# Agents Sdk Typescript Debugging

> Use when troubleshooting an agent built with the Microsoft Agents SDK (@microsoft/agents-hosting and related packages). Trigger on any of these symptoms: build or TypeScript errors, crashes on startup, 401 or auth errors on incoming requests, the bot not responding to messages, .env configuration problems, Azure AD credential failures (AADSTS errors), port conflicts, or the agent not connecting in Teams or the Agents Playground. Use even if the user doesn't mention the SDK by name — trigger on symptoms like "my bot won't start", "getting 401s", or "bot isn't responding."

- Skill: `microsoft/agents-sdk-typescript-debugging` (Agent Skill)
- Install (CLI): `npx skillmds@latest add microsoft/agents-sdk-typescript-debugging`
- Raw SKILL.md: https://api.skillmd.com/api/skills/microsoft/agents-sdk-typescript-debugging/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: DevOps & Infra
- Author: Microsoft (https://skillmd.com/u/microsoft)
- Updated: 2026-09-10
- Page: https://skillmd.com/skills/microsoft/agents-sdk-typescript-debugging

---


# Debugging Agents Built with Microsoft Agents SDK

## Overview

Most agent failures fall into one of three categories: the code doesn't build or start, the configuration is wrong, or the agent isn't reachable. Work through this checklist in order — each step confirms a prerequisite for the next.

## Checklist

You MUST create a task for each of these items and complete them in order:

1. Make sure the code builds successfully.
2. Make sure the application starts and runs without crashing.
3. Make sure the application opens a port and listens for incoming requests.
4. Validate the `.env` configuration.
5. Validate the bot's credentials against Azure AD.
6. Use the Agents Playground to test the agent end-to-end locally.

---

### 1. Build the code

```bash
npm run build
```

Expected: exits with code 0, no errors. Fix any TypeScript or import errors before continuing.

---

### 2. Start the application

Run with debug logging enabled to get detailed output from the SDK internals:

```bash
DEBUG=agents:* npm start
```

Or for anonymous local dev:

```bash
DEBUG=agents:* node ./dist/index.js
```

The `DEBUG=agents:*` flag enables verbose logging across all SDK namespaces. Scope down to reduce noise:

```bash
DEBUG=agents:* npm start                    # everything
DEBUG=agents:authorization:* npm start      # all auth (most useful starting point)
DEBUG=agents:msal npm start                 # token acquisition only
```


#### Auth & connections

| Namespace | What it logs |
|---|---|
| `agents:authorization:connections` | Auth connections loaded at startup (clientId, tenantId, authType); which connection is selected per request |
| `agents:authorization:manager` | Auth handlers configured at startup (type, scopes); which handler is invoked per request |
| `agents:authorization:azurebot` | Azure Bot sign-in flow detail (token exchange, magic code, SSO) |
| `agents:authorization:agentic` | Agentic auth flow detail (token acquisition, OBO) |
| `agents:authorization` | High-level authorization middleware decisions |
| `agents:msal` | MSAL token acquisition (token requests, cache hits, OBO) |
| `agents:jwt-middleware` | Incoming JWT validation |
| `agents:authConfiguration` | Auth configuration loading |

#### Adapter & request handling

| Namespace | What it logs |
|---|---|
| `agents:cloud-adapter` | Incoming request processing, activity dispatch |
| `agents:base-adapter` | Base adapter lifecycle |
| `agents:connector-client` | Outbound calls to the Bot Connector service |
| `agents:user-token-client` | User token client requests |

#### Application & state

| Namespace | What it logs |
|---|---|
| `agents:app` | AgentApplication routing and lifecycle |
| `agents:activity-handler` | ActivityHandler event dispatch |
| `agents:state` | State read/write operations |
| `agents:turnState` | Turn state access |
| `agents:memory-storage` | MemoryStorage read/write |
| `agents:middleware` | Middleware pipeline execution |

#### Streaming, attachments & transcripts

| Namespace | What it logs |
|---|---|
| `agents:streamingResponse` | Streaming response lifecycle |
| `agents:attachmentDownloader` | Attachment download requests |
| `agents:M365AttachmentDownloader` | M365-specific attachment downloads |
| `agents:file-transcript-logger` | File transcript write operations |
| `agents:rest-client` | REST client calls (transcript middleware) |

#### Agent-to-agent

| Namespace | What it logs |
|---|---|
| `agents:agent-client` | Outbound agent client calls and response handling |

Watch for crash output. Common startup errors:

- **`Cannot find module`** — missing `npm install`, or `dist/` not built yet
- **`ERR_MODULE_NOT_FOUND`** — check `"type": "module"` in `package.json` and that imports use `.js` extensions
- **Port already in use** — another process is on port 3978; kill it or set `PORT` in `.env`

If the agent starts cleanly, you should see output like:

```
Server listening on port 3978
```

---

### 3. Confirm the agent is reachable

```bash
curl -s -o /dev/null -w "%{http_code}" \
  -X POST http://localhost:3978/api/messages \
  -H "Content-Type: application/json" \
  -d '{}'
```

| Response | Meaning |
|---|---|
| `401` | Agent is running, auth is active — this is correct for a configured agent |
| `200` | Agent is running with auth disabled (blank `clientId`) — correct for anonymous local dev |
| `000` or connection refused | Agent is not running, wrong port, or crashed on startup |

---

### 4. Validate `.env` configuration

Configuration mistakes are the most common source of failures. Check each area below.

#### 4a. Confirm the file is being loaded

The SDK requires `node --env-file .env` (Node 20+) or a manual `dotenv` call. If you're using `npm start`, check that `package.json` uses `--env-file`:

```json
"start": "node --env-file .env ./dist/index.js"
```

Without `--env-file`, environment variables silently don't load and the agent starts with no auth config.

#### 4b. Check for the correct env var format

The SDK uses the **modern `connections__` format**. Using the legacy flat format (`clientId=`, `clientSecret=`, `tenantId=`) still works but is only for backwards compatibility. Mixing the two formats causes silent misconfiguration.

**Modern format (use this):**
```
connections__serviceConnection__settings__clientId=<your-app-id>
connections__serviceConnection__settings__clientSecret=<your-secret>
connections__serviceConnection__settings__tenantId=<your-tenant-id>
connectionsMap__0__connection=serviceConnection
connectionsMap__0__serviceUrl=*
```

**Legacy format (avoid for new agents):**
```
clientId=<your-app-id>
clientSecret=<your-secret>
tenantId=<your-tenant-id>
```

#### 4c. Check double-underscore separators

A single underscore (`_`) is not the same as a double underscore (`__`). The SDK uses `__` to separate path segments. A typo like `connections_serviceConnection_settings_clientId` will be silently ignored.

#### 4d. Check `connectionsMap` entries

If you have multiple connections, each must have a `connectionsMap` entry. The first entry whose `serviceUrl` pattern matches the incoming request wins. Always include a `serviceUrl=*` fallback as the last entry.

With a single connection, `connectionsMap` can be omitted — the SDK defaults to `serviceUrl=*`.

#### 4e. Check OAuth handler variables

If your agent uses user sign-in (`authorization: { graph: { ... } }`), the OAuth connection name must be set:

```
AgentApplication__UserAuthorization__Handlers__graph__Settings__azureBotOAuthConnectionName=GraphOAuthConnection
```

The handler ID (`graph`) must match the key used in the `authorization` config in your code. A mismatch causes the sign-in flow to fail silently or with a cryptic error.

---

### 5. Validate bot credentials against Azure AD

Once the `.env` looks correct, confirm the credentials actually work by requesting a token:

```bash
curl -s -X POST \
  "https://login.microsoftonline.com/$tenantId/oauth2/v2.0/token" \
  -d "grant_type=client_credentials\
&client_id=$clientId\
&client_secret=$clientSecret\
&scope=https://api.botframework.com/.default" \
  | jq '{token_type, expires_in, error, error_description}'
```

A successful response includes `access_token`. Common errors:

| Error code | Cause |
|---|---|
| `AADSTS700016` | `clientId` not found in tenant — wrong ID or wrong tenant |
| `AADSTS7000215` | Invalid `clientSecret` — expired or incorrect |
| `AADSTS90002` | `tenantId` not found |

---

### 6. Test with Agents Playground

The Agents Playground acts as a mock connector and channel client. It lets you test the full message flow locally without deploying to Azure or configuring a real Bot resource.

**Install:**
```bash
npm install -g agentsplayground
```

**Run against an anonymous agent (no `.env` needed):**
```bash
npm test
```

This assumes your `package.json` has:
```json
"start:anon": "node ./dist/index.js",
"test-tool": "agentsplayground -c emulator",
"test": "npm-run-all -p -r start:anon test-tool"
```

**Run against an authenticated agent:**
```bash
agentsplayground -c msteams \
  --client-id <your-app-id> \
  --client-secret <your-secret> \
  --tenant-id <your-tenant-id>
```

**Channel options** (`-c`): `msteams`, `webchat`, `directline`, `emulator`, `agents`

If the playground connects but messages don't get responses, the agent is running but a message handler may be missing or the route isn't matching. Add a fallback handler to confirm:

```typescript
this.onActivity('message', async (ctx: TurnContext) => {
  await ctx.sendActivity(`Echo: ${ctx.activity.text}`)
}, [], RouteRank.Last)
```

---

### Validate an OAuth connection name

OAuth connection names can only be tested end-to-end through a real sign-in flow:

**Azure Portal → Your Bot Resource → Settings → OAuth Connection Settings → [your connection] → Test Connection**

This confirms the connection name matches, the OAuth app has the right scopes, and the redirect URI (`https://token.botframework.com/.auth/web/redirect`) is registered on the app registration.

## Contributing

If you hit a problem this skill couldn't solve, found a workaround, or noticed something wrong or outdated, that's valuable — please help improve this skill for everyone.

Draft a suggested issue title and body based on the conversation, then ask the user to open it at: https://github.com/microsoft/agents/issues/new

A good issue includes:
- What the user was trying to do
- What went wrong (errors, unexpected behavior)
- What worked — including any workaround found during this conversation
- Relevant code or config snippets