Add LM Studio as a local model provider

[zalex666]

Closes #122

What

Adds LM Studio as a first-class local model provider, alongside Ollama — the second keyless provider riding the OpenAI-compatible /v1 path (OpenAIProvider with a placeholder key, default http://localhost:1234).

Prior art

#184 by @sobiya-22 predates this PR and implements the same core pattern — credit to them for getting there first (I should have searched before opening this one). #206 by @hacksics independently identified the keyless-provider discovery bug before self-closing. On top of the core feature, this PR ships the brand mark, README, a recommended model, and the six correctness fixes below — the first of which is what makes "Detect → models appear" work at the default URL — and is rebased on current main. Zero attachment to which PR lands: if maintainers prefer #184, everything here can be cherry-picked onto it and I'll close this.

Users get the same experience Ollama has today: pick LM Studio in the provider gallery, optionally adjust the server URL, hit Detect, and the models already downloaded in LM Studio appear in the composer picker — gated on a liveness probe so a machine without LM Studio never shows phantom local models.

Why

LM Studio is one of the most popular ways to run local models, exposes the same OpenAI-compatible surface the provider layer already speaks, and (like the compat vendors, per the registry's own note) deserves a first-class listing rather than the "point the OpenAI slot at a custom endpoint with a dummy key" trick.

How

• providers/registry.py — new lmstudio descriptor + builder (OpenAIProvider(api_key="lm-studio", base_url=<root>/v1)). _normalize_ollama_url is generalized to _normalize_local_url(url, default) since both servers want <root>/v1. The keyless verify branch (GET /v1/models, no auth header, friendly 404/401 copy) now covers both local providers via LOCAL_PROVIDERS.
• providers/capabilities.py — lmstudio:* models get the same conservative local-model defaults as Ollama (tools on, vision/parallel-tool-calls off) unless the matrix says otherwise.
• server/manager.py — _ollama_alive/_ollama_models generalized to _local_alive(name)/_local_models(name), driven by a small LOCAL_MODEL_SERVERS table. Ollama keeps its native /api/tags; LM Studio uses the OpenAI-shaped /v1/models (its native REST API is still beta and can sit behind an auth token). Liveness stays cached 30s, now per provider. Picker gating and add-model suggestions follow the table.
• GUI — LM Studio mark vendored from the same MIT lobe-icons set, gallery order next to Ollama, and the keyless "install the app" help copy generalized into a LOCAL_HELP map (it previously hardcoded Ollama as the only keyless provider).
• Recommended model — qwen/qwen3-coder-30b, the LM Studio catalog id of the same verified tool-calling pick the Ollama row recommends.

Review hardening (second commit)

An adversarial review pass (GPT-5.6 Sol @ xhigh via Codex CLI) surfaced five latent issues in the keyless-local-provider flow — all pre-existing Ollama behaviors this change inherited or generalized — fixed here since the PR owns that surface now:

• Keyless Detect persists on every pass. Keyless providers report configured out of the box, so the GUI's dirty/configured gate skipped the first-time save: no stored profile, and set_provider's recommended-model auto-add never ran. _local_models pairs with this by treating a stored empty profile as engaged (is None gate).
• _provider_available (local servers must actually answer; everyone else must be configured) now drives picker gating, model_ready, and the first-working-provider default handoff — a dead local server no longer reads as ready, and no longer blocks a working provider from taking over the default.
• Probes check shape, not just status. Verify and liveness require the provider's list container (data at /v1/models, models at /api/tags) — a random service answering on the port no longer passes as a model server.
• Explicitly-blank endpoint means "back to the default" in verify_provider, instead of resurrecting the stored URL (a passing Test used to validate a config the subsequent save then removed). Blank secrets still fall back to the stored key.
• Liveness cache invalidates on config change, so repointing/removing a local server re-probes immediately instead of serving up-to-30s-stale results.

A follow-up confirmation review verified the fixes.

Tests

Mirrors and extends the existing Ollama coverage: URL normalization, builder base_url/placeholder-key, router prefix-strip for slash-carrying LM Studio ids, capabilities, keyless verify request shape + non-model-server rejection, liveness-gated picker entries (per-provider: one local server being up doesn't surface the other's models), both list-shape parses, cache invalidation, model_ready gating, the first-Detect end-to-end path, and the GUI hook's keyless-Detect save. The Playwright provider fixture gains an lmstudio entry.

• pytest tests/ — 1124 passed, 1 skipped
• tsc --noEmit clean, vitest run — 110 passed
• Live against a real LM Studio server: verify/liveness/discovery, the first-Detect (blank endpoint) flow end-to-end, picker gating and model_ready, and tool-calling round trips on qwen-qwen2.5-coder-7b and qwen/qwen3.6-27b — JIT model load, structured get_weather(city=Paris) call, tool-result fold-in, usage metering, and streaming (a thinking model's reasoning_content flows through the reasoning sidecar correctly)

🤖 Generated with Claude Code