feat: add LM Studio provider support with live model discovery

[dobby-d-elf]

Summary

Add first-class LM Studio support so locally-loaded models appear automatically in the WebUI dropdown and providers card near the top with other configured providers.

Changes

• api/config.py:

  • Resolve merge conflict, keep both _custom_slug_rest_looks_like_host_port (upstream) and new _get_provider_base_url() helper
  • Custom providers now return their configured base_url from config.yaml in resolve_model_provider() instead of None
  • Add "Configured" badge for explicitly configured providers so they appear at top of models dropdown
  • Detect LM Studio via LM_API_KEY + LM_BASE_URL env vars
  • Fetch live loaded models from LM Studio with fallback to direct HTTP requests (handles race where .env hasn't injected yet)

• api/providers.py: fetch live LM Studio model list via hermes_cli for the providers card

• static/style.css: add purple "Configured" badge style (model-opt-badge--configured)

Motivation

LM Studio is a common local inference server. Previously it required manual model listing or didn't integrate cleanly with the provider system. This makes configured LM Studio models appear live in the dropdown alongside cloud providers.

[nesquena-hermes]

Review — split request

Thanks @dobby-d-elf — pulled the branch, ran the model-resolver + custom-provider test suites against it (85/85 pass), and live-verified the LM Studio integration path. There are two conceptually distinct changes bundled here, and the second one has a UX issue that needs to be addressed before this can land. Asking for a split rather than a hold so the LM Studio piece (which is genuinely good) can ship cleanly while the badge piece gets reshaped.

Change 1 — LM Studio live model discovery (LGTM, ready to ship)

The new elif pid == "lmstudio": branch in get_available_models() is exactly the right shape — mirrors the existing nous and ollama-cloud patterns, calls provider_model_ids("lmstudio") first (which is already in hermes_cli.models.CANONICAL_PROVIDERS, verified), falls back to a direct /v1/models HTTP request when the env hasn't been injected, swallows exceptions cleanly, and only emits a group when there are actual models.

The LM_API_KEY + LM_BASE_URL env-var detection at line ~2820 is also correct — matches the canonical agent-side env var names per api/providers.py:116-126 ("lmstudio": "LM_API_KEY").

The 12 LOC addition to api/providers.py for the providers card is consistent with the existing nous block at line ~860 immediately above it.

The new _get_provider_base_url() helper plus the resolve_model_provider() change from return bare_model, provider_hint, None to return bare_model, provider_hint, _get_provider_base_url(provider_hint) is a real improvement — pre-PR, users with providers.X.base_url in config.yaml had to also set it under cfg["model"] for it to flow through resolve_model_provider's @-prefixed branch. Now the providers-section base_url is honored consistently. Verified against tests/test_resolve_model_provider_free_suffix.py + tests/test_model_resolver.py + tests/test_custom_provider_display_name.py — 44/44 still pass.

This part of the diff is ~70 LOC and could ship today on top of v0.51.31.

Change 2 — "Configured" badge ⚠️ UX regression

The new code block at api/config.py:2519-2557 adds a role=configured badge for every model from every provider with an entry in cfg["providers"]. The intent (per the PR description) is "appear at top of models dropdown."

Live-tested with a representative 3-provider config (anthropic, openrouter, lmstudio — fairly common shape):

config = {
    'providers': {
        'anthropic':  {'api_key': '...'},
        'openrouter': {'api_key': '...'},
        'lmstudio':   {'base_url': 'http://localhost:1234/v1'},
    },
}
result = get_available_models()
# configured_model_badges total: 110
#   role=primary:     3 entries (the 3 ID forms of the active model)
#   role=configured: 107 entries  ← this is the new behavior

107 models all marked "Configured" for a 3-provider user. That's effectively the entire openrouter catalog (~60 models × 3 ID forms each = ~180 entries dwarfing the other providers). For most users with anthropic + openrouter configured, this floods the dropdown's pinned section with hundreds of badge-decorated entries — the opposite of "scannable."

This is a primary-surface UX change (the model picker is what the user sees during normal conversation flow), and bunches the entire visible catalog under a single semi-meaningless label. By contrast, the existing Primary and Fallback badges pin specifically the user-selected default + fallback model — a tiny, intentional set.

Two paths forward worth considering instead:

Option A — drop the "Configured" badge entirely. Rely on the existing per-provider grouping in the dropdown (LM Studio gets its own group via Change 1, just like Anthropic / OpenRouter / etc.). The user can scroll to the LM Studio group; there's no pinning needed.

Option B — pin only the active provider's models, not all configured providers. If the goal is "models I'm using right now appear at top," scope the badge to cfg["model"]["provider"] and stop there. That's typically 1-2 providers and 5-15 models, not 107.

Either way is fine. Option A is the simpler change.

Asking for a split

Could you:

1. Drop the "Configured" badge code block at api/config.py:2519-2557 + the model-opt-badge--configured CSS rule from this PR, leaving just the LM Studio integration + the _get_provider_base_url helper / resolve_model_provider change? That's the LGTM portion above and ships cleanly.
2. (Optional, if you still want pinning) Open a separate PR with Option B (scope the badge to active-provider models only) and we can review that on its UX merits independently.

Adding the maintainer-review label while the split lands. Once the badge change is removed, I'll re-review and queue this for the next release batch (probably tomorrow or the day after — there are 2 other PRs ready to ship right now that I don't want to gate on this).

Thanks for the contribution — the LM Studio live-discovery shape itself is well-done and matches the existing nous / ollama-cloud patterns precisely. Just want to keep the dropdown UX intact.

[dobby-d-elf]

Addressed the review request and pushed the split cleanup in a300d9a.

What changed:

• Dropped the backend block that added role=configured badges for every model from every configured provider.
• Removed the corresponding .model-opt-badge--configured CSS rule.
• Left the LM Studio live model discovery and providers card updates intact.
• Left the _get_provider_base_url() / resolve_model_provider() improvement intact.

[nesquena-hermes]

Shipped via stage-337 → master in v0.51.44 (commit f00cb74f). Thanks @dobby-d-elf — your LM Studio integration is now first-class.

Two follow-up fixes landed during stage Opus review and CI testing:

1. The new branch initially only looked at cfg["providers"]["lmstudio"]["base_url"], missing the historical cfg["model"]["base_url"] shape that real users have AND that tests/test_issue1527_lmstudio_base_url_classification was already exercising. 3 pre-existing tests broke on stage. Fix: enhanced _get_provider_base_url() to fall back to cfg["model"]["base_url"] when cfg["model"]["provider"] matches the requested provider id. Negative-case test asserts model.base_url does NOT leak to non-active providers (so a user with model.provider: anthropic + a custom anthropic-proxy URL + providers.openai without base_url still gets None for openai).

2. CI on Python 3.13 (clean editable install, no hermes_cli package) was failing because the outer try/except Exception in the lmstudio branch caught ImportError from from hermes_cli.models import provider_model_ids, silently skipping the urlopen fallback. Restructured into two independent try blocks — hermes_cli tier failure no longer hijacks the urlopen fallback.

7 new regression tests in tests/test_pr1970_lmstudio_base_url_fallback.py lock the two-location lookup, the precedence rule, trailing-slash stripping, the no-leak invariant, and the CI-without-hermes_cli fallback path.

Release: https://github.com/nesquena/hermes-webui/releases/tag/v0.51.44