[pull] main from danny-avila:main

docs/subagent-context.md

@@ -0,0 +1,54 @@
+# Host-owned subagent context
+
+`Run.create({ subagentContext })` lets a host authorize an individual child
+execution, supply its initial messages, partition its tool sessions, and project
+its successful result. The SDK exports `SUBAGENT_CONTEXT_VERSION = 1` so hosts
+can reject an unsupported installation before enabling features that depend on
+execution identity.
+
+`prepare(input)` runs before the child graph is constructed or invoked. Its input
+contains the SDK-owned `executionContext`, `parentThreadId`, `memberAgentIds`, an
+abort `signal`, and a `resumed` marker. The last ancestry entry's `subagentRunId`
+identifies this execution; saved agent IDs alone do not distinguish concurrent
+copies of an agent. Nested children inherit the adapter and receive their full
+ancestry. A graph subagent has one execution identity and multiple member IDs.
+
+Preparation may return:
+
+- `messages`: actual LangChain messages appended after the task description for
+  a fresh child. They may contain multimodal file content. Existing checkpoint
+  messages are preserved on resume without inserting these messages again.
+- `configurable`: host runtime context. SDK run, thread, checkpoint, and execution
+  identities cannot be replaced through this object.
+- `agentSessions`: entries keyed by child member ID, each replacing that member's
+  `codeSessionKey` and `initialSessions`. Omitting `initialSessions` in an entry
+  clears inherited session seeds. Unknown members are rejected. A paused live
+  graph cannot change its session partition when it resumes.
+
+The SDK calls preparation again when a completed execution is retried and when a
+host rebuilds a paused execution, allowing the host to reauthorize access. Make
+authorization idempotent for the execution identity. An ordinary preparation
+failure prevents child execution and produces a generic failure result. Aborting
+the supplied signal instead propagates as an execution error, so callers must
+handle it as cancellation rather than as a normal subagent result.
+
+`complete(input, result)` runs after successful child work. It returns the text
+delivered to the parent and can append durable file references. The SDK retains
+the original completed result if delivery fails, so retrying delivery does not
+repeat model or tool side effects. Completion must be idempotent. Failed and
+cancelled child work does not invoke completion.
+
+The same canonical `executionContext` reaches child tool hooks,
+`ToolExecuteBatchRequest.executionContext`, and the `configurable.executionContext`
+and `metadata.executionContext` of direct tool calls. Root tools have no child
+context. Event-driven hosts should use the batch field as authority and overwrite
+inherited configurable or metadata values before invoking tools and handling
+their artifacts.
+
+The adapter does not implement file authorization, storage, publication, or
+sandbox isolation. In particular, `codeSessionKey` partitions the SDK's transient
+session map; the host must also assign a private runtime workspace and prevent
+inherited runtime hints or direct tool implementations from bypassing that
+workspace. Published files and any private artifact recovery remain the host's
+responsibility. The host must reconstruct its adapter and authorization when resuming. The SDK
+does not replay private artifacts to the host when restoring a checkpoint.

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@librechat/agents",
-  "version": "3.8.5",
+  "version": "3.8.6",
   "reova": {
     "enabled": true,
     "endpoint": "https://telemetry.reo.dev/data"

src/__tests__/stream.eagerEventExecution.test.ts

@@ -238,6 +238,83 @@ describe('ChatModelStreamHandler eager event tool execution', () => {
     expect(graph.toolCallStepIds.has('call_weather')).toBe(true);
   });
 
+  it('captures normalized code-session identity before eager dispatch', async () => {
+    const graph = createGraph({
+      sessions: new Map([
+        [
+          Constants.EXECUTE_CODE,
+          {
+            session_id: 'exec-session',
+            files: [{ id: 'old-file', name: 'report.csv' }],
+            lastUpdated: 1,
+          },
+        ],
+      ]),
+      getAgentContext: jest.fn(
+        (): Partial<AgentContext> => ({
+          provider: Providers.ANTHROPIC,
+          toolDefinitions: [{ name: Constants.EXECUTE_CODE }],
+          graphTools: [],
+          agentId: 'agent_1',
+          codeSessionKey: Constants.EXECUTE_CODE,
+          getCallerCapabilityProjectionSnapshot: jest.fn(() => ({
+            version: 1 as const,
+            directToolNames: [],
+            codeExecutionToolNames: [Constants.EXECUTE_CODE],
+            directOnlyToolNames: [],
+            codeExecutionOnlyToolNames: [Constants.EXECUTE_CODE],
+          })),
+        })
+      ) as unknown as StandardGraph['getAgentContext'],
+    });
+    jest
+      .spyOn(events, 'safeDispatchCustomEvent')
+      .mockImplementation(async (event, data): Promise<void> => {
+        if (event !== GraphEvents.ON_TOOL_EXECUTE) return;
+        const batch = data as t.ToolExecuteBatchRequest;
+        batch.toolCalls[0].codeSessionContext = {
+          session_id: 'new-session',
+          files: [
+            {
+              id: 'new-file',
+              resource_id: 'new-file',
+              name: 'report.csv',
+              storage_session_id: 'new-storage',
+              kind: 'user',
+            },
+          ],
+        };
+        batch.resolve([
+          {
+            toolCallId: 'call_code',
+            status: 'success',
+            content: '',
+          },
+        ]);
+      });
+
+    await new ChatModelStreamHandler().handle(
+      GraphEvents.CHAT_MODEL_STREAM,
+      {
+        chunk: {
+          content: '',
+          tool_calls: [
+            { id: 'call_code', name: Constants.EXECUTE_CODE, args: {} },
+          ],
+          response_metadata: finalToolCallResponseMetadata,
+        } as unknown as t.StreamChunk,
+      },
+      { langgraph_node: 'agent' },
+      graph
+    );
+
+    const record = graph.eagerEventToolExecutions.get('call_code');
+    expect(record?.codeSessionBaselineByName?.get('report.csv')).toBe(
+      'exec-session\0old-file'
+    );
+    expect(record?.request.codeSessionContext?.files?.[0].id).toBe('new-file');
+  });
+
   it('prestarts when subagent callback forwarding can execute tools without a handler registry', async () => {
     const graph = createGraph({
       handlerRegistry: undefined,

src/common/constants.ts

@@ -1,3 +1,6 @@
+/** Host context preparation and trusted subagent execution identity are supported. */
+export const SUBAGENT_CONTEXT_VERSION = 1;
+
 /**
  * Anthropic direct API tool schema overhead multiplier.
  * Empirically calibrated against real MCP tool sets (29 tools).

src/graphs/Graph.ts

@@ -133,10 +133,18 @@ import {
   annotateMessagesForLLM,
   ToolOutputReferenceRegistry,
 } from '@/tools/toolOutputReferences';
+import {
+  prepareProviderRequest,
+  usesNativeOpenAIResponses,
+} from '@/llm/prepareProviderRequest';
 import {
   resolveLangfuseRuntimeScope,
   withLangfuseRuntimeScope,
 } from '@/langfuseRuntimeScope';
+import {
+  ManualSummarizationSkippedError,
+  shouldTriggerSummarization,
+} from '@/summarization';
 import {
   getToolContentCharLength,
   serializeToolContentBounded,
@@ -150,36 +158,28 @@ import {
   PreparedSubagents,
   PreparedSubagentError,
 } from '@/tools/preparedSubagents';
+import {
+  createRemoveAllMessage,
+  messagesStateReducer,
+} from '@/messages/reducer';
+import { createToolHistoryPreparation } from '@/messages/toolHistoryProjection';
 import { ToolNode as CustomToolNode, toolsCondition } from '@/tools/ToolNode';
 import { shouldTraceToolNodeForLangfuse } from '@/langfuseToolOutputTracing';
 import { createLocalCodingToolBundle } from '@/tools/local/LocalCodingTools';
 import { SUBAGENT_REPLAY_CONTROLLER } from '@/tools/subagent/SubagentReplay';
 import { applyGraphRuntimeConfig } from '@/graphs/applyGraphRuntimeConfig';
 import { isFadingTier, isInformativeFadingTier } from '@/messages/fading';
 import { createContextPressureMeter } from '@/llm/contextPressureMeter';
-import { createToolHistoryPreparation } from '@/messages/toolHistoryProjection';
 import { safeDispatchCustomEvent, emitAgentLog } from '@/utils/events';
-import {
-  prepareProviderRequest,
-  usesNativeOpenAIResponses,
-} from '@/llm/prepareProviderRequest';
 import { createCloudflareCodingToolBundle } from '@/tools/cloudflare';
 import { calculateMaxToolCallInputChars } from '@/utils/truncation';
 import { prepareToolsForPromptCache } from '@/llm/promptCacheTools';
 import { providerRequiresStrictAlternation } from '@/llm/providers';
 import { buildSubagentToolParams } from '@/tools/SubagentTool';
 import { initializeLangfuseTracing } from '@/instrumentation';
-import {
-  ManualSummarizationSkippedError,
-  shouldTriggerSummarization,
-} from '@/summarization';
 import { isRunStepResumeState } from '@/tools/runStepResume';
 import { resolveLocalToolsForBinding } from '@/tools/local';
 import { createSummarizeNode } from '@/summarization/node';
-import {
-  createRemoveAllMessage,
-  messagesStateReducer,
-} from '@/messages/reducer';
 import { getTruncationStopReason } from '@/llm/truncation';
 import { createSchemaOnlyTools } from '@/tools/schema';
 import { AgentContext } from '@/agents/AgentContext';
@@ -1356,7 +1356,8 @@ export class StandardGraph extends Graph<t.BaseGraphState, t.GraphNode> {
   /** See {@link t.StandardGraphInput.subagentTasks}. */
   subagentTasks: t.SubagentTaskConfig | undefined;
   /** See {@link t.StandardGraphInput.subagentExecutionContext}. */
-  private readonly subagentExecutionContext?: t.SubagentExecutionContext;
+  readonly subagentExecutionContext?: t.SubagentExecutionContext;
+  private readonly subagentContext?: t.SubagentContextAdapter;
   /** See {@link t.StandardGraphInput.preemption}. */
   preemption?: t.StreamPreemption;
   /**
@@ -1534,6 +1535,7 @@ export class StandardGraph extends Graph<t.BaseGraphState, t.GraphNode> {
       subagentTasks,
       subagentScope,
       subagentExecutionContext,
+      subagentContext,
       preemption,
       streamLimits,
       toolExecution,
@@ -1560,6 +1562,7 @@ export class StandardGraph extends Graph<t.BaseGraphState, t.GraphNode> {
     this.subagentTasks = subagentTasks;
     this.subagentScope = subagentScope === true;
     this.subagentExecutionContext = subagentExecutionContext;
+    this.subagentContext = subagentContext;
     this.preemption = preemption;
     this.streamLimits = resolveStreamLimits(streamLimits);
     this.toolExecution = toolExecution;
@@ -2881,6 +2884,7 @@ export class StandardGraph extends Graph<t.BaseGraphState, t.GraphNode> {
         // run_id); `executingAgentId` always identifies the owning agent.
         agentId: this.subagentScope ? agentContext?.agentId : undefined,
         executingAgentId: agentContext?.agentId,
+        executionContext: this.subagentExecutionContext,
         executingAgentName: agentContext?.name,
         rootAgentId: this.defaultAgentId,
         rootAgentName: this.agentContexts.get(this.defaultAgentId)?.name,
@@ -2965,6 +2969,7 @@ export class StandardGraph extends Graph<t.BaseGraphState, t.GraphNode> {
       // hooks can attribute the batch even at the top level.
       agentId: this.subagentScope ? agentContext?.agentId : undefined,
       executingAgentId: agentContext?.agentId,
+      executionContext: this.subagentExecutionContext,
       executingAgentName: agentContext?.name,
       rootAgentId: this.defaultAgentId,
       rootAgentName: this.agentContexts.get(this.defaultAgentId)?.name,
@@ -4627,7 +4632,9 @@ export class StandardGraph extends Graph<t.BaseGraphState, t.GraphNode> {
                       fallbackProvider,
                       fallbackClientOptions
                     );
-                    let projection: ReturnType<typeof measureProviderPayload> | undefined;
+                    let projection:
+                      | ReturnType<typeof measureProviderPayload>
+                      | undefined;
                     const request = prepareProviderRequest({
                       model: fallbackModel,
                       messages: cached,
@@ -4644,7 +4651,9 @@ export class StandardGraph extends Graph<t.BaseGraphState, t.GraphNode> {
                       },
                     });
                     if (projection == null) {
-                      throw new Error('Fallback preparation did not measure its payload');
+                      throw new Error(
+                        'Fallback preparation did not measure its payload'
+                      );
                     }
                     return { request, projection };
                   };
@@ -4664,8 +4673,7 @@ export class StandardGraph extends Graph<t.BaseGraphState, t.GraphNode> {
                   if (!projection.fits) {
                     const compacted = compactSyntheticProviderContext(
                       servingFallbackMessages,
-                      (candidate) =>
-                        prepareFallback(candidate).projection
+                      (candidate) => prepareFallback(candidate).projection
                     );
                     if (compacted !== servingFallbackMessages) {
                       preparedFallbackRequest = prepareFallback(compacted);
@@ -5025,6 +5033,7 @@ export class StandardGraph extends Graph<t.BaseGraphState, t.GraphNode> {
         runId,
         threadId: configurable?.thread_id as string | undefined,
         agentId: this.subagentScope ? agentId : undefined,
+        executionContext: this.subagentExecutionContext,
         executingAgentId: agentId,
         sealCount: this.preemptSealCount,
       },
@@ -5247,6 +5256,7 @@ export class StandardGraph extends Graph<t.BaseGraphState, t.GraphNode> {
           tokenCounter: agentContext.tokenCounter,
           usageSink: this.subagentUsageSink,
           taskConfig: this.subagentTasks,
+          subagentContext: this.subagentContext,
           streamLimits: this.streamLimits,
           humanInTheLoop: this.humanInTheLoop,
           checkpointer: this.compileOptions?.checkpointer,
@@ -5343,6 +5353,9 @@ export class StandardGraph extends Graph<t.BaseGraphState, t.GraphNode> {
               });
             }
             const result = await executor.execute(executeParams);
+            if (result.retryableDelivery === true) {
+              throw new Error(result.content);
+            }
             return result.content;
           },
           buildSubagentToolParams(executableConfigs, {
@@ -5527,6 +5540,7 @@ export class StandardGraph extends Graph<t.BaseGraphState, t.GraphNode> {
             },
             runId: this.runId,
             isMultiAgent: this.isMultiAgentGraph(),
+            getSubagentExecutionContext: () => this.subagentExecutionContext,
             hookRegistry: this.hookRegistry,
             getToolsForBinding: (
               provider: t.ProviderName,

src/hooks/__tests__/compactHooks.test.ts

@@ -24,6 +24,21 @@ const callerConfig = {
   version: 'v2' as const,
 };
 
+const childExecutionContext: t.SubagentExecutionContext = {
+  rootRunId: 'root-run',
+  hookSessionId: 'compact-run',
+  depth: 1,
+  ancestry: [
+    {
+      subagentRunId: 'child-run',
+      subagentType: 'worker',
+      subagentKind: 'agent',
+      subagentAgentId: 'worker',
+      parentRunId: 'root-run',
+    },
+  ],
+};
+
 let getChatModelClassSpy: jest.SpyInstance;
 const originalGetChatModelClass = providers.getChatModelClass;
 
@@ -145,6 +160,29 @@ describe('Compaction hook integration', () => {
     });
   });
 
+  it('carries child execution identity through compact hooks', async () => {
+    const registry = new HookRegistry();
+    let pre: PreCompactHookInput | undefined;
+    let post: PostCompactHookInput | undefined;
+    registry.register('PreCompact', {
+      hooks: [async (input) => ((pre = input), {})],
+    });
+    registry.register('PostCompact', {
+      hooks: [async (input) => ((post = input), {})],
+    });
+    const run = await createCompactingRun(tokenCounter, registry);
+    Object.defineProperty(run.Graph, 'subagentExecutionContext', {
+      value: childExecutionContext,
+    });
+    expect(run.Graph?.subagentExecutionContext).toEqual(childExecutionContext);
+    run.Graph!.overrideTestModel(['Final answer after compaction.']);
+
+    await run.processStream(buildConversation(), callerConfig);
+
+    expect(pre?.executionContext).toEqual(childExecutionContext);
+    expect(post?.executionContext).toEqual(childExecutionContext);
+  });
+
   describe('PostCompact', () => {
     it('fires with summary text after compaction (legacy retainRecent.turns=0 shape)', async () => {
       const registry = new HookRegistry();