					tags: ["librechat", "agent"],
-					traceIdSeed: langfuse?.deterministicTraceId === true ? this.runId : void 0
+					traceIdSeed: langfuse?.deterministicTraceId === true ? this.runId : void 0,
+					runId: this.langfuseScopeRunId,
+					toolOutputTracing: hasToolOutputTracingConfig(this.langfuse, agentContext.langfuse) ? resolveToolOutputTracingConfig(this.langfuse, agentContext.langfuse) : void 0
 				});
 				if (langfuseHandler != null) invokeConfig = {
 					...invokeConfig,
@@ -1653,7 +1702,9 @@ var StandardGraph = class StandardGraph extends Graph {
 				if (preInvokeContextOverflowError != null) throw preInvokeContextOverflowError;
 				result = await withLangfuseRuntimeScope(resolveLangfuseRuntimeScope({
 					runLangfuse: this.langfuse,
-					langfuseOverlay: agentContext.langfuse
+					langfuseOverlay: agentContext.langfuse,
+					runId: this.langfuseScopeRunId,
+					agentId
 				}), () => attemptInvoke({
 					model: this.overrideModel ?? model,
 					messages: finalMessages,
@@ -1746,7 +1797,9 @@ var StandardGraph = class StandardGraph extends Graph {
 				try {
 					result = await withLangfuseRuntimeScope(resolveLangfuseRuntimeScope({
 						runLangfuse: this.langfuse,
-						langfuseOverlay: agentContext.langfuse
+						langfuseOverlay: agentContext.langfuse,
+						runId: this.langfuseScopeRunId,
+						agentId
 					}), () => tryFallbackProviders({
 						fallbacks,
 						tools: agentContext.tools,