andreiborza · Oct 8, 2026 · Oct 7, 2026 · Oct 7, 2026 · Oct 7, 2026 · Oct 7, 2026
@@ -0,0 +1,61 @@
+import { BaseCallbackHandler } from '@langchain/core/callbacks/base';
+import { RunnableLambda } from '@langchain/core/runnables';
+import { TypeSafeClassifier } from '@langchain/typesafe';
+// Sets up the AsyncLocalStorage that passes a parent run's config to its children, as it is inside `createAgent`.
+import '@langchain/langgraph';
+import * as Sentry from '@sentry/node';
+import express from 'express';
+
+function startMockTypeSafeServer() {
+  const app = express();
+  app.use(express.json());
+
+  app.post('/v1/systemone', (req, res) => {
+    res.json({
+      model: 'jev-1.13',
+      answers: { urgent: { type: 'noul', noul: 0.9 } },
+      usage: { input_tokens: 30, output_tokens: 2 },
+    });
+  });
+
+  return new Promise(resolve => {
+    const server = app.listen(0, () => {
+      resolve(server);
+    });
+  });
+}
+
+// Stands in for a tracer such as LangSmith, which the user passes to the parent run.
+class RecordingHandler extends BaseCallbackHandler {
+  name = 'RecordingHandler';
+  runs = [];
+
+  handleChainStart(chain, _inputs, _runId, parentRunId) {
+    this.runs.push(`${chain.id.at(-1)}:${parentRunId ? 'child' : 'root'}`);
+  }
+}
+
+async function run() {
+  const server = await startMockTypeSafeServer();
+  const baseUrl = `http://localhost:${server.address().port}`;
+
+  await Sentry.startSpan({ op: 'function', name: 'main' }, async span => {
+    const classifier = new TypeSafeClassifier({
+      apiKey: 'mock-api-key',
+      baseUrl,
+      questions: { urgent: { type: 'noul', instructions: 'Is this urgent?' } },
+    });
+
+    // Like the `@langchain/typesafe` middlewares, call the classifier without a config.
+    const parent = RunnableLambda.from(input => classifier.invoke(input)).withConfig({ runName: 'parent' });
+    const recorder = new RecordingHandler();
+    await parent.invoke('My payouts have been failing.', { callbacks: [recorder] });
+
+    span.setAttribute('test.recorded_runs', recorder.runs.join(','));
+  });
+
+  await Sentry.flush(2000);
+  server.close();
+}
+
+run();
@@ -0,0 +1,43 @@
+import { HumanMessage } from '@langchain/core/messages';
+import { TypeSafeClassifier } from '@langchain/typesafe';
+import * as Sentry from '@sentry/node';
+import express from 'express';
+
+function startMockTypeSafeServer() {
+  const app = express();
+  app.use(express.json());
+
+  app.post('/v1/systemone', (req, res) => {
+    res.json({
+      model: 'jev-1.13',
+      answers: { urgent: { type: 'noul', noul: 0.9 } },
+      usage: { input_tokens: 30, output_tokens: 2 },
+    });
+  });
+
+  return new Promise(resolve => {
+    const server = app.listen(0, () => {
+      resolve(server);
+    });
+  });
+}
+
+async function run() {
+  const server = await startMockTypeSafeServer();
+  const baseUrl = `http://localhost:${server.address().port}`;
+
+  await Sentry.startSpan({ op: 'function', name: 'main' }, async () => {
+    const classifier = new TypeSafeClassifier({
+      apiKey: 'mock-api-key',
+      baseUrl,
+      questions: { urgent: { type: 'noul', instructions: 'Is this urgent?' } },
+    });
+
+    await classifier.invoke('My payouts have been failing.');
+    await classifier.invoke(new HumanMessage('My card was charged twice.'));
+  });
+
+  server.close();
+}
+
+run();
@@ -3,6 +3,7 @@ import {
   GEN_AI_CONVERSATION_ID,
   GEN_AI_INPUT_MESSAGES,
   GEN_AI_OPERATION_NAME,
+  GEN_AI_OUTPUT_MESSAGES,
   GEN_AI_PROVIDER_NAME,
   GEN_AI_REQUEST_MAX_TOKENS,
   GEN_AI_REQUEST_MODEL,
@@ -18,6 +19,7 @@ import {
   SENTRY_OP,
   SENTRY_ORIGIN,
 } from '@sentry/conventions/attributes';
+import { GEN_AI_EVALUATE } from '@sentry/conventions/op';
 import { GEN_AI_RESPONSE_STOP_REASON_ATTRIBUTE } from '../../../../../../packages/server-utils/src/ai/core/gen-ai-attributes';
 import { cleanupChildProcesses, createEsmAndCjsTests } from '../../../../utils/runner';
 import { createEsmTests } from '../../../../utils/runner/createEsmAndCjsTests';
@@ -233,6 +235,88 @@ describe('LangChain integration (v1)', () => {
     },
   );
 
+  createEsmAndCjsTests(
+    __dirname,
+    'scenario-typesafe-classifier.mjs',
+    'instrument-with-pii.mjs',
+    (createRunner, test) => {
+      test('records a TypeSafeClassifier run as a gen_ai.evaluate span', async () => {
+        const runner = createRunner().ignore('event');
+        const spansPromise = runner.collectStreamedSpansUntilSegment('main');
+
+        await runner.start().completed();
+
+        const spans = (await spansPromise).filter(
+          span => span.attributes[SENTRY_ORIGIN]?.value === 'auto.ai.langchain',
+        );
+        expect(spans.map(span => span.name)).toEqual(['evaluate jev-latest', 'evaluate jev-latest']);
+
+        const [evaluateSpan, messageEvaluateSpan] = spans.sort((a, b) => a.start_timestamp - b.start_timestamp);
+        expect(evaluateSpan.attributes[SENTRY_OP].value).toBe(GEN_AI_EVALUATE);
+        expect(evaluateSpan.attributes[SENTRY_ORIGIN].value).toBe('auto.ai.langchain');
+        expect(evaluateSpan.attributes[GEN_AI_OPERATION_NAME].value).toBe('evaluate');
+        expect(evaluateSpan.attributes[GEN_AI_PROVIDER_NAME].value).toBe('typesafe');
+        expect(evaluateSpan.attributes[GEN_AI_REQUEST_MODEL].value).toBe('jev-latest');
+        expect(evaluateSpan.attributes[GEN_AI_RESPONSE_MODEL].value).toBe('jev-1.13');
+        expect(evaluateSpan.attributes[GEN_AI_USAGE_INPUT_TOKENS].value).toBe(30);
+        expect(evaluateSpan.attributes[GEN_AI_USAGE_OUTPUT_TOKENS].value).toBe(2);
+        expect(evaluateSpan.attributes[GEN_AI_USAGE_TOTAL_TOKENS].value).toBe(32);
+        expect(JSON.parse(evaluateSpan.attributes[GEN_AI_INPUT_MESSAGES].value)).toEqual([
+          {
+            type: 'evaluation',
+            state: 'My payouts have been failing.',
+            questions: { urgent: { type: 'noul', instructions: 'Is this urgent?' } },
+          },
+        ]);
+        // A message is recorded as the transcript line the classifier sends, not as LangChain's serialized form.
+        expect(JSON.parse(messageEvaluateSpan!.attributes[GEN_AI_INPUT_MESSAGES].value)).toEqual([
+          {
+            type: 'evaluation',
+            state: 'user: My card was charged twice.',
+            questions: { urgent: { type: 'noul', instructions: 'Is this urgent?' } },
+          },
+        ]);
+        expect(JSON.parse(evaluateSpan.attributes[GEN_AI_OUTPUT_MESSAGES].value)).toEqual([
+          { type: 'evaluation', answers: { urgent: { type: 'noul', noul: 0.9 } } },
+        ]);
+      });
+    },
+    {
+      additionalDependencies: {
+        langchain: '^1.0.0',
+        '@langchain/core': '^1.0.0',
+        '@langchain/typesafe': '^0.0.2',
+      },
+    },
+  );
+
+  createEsmAndCjsTests(
+    __dirname,
+    'scenario-typesafe-classifier-inherited-callbacks.mjs',
+    'instrument-with-pii.mjs',
+    (createRunner, test) => {
+      test('keeps the parent run callbacks for a TypeSafeClassifier called without config', async () => {
+        const runner = createRunner().ignore('event');
+        const spansPromise = runner.collectStreamedSpansUntilSegment('main');
+
+        await runner.start().completed();
+
+        const spans = await spansPromise;
+        const segment = spans.find(span => span.is_segment && span.name === 'main');
+        expect(segment!.attributes['test.recorded_runs'].value).toBe('RunnableLambda:root,TypeSafeClassifier:child');
+        expect(spans.filter(span => span.attributes[SENTRY_OP]?.value === GEN_AI_EVALUATE)).toHaveLength(1);
+      });
+    },
+    {
+      additionalDependencies: {
+        langchain: '^1.0.0',
+        '@langchain/core': '^1.0.0',
+        '@langchain/typesafe': '^0.0.2',
+        '@langchain/langgraph': '^1.0.0',
+      },
+    },
+  );
+
   createEsmTests(
     __dirname,
     'scenario-openai-before-langchain.mjs',

@@ -15,6 +15,11 @@ import {
 import { GEN_AI_CHAT, GEN_AI_EXECUTE_TOOL, GEN_AI_INVOKE_AGENT } from '@sentry/conventions/op';
 import { resolveAIRecordingOptions } from '../core/utils';
 import { LANGCHAIN_ORIGIN } from './constants';
+import {
+  addTypeSafeClassifierResponseAttributes,
+  isTypeSafeClassifier,
+  startTypeSafeClassifierSpan,
+} from './typesafe-classifier';
 import type {
   LangChainCallbackHandler,
   LangChainLLMResult,
@@ -43,6 +48,7 @@ export function createLangChainCallbackHandler(options: LangChainOptions = {}):
 
   // Internal state - single instance tracks all spans
   const spanMap = new Map<string, Span>();
+  const evaluateRunIds = new Set<string>();
 
   /**
    * Exit a span and clean up
@@ -51,8 +57,9 @@ export function createLangChainCallbackHandler(options: LangChainOptions = {}):
     const span = spanMap.get(runId);
     if (span?.isRecording()) {
       span.end();
-      spanMap.delete(runId);
     }
+    spanMap.delete(runId);
+    evaluateRunIds.delete(runId);
   };
 
   /**
@@ -215,6 +222,13 @@ export function createLangChainCallbackHandler(options: LangChainOptions = {}):
       _runType?: string,
       runName?: string,
     ) {
+      // A Jev call is a real model call, so it is recorded inside an agent too.
+      if (isTypeSafeClassifier(chain)) {
+        spanMap.set(runId, startTypeSafeClassifierSpan(chain, inputs, metadata, recordInputs));
+        evaluateRunIds.add(runId);
+        return;
+      }
+
       // Skip chain spans when inside an agent context (createReactAgent).
       // The agent already creates an invoke_agent span; internal chain steps
       // (ChannelWrite, Branch, prompt, etc.) are noise.
@@ -264,14 +278,16 @@ export function createLangChainCallbackHandler(options: LangChainOptions = {}):
     handleChainEnd(outputs: unknown, runId: string) {
       const span = spanMap.get(runId);
       if (span?.isRecording()) {
-        // Add outputs if recordOutputs is enabled
-        if (recordOutputs) {
+        if (evaluateRunIds.has(runId)) {
+          addTypeSafeClassifierResponseAttributes(span, outputs, recordOutputs);
+        } else if (recordOutputs) {
           span.setAttributes({
             'langchain.chain.outputs': JSON.stringify(outputs),
           });
         }
-        exitSpan(runId);
       }
+      // Also for a sampled-out span, which would otherwise stay tracked.
+      exitSpan(runId);
     },
 
     // Chain Error Handler
@@ -281,8 +297,8 @@ export function createLangChainCallbackHandler(options: LangChainOptions = {}):
       const span = spanMap.get(runId);
       if (span?.isRecording()) {
         span.setStatus({ code: SPAN_STATUS_ERROR, message: 'internal_error' });
-        exitSpan(runId);
       }
+      exitSpan(runId);
     },
 
     // Tool Start Handler

@@ -0,0 +1,86 @@
+import { SENTRY_ORIGIN } from '@sentry/conventions/attributes';
+import type { Span } from '@sentry/core';
+import { isObjectLike } from '@sentry/core';
+import { addResponseAttributes, startEvaluateSpan } from '../typesafe';
+import { LANGCHAIN_ORIGIN } from './constants';
+import { getAgentNameFromMetadata, getConversationIdFromMetadata } from './utils';
+
+// `TypeSafeClassifier` from `@langchain/typesafe` calls Jev with `fetch`, not through `@typesafe-ai/sdk`,
+// so the TypeSafe integration does not see it. Its serialized id is `[...lc_namespace, lc_name()]`; only
+// the class name is matched, as the namespace is the part that changes between versions.
+const TYPESAFE_CLASSIFIER_NAME = 'TypeSafeClassifier';
+
+/** The package's default, used when the classifier is constructed without a `model`. */
+const DEFAULT_TYPESAFE_CLASSIFIER_MODEL = 'jev-latest';
+
+// The `state` the classifier sends to Jev (messages rendered as transcript lines), keyed by the input
+// passed to `invoke()`. LangChain hands that same object to `handleChainStart`, so the span records what
+// Jev received rather than LangChain's serialized messages.
+const wireStates = new WeakMap<object, unknown>();
+
+/** Record the `state` a `TypeSafeClassifier` sends for `input`, using the classifier's own serializer. */
+export function recordTypeSafeClassifierState(classifier: unknown, input: unknown): void {
+  // A string state is sent as is, so there is nothing to record.
+  if (!isObjectLike(input) || !isObjectLike(classifier) || typeof classifier.payload !== 'function') {
+    return;
+  }
+
+  try {
+    const body: unknown = JSON.parse(classifier.payload(input));
+    if (isObjectLike(body)) {
+      wireStates.set(input, body.state);
+    }
+  } catch {
+    // The classifier rejects the same input itself; the span keeps LangChain's form of it.
+  }
+}
+
+// Typed loosely: LangChain's `Serialized` union does not match our handler's chain type.
+export function isTypeSafeClassifier(chain: unknown): boolean {
+  return isObjectLike(chain) && Array.isArray(chain.id) && chain.id[chain.id.length - 1] === TYPESAFE_CLASSIFIER_NAME;
+}
+
+/** Start an `evaluate` span for a `TypeSafeClassifier` run, from its serialized constructor arguments. */
+export function startTypeSafeClassifierSpan(
+  chain: unknown,
+  inputs: Record<string, unknown>,
+  metadata: Record<string, unknown> | undefined,
+  recordInputs: boolean,
+): Span {
+  const kwargs = isObjectLike(chain) && isObjectLike(chain.kwargs) ? chain.kwargs : {};
+  const model = typeof kwargs.model === 'string' ? kwargs.model : DEFAULT_TYPESAFE_CLASSIFIER_MODEL;
+
+  return startEvaluateSpan({ model, state: getState(inputs), questions: kwargs.questions }, undefined, recordInputs, {
+    [SENTRY_ORIGIN]: LANGCHAIN_ORIGIN,
+    ...getAgentNameFromMetadata(metadata),
+    ...getConversationIdFromMetadata(metadata),
+  });
+}
+
+/** The classifier result reports usage in camelCase, the TypeSafe helpers read the API's snake_case. */
+export function addTypeSafeClassifierResponseAttributes(span: Span, outputs: unknown, recordOutputs: boolean): void {
+  if (!isObjectLike(outputs)) {
+    return;
+  }
+
+  const usage = isObjectLike(outputs.usage) ? outputs.usage : {};
+  addResponseAttributes(
+    span,
+    {
+      model: outputs.model,
+      answers: outputs.answers,
+      usage: { input_tokens: usage.inputTokens, output_tokens: usage.outputTokens },
+    },
+    recordOutputs,
+  );
+}
+
+function getState(inputs: Record<string, unknown>): unknown {
+  // LangChain hands a string or array input to callbacks wrapped as `{ input }`.
+  const keys = Object.keys(inputs);
+  const input = inputs.input;
+  const state =
+    keys.length === 1 && keys[0] === 'input' && (typeof input === 'string' || Array.isArray(input)) ? input : inputs;
+
+  return isObjectLike(state) && wireStates.has(state) ? wireStates.get(state) : state;
+}
@@ -564,6 +564,22 @@ export function extractToolDefinitions(extraParams?: Record<string, unknown>): s
   return JSON.stringify(toolDefs);
 }
 
+// A run started without config inherits the parent run's config, which LangChain keeps in a global
+// AsyncLocalStorage. `ensureConfig` lets an explicit `callbacks` key replace the inherited one.
+const LANGCHAIN_ALS_KEY = Symbol.for('ls:tracing_async_local_storage');
+const LANGCHAIN_CHILD_CONFIG_KEY = Symbol.for('lc:child_config');
+
+/** The callbacks LangChain gives a run started without config: the ones of the parent run, if any. */
+export function getInheritedLangChainCallbacks(): unknown {
+  const storage = (globalThis as Record<symbol, unknown>)[LANGCHAIN_ALS_KEY] as
+    | { getStore?: () => unknown }
+    | undefined;
+  const store = typeof storage?.getStore === 'function' ? storage.getStore() : undefined;
+  const extra = isObjectLike(store) ? (store.extra as Record<symbol, unknown> | undefined) : undefined;
+  const config = isObjectLike(extra) ? extra[LANGCHAIN_CHILD_CONFIG_KEY] : undefined;
+  return isObjectLike(config) ? config.callbacks : undefined;
+}
+
 /** Duck-types a LangChain `CallbackManager` (avoids coupling to a specific `@langchain/core` resolution). */
 function isCallbackManager(value: unknown): value is {
   addHandler: (handler: unknown, inherit?: boolean) => void;