@librechat/agents 3.1.75 → 3.1.77-dev.1

package/dist/cjs/run.cjs

@@ -5,6 +5,8 @@ var langchain = require('@langfuse/langchain');
 var prompts = require('@langchain/core/prompts');
 var runnables = require('@langchain/core/runnables');
 var openai = require('@langchain/openai');
+var langgraph = require('@langchain/langgraph');
+var messages = require('@langchain/core/messages');
 var base = require('@langchain/core/callbacks/base');
 var title = require('./utils/title.cjs');
 var tokens = require('./utils/tokens.cjs');
@@ -35,6 +37,7 @@ class Run {
     tokenCounter;
     handlerRegistry;
     hookRegistry;
+    humanInTheLoop;
     toolOutputReferences;
     indexTokenCountMap;
     calibrationRatio = 1;
@@ -43,6 +46,15 @@ class Run {
     returnContent = false;
     skipCleanup = false;
     _streamResult;
+    /**
+     * Captured interrupt payload typed as `unknown` because the SDK
+     * does not validate the runtime shape — custom graph nodes can
+     * raise interrupts with arbitrary payloads (not just the SDK's
+     * `HumanInterruptPayload` union). The public `getInterrupt<T>()`
+     * lets callers assert the type they expect.
+     */
+    _interrupt;
+    _haltedReason;
     constructor(config) {
         const runId = config.runId ?? '';
         if (!runId) {
@@ -62,6 +74,7 @@ class Run {
         }
         this.handlerRegistry = handlerRegistry;
         this.hookRegistry = config.hooks;
+        this.humanInTheLoop = config.humanInTheLoop;
         this.toolOutputReferences = config.toolOutputReferences;
         if (!config.graphConfig) {
             throw new Error('Graph config not provided');
@@ -123,8 +136,9 @@ class Run {
             calibrationRatio: this.calibrationRatio,
         });
         /** Propagate compile options from graph config */
-        standardGraph.compileOptions = config.compileOptions;
+        standardGraph.compileOptions = this.applyHITLCheckpointerFallback(config.compileOptions);
         standardGraph.hookRegistry = this.hookRegistry;
+        standardGraph.humanInTheLoop = this.humanInTheLoop;
         standardGraph.toolOutputReferences = this.toolOutputReferences;
         this.Graph = standardGraph;
         return standardGraph.createWorkflow();
@@ -139,14 +153,180 @@ class Run {
             indexTokenCountMap: this.indexTokenCountMap,
             calibrationRatio: this.calibrationRatio,
         });
-        if (compileOptions != null) {
-            multiAgentGraph.compileOptions = compileOptions;
-        }
+        multiAgentGraph.compileOptions =
+            this.applyHITLCheckpointerFallback(compileOptions);
         multiAgentGraph.hookRegistry = this.hookRegistry;
+        multiAgentGraph.humanInTheLoop = this.humanInTheLoop;
         multiAgentGraph.toolOutputReferences = this.toolOutputReferences;
         this.Graph = multiAgentGraph;
         return multiAgentGraph.createWorkflow();
     }
+    /**
+     * When the host opted into HITL via `humanInTheLoop: { enabled: true }`
+     * and did not supply a checkpointer, install an in-memory `MemorySaver`
+     * so `interrupt()` can persist checkpoints and `Command({ resume })`
+     * can rebuild state. The fallback is intentionally process-local:
+     * production hosts that need durable resumption across processes /
+     * restarts must provide their own checkpointer (Redis, Postgres, etc.)
+     * on `compileOptions.checkpointer`.
+     *
+     * No-op when HITL is off (the default — omitted, or
+     * `{ enabled: false }`) or the host already supplied a checkpointer
+     * of their own. See `HumanInTheLoopConfig` JSDoc for the rationale
+     * behind the default-off stance.
+     */
+    applyHITLCheckpointerFallback(compileOptions) {
+        if (this.humanInTheLoop?.enabled !== true) {
+            return compileOptions;
+        }
+        if (compileOptions?.checkpointer != null) {
+            return compileOptions;
+        }
+        return {
+            ...(compileOptions ?? {}),
+            checkpointer: new langgraph.MemorySaver(),
+        };
+    }
+    /**
+     * Run RunStart + UserPromptSubmit hooks before the graph stream
+     * begins, accumulate any `additionalContext` strings into the input
+     * messages, and short-circuit when a hook signals the run should not
+     * proceed (deny / ask decision on the prompt, or `preventContinuation`
+     * on either hook).
+     *
+     * Returns `true` when the caller should bail with `undefined` (run
+     * was halted before any model call); returns `false` to proceed
+     * into the stream loop.
+     *
+     * ## Side effects
+     *
+     * On the success path:
+     *   - Mutates `stateInputs.messages` in place to append a
+     *     consolidated `HumanMessage` carrying any hook
+     *     `additionalContext` strings. Safe because the host owns the
+     *     array and `processStream` is the only consumer until LangGraph
+     *     reads it.
+     *
+     * On the halt path (returning `true`):
+     *   - Sets `this._haltedReason` so callers (and the eventual host)
+     *     can distinguish a hook-driven halt from a natural completion.
+     *   - Calls `registry.clearSession(this.id)` and
+     *     `registry.clearHaltSignal(this.id)` because no resume is
+     *     expected from a pre-stream halt — the run never entered the
+     *     graph, so the session/halt state for this run would otherwise
+     *     leak to the next `processStream` invocation on the same
+     *     registry. Other concurrent runs on the same registry are
+     *     untouched (halt signals are scoped per session id).
+     *   - Sets `config.callbacks = undefined` to drop the callback
+     *     references the caller built (langfuse handler, custom event
+     *     handler, etc.) since they won't be exercised. Mirrors the
+     *     equivalent cleanup the `processStream` `finally` block does
+     *     on the natural-completion path.
+     */
+    async runPreStreamHooks(stateInputs, threadId, config) {
+        const registry = this.hookRegistry;
+        /**
+         * Defensive guard: `processStream` already validated `this.Graph`
+         * before calling this helper, but TypeScript can't propagate that
+         * narrowing across method boundaries. The check keeps the body
+         * free of `this.Graph!` non-null assertions.
+         */
+        if (registry == null || this.Graph == null) {
+            return false;
+        }
+        const preStreamContexts = [];
+        const runStartResult = await executeHooks.executeHooks({
+            registry,
+            input: {
+                hook_event_name: 'RunStart',
+                runId: this.id,
+                threadId,
+                agentId: this.Graph.defaultAgentId,
+                messages: stateInputs.messages,
+            },
+            sessionId: this.id,
+        });
+        for (const ctx of runStartResult.additionalContexts) {
+            preStreamContexts.push(ctx);
+        }
+        /**
+         * Honor `preventContinuation` from RunStart before the stream
+         * starts. Mid-flight halts (from tool/compact/subagent hooks)
+         * route through `HookRegistry.haltRun` and are polled by the
+         * stream loop in `processStream` — different mechanism, same
+         * intent.
+         */
+        if (runStartResult.preventContinuation === true) {
+            this._haltedReason = runStartResult.stopReason ?? 'preventContinuation';
+            registry.clearSession(this.id);
+            registry.clearHaltSignal(this.id);
+            config.callbacks = undefined;
+            return true;
+        }
+        const lastHuman = findLastMessageOfType(stateInputs.messages, 'human');
+        if (lastHuman != null) {
+            const promptResult = await executeHooks.executeHooks({
+                registry,
+                input: {
+                    hook_event_name: 'UserPromptSubmit',
+                    runId: this.id,
+                    threadId,
+                    agentId: this.Graph.defaultAgentId,
+                    prompt: extractPromptText(lastHuman),
+                    // attachments: not yet wired — Phase 2 will extract
+                    // non-text content blocks (images, files) from messages
+                },
+                sessionId: this.id,
+            });
+            if (promptResult.decision === 'deny' ||
+                promptResult.decision === 'ask' ||
+                promptResult.preventContinuation === true) {
+                /**
+                 * Always set `_haltedReason` so the host can call
+                 * `getHaltReason()` and distinguish a hook-blocked prompt
+                 * from a natural empty-output completion. Three signals can
+                 * land here, each with its own canonical reason string when
+                 * the hook didn't supply one.
+                 */
+                if (promptResult.preventContinuation === true) {
+                    this._haltedReason = promptResult.stopReason ?? 'preventContinuation';
+                }
+                else if (promptResult.decision === 'deny') {
+                    this._haltedReason = promptResult.reason ?? 'prompt_denied';
+                }
+                else {
+                    this._haltedReason =
+                        promptResult.reason ?? 'prompt_requires_approval';
+                }
+                registry.clearSession(this.id);
+                registry.clearHaltSignal(this.id);
+                config.callbacks = undefined;
+                return true;
+            }
+            for (const ctx of promptResult.additionalContexts) {
+                preStreamContexts.push(ctx);
+            }
+        }
+        if (preStreamContexts.length > 0) {
+            /**
+             * Wraps the joined hook contexts as a `HumanMessage` even though
+             * the intent is system-level guidance. Using a `SystemMessage`
+             * mid-conversation is rejected by Anthropic and Google providers
+             * (system messages must be the leading entry), so the LangChain
+             * convention — also used by `ToolNode.convertInjectedMessages`
+             * — is `HumanMessage` carrying `additional_kwargs.role` as a
+             * marker for hosts inspecting state. The model still sees a
+             * user-role message; the `role: 'system'` field is metadata
+             * only. Hosts that want a true system message should compose
+             * it into the agent's `instructions` config instead.
+             */
+            stateInputs.messages.push(new messages.HumanMessage({
+                content: preStreamContexts.join('\n\n'),
+                additional_kwargs: { role: 'system', source: 'hook' },
+            }));
+        }
+        return false;
+    }
     static async create(config) {
         /** Create tokenCounter if indexTokenCountMap is provided but tokenCounter is not */
         if (config.indexTokenCountMap && !config.tokenCounter) {
@@ -205,12 +385,37 @@ class Run {
         if (!this.Graph) {
             throw new Error('Graph not initialized. Make sure to use Run.create() to instantiate the Run.');
         }
+        /**
+         * `Command` inputs (currently only `Command({ resume })`) are
+         * resume-mode invocations: LangGraph rebuilds graph state from the
+         * checkpointer, so we skip RunStart / UserPromptSubmit hooks (no
+         * new prompt to evaluate) and read run-state from the Graph wrapper
+         * instead of `inputs.messages`.
+         */
+        const isResume = inputs instanceof langgraph.Command;
+        const stateInputs = isResume ? undefined : inputs;
         const config = {
             recursionLimit: 50,
             ...callerConfig,
             configurable: { ...callerConfig.configurable },
         };
-        this.Graph.resetValues(streamOptions?.keepContent);
+        /**
+         * Skip `resetValues` on resume — we're continuing an in-flight
+         * run, not starting a fresh one. Resetting would wipe the
+         * sidecars (`toolCallStepIds`, `stepKeyIds`, accumulated
+         * `messages`, etc.) the resumed `ToolNode` needs to dispatch
+         * tool completions with the correct step ids and re-resolve
+         * `{{tool<i>turn<n>}}` references. Pairs with the
+         * `awaitingResume` gate on `clearHeavyState` in the `finally`
+         * block so the sidecars survive both ends of the interrupt
+         * boundary.
+         */
+        if (!isResume) {
+            this.Graph.resetValues(streamOptions?.keepContent);
+        }
+        this._interrupt = undefined;
+        this._haltedReason = undefined;
+        this.hookRegistry?.clearHaltSignal(this.id);
         /** Custom event callback to intercept and handle custom events */
         const customEventCallback = this.createCustomEventCallback();
         const baseCallbacks = config.callbacks ?? [];
@@ -250,41 +455,18 @@ class Run {
             run_id: this.id,
         });
         const threadId = config.configurable.thread_id;
-        if (this.hookRegistry != null) {
-            await executeHooks.executeHooks({
-                registry: this.hookRegistry,
-                input: {
-                    hook_event_name: 'RunStart',
-                    runId: this.id,
-                    threadId,
-                    agentId: this.Graph.defaultAgentId,
-                    messages: inputs.messages,
-                },
-                sessionId: this.id,
-            });
-            const lastHuman = findLastMessageOfType(inputs.messages, 'human');
-            if (lastHuman != null) {
-                const promptResult = await executeHooks.executeHooks({
-                    registry: this.hookRegistry,
-                    input: {
-                        hook_event_name: 'UserPromptSubmit',
-                        runId: this.id,
-                        threadId,
-                        agentId: this.Graph.defaultAgentId,
-                        prompt: extractPromptText(lastHuman),
-                        // attachments: not yet wired — Phase 2 will extract
-                        // non-text content blocks (images, files) from messages
-                    },
-                    sessionId: this.id,
-                });
-                if (promptResult.decision === 'deny' ||
-                    promptResult.decision === 'ask') {
-                    this.hookRegistry.clearSession(this.id);
-                    config.callbacks = undefined;
-                    return undefined;
-                }
+        if (this.hookRegistry != null && stateInputs != null) {
+            const shouldHalt = await this.runPreStreamHooks(stateInputs, threadId, config);
+            if (shouldHalt) {
+                return undefined;
             }
         }
+        /**
+         * `streamEvents` accepts both state inputs and `Command` (resume) at
+         * runtime, but our `CompiledStateWorkflow` type narrows the first
+         * arg to `BaseGraphState`. Cast on the call so the resume path
+         * type-checks without widening the wrapper for every caller.
+         */
         const stream = this.graphRunnable.streamEvents(inputs, config, {
             raiseError: true,
             /**
@@ -297,6 +479,16 @@ class Run {
              */
             ignoreCustomEvent: true,
         });
+        /**
+         * Tracks whether the stream loop threw. Used by the `finally`
+         * block to decide whether to honor the interrupt-preservation
+         * guard for session hooks: a captured `_interrupt` is only
+         * meaningful if the stream completed cleanly. If the loop errored
+         * after stashing an interrupt (e.g. a downstream handler throws
+         * after the interrupt event landed), the interrupt is stale —
+         * preserving session hooks would leak them into the next run.
+         */
+        let streamThrew = false;
         try {
             for await (const event of stream) {
                 const { data, metadata, ...info } = event;
@@ -305,12 +497,74 @@ class Run {
                 if (eventName === _enum.GraphEvents.ON_CUSTOM_EVENT) {
                     continue;
                 }
+                /**
+                 * Detect interrupts surfaced by LangGraph as a synthetic
+                 * `__interrupt__` field on the streamed chunk and stash the
+                 * first one for the host to read via `run.getInterrupt()`
+                 * once the stream drains. Captured as `unknown` because the
+                 * SDK does not validate the runtime payload shape — the
+                 * built-in ToolNode raises a `HumanInterruptPayload`
+                 * (`tool_approval` / `ask_user_question`), but custom nodes
+                 * can pass any payload to `interrupt()`. Callers narrow with
+                 * the `isToolApprovalInterrupt` / `isAskUserQuestionInterrupt`
+                 * guards or assert via `getInterrupt<T>()`.
+                 */
+                if (this._interrupt == null &&
+                    data.chunk != null &&
+                    langgraph.isInterrupted(data.chunk)) {
+                    const interrupts = data.chunk[langgraph.INTERRUPT];
+                    if (interrupts.length > 0) {
+                        const first = interrupts[0];
+                        /**
+                         * Capture the interrupt unconditionally — `interrupt(null)`
+                         * and `interrupt(undefined)` are valid pauses (a custom
+                         * node may want to pause without metadata) and the host
+                         * still needs to know the run is awaiting resume. Gating
+                         * on `payload != null` would silently downgrade a paused
+                         * run to "completed" and let the `Stop` hook fire,
+                         * breaking host resume handling.
+                         */
+                        this._interrupt = {
+                            interruptId: first.id ?? '',
+                            threadId,
+                            payload: first.value,
+                        };
+                    }
+                }
                 const handler = this.handlerRegistry?.getHandler(eventName);
                 if (handler) {
                     await handler.handle(eventName, data, metadata, this.Graph);
                 }
+                /**
+                 * Mid-flight halt: any hook (PreToolUse, PostToolUse,
+                 * PostToolBatch, SubagentStart/Stop, PreCompact, PostCompact)
+                 * that returned `preventContinuation: true` raises a halt
+                 * signal on the registry via `executeHooks`. We poll between
+                 * stream events and break out as soon as one is set so the
+                 * graph doesn't take another model turn after the halting
+                 * operation completes.
+                 *
+                 * Limitation: the current step (in-flight model call, ongoing
+                 * tool batch) is not aborted — only the next step is skipped.
+                 * This matches Claude Code's `continue: false` semantic where
+                 * the active operation finishes before halting takes effect.
+                 */
+                const haltSignal = this.hookRegistry?.getHaltSignal(this.id);
+                if (haltSignal != null) {
+                    this._haltedReason = haltSignal.reason;
+                    break;
+                }
             }
-            if (this.hookRegistry?.hasHookFor('Stop', this.id) === true) {
+            /**
+             * Skip the Stop hook when the run paused on a HITL interrupt
+             * (still pending human input) or was halted by a hook (the host
+             * already chose to stop, so a Stop hook firing now would be
+             * misleading). The host fires Stop on the resumed-and-completed
+             * run instead.
+             */
+            if (this._interrupt == null &&
+                this._haltedReason == null &&
+                this.hookRegistry?.hasHookFor('Stop', this.id) === true) {
                 await executeHooks.executeHooks({
                     registry: this.hookRegistry,
                     input: {
@@ -318,7 +572,7 @@ class Run {
                         runId: this.id,
                         threadId,
                         agentId: this.Graph.defaultAgentId,
-                        messages: this.Graph.getRunMessages() ?? inputs.messages,
+                        messages: this.Graph.getRunMessages() ?? stateInputs?.messages ?? [],
                         stopHookActive: false, // will be true when stop is triggered by a hook (Phase 2)
                     },
                     sessionId: this.id,
@@ -328,6 +582,7 @@ class Run {
             }
         }
         catch (err) {
+            streamThrew = true;
             if (this.hookRegistry?.hasHookFor('StopFailure', this.id) === true) {
                 const runMessages = this.Graph.getRunMessages() ?? [];
                 await executeHooks.executeHooks({
@@ -348,7 +603,34 @@ class Run {
             throw err;
         }
         finally {
-            this.hookRegistry?.clearSession(this.id);
+            /**
+             * Preserve session-scoped hooks when the run paused on a HITL
+             * interrupt — the very next call will be `Run.resume()`, which
+             * needs the same policy hooks (e.g., the `PreToolUse` matcher
+             * that triggered the interrupt) to fire on the re-executed node
+             * and uphold the approval flow. Clearing here would leak the
+             * approval gate on resume. The session is cleared instead at
+             * natural completion, error (including errors that happen AFTER
+             * an interrupt was captured — those interrupts are stale), or
+             * hook-driven halt (including hooks that returned BOTH `ask`
+             * and `preventContinuation` — the halt wins, no resume is
+             * expected, sessions must drop). Every state where no resume
+             * is expected clears.
+             */
+            if (this._interrupt == null ||
+                this._haltedReason != null ||
+                streamThrew) {
+                this.hookRegistry?.clearSession(this.id);
+            }
+            /**
+             * Drop any halt signal raised mid-stream for this run so a
+             * subsequent `processStream` / `resume` starts with clean state.
+             * The Run captured `_haltedReason` already; the registry entry
+             * for this `sessionId` would otherwise spuriously trip the next
+             * loop. Other concurrent runs sharing this registry are
+             * unaffected — their entries live under their own session ids.
+             */
+            this.hookRegistry?.clearHaltSignal(this.id);
             /**
              * Break the reference chain that keeps heavy data alive via
              * LangGraph's internal `__pregel_scratchpad.currentTaskInput` →
@@ -377,13 +659,89 @@ class Run {
                 ? this.Graph.getContentParts()
                 : undefined;
             this.calibrationRatio = this.Graph.getCalibrationRatio();
-            if (!this.skipCleanup) {
+            /**
+             * Skip `clearHeavyState()` when the run paused on a clean HITL
+             * interrupt awaiting resume — `Run.resume()` re-enters the same
+             * `ToolNode` instance and needs the sidecars `clearHeavyState`
+             * would wipe (`toolCallStepIds` for completion-event step ids,
+             * the `_toolOutputRegistry` for `{{tool<i>turn<n>}}`
+             * substitutions, `sessions` for code-env continuity, plus the
+             * `hookRegistry` and `humanInTheLoop` config the interrupt
+             * branch itself relies on). Without preservation, the resumed
+             * tool completion would dispatch `ON_RUN_STEP_COMPLETED` with
+             * an empty step id and downstream stream consumers would drop
+             * the result.
+             *
+             * The natural-completion / error / hook-driven-halt paths still
+             * clean up — `_haltedReason != null` or `streamThrew` mean no
+             * resume is expected. Cross-process resume (host rebuilds the
+             * Run from scratch) is a separate concern; see
+             * `HumanInTheLoopConfig` JSDoc.
+             */
+            const awaitingResume = this._interrupt != null && this._haltedReason == null && !streamThrew;
+            if (!this.skipCleanup && !awaitingResume) {
                 this.Graph.clearHeavyState();
             }
             this._streamResult = result;
         }
         return this._streamResult;
     }
+    /**
+     * Returns the pending interrupt captured during the most recent
+     * `processStream` (or `resume`) invocation. `undefined` when the run
+     * either has not been streamed yet or completed without pausing.
+     *
+     * Hosts call this immediately after `processStream` returns to decide
+     * whether the run is awaiting human input. Persist the returned
+     * descriptor (alongside `thread_id` and the agent run config) so a
+     * later `resume(decisions)` can rebuild the run.
+     *
+     * The default `TPayload` is the SDK's `HumanInterruptPayload` union
+     * (`tool_approval` / `ask_user_question`), suitable for the common
+     * case where interrupts come from the built-in ToolNode or
+     * `askUserQuestion()` helper. Hosts that raise custom interrupts
+     * from custom graph nodes pass their own type — the SDK does not
+     * validate the runtime shape, it just transports whatever the
+     * `interrupt()` call carried. When in doubt, narrow with the
+     * `isToolApprovalInterrupt` / `isAskUserQuestionInterrupt` type
+     * guards (which accept `unknown`) before reading variant-specific
+     * fields.
+     */
+    getInterrupt() {
+        return this._interrupt;
+    }
+    /**
+     * Returns the reason a hook halted the run via
+     * `preventContinuation: true`, or `undefined` if no hook halted.
+     *
+     * Hosts inspect this after `processStream` returns to distinguish a
+     * natural completion (`undefined`) from a hook-driven halt (a
+     * truthy string). Independent from `getInterrupt()` — a halted run
+     * has no interrupt; an interrupted run has no halt reason.
+     */
+    getHaltReason() {
+        return this._haltedReason;
+    }
+    /**
+     * Resume a paused HITL run with the value the user (or whatever
+     * decided the interrupt) supplied. The default `TResume` covers the
+     * `tool_approval` interrupt (the common case): an array of decisions
+     * in `action_requests` order, or a record keyed by `tool_call_id`.
+     *
+     * For other interrupt types (e.g., `ask_user_question` →
+     * `AskUserQuestionResolution`, or any custom interrupt a host raises
+     * from a custom node), pass the type parameter and the SDK forwards
+     * the value through unchanged. LangGraph delivers it as the return
+     * value of the original `interrupt()` call inside the paused node.
+     *
+     * The host MUST construct this Run with the same `thread_id` and the
+     * same checkpointer as the original paused run; LangGraph rebuilds
+     * graph state from the checkpoint and re-enters the interrupted node
+     * from the start.
+     */
+    async resume(resumeValue, callerConfig, streamOptions) {
+        return this.processStream(new langgraph.Command({ resume: resumeValue }), callerConfig, streamOptions);
+    }
     createSystemCallback(clientCallbacks, key) {
         return ((...args) => {
             const clientCallback = clientCallbacks[key];