fix: route slash-based custom provider models correctly

[smurmann]

I faced an issue where using a gamma4 through lmstudio on mac via hermes cli chat was fine but through the web UI had issues. I asked my AI to provide a fix for it and this is the PR for it. I'm no expert of this tech stack but I can confirm that the fix worked, so i figured to make a PR.

Here's the AI's report

Summary:

• check custom_providers for an exact model match before slash-based provider/model heuristics
• prevent slash-containing custom endpoint models from being misrouted as OpenRouter models
• add a regression test covering custom slash-based model ids

Repro:

• configure a custom provider with model google/gemma-4-26b-a4b
• select that model in Hermes WebUI
• without this fix, the resolver can treat it as an OpenRouter/provider-path style id instead of routing to the named custom endpoint

Validation:

• pytest -q /root/hermes-webui/tests/test_model_resolver.py

[nesquena-hermes]

Thanks for digging into this and sending a fix! The root cause is real and the fix is correct.

What the bug is

When config_provider is set to openrouter, resolve_model_provider() currently returns early with (model_id, 'openrouter', config_base_url) for any slash-containing model ID. This means a custom provider entry like google/gemma-4-26b-a4b gets silently misrouted to OpenRouter instead of the configured local endpoint.

Fix assessment

The fix inserts a custom-providers lookup before the slash-heuristic block, so named entries in config.yaml -> custom_providers win regardless of the slash format. The logic is sound:

# Custom providers declared in config.yaml should win over slash-based
# OpenRouter heuristics. Their model IDs commonly contain '/' too.
custom_providers = cfg.get('custom_providers', [])
if isinstance(custom_providers, list):
    for entry in custom_providers:
        ...
        if entry_model and entry_name and model_id == entry_model:
            provider_hint = 'custom:' + entry_name.lower().replace(' ', '-')
            return model_id, provider_hint, entry_base_url or None

• ✅ Exact model-ID match guards against false positives
• ✅ entry_base_url or None correctly handles missing base_url
• ✅ entry_name.lower().replace(' ', '-') normalisation is consistent with how custom providers are resolved downstream
• ✅ Test coverage added (test_custom_provider_model_with_slash_routes_to_named_custom_provider) with a realistic repro case

One minor note: the api_key field in custom_providers entries isn't extracted here, but that matches the existing pattern — resolve_model_provider only returns (model, provider, base_url), and the key lookup happens separately in resolve_runtime_provider. No issue.

No blockers

Clean fix, well-tested. Glad you confirmed it solved the problem against a real LM Studio + google/gemma-4-26b-a4b setup.

[nesquena]

Full Review: PR #189 — custom provider slash model routing

Thanks @smurmann for finding and fixing this! Real bug — slash-containing model IDs from custom providers (like google/gemma-4-26b-a4b) were being misrouted as OpenRouter models.

Security Audit

Clean. Python-only changes to model routing logic. No external resources, no injection vectors, no XSS.

Code Review

The fix is correct — custom_providers entries from config.yaml are checked for exact model match before the slash-based OpenRouter heuristic kicks in. This prevents models like google/gemma-4-26b-a4b from being sent to OpenRouter when they belong to a local LM Studio endpoint.

The custom:<name> provider hint format is a reasonable convention for routing through named custom endpoints. The test covers the key scenario well.

One minor note: If custom_providers grows large, the linear scan per model resolution could be slow. For now this is fine — most users have 0-3 custom providers.

Agent Review Verification

Agent approved, no fixes needed. Confirmed — the code matches the agent's description.

Tests

507 passed, 0 failed, 41 skipped. No regressions. New test test_custom_provider_model_with_slash_routes_to_named_custom_provider is well-structured.

Merge Order Note

This PR and #191 both touch api/config.py and test_model_resolver.py. They modify different functions (this one: resolve_model_provider(), #191: get_available_models()), but the test file may need a trivial rebase after whichever merges first.

Verdict

Approved. Ready to merge.

[nesquena-hermes]

Rebased cleanly onto current master (post v0.41.0). All 15 model resolver tests pass including the new regression test for the slash-routing fix.

The fix is correct: custom provider entries in config.yaml → custom_providers now win over the OpenRouter slash-heuristic, so model IDs like google/gemma-4-26b-a4b route to their configured local endpoint rather than silently going to OpenRouter. Exact model-ID matching avoids false positives. Ready to merge.

[nesquena-hermes]

Thank you, @smurmann!

Custom provider model routing fix is live and working. That slash-heuristic was a real footgun for anyone running Ollama or LM Studio with multi-segment model IDs. You're in the Contributors section — thank you!