fix(#3750): preserve LM Studio auth on reasoning probe (self-rebased #3837 + credential hardening) (#4202)

* fix(#3750): preserve LM Studio auth on reasoning probe

Rebased rodboev's #3837 onto current master (master reworked
resolve_model_reasoning_efforts via #3431 with a combined hermes_cli
import + _nested_route_reasoning_denied guard). Threaded the api_key
resolution + the _lmstudio_model_reasoning_options wrapper + the
HTTP-fallback probe into master's current structure; intent preserved
byte-identically (+164/-24, same as the original PR diff).

Co-authored-by: Rod Boev <rod.boev@gmail.com>

* fix(#3837): gate LM Studio probe credential to configured endpoint + no-redirect opener

Codex security review (CORE) on the rebased #3837: GET /api/reasoning accepts
a caller-supplied base_url, and the new probe attached the configured LM Studio
API key as a Bearer token; urllib also re-sends Authorization across redirects.
Together that could exfiltrate the stored credential to an attacker-controlled
host. (master only escalates here because master probed WITHOUT a key.)

Fixes:
1. Only forward the LM Studio credential when the probe target normalizes to
   the configured LM Studio base URL (reuses _normalize_base_url_for_match);
   a caller-supplied non-matching base_url is probed keyless.
2. _lmstudio_reasoning_probe_options_fallback uses a no-redirect opener
   (_NoRedirectHandler) so a 3xx can't forward the Authorization header onward.
3. Two regression tests: caller-supplied base_url drops the key; redirect is
   not followed (credential never reaches the redirect target).

Co-authored-by: Rod Boev <rod.boev@gmail.com>

* fix(#3837): route ALL credentialed LM Studio probes through the no-redirect probe

Codex re-gate caught that the first fix only hardened the built-in fallback;
in production hermes_cli IS importable, so the preferred path called
hermes_cli.models.lmstudio_model_reasoning_options(..., api_key=...) which uses
plain urlopen and still follows redirects with the credential.

Fix: _lmstudio_model_reasoning_options now short-circuits to the built-in
no-redirect probe whenever api_key is set; the hermes_cli path is used ONLY for
keyless probes (no credential to leak). Reworked tests accordingly:
- test_credentialed_probe_never_calls_hermes_cli: credentialed probe bypasses CLI
- test_keyless_probe_logs_signature_mismatch_before_fallback: keyless CLI degrade
- test_credentialed_probe_does_not_follow_redirects_end_to_end: E2E through
  resolve_model_reasoning_efforts with real hermes_cli; redirect target never hit

Verified both security tests fail without the guard and pass with it.

Co-authored-by: Rod Boev <rod.boev@gmail.com>

* chore(release): stamp v0.51.420 (Release OG) in CHANGELOG

---------

Co-authored-by: nesquena-hermes <agent@nesquena-hermes>
Co-authored-by: Rod Boev <rod.boev@gmail.com>

CHANGELOG.md

@@ -3,6 +3,12 @@
 
 ## [Unreleased]
 
+## [v0.51.420] — 2026-06-14 — Release OG (LM Studio reasoning-probe auth, #3750/#3837)
+
+### Fixed
+
+- **LM Studio reasoning-effort probes now keep their auth and work without the bundled CLI.** When the active model is an LM Studio provider, the WebUI reasoning-effort probe now resolves an LM Studio API key (active model key → `providers.lmstudio.api_key` → `LM_API_KEY` → `LMSTUDIO_API_KEY`) and forwards it as a Bearer token, so authenticated LM Studio servers no longer 401 the probe and silently report "no reasoning support." If `hermes_cli` is unavailable (or its `lmstudio_model_reasoning_options` signature changes), the probe falls back to a built-in HTTP query of the LM Studio `/api/v1/models` capabilities endpoint instead of skipping reasoning detection entirely. The configured LM Studio credential is only ever sent to the configured LM Studio endpoint (a caller-supplied `base_url` that does not match is probed without the key), and the probe refuses HTTP redirects so the credential can never be forwarded to another host. (#3750, #3837)
+
 ## [v0.51.419] — 2026-06-14 — Release OF (settled-stream message-count consistency, #4192)
 
 ### Fixed

api/config.py

@@ -21,6 +21,8 @@ import sys
 import threading
 import time
 import traceback
+import urllib.error
+import urllib.request
 import uuid
 from pathlib import Path
 from urllib.parse import parse_qs, urlparse
@@ -2623,6 +2625,184 @@ def _models_dev_reasoning_efforts(model_id: str, provider_id: str) -> list[str]
     return None
 
 
+class _NoRedirectHandler(urllib.request.HTTPRedirectHandler):
+    """urllib redirect handler that refuses to follow any redirect.
+
+    Used by the LM Studio reasoning probe so a 3xx from the probe URL can never
+    forward the ``Authorization`` header (the configured LM Studio key) to a
+    redirected, possibly attacker-controlled host. ``redirect_request``
+    returning ``None`` makes urllib raise the original 3xx as an ``HTTPError``,
+    which the probe swallows. (#3837 security review)
+    """
+
+    def redirect_request(self, req, fp, code, msg, headers, newurl):
+        return None
+
+
+def _get_lmstudio_reasoning_probe_api_key() -> str | None:
+    """Resolve the LM Studio key for reasoning probes with WebUI precedence."""
+    config_data = cfg
+    model_cfg = config_data.get("model") or {}
+    if isinstance(model_cfg, dict):
+        active_provider = str(model_cfg.get("provider") or "").strip().lower()
+        model_key = str(model_cfg.get("api_key") or "").strip()
+        if active_provider == "lmstudio" and model_key:
+            return model_key
+
+    providers_cfg = config_data.get("providers") or {}
+    if isinstance(providers_cfg, dict):
+        lmstudio_cfg = providers_cfg.get("lmstudio") or {}
+        if isinstance(lmstudio_cfg, dict):
+            config_key = str(lmstudio_cfg.get("api_key") or "").strip()
+            if config_key:
+                return config_key
+
+    env_key = str(os.getenv("LM_API_KEY") or "").strip()
+    if env_key:
+        return env_key
+
+    legacy_env_key = str(os.getenv("LMSTUDIO_API_KEY") or "").strip()
+    if legacy_env_key:
+        return legacy_env_key
+
+    return None
+
+
+def _lmstudio_reasoning_probe_options_fallback(
+    model: str,
+    base_url: str | None,
+    *,
+    api_key: str | None = None,
+    timeout: float = 5.0,
+) -> list[str]:
+    """Query LM Studio reasoning options without relying on hermes_cli."""
+    server_root = str(base_url or "").strip().rstrip("/")
+    if server_root.endswith("/v1"):
+        server_root = server_root[:-3].rstrip("/")
+    if not server_root or not model:
+        return []
+
+    headers = {
+        "Accept": "application/json",
+        "User-Agent": "hermes-webui-reasoning-probe",
+    }
+    if api_key:
+        headers["Authorization"] = f"Bearer {api_key}"
+
+    request = urllib.request.Request(
+        server_root + "/api/v1/models",
+        headers=headers,
+        method="GET",
+    )
+    # SECURITY: never follow redirects on this probe. urllib re-sends request
+    # headers (including Authorization: Bearer <lmstudio key>) to the redirect
+    # target, so a 3xx from the probe URL could exfiltrate the configured
+    # LM Studio credential to an attacker-controlled host. A no-redirect opener
+    # turns any 3xx into an HTTPError we swallow below. (#3837 security review)
+    opener = urllib.request.build_opener(_NoRedirectHandler)
+    try:
+        with opener.open(request, timeout=timeout) as response:  # nosec B310
+            payload = json.loads(response.read().decode("utf-8", errors="replace"))
+    except urllib.error.HTTPError as exc:
+        logger.debug(
+            "LM Studio reasoning probe at %s failed with HTTP %s",
+            server_root,
+            exc.code,
+        )
+        return []
+    except Exception as exc:
+        logger.debug("LM Studio reasoning probe at %s failed: %s", server_root, exc)
+        return []
+
+    raw_models = payload.get("models") if isinstance(payload, dict) else None
+    if not isinstance(raw_models, list):
+        logger.debug(
+            "LM Studio reasoning probe at %s returned malformed payload",
+            server_root,
+        )
+        return []
+
+    for raw in raw_models:
+        if not isinstance(raw, dict):
+            continue
+        if raw.get("key") != model and raw.get("id") != model:
+            continue
+        caps = raw.get("capabilities")
+        reasoning = caps.get("reasoning") if isinstance(caps, dict) else None
+        opts = reasoning.get("allowed_options") if isinstance(reasoning, dict) else None
+        if isinstance(opts, list):
+            return [str(opt).strip().lower() for opt in opts if isinstance(opt, str)]
+        return []
+    return []
+
+
+def _lmstudio_model_reasoning_options(
+    model: str,
+    base_url: str | None,
+    *,
+    api_key: str | None = None,
+    timeout: float = 5.0,
+) -> list[str]:
+    """Prefer hermes_cli, but keep WebUI reasoning probes working without it.
+
+    SECURITY: when an ``api_key`` is being sent, always use the built-in
+    no-redirect fallback probe rather than ``hermes_cli``. The bundled CLI probe
+    uses a plain ``urllib.request.urlopen`` that follows redirects and re-sends
+    the ``Authorization`` header to the redirect target, which could leak the
+    configured LM Studio credential to another host. We can only guarantee
+    redirect safety for code we control, so credentialed probes go through
+    ``_lmstudio_reasoning_probe_options_fallback`` (no-redirect opener). Keyless
+    probes have no credential to leak, so they may use the richer CLI path.
+    (#3837 security review)
+    """
+    if api_key:
+        return _lmstudio_reasoning_probe_options_fallback(
+            model,
+            base_url,
+            api_key=api_key,
+            timeout=timeout,
+        )
+
+    try:
+        from hermes_cli.models import (
+            lmstudio_model_reasoning_options as _cli_lmstudio_model_reasoning_options,
+        )
+    except Exception:
+        return _lmstudio_reasoning_probe_options_fallback(
+            model,
+            base_url,
+            api_key=api_key,
+            timeout=timeout,
+        )
+
+    try:
+        return _cli_lmstudio_model_reasoning_options(
+            model,
+            base_url,
+            api_key=api_key,
+            timeout=timeout,
+        )
+    except (TypeError, AttributeError):
+        logger.warning(
+            "hermes_cli.lmstudio_model_reasoning_options has an unexpected signature; "
+            "falling back to the built-in LM Studio reasoning probe",
+            exc_info=True,
+        )
+        return _lmstudio_reasoning_probe_options_fallback(
+            model,
+            base_url,
+            api_key=api_key,
+            timeout=timeout,
+        )
+    except Exception:
+        return _lmstudio_reasoning_probe_options_fallback(
+            model,
+            base_url,
+            api_key=api_key,
+            timeout=timeout,
+        )
+
+
 def resolve_model_reasoning_efforts(
     model_id: str | None = None,
     provider_id: str | None = None,
@@ -2650,34 +2830,47 @@ def resolve_model_reasoning_efforts(
     if _nested_route_reasoning_denied(hinted_model):
         return []
 
-    try:
-        from hermes_cli.models import (
-            github_model_reasoning_efforts,
-            lmstudio_model_reasoning_options,
-        )
-    except Exception:
-        if provider in {"copilot", "github-copilot"}:
+    if provider in {"copilot", "github-copilot"}:
+        try:
+            from hermes_cli.models import github_model_reasoning_efforts
+        except Exception:
             return _heuristic_reasoning_efforts(hinted_model, provider)
-    else:
-        if provider in {"copilot", "github-copilot"}:
-            return _filter_reasoning_efforts_for_provider(
-                github_model_reasoning_efforts(hinted_model), hinted_model, provider
-            )
+        return _filter_reasoning_efforts_for_provider(
+            github_model_reasoning_efforts(hinted_model), hinted_model, provider
+        )
 
-        if provider == "lmstudio":
-            probe_base = resolved_base_url or _get_provider_base_url(provider)
-            opts = lmstudio_model_reasoning_options(hinted_model, probe_base)
-            normalized = [str(opt).strip().lower() for opt in opts if str(opt).strip()]
-            if not normalized or set(normalized).issubset({"off"}):
-                return []
-            level_opts = [opt for opt in normalized if opt in VALID_REASONING_EFFORTS]
-            if level_opts:
-                return _filter_reasoning_efforts_for_provider(
-                    level_opts, hinted_model, provider
-                )
-            if set(normalized).issubset({"off", "on"}):
-                return []
+    if provider == "lmstudio":
+        configured_base = _get_provider_base_url(provider)
+        probe_base = resolved_base_url or configured_base
+        # SECURITY: only forward the configured LM Studio credential when the
+        # probe target is the configured LM Studio endpoint. /api/reasoning
+        # accepts a caller-supplied base_url, so a request could otherwise point
+        # the probe at an arbitrary host and harvest the stored key. When the
+        # caller supplies a base_url that does not normalize to the configured
+        # one, probe it WITHOUT a key. (#3837 security review)
+        probe_key: str | None = None
+        if not resolved_base_url or (
+            configured_base
+            and _normalize_base_url_for_match(probe_base)
+            == _normalize_base_url_for_match(configured_base)
+        ):
+            probe_key = _get_lmstudio_reasoning_probe_api_key()
+        opts = _lmstudio_model_reasoning_options(
+            hinted_model,
+            probe_base,
+            api_key=probe_key,
+        )
+        normalized = [str(opt).strip().lower() for opt in opts if str(opt).strip()]
+        if not normalized or set(normalized).issubset({"off"}):
             return []
+        level_opts = [opt for opt in normalized if opt in VALID_REASONING_EFFORTS]
+        if level_opts:
+            return _filter_reasoning_efforts_for_provider(
+                level_opts, hinted_model, provider
+            )
+        if set(normalized).issubset({"off", "on"}):
+            return []
+        return []
 
     # _models_dev_reasoning_efforts already applies the provider/model filter
     # internally, so it is returned as-is here (filtering again would be

tests/test_issue3750_lmstudio_probe_auth.py

@@ -0,0 +1,626 @@
+"""Regression tests for #3750 LM Studio reasoning probes dropping auth."""
+
+from __future__ import annotations
+
+import json
+import sys
+import threading
+from http.server import BaseHTTPRequestHandler, HTTPServer
+
+import pytest
+
+import api.config as config
+import api.onboarding as onboarding
+import api.profiles as profiles
+
+
+_API_KEY_ENV_VARS = (
+    "LM_API_KEY",
+    "LMSTUDIO_API_KEY",
+)
+
+
+class _LmStudioProbeServer:
+    def __init__(self):
+        self.requests: list[dict[str, str | None]] = []
+        self._server = HTTPServer(("127.0.0.1", 0), self._handler())
+        self._thread = threading.Thread(target=self._server.serve_forever, daemon=True)
+        self._thread.start()
+
+    @property
+    def base_v1(self) -> str:
+        return f"http://127.0.0.1:{self._server.server_address[1]}/v1"
+
+    def close(self) -> None:
+        self._server.shutdown()
+        self._server.server_close()
+        self._thread.join(timeout=5)
+
+    def _handler(self):
+        parent = self
+
+        class Handler(BaseHTTPRequestHandler):
+            def do_GET(self):
+                auth = self.headers.get("Authorization")
+                parent.requests.append(
+                    {
+                        "path": self.path,
+                        "authorization": auth,
+                    }
+                )
+
+                if self.path == "/api/v1/models":
+                    if auth:
+                        self.send_response(200)
+                        self.send_header("Content-Type", "application/json")
+                        self.end_headers()
+                        self.wfile.write(
+                            json.dumps(
+                                {
+                                    "models": [
+                                        {
+                                            "key": "auth-model",
+                                            "capabilities": {
+                                                "reasoning": {
+                                                    "allowed_options": ["low", "medium", "high"]
+                                                }
+                                            },
+                                        }
+                                    ]
+                                }
+                            ).encode()
+                        )
+                        return
+
+                    self.send_response(401)
+                    self.send_header("Content-Type", "application/json")
+                    self.end_headers()
+                    self.wfile.write(json.dumps({"error": "unauthorized"}).encode())
+                    return
+
+                if self.path == "/v1/models":
+                    if auth:
+                        self.send_response(200)
+                        self.send_header("Content-Type", "application/json")
+                        self.end_headers()
+                        self.wfile.write(json.dumps({"data": [{"id": "auth-model"}]}).encode())
+                        return
+
+                    self.send_response(401)
+                    self.send_header("Content-Type", "application/json")
+                    self.end_headers()
+                    self.wfile.write(
+                        json.dumps({"error": {"message": "unauthorized"}}).encode()
+                    )
+                    return
+
+                self.send_response(404)
+                self.end_headers()
+
+            def log_message(self, fmt, *args):
+                return None
+
+        return Handler
+
+
+@pytest.fixture(autouse=True)
+def _isolate_config(monkeypatch, tmp_path):
+    old_cfg = dict(config.cfg)
+    old_mtime = config._cfg_mtime
+    old_path = config._cfg_path
+    old_fp = config._cfg_fingerprint
+    monkeypatch.setattr(profiles, "get_active_hermes_home", lambda: tmp_path)
+    monkeypatch.setenv("BROWSER", "echo")
+    for var in _API_KEY_ENV_VARS:
+        monkeypatch.delenv(var, raising=False)
+    config.invalidate_models_cache()
+    yield
+    config.cfg.clear()
+    config.cfg.update(old_cfg)
+    config._cfg_mtime = old_mtime
+    config._cfg_path = old_path
+    config._cfg_fingerprint = old_fp
+    config.invalidate_models_cache()
+
+
+@pytest.fixture
+def lmstudio_probe_server():
+    server = _LmStudioProbeServer()
+    try:
+        yield server
+    finally:
+        server.close()
+
+
+def _write_config(tmp_path, monkeypatch, text: str) -> None:
+    cfgfile = tmp_path / "config.yaml"
+    cfgfile.write_text(text, encoding="utf-8")
+    monkeypatch.setattr(config, "_get_config_path", lambda: cfgfile)
+    config.reload_config()
+    config.invalidate_models_cache()
+
+
+def test_reasoning_probe_uses_config_key_before_env_aliases(
+    tmp_path,
+    monkeypatch,
+    lmstudio_probe_server,
+):
+    _write_config(
+        tmp_path,
+        monkeypatch,
+        f"""
+model:
+  provider: lmstudio
+  default: auth-model
+  base_url: {lmstudio_probe_server.base_v1}
+providers:
+  lmstudio:
+    api_key: config-token
+agent:
+  reasoning_effort: medium
+display:
+  show_reasoning: true
+""",
+    )
+    monkeypatch.setenv("LM_API_KEY", "env-token")
+    monkeypatch.setenv("LMSTUDIO_API_KEY", "legacy-token")
+
+    status = config.get_reasoning_status()
+
+    assert status["supports_reasoning_effort"] is True
+    assert status["supported_efforts"] == ["low", "medium", "high"]
+    assert lmstudio_probe_server.requests[0] == {
+        "path": "/api/v1/models",
+        "authorization": "Bearer config-token",
+    }
+
+
+def test_reasoning_probe_honors_active_model_api_key(
+    tmp_path,
+    monkeypatch,
+    lmstudio_probe_server,
+):
+    _write_config(
+        tmp_path,
+        monkeypatch,
+        f"""
+model:
+  provider: lmstudio
+  default: auth-model
+  base_url: {lmstudio_probe_server.base_v1}
+  api_key: model-token
+providers:
+  lmstudio:
+    api_key: provider-token
+agent:
+  reasoning_effort: medium
+display:
+  show_reasoning: true
+""",
+    )
+
+    status = config.get_reasoning_status()
+
+    assert status["supports_reasoning_effort"] is True
+    assert status["supported_efforts"] == ["low", "medium", "high"]
+    assert lmstudio_probe_server.requests[0] == {
+        "path": "/api/v1/models",
+        "authorization": "Bearer model-token",
+    }
+
+
+def test_reasoning_probe_falls_back_without_hermes_cli(
+    tmp_path,
+    monkeypatch,
+    lmstudio_probe_server,
+):
+    blocked_modules = [
+        name
+        for name in list(sys.modules)
+        if name == "hermes_cli" or name.startswith("hermes_cli.")
+    ]
+    for name in blocked_modules:
+        monkeypatch.delitem(sys.modules, name, raising=False)
+
+    class _Blocker:
+        def find_module(self, name, path=None):
+            if name == "hermes_cli" or name.startswith("hermes_cli."):
+                return self
+            return None
+
+        def load_module(self, name):
+            raise ImportError(f"hermes_cli blocked for test: {name}")
+
+    blocker = _Blocker()
+    sys.meta_path.insert(0, blocker)
+    try:
+        _write_config(
+            tmp_path,
+            monkeypatch,
+            f"""
+model:
+  provider: lmstudio
+  default: auth-model
+  base_url: {lmstudio_probe_server.base_v1}
+providers:
+  lmstudio:
+    api_key: config-token
+agent:
+  reasoning_effort: medium
+display:
+  show_reasoning: true
+""",
+        )
+
+        status = config.get_reasoning_status()
+
+        assert status["supports_reasoning_effort"] is True
+        assert status["supported_efforts"] == ["low", "medium", "high"]
+        assert lmstudio_probe_server.requests[0] == {
+            "path": "/api/v1/models",
+            "authorization": "Bearer config-token",
+        }
+    finally:
+        try:
+            sys.meta_path.remove(blocker)
+        except ValueError:
+            pass
+
+
+def test_reasoning_probe_stays_keyless_when_no_key_is_configured(
+    tmp_path,
+    monkeypatch,
+    lmstudio_probe_server,
+):
+    _write_config(
+        tmp_path,
+        monkeypatch,
+        f"""
+model:
+  provider: lmstudio
+  default: auth-model
+  base_url: {lmstudio_probe_server.base_v1}
+agent:
+  reasoning_effort: medium
+display:
+  show_reasoning: true
+""",
+    )
+
+    status = config.get_reasoning_status()
+
+    assert status["supports_reasoning_effort"] is False
+    assert status["supported_efforts"] == []
+    assert lmstudio_probe_server.requests[0] == {
+        "path": "/api/v1/models",
+        "authorization": None,
+    }
+
+
+def test_onboarding_probe_remains_authorized_control(
+    lmstudio_probe_server,
+):
+    result = onboarding.probe_provider_endpoint(
+        "lmstudio",
+        lmstudio_probe_server.base_v1,
+        api_key="control-token",
+    )
+
+    assert result["ok"] is True
+    assert lmstudio_probe_server.requests[0] == {
+        "path": "/v1/models",
+        "authorization": "Bearer control-token",
+    }
+
+
+def test_credentialed_probe_never_calls_hermes_cli(
+    tmp_path,
+    monkeypatch,
+    lmstudio_probe_server,
+):
+    """When a credential is configured, the probe must use the built-in
+    no-redirect fallback and NEVER the bundled hermes_cli probe (which follows
+    redirects and could forward the key). (#3837 security review)
+    """
+    _write_config(
+        tmp_path,
+        monkeypatch,
+        f"""
+model:
+  provider: lmstudio
+  default: auth-model
+  base_url: {lmstudio_probe_server.base_v1}
+providers:
+  lmstudio:
+    api_key: config-token
+agent:
+  reasoning_effort: medium
+display:
+  show_reasoning: true
+""",
+    )
+
+    cli_calls: list[tuple] = []
+
+    class _CliModule:
+        @staticmethod
+        def lmstudio_model_reasoning_options(model, base_url, api_key=None, timeout=5.0):
+            cli_calls.append((model, base_url, api_key))
+            return ["low", "medium", "high"]
+
+    monkeypatch.setitem(sys.modules, "hermes_cli", type("HermesCli", (), {})())
+    monkeypatch.setitem(sys.modules, "hermes_cli.models", _CliModule)
+
+    status = config.get_reasoning_status()
+
+    assert status["supports_reasoning_effort"] is True
+    assert status["supported_efforts"] == ["low", "medium", "high"]
+    # The credential reached the configured endpoint via the built-in probe …
+    assert lmstudio_probe_server.requests[0] == {
+        "path": "/api/v1/models",
+        "authorization": "Bearer config-token",
+    }
+    # … and the redirect-following hermes_cli probe was never invoked.
+    assert cli_calls == [], "credentialed probe must bypass hermes_cli"
+
+
+def test_keyless_probe_logs_signature_mismatch_before_fallback(
+    tmp_path,
+    monkeypatch,
+    lmstudio_probe_server,
+):
+    """Keyless probes still prefer hermes_cli; an incompatible CLI signature
+    logs a warning and degrades to the built-in probe. (#3750)
+    """
+    _write_config(
+        tmp_path,
+        monkeypatch,
+        f"""
+model:
+  provider: lmstudio
+  default: auth-model
+  base_url: {lmstudio_probe_server.base_v1}
+agent:
+  reasoning_effort: medium
+display:
+  show_reasoning: true
+""",
+    )
+
+    captured: list[tuple[str, bool]] = []
+
+    class _CliModule:
+        @staticmethod
+        def lmstudio_model_reasoning_options(model, base_url):
+            raise TypeError("unexpected keyword argument: api_key")
+
+    monkeypatch.setitem(sys.modules, "hermes_cli", type("HermesCli", (), {})())
+    monkeypatch.setitem(sys.modules, "hermes_cli.models", _CliModule)
+    monkeypatch.setattr(
+        config.logger,
+        "warning",
+        lambda msg, *args, **kwargs: captured.append((str(msg), bool(kwargs.get("exc_info")))),
+    )
+
+    # No key configured -> CLI path is attempted, raises TypeError, warning is
+    # logged, then the built-in (keyless) probe runs against the auth-requiring
+    # test server, which 401s -> []. The point is the warning + degrade path.
+    status = config.get_reasoning_status()
+
+    assert status["supports_reasoning_effort"] is False
+    assert status["supported_efforts"] == []
+    assert lmstudio_probe_server.requests[-1]["authorization"] is None
+    assert captured, "TypeError fallback must log a warning before probing directly"
+    assert "unexpected signature" in captured[0][0]
+    assert captured[0][1] is True
+
+
+def test_reasoning_probe_drops_key_for_caller_supplied_base_url(
+    tmp_path,
+    monkeypatch,
+    lmstudio_probe_server,
+):
+    """A caller-supplied base_url that is NOT the configured LM Studio endpoint
+    must be probed WITHOUT the stored credential. /api/reasoning takes a
+    base_url query param, so attaching the key to an arbitrary host would leak
+    it. (#3837 security review)
+    """
+    _write_config(
+        tmp_path,
+        monkeypatch,
+        f"""
+model:
+  provider: lmstudio
+  default: auth-model
+  base_url: {lmstudio_probe_server.base_v1}
+providers:
+  lmstudio:
+    api_key: config-token
+agent:
+  reasoning_effort: medium
+display:
+  show_reasoning: true
+""",
+    )
+
+    # Point the probe at the SAME server but via an unconfigured base_url. The
+    # configured base_url is the only one allowed to carry the key. The probe
+    # server returns 401 without auth, so reasoning is reported unsupported and,
+    # crucially, no Bearer token reaches the caller-supplied URL.
+    efforts = config.resolve_model_reasoning_efforts(
+        "auth-model",
+        provider_id="lmstudio",
+        base_url=lmstudio_probe_server.base_v1.replace("127.0.0.1", "localhost"),
+    )
+
+    assert efforts == []
+    assert lmstudio_probe_server.requests, "probe should still be attempted keyless"
+    assert lmstudio_probe_server.requests[0]["authorization"] is None
+
+
+def test_reasoning_probe_does_not_follow_redirects_with_credential(
+    tmp_path,
+    monkeypatch,
+):
+    """A 3xx from the configured probe URL must NOT forward the Authorization
+    header to the redirect target. urllib re-sends request headers across
+    redirects, so the probe uses a no-redirect opener. (#3837 security review)
+    """
+    captured: list[dict[str, str | None]] = []
+
+    class _Target(BaseHTTPRequestHandler):
+        def do_GET(self):
+            captured.append(
+                {"host": "target", "authorization": self.headers.get("Authorization")}
+            )
+            self.send_response(200)
+            self.send_header("Content-Type", "application/json")
+            self.end_headers()
+            self.wfile.write(json.dumps({"models": []}).encode())
+
+        def log_message(self, fmt, *args):
+            return None
+
+    class _Redirector(BaseHTTPRequestHandler):
+        def do_GET(self):
+            captured.append(
+                {"host": "redirector", "authorization": self.headers.get("Authorization")}
+            )
+            self.send_response(302)
+            self.send_header(
+                "Location", f"http://127.0.0.1:{target.server_address[1]}/api/v1/models"
+            )
+            self.end_headers()
+
+        def log_message(self, fmt, *args):
+            return None
+
+    target = HTTPServer(("127.0.0.1", 0), _Target)
+    target_thread = threading.Thread(target=target.serve_forever, daemon=True)
+    target_thread.start()
+    redir = HTTPServer(("127.0.0.1", 0), _Redirector)
+    redir_thread = threading.Thread(target=redir.serve_forever, daemon=True)
+    redir_thread.start()
+    try:
+        result = config._lmstudio_reasoning_probe_options_fallback(
+            "auth-model",
+            f"http://127.0.0.1:{redir.server_address[1]}/v1",
+            api_key="secret-token",
+        )
+    finally:
+        redir.shutdown()
+        redir.server_close()
+        redir_thread.join(timeout=5)
+        target.shutdown()
+        target.server_close()
+        target_thread.join(timeout=5)
+
+    # The probe stops at the redirector (no-redirect opener) and never reaches
+    # the redirect target, so the credential is never forwarded onward.
+    assert result == []
+    hosts_hit = {c["host"] for c in captured}
+    assert "target" not in hosts_hit, "redirect must not be followed"
+    assert all(c["host"] == "redirector" for c in captured)
+
+
+def test_credentialed_probe_does_not_follow_redirects_end_to_end(
+    tmp_path,
+    monkeypatch,
+):
+    """End-to-end through resolve_model_reasoning_efforts() with the REAL
+    hermes_cli importable: when the CONFIGURED LM Studio endpoint returns a 302,
+    the credentialed probe must not forward the key to the redirect target.
+    This is the production path (hermes_cli present) — credentialed probes route
+    through the built-in no-redirect probe, so the redirect is refused.
+    (#3837 security review)
+    """
+    captured: list[dict[str, str | None]] = []
+
+    class _Target(BaseHTTPRequestHandler):
+        def do_GET(self):
+            captured.append(
+                {"host": "target", "authorization": self.headers.get("Authorization")}
+            )
+            self.send_response(200)
+            self.send_header("Content-Type", "application/json")
+            self.end_headers()
+            self.wfile.write(
+                json.dumps(
+                    {
+                        "models": [
+                            {
+                                "key": "auth-model",
+                                "capabilities": {
+                                    "reasoning": {"allowed_options": ["low", "high"]}
+                                },
+                            }
+                        ]
+                    }
+                ).encode()
+            )
+
+        def log_message(self, fmt, *args):
+            return None
+
+    class _Redirector(BaseHTTPRequestHandler):
+        def do_GET(self):
+            captured.append(
+                {"host": "redirector", "authorization": self.headers.get("Authorization")}
+            )
+            self.send_response(302)
+            self.send_header(
+                "Location", f"http://127.0.0.1:{target.server_address[1]}/api/v1/models"
+            )
+            self.end_headers()
+
+        def log_message(self, fmt, *args):
+            return None
+
+    target = HTTPServer(("127.0.0.1", 0), _Target)
+    target_thread = threading.Thread(target=target.serve_forever, daemon=True)
+    target_thread.start()
+    redir = HTTPServer(("127.0.0.1", 0), _Redirector)
+    redir_thread = threading.Thread(target=redir.serve_forever, daemon=True)
+    redir_thread.start()
+
+    # Sanity: this test only proves the redirect guard if hermes_cli is
+    # importable (the production path). If it isn't, the keyless/credentialed
+    # split still routes credentialed probes through the no-redirect fallback,
+    # so the assertion holds either way.
+    redir_base = f"http://127.0.0.1:{redir.server_address[1]}/v1"
+    _write_config(
+        tmp_path,
+        monkeypatch,
+        f"""
+model:
+  provider: lmstudio
+  default: auth-model
+  base_url: {redir_base}
+providers:
+  lmstudio:
+    api_key: secret-token
+agent:
+  reasoning_effort: medium
+display:
+  show_reasoning: true
+""",
+    )
+    try:
+        efforts = config.resolve_model_reasoning_efforts(
+            "auth-model",
+            provider_id="lmstudio",
+            base_url=redir_base,
+        )
+    finally:
+        redir.shutdown()
+        redir.server_close()
+        redir_thread.join(timeout=5)
+        target.shutdown()
+        target.server_close()
+        target_thread.join(timeout=5)
+
+    assert efforts == []
+    hosts_hit = {c["host"] for c in captured}
+    assert "target" not in hosts_hit, "credentialed redirect must not be followed"
+    # The credential only ever touched the configured (redirector) endpoint.
+    assert all(c["host"] == "redirector" for c in captured)