Python
Add LM Studio as a local server provider
Rohit C Prasad45d58dc9a5d6
diff --git a/coworker/providers/capabilities.py b/coworker/providers/capabilities.pyindex c3514148..f9285fa5 100644--- a/coworker/providers/capabilities.py+++ b/coworker/providers/capabilities.py@@ -25,7 +25,7 @@ def capabilities_for(model: str) -> ModelCapabilities:# Ollama (local) models vary widely and many fake/mishandle parallel tool calls — assume# tools work (we only point at tool-capable models) but stay conservative otherwise.# Vision is detected from common model naming conventions (-vl, vision, llava, etc.).- if provider in ("ollama", "llamacpp", "vllm"):+ if provider in ("ollama", "llamacpp", "vllm", "lmstudio"):_vision_patterns = ("-vl", "vision", "llava", "bakllava", "cogvlm", "minicpm-v")has_vision = any(p in name for p in _vision_patterns)return ModelCapabilities(diff --git a/coworker/providers/local_server.py b/coworker/providers/local_server.pyindex f33f9154..51f20375 100644--- a/coworker/providers/local_server.py+++ b/coworker/providers/local_server.py@@ -1,6 +1,6 @@-"""llama.cpp and vLLM: servers the user starts, reached through their OpenAI-compatible-`/v1`. One module for both, since OpenWorker treats them the same way and the two-differ only in what they report about themselves.+"""llama.cpp, vLLM and LM Studio: servers the user starts, reached through their+OpenAI-compatible `/v1`. One module for all three, since OpenWorker treats them the same+way and they differ only in what they report about themselves.- llama.cpp `llama-server`: `localhost:8080` by default; `/v1/models` lists the loadedmodel (its id is the file path unless the server was started with `--alias`), with@@ -11,8 +11,12 @@ differ only in what they report about themselves.`max_model_len`. Tool calls work only when the server was started with`--enable-auto-tool-choice --tool-call-parser <name>`; the harness cannot turn them on,but it can find out, and say so, before a session breaks.+- LM Studio: `localhost:1234` by default (`lms server start` or the Developer tab); its+ `/v1/models` lists every downloaded model, and its own `/api/v0/models` adds each one's+ `max_context_length` and whether it is loaded. Tool calls go through the model's chat+ template, as in llama.cpp.-Neither tells the harness what the machine is; that is local_machine.py. Both may take an+None of them tells the harness what the machine is; that is local_machine.py. Both may take anAPI key; without one any value is accepted, and OpenWorker sends a placeholder."""@@ -31,6 +35,7 @@ from .recommended import recommendation_forLOCAL_SERVERS: dict[str, dict[str, str]] = {"llamacpp": {"title": "llama.cpp", "default_base": "http://localhost:8080"},"vllm": {"title": "vLLM", "default_base": "http://localhost:8000"},+ "lmstudio": {"title": "LM Studio", "default_base": "http://localhost:1234"},}PLACEHOLDER_KEY = "none"CACHE_SECONDS = 60.0@@ -122,6 +127,8 @@ def _fetch(name: str, base: str, api_key: Optional[str]) -> list[dict[str, Any]]if not isinstance(items, list):return []props: dict[str, Any] = {}+ if name == "lmstudio":+ props = _lmstudio_catalog(base, headers)if name == "llamacpp":try:got = httpx.get(base[: -len("/v1")] + "/props", headers=headers, timeout=4.0).json()@@ -142,6 +149,20 @@ def _fetch(name: str, base: str, api_key: Optional[str]) -> list[dict[str, Any]]return rows+def _lmstudio_catalog(base: str, headers: dict[str, str]) -> dict[str, Any]:+ """LM Studio's own listing (`/api/v0/models`), keyed by id: the trained window+ (`max_context_length`) and the load state the OpenAI-compatible listing leaves out.+ Empty when the server does not answer it."""+ try:+ got = httpx.get(base[: -len("/v1")] + "/api/v0/models", headers=headers, timeout=4.0).json()+ except (httpx.HTTPError, ValueError, OSError):+ return {}+ items = got.get("data") if isinstance(got, dict) else None+ if not isinstance(items, list):+ return {}+ return {str(it["id"]): it for it in items if isinstance(it, dict) and it.get("id")}++def _vllm_tools_on(base: str, headers: dict[str, str], model: str) -> Optional[bool]:"""Ask vLLM for a one-token reply with a tool offered. A server started without toolcalling refuses the request and names the flag; one with it answers."""@@ -171,6 +192,13 @@ def _row(name: str, item: dict[str, Any], props: dict[str, Any], tools: Optionalcontext = _int((props.get("default_generation_settings") or {}).get("n_ctx"))context_max = _int(meta.get("n_ctx_train"))size = _int(meta.get("size"))+ elif name == "lmstudio":+ # LM Studio sizes the window when it loads the model (JIT, on first request, by+ # default); until then the trained window is the only figure it reports.+ own = props.get(served_id) if isinstance(props.get(served_id), dict) else {}+ context_max = _int(own.get("max_context_length"))+ context = _int(own.get("loaded_context_length")) or context_max+ size = Noneelse:context = _int(item.get("max_model_len"))context_max = context@@ -205,8 +233,8 @@ def _int(value: Any) -> Optional[int]:def thinking_as_template_kwargs(settings: dict[str, Any]) -> dict[str, Any]:- """The thinking switch travels as Ollama's `think` field (model_config.py); llama.cpp- and vLLM read it from the chat template's `enable_thinking` instead. Rewrites one into+ """The thinking switch travels as Ollama's `think` field (model_config.py); llama.cpp,+ vLLM and LM Studio read it from the chat template's `enable_thinking` instead. Rewrites one intothe other and leaves everything else alone."""extra = settings.get("extra_body")if not isinstance(extra, dict) or "think" not in extra:diff --git a/coworker/providers/registry.py b/coworker/providers/registry.pyindex 6df37aa5..6dc5a6df 100644--- a/coworker/providers/registry.py+++ b/coworker/providers/registry.py@@ -211,7 +211,7 @@ def _build_ollama(profile: dict[str, Any], secrets: Any) -> ProviderClient:def _local_server(name: str):- """Builder for llama.cpp and vLLM (providers/local_server.py): the user's own server,+ """Builder for llama.cpp, vLLM and LM Studio (providers/local_server.py): the user's own server,reached through its `/v1`, with an optional key. Without a key the server takes anyvalue, so a placeholder goes out."""from .local_server import PLACEHOLDER_KEY, remember_server, thinking_as_template_kwargs, v1_base@@ -861,6 +861,34 @@ DESCRIPTORS: list[ProviderDescriptor] = [blurb="A vLLM server, usually on a Linux machine with an NVIDIA GPU.",kind="local",),+ ProviderDescriptor(+ name="lmstudio",+ title="LM Studio",+ needs_key=False,+ fields=[+ ProviderField(+ "base_url",+ "Server address",+ secret=False,+ required=False,+ placeholder="http://localhost:1234",+ help="Where LM Studio's local server is listening (the Developer tab, or "+ "`lms server start`). The /v1 path is added automatically.",+ ),+ ProviderField(+ "api_key",+ "API key (only if the server was started with one)",+ secret=True,+ required=False,+ ),+ ],+ build=_local_server("lmstudio"),+ # The same verified tool-calling pick as Ollama's, under LM Studio's catalog id+ # (lmstudio.ai/models/qwen/qwen3-coder-30b, read 2026-10-25) — download it in-app.+ recommended_model="qwen/qwen3-coder-30b",+ blurb="The LM Studio app's local server, with whatever models you downloaded there.",+ kind="local",+ ),]_BY_NAME = {d.name: d for d in DESCRIPTORS}@@ -1188,7 +1216,7 @@ def verify_provider_key(elif name == "ollama":base = _normalize_ollama_url(base_url)resp = httpx.get(base.rstrip("/") + "/models", timeout=timeout)- elif name in ("llamacpp", "vllm"):+ elif name in ("llamacpp", "vllm", "lmstudio"):from .local_server import v1_baseheaders = {"Authorization": f"Bearer {key}"} if key else {}diff --git a/coworker/server/manager.py b/coworker/server/manager.pyindex acb8bc9e..12a4939f 100644--- a/coworker/server/manager.py+++ b/coworker/server/manager.py@@ -4288,7 +4288,7 @@ class SessionManager:topped up with the compat-vendor extras the matrix doesn't vouch for."""if name == "ollama":return [m.split(":", 1)[-1] for m in self._ollama_models()]- if name in ("llamacpp", "vllm"):+ if name in ("llamacpp", "vllm", "lmstudio"):return [m.split(":", 1)[-1] for m in self._local_server_models(name)]from ..providers.matrix import models_for_provider@@ -4351,7 +4351,7 @@ class SessionManager:if rec and rec in suggested:# OpenAI models stay bare (the router's default); others carry their prefix.added = rec if name == "openai" else f"{name}:{rec}"- elif name in ("ollama", "llamacpp", "vllm") and suggested:+ elif name in ("ollama", "llamacpp", "vllm", "lmstudio") and suggested:# A local server offers whatever the user loaded — the recommended model may# not be one.added = f"{name}:{suggested[0]}"@@ -4590,7 +4590,7 @@ class SessionManager:return {"provider": name, "models": [], "error": "no local model list for this provider"}def _local_server_alive(self, name: str) -> bool:- """Best-effort liveness of a llama.cpp or vLLM server, cached 30 s like Ollama's."""+ """Best-effort liveness of a llama.cpp, vLLM or LM Studio server, cached 30 s like Ollama's."""import timefrom ..providers import local_server@@ -4607,7 +4607,7 @@ class SessionManager:return alivedef _local_server_models(self, name: str) -> list[str]:- """The model ids a llama.cpp or vLLM server serves, as `<name>:<id>`; empty when+ """The model ids a llama.cpp, vLLM or LM Studio server serves, as `<name>:<id>`; empty whenit does not answer or was started without tool calling."""from ..providers import local_server@@ -4631,7 +4631,7 @@ class SessionManager:provider = self._model_provider(model)if provider == "ollama":return self._ollama_alive()- if provider in ("llamacpp", "vllm"):+ if provider in ("llamacpp", "vllm", "lmstudio"):return self._local_server_alive(provider)return self._provider_configured(provider)@@ -4864,7 +4864,7 @@ class SessionManager:# Pulled Ollama models are offered live while Ollama answers — no need to add each one.if self._ollama_alive():selectable = list(dict.fromkeys([*selectable, *self._ollama_models()]))- for local in ("llamacpp", "vllm"):+ for local in ("llamacpp", "vllm", "lmstudio"):if self.secrets.get(f"provider:{local}") and self._local_server_alive(local):selectable = list(dict.fromkeys([*selectable, *self._local_server_models(local)]))if self.model not in selectable:diff --git a/docs/local-servers.md b/docs/local-servers.mdindex e904fe9e..f5a4addb 100644--- a/docs/local-servers.md+++ b/docs/local-servers.md@@ -1,8 +1,8 @@-# llama.cpp and vLLM+# llama.cpp, vLLM and LM Studio-OpenWorker can use a model served by [llama.cpp](https://github.com/ggml-org/llama.cpp)-or [vLLM](https://github.com/vllm-project/vllm), on this computer or on another machine.-Each is its own provider in Settings ▸ Models & Keys, under "On your own hardware", next+OpenWorker can use a model served by [llama.cpp](https://github.com/ggml-org/llama.cpp),+[vLLM](https://github.com/vllm-project/vllm) or [LM Studio](https://lmstudio.ai), on this+computer or on another machine. Each is its own provider in Settings ▸ Models & Keys, under "On your own hardware", nextto Ollama. Give it the server's address and, only if the server was started with one, itsAPI key. OpenWorker reads the rest from the server.@@ -44,9 +44,22 @@ vllm serve nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 \The default address is `http://localhost:8000`.+## LM Studio++Download a model in the LM Studio app, then start its server from the Developer tab or+with `lms server start`. Every downloaded model is listed; LM Studio loads one when the+first request names it.++- The model's id is its LM Studio catalog name, for example `qwen/qwen3-coder-30b`.+- The context size is the one LM Studio loaded the model with (its own listing reports+ it); until the model is loaded, the trained window is shown instead.+- A key is needed only if you turned one on in the app.++The default address is `http://localhost:1234`.+## In the CLI-The model id is `llamacpp:<id>` or `vllm:<id>`:+The model id is `llamacpp:<id>`, `vllm:<id>` or `lmstudio:<id>`:```bashopenworker run --model llamacpp:my-model --prompt "..."@@ -55,5 +68,6 @@ openworker run --model llamacpp:my-model --prompt "..."## Per-model settingsLongest reply, sampling, thinking and auto-compaction are set per model as for any other-provider; see [model settings](model-settings.md). The context size of these two servers-is set when the server starts, so OpenWorker shows it rather than setting it.+provider; see [model settings](model-settings.md). The context size of these servers is+set when the server starts (or, in LM Studio, when it loads the model), so OpenWorker+shows it rather than setting it.diff --git a/surfaces/gui/src/locales/en.json b/surfaces/gui/src/locales/en.jsonindex 51709c75..c1fecfe0 100644--- a/surfaces/gui/src/locales/en.json+++ b/surfaces/gui/src/locales/en.json@@ -1621,7 +1621,9 @@"llamacpp_desc": "No API key needed unless your llama-server was started with one.","install_llamacpp": "Install llama.cpp","vllm_desc": "vLLM runs on Linux with an NVIDIA GPU. A key is needed only if the server was started with one.",- "install_vllm": "Install vLLM"+ "install_vllm": "Install vLLM",+ "lmstudio_desc": "The LM Studio app serves the models you downloaded there. A key is needed only if you turned one on in the app.",+ "install_lmstudio": "Install LM Studio"},"workspace_trust": {"title": "Trust this workspace’s commands?",diff --git a/surfaces/gui/src/locales/zh.json b/surfaces/gui/src/locales/zh.jsonindex 534f68c9..fa111121 100644--- a/surfaces/gui/src/locales/zh.json+++ b/surfaces/gui/src/locales/zh.json@@ -1601,7 +1601,9 @@"llamacpp_desc": "无需 API 密钥,除非你的 llama-server 启动时设置了密钥。","install_llamacpp": "安装 llama.cpp","vllm_desc": "vLLM 运行在带 NVIDIA GPU 的 Linux 上。仅当服务器启动时设置了密钥才需要填写。",- "install_vllm": "安装 vLLM"+ "install_vllm": "安装 vLLM",+ "lmstudio_desc": "LM Studio 应用会提供你在其中下载的模型。仅当你在应用中开启了密钥才需要填写。",+ "install_lmstudio": "安装 LM Studio"},"workspace_trust": {"title": "信任此工作区的命令?",diff --git a/surfaces/gui/src/providers/ProviderSetup.tsx b/surfaces/gui/src/providers/ProviderSetup.tsxindex 941799b9..1227b51f 100644--- a/surfaces/gui/src/providers/ProviderSetup.tsx+++ b/surfaces/gui/src/providers/ProviderSetup.tsx@@ -28,6 +28,7 @@ const LOCAL_HELP: Record<string, { desc: string; link: string; url: string }> =ollama: { desc: "provider.no_key_needed_desc", link: "provider.install_ollama", url: "https://ollama.com/download" },llamacpp: { desc: "provider.llamacpp_desc", link: "provider.install_llamacpp", url: "https://github.com/ggml-org/llama.cpp#quick-start" },vllm: { desc: "provider.vllm_desc", link: "provider.install_vllm", url: "https://docs.vllm.ai/en/latest/getting_started/installation/" },+ lmstudio: { desc: "provider.lmstudio_desc", link: "provider.install_lmstudio", url: "https://lmstudio.ai/download" },};export const KEY_HELP: Record<string, { url: string; label: string; labelKey?: string }> = {diff --git a/surfaces/gui/src/providers/logos.ts b/surfaces/gui/src/providers/logos.tsindex 487bb314..7e2995e9 100644--- a/surfaces/gui/src/providers/logos.ts+++ b/surfaces/gui/src/providers/logos.ts@@ -11,9 +11,11 @@ import gemini from "./logos/gemini.svg";import byteplus from "./logos/byteplus.svg";import volcengine from "./logos/volcengine.svg";import ollama from "./logos/ollama.svg";-// llama.cpp: selfhst/icons (CC BY 4.0); vLLM: homarr-labs/dashboard-icons (Apache 2.0).+// llama.cpp: selfhst/icons (CC BY 4.0); vLLM: homarr-labs/dashboard-icons (Apache 2.0);+// LM Studio: lobe-icons (MIT).import llamacpp from "./logos/llamacpp.svg";import vllm from "./logos/vllm.svg";+import lmstudio from "./logos/lmstudio.svg";import bedrock from "./logos/bedrock.svg";import vertex from "./logos/vertex.svg";import openrouter from "./logos/openrouter.svg";@@ -49,6 +51,7 @@ export const PROVIDER_LOGOS: Record<string, string> = {ollama,llamacpp,vllm,+ lmstudio,bedrock,vertex,openrouter,@@ -83,6 +86,7 @@ export const PROVIDER_ORDER = ["ollama","llamacpp","vllm",+ "lmstudio","bedrock","vertex","openrouter",diff --git a/surfaces/gui/src/providers/logos/lmstudio.svg b/surfaces/gui/src/providers/logos/lmstudio.svgnew file mode 100644index 00000000..ea0816b6--- /dev/null+++ b/surfaces/gui/src/providers/logos/lmstudio.svg@@ -0,0 +1 @@+<svg fill="currentColor" fill-rule="evenodd" height="1em" style="flex:none;line-height:1" viewBox="0 0 24 24" width="1em" xmlns="http://www.w3.org/2000/svg"><title>LM Studio</title><path d="M2.84 2a1.273 1.273 0 100 2.547h14.107a1.273 1.273 0 100-2.547H2.84zM7.935 5.33a1.273 1.273 0 000 2.548H22.04a1.274 1.274 0 000-2.547H7.935zM3.624 9.935c0-.704.57-1.274 1.274-1.274h14.106a1.274 1.274 0 010 2.547H4.898c-.703 0-1.274-.57-1.274-1.273zM1.273 12.188a1.273 1.273 0 100 2.547H15.38a1.274 1.274 0 000-2.547H1.273zM3.624 16.792c0-.704.57-1.274 1.274-1.274h14.106a1.273 1.273 0 110 2.547H4.898c-.703 0-1.274-.57-1.274-1.273zM13.029 18.849a1.273 1.273 0 100 2.547h9.698a1.273 1.273 0 100-2.547h-9.698z" fill-opacity=".3"></path><path d="M2.84 2a1.273 1.273 0 100 2.547h10.287a1.274 1.274 0 000-2.547H2.84zM7.935 5.33a1.273 1.273 0 000 2.548H18.22a1.274 1.274 0 000-2.547H7.935zM3.624 9.935c0-.704.57-1.274 1.274-1.274h10.286a1.273 1.273 0 010 2.547H4.898c-.703 0-1.274-.57-1.274-1.273zM1.273 12.188a1.273 1.273 0 100 2.547H11.56a1.274 1.274 0 000-2.547H1.273zM3.624 16.792c0-.704.57-1.274 1.274-1.274h10.286a1.273 1.273 0 110 2.547H4.898c-.703 0-1.274-.57-1.274-1.273zM13.029 18.849a1.273 1.273 0 100 2.547h5.78a1.273 1.273 0 100-2.547h-5.78z"></path></svg>\ No newline at end of filediff --git a/tests/test_local_server.py b/tests/test_local_server.pyindex d896372d..889809ee 100644--- a/tests/test_local_server.py+++ b/tests/test_local_server.py@@ -1,4 +1,4 @@-"""llama.cpp and vLLM as providers (providers/local_server.py, registry.py): the user's+"""llama.cpp, vLLM and LM Studio as providers (providers/local_server.py, registry.py): the user'sown server through its /v1, what each reports about its model, and how the server'smodel lists, liveness and adoption treat them."""@@ -29,10 +29,10 @@ def _fake_http(monkeypatch, handler):def test_descriptors_carry_a_kind_for_grouping():kinds = {d.name: d.kind for d in provider_descriptors()}- assert kinds["ollama"] == kinds["llamacpp"] == kinds["vllm"] == "local"+ assert kinds["ollama"] == kinds["llamacpp"] == kinds["vllm"] == kinds["lmstudio"] == "local"assert kinds["openai-codex"] == "subscription"assert kinds["openai"] == kinds["anthropic"] == "api_key"- for name in ("llamacpp", "vllm"):+ for name in ("llamacpp", "vllm", "lmstudio"):d = get_descriptor(name)assert d is not None and not d.needs_keyassert [f.key for f in d.fields] == ["base_url", "api_key"]@@ -44,6 +44,8 @@ def test_v1_base_takes_the_root_or_an_explicit_v1():assert local_server.v1_base("llamacpp", None) == "http://localhost:8080/v1"assert local_server.v1_base("llamacpp", "http://box.local:9000/") == "http://box.local:9000/v1"assert local_server.v1_base("vllm", "http://box.local:8000/v1") == "http://box.local:8000/v1"+ assert local_server.v1_base("lmstudio", None) == "http://localhost:1234/v1"+ assert local_server.v1_base("lmstudio", "http://127.0.0.1:1234/v1/") == "http://127.0.0.1:1234/v1"def test_the_builder_sends_a_placeholder_key_when_none_is_set():@@ -52,6 +54,8 @@ def test_the_builder_sends_a_placeholder_key_when_none_is_set():assert client._api_key == "none"client = get_descriptor("vllm").build({"base_url": "http://gpu:8000", "api_key": "s3"}, None)assert client._base_url == "http://gpu:8000/v1" and client._api_key == "s3"+ client = get_descriptor("lmstudio").build({}, None)+ assert client._base_url == "http://localhost:1234/v1" and client._api_key == "none"def test_llamacpp_facts_come_from_models_and_props(monkeypatch):@@ -77,6 +81,49 @@ def test_llamacpp_facts_come_from_models_and_props(monkeypatch):assert local_server.context_window_for("ollama:x") is None+def test_lmstudio_facts_add_the_window_from_its_own_listing(monkeypatch):+ """LM Studio's /v1/models names every downloaded model; its /api/v0/models carries+ the trained window and the load state, which the row reads when they answer."""+ def handler(request: httpx.Request) -> httpx.Response:+ if request.url.path == "/v1/models":+ return httpx.Response(200, json={"data": [+ {"id": "qwen/qwen3-coder-30b", "object": "model", "owned_by": "organization_owner"},+ {"id": "text-embedding-nomic-embed-text-v1.5", "object": "model", "owned_by": "organization_owner"},+ ]})+ if request.url.path == "/api/v0/models":+ return httpx.Response(200, json={"data": [+ {"id": "qwen/qwen3-coder-30b", "type": "llm", "state": "loaded",+ "max_context_length": 262144, "loaded_context_length": 32768},+ {"id": "text-embedding-nomic-embed-text-v1.5", "type": "embeddings", "state": "not-loaded",+ "max_context_length": 2048},+ ]})+ return httpx.Response(404)++ _fake_http(monkeypatch, handler)+ rows = local_server.model_facts("lmstudio", "http://localhost:1234")+ by_id = {r["name"]: r for r in rows}+ coder = by_id["qwen/qwen3-coder-30b"]+ assert coder["model"] == "lmstudio:qwen/qwen3-coder-30b"+ assert coder["context_max"] == 262144 and coder["context"] == 32768+ assert coder["context_from"] == "server" and coder["tools"] is None+ # Not loaded yet: only the trained window is known, and it stands in for the window.+ embed = by_id["text-embedding-nomic-embed-text-v1.5"]+ assert embed["context_max"] == embed["context"] == 2048+ assert local_server.context_window_for("lmstudio:qwen/qwen3-coder-30b") == 32768+++def test_lmstudio_facts_survive_a_missing_own_listing(monkeypatch):+ def handler(request: httpx.Request) -> httpx.Response:+ if request.url.path == "/v1/models":+ return httpx.Response(200, json={"data": [{"id": "qwen/qwen3-coder-30b", "object": "model"}]})+ return httpx.Response(404)++ _fake_http(monkeypatch, handler)+ (row,) = local_server.model_facts("lmstudio", "http://localhost:1234")+ assert row["model"] == "lmstudio:qwen/qwen3-coder-30b"+ assert row["context"] is None and row["context_max"] is None++def test_vllm_facts_probe_whether_tool_calling_is_on(monkeypatch):calls = []@@ -119,7 +166,7 @@ def test_thinking_switch_reaches_llama_cpp_and_vllm_as_template_kwargs(monkeypatsent = {}monkeypatch.setattr(openai_provider.OpenAIProvider, "complete", lambda self, **kw: sent.update(kw) or "ok")- for name in ("llamacpp", "vllm"):+ for name in ("llamacpp", "vllm", "lmstudio"):client = get_descriptor(name).build({}, None)client.complete(model="m", messages=[], max_tokens=10, extra_body={"think": False})assert sent["extra_body"] == {"chat_template_kwargs": {"enable_thinking": False}}diff --git a/tests/test_matrix_current_models.py b/tests/test_matrix_current_models.pyindex ea2516da..a7747686 100644--- a/tests/test_matrix_current_models.py+++ b/tests/test_matrix_current_models.py@@ -82,11 +82,12 @@ _KNOWN_UNVERIFIED_RECOMMENDATIONS = {"minimax:MiniMax-M2.5", "meta:muse-spark-1.def test_every_recommended_model_is_in_the_matrix():"""The Settings pane suggests each provider's `recommended_model`; every one must be a- curated row (Ollama excepted: a user-local model name, not a vendor catalog id)."""+ curated row (local servers excepted: Ollama and LM Studio recommend a model the user+ downloads, named as that app names it, not a vendor catalog id)."""missing = [_matrix_key(d.name, d.recommended_model)for d in DESCRIPTORS- if d.name != "ollama" and d.recommended_model+ if d.kind != "local" and d.recommended_modeland _matrix_key(d.name, d.recommended_model) not in MATRIX]assert not missing, f"recommended models without a matrix row: {missing}"@@ -99,7 +100,7 @@ def test_recommended_models_without_a_window_are_exactly_the_known_set():without = {_matrix_key(d.name, d.recommended_model)for d in DESCRIPTORS- if d.name != "ollama" and d.recommended_model+ if d.kind != "local" and d.recommended_modeland _matrix_key(d.name, d.recommended_model) not in windows}assert without == _KNOWN_UNVERIFIED_RECOMMENDATIONS