Use the Qwen Code Adapter | NVIDIA NeMo Fabric

Use the nvidia.fabric.qwen adapter to run tasks with Qwen Code. Each NVIDIA NeMo Fabric runtime uses @qwen-code/sdk to start one bundled Qwen CLI child process and reuses one live session across ordered invocations. Qwen Code runs as a separate process; it is not embedded in the adapter process.

Install the Adapter

Install Node.js 22.19 or later. Then install the adapter and its exact-pinned, consumer-managed SDK in the project that owns the NeMo Fabric configuration:

npm install --save-exact nemo-fabric-adapters-qwen @qwen-code/sdk@0.1.16

The SDK is an optional peer and bundles the Qwen CLI. Installing the adapter alone does not install it. Starting the adapter without the compatible SDK returns qwen_sdk_missing.

For source development, run these commands from the repository root:

just install-typescript-qwen
npm run build --prefix adapter-contract/typescript
npm run build --prefix adapters/typescript --workspace nemo-fabric-adapters-common
npm run build --prefix adapters/typescript --workspace nemo-fabric-adapters-qwen

Configure the Adapter

Select the Qwen Code integration and point discovery at the installed adapter descriptor:

from nemo_fabric import DiscoveryConfig, FabricConfig, HarnessConfig
from nemo_fabric import MetadataConfig, ModelConfig
config = FabricConfig(
    metadata=MetadataConfig(name="qwen-review"),
    discovery=DiscoveryConfig(
        local_paths=[
            "./node_modules/nemo-fabric-adapters-qwen/qwen.fabric-adapter.json"
        ]
    ),
    harness=HarnessConfig(
        adapter_id="nvidia.fabric.qwen",
        settings={"permission_mode": "default"},
    ),
    models={
        "default": ModelConfig(
            provider="openai",
            model="nvidia/nemotron-3-nano-omni-30b-a3b-reasoning",
            api_key_env="NVIDIA_API_KEY",
            base_url="https://integrate.api.nvidia.com/v1",
        )
    },
)

Set NVIDIA_API_KEY before starting the runtime. For a source build, set discovery.local_paths to adapters/typescript/qwen/qwen.fabric-adapter.json instead. The build commands above create the descriptor’s dist/cli.js runner.

Supported Configuration

The adapter supports the following normalized configuration:

Models: Selects the default role, or the only configured role. It maps model, api_key_env, base_url, temperature, and top_p for the openai provider.
Instructions: Supports instructions.system with mode: replace or mode: append.
Tools: Maps Qwen-native tool names from tools.blocked. Tool definitions and tools.enabled are unsupported.
Skills: Each skills.paths entry must be a directory containing SKILL.md.
MCP: Supports stdio and streamable HTTP servers, including per-server allowed and blocked tool filters.
Harness settings: Maps permission_mode to Qwen Code’s default, plan, auto-edit, auto, or yolo approval mode.

Model Endpoints

The adapter requires models.<role>.provider: openai. When base_url is set, it must identify an OpenAI-compatible Chat Completions endpoint.

HTTPS is required by default, except for loopback HTTP. For a trusted Docker or Kubernetes gateway, set harness.settings.insecure_http_model_hosts to the exact hostnames that can receive HTTP model requests. Entries contain only hostnames or IPv4 addresses, without schemes, ports, paths, or wildcards. Matching ignores case and a trailing DNS dot, and permits any port or path on the listed host. The adapter does not resolve DNS to infer trust. Allowlisted HTTP endpoints receive model credentials and prompt content over an unencrypted connection. This setting applies only to model endpoints; MCP endpoint restrictions are unchanged.

For example, allow the gateway hosts used by your deployment:

harness:
  settings:
    insecure_http_model_hosts:
      - host.docker.internal
      - release-nhx-api
      - gateway.namespace.svc.cluster.local

The adapter reads the variable named by api_key_env from NeMo Fabric environment.env first and then the parent process environment. It does not write the credential to Qwen settings.

MCP Configuration

Stdio servers map url to the executable and may define args and env, but not HTTP headers. Streamable HTTP servers may define custom_headers, but not process arguments or environment variables. Remote MCP endpoints require HTTPS except for loopback development endpoints.

Header values may reference ${NAME}. Resolution checks NeMo Fabric environment.env first and then the parent process environment. Unresolved references fail startup. allowed_tools and blocked_tools map to Qwen Code’s per-server filters. Normalized MCP authentication is unsupported.

Only explicitly configured server names are enabled. The pinned SDK discovers external MCP servers on the first prompt and does not include them in mcpServerStatus(). The adapter captures the SDK’s MCP startup diagnostic and returns qwen_mcp_unavailable instead of accepting a result produced without every explicitly configured server.

Runtime Behavior

The adapter accepts plain-text input and returns the terminal assistant text in output.response. Qwen Code reports cumulative usage for a live query, so the adapter returns the per-invocation difference. Create a new runtime to change the selected model, workspace, system instructions, skills, MCP servers, tool policy, or permission mode.

Input totals include cached tokens, and output totals include reasoning tokens under the adapter’s fixed OpenAI authentication path. Cache counters are reported separately as a breakdown and are not added again. Missing or invalid counters, cumulative-counter resets, and all-zero usage observations remain unknown rather than being converted to zero. A valid snapshot establishes a baseline for later invocation deltas.

These are the pinned SDK’s provider-reported observations, not independently verified billing totals. The SDK normalizes missing provider usage to zero and can omit usage from failed requests. An all-zero observation cannot distinguish missing data from real zero consumption, and partial omissions inside a nonzero summary cannot be detected through the public result.

The adapter gives the child process isolated temporary Qwen home and runtime directories. It disables Qwen Code’s optional product-usage analytics and telemetry. This does not affect model-provider metering or token usage reported by NeMo Fabric. The adapter also disables extensions, ambient skills, and ambient MCP servers. Explicit NeMo Fabric environment.env values and the selected model credential are forwarded to the child process.

The adapter disables Qwen’s automatic CLI relaunch, which sizes the replacement process’s heap from host RAM. The SDK-owned child retains Node.js’s default memory sizing for its container and any explicit NODE_OPTIONS heap setting forwarded through environment.env.

In SDK mode, set harness.settings.permission_mode to control Qwen Code’s permission policy. The default value, default, denies tool calls that need interactive approval. Use auto for Qwen’s classifier-mediated policy. Use yolo only inside an appropriate task sandbox because it bypasses edit approval. Qwen’s permission policy is not a filesystem sandbox.

models.max_tokens, provider-specific model settings, runtime.max_turns, tools.enabled, native streaming, cancellation, service mode, and Relay telemetry are unsupported.

Understand the Runtime Lifecycle

The adapter runs as a persistent Node.js process, but the Qwen Code harness does not run in-process. start creates isolated configuration and asks @qwen-code/sdk to launch its bundled Qwen CLI as a child process. The SDK exchanges JSON-lines messages with the CLI over stdin and stdout. stop closes the live query, terminates the child process, and removes the temporary profile.

To try the source adapter, use the code-review agent Qwen variant.

Runtime Health

Fabric.check() supports live, active, and ready. Each level obtains a fresh native Query.supportedCommands() response from the owned CLI query. Readiness establishes initialized CLI responsiveness.

Refer to the SDK health-check overview for shared behavior and a comparison of native evidence across adapters.