OpenClaw 2026.5.18 bundled in NemoClaw v0.0.50 has critical regressions: Telegram silent failure + web_fetch SSRF blocks host.openshell.internal

[convexchaos]

Description

Environment

• NemoClaw: v0.0.50
• OpenClaw: 2026.5.18
• Hardware: NVIDIA DGX Spark (aarch64)
• Inference: vLLM via compatible-endpoint

Issues

1. Telegram channel silently fails to start

After onboard with Telegram configured, the bot never connects.
No errors in logs — gateway starts, "starting channels and sidecars..." logs, then nothing for Telegram.
Bot token is registered but the bridge never initializes.
Fixed in OpenClaw 2026.5.22 (Telegram polling stall fix).

2. web_fetch SSRF blocks host.openshell.internal

Agents cannot reach local services via web_fetch using host.openshell.internal even when the local-inference policy explicitly allows the endpoint with allowed_ips configured.
Error: "Blocked hostname or private/internal/special-use IP address"
This breaks any agent architecture that uses local APIs (Gmail, custom tools, etc.).
Related: openclaw/openclaw#39604

Request

Please bundle OpenClaw 2026.5.22 in the next NemoClaw release (v0.0.51). Both issues are fixed in that version.

Reproduction Steps

Run curl -fsSL https://www.nvidia.com/nemoclaw.sh | bash

Environment

DGX Spark running NemoClaw v0.0.50 which comes with OpenClaw v2026.5.18

Debug Output

Logs

Checklist

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

[wscurran]

✨ Thanks for submitting this detailed issue about the OpenClaw 2026.5.18 regressions in NemoClaw v0.0.50, including the Telegram channel silently failing to start and the web_fetch SSRF blocking host.openshell.internal. This identifies bugs where the Telegram bridge never initializes and the local-inference policy fails to allow the endpoint with allowed_ips configured, breaking agent architectures that use local APIs.