Bug Description

is_local_endpoint() in agent/model_metadata.py only matches IP addresses, localhost, and _LOCAL_HOSTS. When running Hermes behind a local LLM proxy (e.g. LiteLLM, Ollama) accessed via Docker/Podman DNS name (like hermes-litellm or ollama), the function returns False.

This causes the stale stream timeout (180s default) to NOT be auto-disabled (the fix from #6368), leading to the infinite retry loop described in #7069 — the local model is still prefilling but the agent kills the connection and reconnects.

Root Cause

The check at line ~4951 in run_agent.py:

if _stream_stale_timeout_base == 180.0 and self.base_url and is_local_endpoint(self.base_url):
    _stream_stale_timeout = float("inf")

...never triggers because is_local_endpoint("http://hermes-litellm:4000/v1") returns False — it's not an IP, not in _LOCAL_HOSTS, and the hostname can't be parsed as an IP address.

Proposed Fix

Three additions to is_local_endpoint():

1. Unqualified hostnames (no dots) → always local. A hostname like ollama, litellm, or hermes-litellm is always a local network name (Docker DNS, mDNS, /etc/hosts). This covers the most common case with zero configuration.

2. DNS resolution fallback — resolve the hostname to an IP with socket.gethostbyname(), then check if it's private. This catches qualified local hostnames like ollama.local or litellm.internal.

3. Configurable model.local_endpoints — a list of hostnames in config.yaml to explicitly mark as local. Useful for edge cases where DNS resolution isn't available or the hostname is unusual.

Related

• [Bug]: Infinite retry loop caused by stream stale timeout during local LLM prefill phase #7069 — Infinite retry loop caused by stream stale timeout during local LLM prefill phase
• fix(agent): disable stale stream timeout for local providers #6368 — fix: disable stale stream timeout for local providers (merged, but incomplete for Docker DNS)