# [Bug/Question] Runtime Context Not Working in Deployed LangSmith Agent - Headers and Config Stripped

LangSmith Product Help Deployment
, ,
1

Summary

I’m unable to pass runtime configuration (specifically x-nxty-tenant-id) to my deployed LangGraph agent on LangSmith. Both HTTP headers and request body fields are being stripped/ignored, but the same code works almost perfectly in local development.

Similar issue: Manager graph does not get the header despite adding the custom heade in langgraph.json - #14 by mariam-eissa

Environment

  • LangGraph Version: Latest (as of Jan 2026)
  • LangChain Version: 1.2.6
  • Node Version: 20
  • Deployment: LangSmith Cloud (deployed via GitHub integration)
  • Deployment ID: afc7591d-cb78-4aa0-85b8-f97e35c5b110
  • Agent Type: createAgent from langchain package with custom middleware

Expected Behavior

The middleware should receive tenant ID via runtime.context or runtime.configurable from:

  1. HTTP header x-nxty-tenant-id (configured in langgraph.json)
  2. Request body context field (with contextSchema defined)
  3. Request body config field (for HTTP API)

Actual Behavior

Local Development (langgraph dev)

  • :white_check_mark: Works: context field in request body → accessible in runtime.context
  • :cross_mark: Doesn’t work: HTTP header x-nxty-tenant-id → not in runtime.configurable

Deployed on LangSmith

  • :cross_mark: Doesn’t work: context field in request body → runtime.context is empty {}
  • :cross_mark: Doesn’t work: config field in request body → runtime.configurable is empty {}
  • :cross_mark: Doesn’t work: HTTP header x-nxty-tenant-id → runtime.configurable is empty {}

Configuration Files

langgraph.json

{
  "node_version": "20",
  "graphs": {
    "agent": "./src/agent/react-agent/index.ts:agent"
  },
  "http": {
    "configurable_headers": {
      "includes": ["x-nxty-tenant-id"]
    }
  },
  "env": ".env",
  "dependencies": ["."],
  "image_distro": "wolfi"
}

Agent Configuration (src/agent/react-agent/index.ts)

import { createAgent } from 'langchain'
import * as z from 'zod'

const contextSchema = z.object({
  'x-nxty-tenant-id': z.string().optional(),
})

const stateSchema = z.object({
  messages: z.array(z.any()),
})

export const agent = createAgent({
  model: 'gpt-4o-mini',
  tools: [...],
  stateSchema,
  contextSchema,  // Defined to support runtime context
  middleware: [configureModelMiddleware],
  checkpointer: new MemorySaver(),
})

Middleware (src/agent/react-agent/middleware/configure-model.ts)

import { createMiddleware } from 'langchain'

export const configureModelMiddleware = createMiddleware({
  name: 'ConfigureModel',
  stateSchema,
  wrapModelCall: async (request, handler) => {
    console.log('[ConfigureModel] Runtime:', JSON.stringify(request.runtime, null, 2))
    
    // Try all possible sources
    const tenantIdFromContext = request.runtime.context?.['x-nxty-tenant-id']
    const tenantIdFromConfigurable = request.runtime.configurable?.['x-nxty-tenant-id']
    
    console.log('tenantIdFromContext:', tenantIdFromContext)
    console.log('tenantIdFromConfigurable:', tenantIdFromConfigurable)
    
    // Both are undefined in deployed environment!
    
    return handler(request)
  },
})

Test Cases

Local Development (Works Partially)

# ✅ Works: context in body
curl -X POST http://localhost:2024/threads/test-123/runs/stream \
  -H "Content-Type: application/json" \
  -d '{
    "input": {"messages": [{"role": "user", "content": "hello"}]},
    "context": {"x-nxty-tenant-id": "test-tenant"},
    "assistant_id": "agent"
  }'

# Logs show: runtime.context = {"x-nxty-tenant-id": "test-tenant"} ✅

# ❌ Doesn't work: HTTP header
curl -X POST http://localhost:2024/threads/test-123/runs/stream \
  -H "Content-Type: application/json" \
  -H "x-nxty-tenant-id: test-tenant" \
  -d '{
    "input": {"messages": [{"role": "user", "content": "hello"}]},
    "assistant_id": "agent"
  }'

# Logs show: runtime.configurable = undefined ❌

Deployed on LangSmith (Nothing Works)

# ❌ Attempt 1: context field
curl -X POST https://my-app.us.langgraph.app/threads/test-123/runs/stream \
  -H "Content-Type: application/json" \
  -H "x-api-key: lsv2_pt_..." \
  -d '{
    "input": {"messages": [{"role": "user", "content": "hello"}]},
    "context": {"x-nxty-tenant-id": "test-tenant"},
    "assistant_id": "agent"
  }'

# Logs show: runtime.context = {} ❌

# ❌ Attempt 2: config field (as per docs)
curl -X POST https://my-app.us.langgraph.app/threads/test-123/runs/stream \
  -H "Content-Type: application/json" \
  -H "x-api-key: lsv2_pt_..." \
  -d '{
    "input": {"messages": [{"role": "user", "content": "hello"}]},
    "config": {"x-nxty-tenant-id": "test-tenant"},
    "assistant_id": "agent"
  }'

# Logs show: runtime.configurable = undefined ❌

# ❌ Attempt 3: HTTP header (as per langgraph.json config)
curl -X POST https://my-app.us.langgraph.app/threads/test-123/runs/stream \
  -H "Content-Type: application/json" \
  -H "x-api-key: lsv2_pt_..." \
  -H "x-nxty-tenant-id: test-tenant" \
  -d '{
    "input": {"messages": [{"role": "user", "content": "hello"}]},
    "assistant_id": "agent"
  }'

# Logs show: runtime.configurable = undefined ❌

Observations

  1. Local dev: contextSchema works for context field, but configurable_headers doesn’t work for HTTP headers
  2. Deployed: Neither contextSchema nor configurable_headers work - both runtime.context and runtime.configurable are always empty objects or undefined
  3. The headers ARE being sent (verified with curl verbose mode)
  4. The LangSmith deployment logs show the middleware is being called, but runtime objects are empty

Questions

  1. Is there a known issue with configurable_headers in LangSmith deployments? The documentation suggests this should work, but it appears to be stripped before reaching the middleware.

  2. Why doesn’t contextSchema work in deployed environments? The docs show examples using context parameter with SDK, but it’s unclear if this should work with HTTP API.

  3. Is there a different approach for passing runtime configuration to deployed agents? We need to support multi-tenant deployments where each request needs tenant-specific configuration.

  4. Are there any LangSmith-specific configurations needed? Maybe something in the deployment settings or environment variables?

Workarounds Attempted

  • :white_check_mark: Using assistants API with pre-configured assistants (works but not scalable for multi-tenant)
  • :cross_mark: Environment variables (can’t be dynamic per-request)
  • :cross_mark: State-based approach (state is for conversation-level data, not request-level config)
  • :cross_mark: Custom HTTP middleware via webapp.py (Python only, we’re using TypeScript/JavaScript)

Documentation References

Request

Could someone from the LangChain team or community provide guidance on:

  1. How to properly pass runtime configuration to deployed LangSmith agents
  2. Whether this is a bug or a configuration issue on our end
  3. Any workarounds or best practices for multi-tenant deployments

This is blocking our production deployment as we need per-request tenant isolation for security and data separation.

Thank you!

Additional Context: We’re building a multi-tenant SaaS where each request needs to load tenant-specific secrets (API keys, database connections, etc.) based on the incoming tenant ID. The current behavior makes it impossible to implement this pattern securely.

2

I was able to get this data eventually by simply calling getConfig()

But this puzzles me even more. How does getConfig know what thread am I in right know and which headers were retrieved in last request? What happens if multiple clients are using my agent and I have multiple threads running in parallel at the same time? Can I get into some race condition?