Skip to content

Nemotron 3 Ultra returns HTTP 503 through NemoClaw managed inference / Deep Agents Code but succeeds via direct NVIDIA API #10551

Description

@Proovetech

Investigation Summary

NemoClaw sandbox using nvidia/nemotron-3-ultra-550b-a55b reports managed inference as unhealthy with HTTP 503 from https://inference.local/v1/models, although route reachability is reported as reachable.
Deep Agents Code execution failed with status=agent_failure, response=null, response_bytes=0, correlation ID 01a04478-c0f4-7493-bf38-1452116f0adb.
The same ~111,564-character engineering-review prompt succeeds when sent directly to NVIDIA https://integrate.api.nvidia.com/v1/chat/completions using the same Ultra model, returning a complete ~22 KB response.
A separately onboarded nvidia/nemotron-3-super-120b-a12b Deep Agents Code sandbox reports managed inference as healthy.
This isolates the apparent fault to the NemoClaw/OpenShell/Deep Agents Code managed-inference path for Ultra rather than the NVIDIA credential, model endpoint or prompt. Exact underlying cause remains unknown.

Description

We are evaluating NemoClaw for a controlled software-engineering and adversarial code-review workflow.

A Deep Agents Code sandbox configured for:

nvidia/nemotron-3-ultra-550b-a55b

is created successfully, but its status reports:

Model: nvidia/nemotron-3-ultra-550b-a55b
Provider: nvidia-prod
Inference: unhealthy (https://inference.local/v1/models)
Inference gateway did not serve an inference request:
sandbox inference invocation probe returned HTTP 503.
Inference (route reachability): reachable

An attempted agent invocation also failed with:

managed non-interactive error:
error_class=unknown
category=unknown
retryable=false
correlation_id=01a04478-c0f4-7493-bf38-1452116f0adb

Unexpected error

status=agent_failure
exit_code=1
response=null
response_bytes=0

To isolate the fault, we bypassed NemoClaw and Deep Agents Code and sent the same engineering-review workload directly to NVIDIA's OpenAI-compatible endpoint:

https://integrate.api.nvidia.com/v1/chat/completions

using:

model: nvidia/nemotron-3-ultra-550b-a55b

The prompt was approximately 111,564 characters. The direct request succeeded and returned a complete substantive response of approximately 22 KB.

This demonstrates that:

the NVIDIA credential is valid;
the Ultra model endpoint is operational;
the prompt is accepted by the model;
the model can return a full response normally;
the failure appears specific to the NemoClaw managed-inference / inference.local / Deep Agents Code path.

For comparison, we onboarded a separate Deep Agents Code sandbox using:

nvidia/nemotron-3-super-120b-a12b

That sandbox reports:

Inference: healthy
Inference (route reachability): reachable

Environment:

OpenShell: 0.0.101
LangChain Deep Agents Code: v0.1.34
Harness: langchain-deepagents-code
Runtime: Docker
Host GPU: none
Sandbox GPU: disabled
Cloud NVIDIA inference

Expected behaviour:

A Deep Agents Code sandbox configured for Nemotron 3 Ultra should be able to invoke the model through NemoClaw managed inference if the model is supported.

Actual behaviour:

The managed route is reachable, but the invocation probe returns HTTP 503 and Deep Agents Code can terminate with an agent failure and zero response bytes.

We would particularly appreciate confirmation of:

whether nvidia/nemotron-3-ultra-550b-a55b is currently supported through NemoClaw managed inference with Deep Agents Code;
whether the correlation ID above can be traced;
whether there is a known compatibility issue involving OpenShell 0.0.101, Deep Agents Code v0.1.34 and Nemotron 3 Ultra;
whether additional gateway/provider configuration is required for Ultra;
and what additional sanitized logs would be useful.

Separately, we also encountered a 32,768-byte CLI command-argument limit when passing a large review prompt directly as an argument. We worked around that using stdin/file transfer, so we do not believe that is the primary fault reported here.

No API keys or credentials are included in this report.

Reproduction Steps

  1. Onboard a NemoClaw LangChain Deep Agents Code sandbox using NVIDIA cloud inference and select:
    nvidia/nemotron-3-ultra-550b-a55b

  2. Confirm the sandbox reaches Ready state.

  3. Run:
    nemoclaw evistate-code status

  4. Observe:
    Inference: unhealthy (https://inference.local/v1/models)
    Inference gateway did not serve an inference request:
    sandbox inference invocation probe returned HTTP 503.
    Inference (route reachability): reachable

  5. Attempt a Deep Agents Code task against the sandbox.

  6. Observe an agent failure. One affected execution returned:
    error_class=unknown
    category=unknown
    retryable=false
    correlation_id=01a04478-c0f4-7493-bf38-1452116f0adb
    status=agent_failure
    exit_code=1
    response=null
    response_bytes=0

  7. Bypass NemoClaw/Deep Agents Code and submit the same workload directly to:
    https://integrate.api.nvidia.com/v1/chat/completions

    with model:
    nvidia/nemotron-3-ultra-550b-a55b

  8. Observe that the direct NVIDIA API request succeeds and returns a complete model response (~22 KB for our ~111,564-character test prompt).

  9. As a comparison, onboard a fresh Deep Agents Code sandbox using:
    nvidia/nemotron-3-super-120b-a12b

  10. Run its status command and observe:
    Inference: healthy
    Inference (route reachability): reachable

No API key is required to reproduce from this report; use a normal valid NVIDIA credential.

Environment

Host:

  • Linux Google Cloud VM
  • x86_64
  • Docker container runtime
  • 4 vCPU
  • ~15.6 GiB RAM
  • No swap
  • No NVIDIA GPU detected
  • Sandbox GPU: disabled
  • NVIDIA cloud inference

NemoClaw / sandbox:

  • OpenShell: 0.0.101
  • LangChain Deep Agents Code: v0.1.34
  • Harness: langchain-deepagents-code
  • Runtime: terminal
  • Policies: npm, pypi, huggingface, brew, github

Affected sandbox:

  • Name: evistate-code
  • Model: nvidia/nemotron-3-ultra-550b-a55b
  • Provider reported by status: nvidia-prod
  • Inference route: https://inference.local/v1/models
  • Route reachability: reachable
  • Invocation probe: HTTP 503

Comparison sandbox:

  • Name: evistate-super
  • Model: nvidia/nemotron-3-super-120b-a12b
  • Provider reported by status: nvidia-prod
  • Inference: healthy
  • Route reachability: reachable

NemoClaw onboarding uses NVIDIA Build/cloud inference.

API credentials intentionally omitted.

Debug Output

nemoclaw debug --quick was run against the affected evistate-code sandbox.

Relevant output:

System:

  • Linux x86_64 Google Cloud VM
  • RAM: 15,987 MB
  • Swap: 0
  • No local NVIDIA GPU

OpenShell:

  • Gateway: nemoclaw
  • Server: https://127.0.0.1:8080
  • Status: Connected
  • Authentication: Authenticated (mTLS transport)
  • Version: 0.0.101

Affected sandbox:

  • Name: evistate-code
  • Phase: Ready
  • Sandbox ID: 391dbd6a-bfb7-4d03-bb3c-559603f108a9
  • Revision: 4

Managed inference policy explicitly permits:

  • inference.local:443
  • POST /v1/chat/completions
  • POST /v1/completions
  • POST /v1/embeddings
  • GET /v1/models
  • GET /v1/models/**

Relevant router diagnostics:

[sandbox] NET:OPEN ALLOWED inference.local:443
[sandbox] [openshell_router] routing proxy inference request endpoint=https://integrate.api.nvidia.com/v1 method=GET path=/v1/models

and during agent execution:

[sandbox] NET:OPEN ALLOWED inference.local:443
[sandbox] [openshell_router] routing proxy inference request (streaming) endpoint=https://integrate.api.nvidia.com/v1 method=POST path=/v1/chat/completions

Multiple subsequent streaming POST /v1/chat/completions requests are visible in the debug output.

The sandbox and gateway therefore appear operational and network policy permits managed inference; the failure occurs after the managed inference request is routed toward the NVIDIA endpoint.

No API credentials are included in this extract. Full nemoclaw debug --quick output is available if maintainers require it.

Logs

Affected Deep Agents Code execution:

managed non-interactive error:
error_class=unknown
category=unknown
retryable=false
correlation_id=01a04478-c0f4-7493-bf38-1452116f0adb

Unexpected error
correlation_id=01a04478-c0f4-7493-bf38-1452116f0adb

status=agent_failure
exit_code=1
response=null
response_bytes=0

Affected sandbox status separately reported:

Model:    nvidia/nemotron-3-ultra-550b-a55b
Provider: nvidia-prod
Inference: unhealthy (https://inference.local/v1/models)
  Inference gateway did not serve an inference request:
  sandbox inference invocation probe returned HTTP 503.
Inference (route reachability): reachable

Isolation test:
The same Nemotron 3 Ultra model and substantive workload succeeded when called directly through NVIDIA's integrate.api.nvidia.com/v1/chat/completions endpoint.  

Related issues reviewed before filing include #5744, #9895 and #6913; none appears to reproduce this specific Deep Agents Code + NVIDIA Endpoints + Nemotron 3 Ultra managed-inference failure.

Checklist

  • I confirmed this bug is reproducible
  • I searched existing issues and this is not a duplicate

Metadata

Metadata

Assignees

No one assigned

    Labels

    area: inferenceInference routing, serving, model selection, or outputsarea: local-modelsLocal model providers, downloads, launch, or connectivityarea: providersInference provider integrations and provider behaviorarea: sandboxOpenShell sandbox lifecycle, runtime, config, or recoveryintegration: dcodeLangChain Deep Code integration behaviorprovider: nvidiaNVIDIA inference endpoint, NIM, or NVIDIA provider behavior

    Type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions