Description

Description

After fresh install + onboard with cloud OpenAI provider on multiple Linux hosts (Brev cloud + on-prem Ubuntu), OpenClaw TUI inside the sandbox returns "Connection error" when sending any prompt. Sandbox logs show OpenClaw's url-fetch security guard is blocking https://inference.local/v1/responses (the Gateway's own inference proxy route) as a "private/internal/special-use IP address", which prevents both cloud（OpenAI) from working through the TUI.

Device:        Brev cloud (host: brev-ydoa5pmhb), NVIDIA GPU 1x 80GB VRAM
OS:            Ubuntu (version not captured — typical Brev base)
Architecture:  x86_64
Node.js:       v22.22.2
npm:           10.9.7
Docker:        not captured (preflight: "Docker is running", "Container runtime: docker")
OpenShell CLI: openshell 0.0.36
NemoClaw:      v0.0.26
OpenClaw:      2026.4.9 (0512059)

Environment2:
Host:          local-glennz@2u1g-b650-0082 (Ubuntu host with Docker CE)
NemoClaw:      v0.0.26
OpenShell:     0.0.36
OpenClaw:      2026.4.9 (0512059)
Sandbox name:  lynn0426
Onboard:       OpenAI provider, gpt-5.4-mini, default Balanced policy preset
                   (local-inference NOT enabled)

1. On a fresh Linux host with NVIDIA GPU and Docker running, run:
   curl -fsSL https://www.nvidia.com/nemoclaw.sh | NEMOCLAW_INSTALL_TAG=v0.0.26 bash
2. In the onboard wizard, choose:
   - Provider: 2 (OpenAI)
   - Model: 2 (gpt-5.4-mini)
   - API key: valid OpenAI key
   - Sandbox name: lynn0427
   - Web search: enable Brave (optional, not required to repro)
   - Telegram: enable (optional, not required to repro)
   - Policy tier: Balanced (default presets — local-inference is NOT enabled)
3. Wait for sandbox build + Gateway ready.
4. Run: nemoclaw lynn0427 connect
5. Inside sandbox: openclaw tui
6. Type any prompt (e.g. "hello") and press Enter.

TUI displays a valid agent reply via the configured cloud provider (OpenAI gpt-5.4-mini), confirming inference flows TUI -> Gateway (inference.local route) -> OpenAI API.

TUI returns "Connection error." repeatedly for every prompt. Status bar shows: "connected | error" / "agent main | session main (openclaw-tui) | openai/gpt-5.4-mini | tokens ?/131k".

Sandbox logs (nemoclaw lynn0427 logs --follow) show:

[security] blocked URL fetch (url-fetch) reason=Blocked hostname or private/internal/special-use IP address
[agent] embedded run agent end: runId=2d399f3d-... isError=true model=gpt-5.4-mini provider=openai error=LLM request failed: network connection error. rawError=Connection error.

[agent] embedded run failover decision: runId=... stage=assistant decision=surface_error reason=timeout provider=openai/gpt-5.4-mini profile=-

The block + agent error pattern repeats for every prompt. Inference never succeeds.

Excerpt from nemoclaw lynn0427 logs --follow:

2026-04-27T08:51:24.478+00:00 [security] blocked URL fetch (url-fetch) reason=Blocked hostname or private/internal/special-use IP address
2026-04-27T08:51:24.971+00:00 [security] blocked URL fetch (url-fetch) reason=Blocked hostname or private/internal/special-use IP address

2026-04-27T08:51:25.824+00:00 [security] blocked URL fetch (url-fetch) reason=Blocked hostname or private/internal/special-use IP address

2026-04-27T08:51:25.844+00:00 [agent] embedded run agent end: runId=2d399f3d-e86b-4905-acf2-8964f90e3611 isError=true model=gpt-5.4-mini provider=openai error=LLM request failed: network connection error. rawError=Connection error.

2026-04-27T08:51:43.846+00:00 [agent] embedded run failover decision: runId=2d399f3d-e86b-4905-acf2-8964f90e3611 stage=assistant decision=surface_error reason=timeout provider=openai/gpt-5.4-mini profile=-

Also seen continuously (likely cosmetic / sandbox-locked syscall, may be unrelated):

[guard] os.networkInterfaces() failed: A system error occurred: uv_interface_addresses returned Unknown system error 1 (Unknown system error 1) — returning empty (mDNS disabled)

Onboard policy step (Balanced tier, default selection — local-inference NOT enabled):

[✓] npm  [✓] pypi  [✓] huggingface  [✓] brew  [✓] brave  [✓] telegram

[ ] discord  [ ] github  [ ] jira  [ ] local-inference  [ ] outlook  [ ] slack

  2026-04-27T09:26:22.747+00:00 [security] blocked URL fetch (url-fetch)

    reason=Blocked hostname or private/internal/special-use IP address
  2026-04-27T09:26:24.053+00:00 [agent] embedded run agent end:
    runId=9c6de9ae-c6a7-4851-ad30-e20d5a660235 isError=true
    model=gpt-5.4-mini provider=openai
    error=LLM request failed: network connection error
  2026-04-27T09:26:41.851+00:00 [agent] embedded run failover decision: ...
    decision=surface_error reason=timeout provider=openai/gpt-5.4-mini

---

Suggested Fix

inference.local is the Gateway's own routing hostname (set during onboard: "Route: inference.local / Provider: openai-api / Model: gpt-5.4-mini"). All TUI inference traffic — including cloud provider calls — flows TUI -> inference.local -> Gateway proxy -> external API. The url-fetch guard appears to resolve inference.local through the sandbox DNS proxy (10.200.0.1) and classify the result as a private/internal address, then deny it.

Possible fixes (engineering decision):
(a) Whitelist the Gateway's managed inference.local route in the url-fetch guard by default.
(b) Auto-enable the local-inference policy preset whenever any inference provider is configured (cloud or local), since inference.local is required regardless of provider type.
(c) Surface a more actionable error in the TUI ("blocked by sandbox policy: enable local-inference preset") instead of generic "Connection error.".

Reproduces on default Balanced policy preset with the most common cloud provider choice.
Breaks the primary install → onboard → TUI → chat user journey end-to-end.
Affects every cloud provider option in the wizard (1, 2, 3, 4, 5, 6) since all cloud inference traffic routes through inference.local.

Bug Details

[NVB#6115797]