fix: support Ollama Cloud and remote Ollama endpoints

[kshitijk4poor]

Summary

Hermes assumed Ollama servers are always local (localhost / RFC-1918 IPs). Users pointing at Ollama Cloud or self-hosted remote Ollama instances found that:

• Context length detection fell back to 128K default instead of querying /api/show
• Server type detection (/api/tags) was skipped entirely
• Model auto-detection only worked for localhost / 127.0.0.1

The workaround was to run Ollama in a local Docker container, pull cloud models there, and point hermes at the Docker IP — which worked only because private IPs pass the is_local_endpoint() check.

This PR removes that locality gate so any custom endpoint (not just local ones) gets full Ollama-aware probing.

Changes

agent/model_metadata.py

• Rename detect_local_server_type() → detect_server_type() — now accepts _is_local param to avoid redundant is_local_endpoint() calls
• Rename _query_local_context_length() → _query_server_context_length() — threads locality check through to avoid double computation
• Remove is_local_endpoint() gate from context length probing at step 3 — any custom (non-known-provider) endpoint is now probed
• Remove redundant step 9 duplicate probe (step 3 already covers the same URL)
• Use longer httpx timeouts for remote endpoints (2s→5s for detection, 3s→8s for context queries)

cli.py

• Broaden model auto-detect from localhost/127.0.0.1 check to _is_known_provider_base_url() — excludes all known providers (OpenRouter, OpenAI, Anthropic, etc.), not just OpenRouter

hermes_cli/runtime_provider.py

• Same broadening in _get_model_config() using _is_known_provider_base_url()

Tests

• Updated all references to renamed functions
• Added test_remote_custom_endpoint_queries_server — verifies cloud Ollama URLs ARE probed
• Added test_known_provider_endpoint_does_not_query_server — verifies known providers are NOT probed
• Updated existing test to mock the new _query_server_context_length call path

Test plan

• python -m pytest tests/agent/test_model_metadata.py tests/test_model_metadata_local_ctx.py -q — 95 tests pass
• Full suite python -m pytest tests/ -q — 6136 passed, 0 new failures (4 pre-existing unrelated)
• Point base_url at a remote Ollama instance (cloud or VPS) and verify context length detection works
• Point base_url at local Ollama (localhost:11434) and verify existing behavior is unchanged
• Verify known providers (OpenRouter, OpenAI, Anthropic) are NOT probed with /api/tags

[krisiliev-dev]

Any updates on this PR? I am running into the issue mentioned above and I am not sure how to fix it.

[kshitijk4poor]

Closing — this is our old PR from before Ollama Cloud was implemented. Ollama Cloud is now fully supported as a first-class provider on main (added in PR #10782). Provider registration, model fetching with caching, and credential resolution are all in place.

Note: there's a known bug where the /model TUI picker shows 0 models for Ollama Cloud — tracked in #10977 with a fix in PR #10964.

[dorusone]

Hermes agent 0.7. does not connect to een started and working local Ollama tried everything

[kshitijk4poor]

Hermes agent 0.7. does not connect to een started and working local Ollama tried everything

hey please update the latest hermes, this is fixed!