On macOS, an app without Local Network permission gets "[Errno 65] No route
to host" for every host on the local network, surfaced by the OpenAI client as
a bare "Connection error." Some launchers (Comfy Desktop 1.1.3) never request
the permission, so it cannot be granted in System Settings at all.
- network.py: detect the denial (darwin + EHOSTUNREACH anywhere in the
exception chain + a LAN target) and raise LocalNetworkBlockedError with a
fix built from the user's own server_url. Wired into Connect, the
/lmstudio/test and /lmstudio/models routes, Text Gen and Image To Text.
Other errors and platforms are untouched.
- scripts/lan_forward.py: stdlib-only relay from 127.0.0.1 (loopback is
exempt) to the LM Studio host. --target is required; no built-in address.
Probes the target on startup and explains failures.
- README: troubleshooting section. launchd agents are blocked the same way,
so the start-at-login recipe is a .command file opened by Terminal.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Replace the binary "Enable Reasoning" toggle with a tri-state
Thinking / Reasoning control (auto / on / off) that drives LM Studio's
template-level reasoning switch via chat_template_kwargs.enable_thinking
(nested — a top-level field never reaches the Jinja template):
- off -> suppress: enable_thinking:false + a closed <think></think>
assistant prefill, so a reasoning model answers directly instead
of burning its whole budget on a hidden think trace.
- on -> force: enable_thinking:true.
- auto -> force for families that stay silent by default (Gemma), otherwise
send nothing (Qwen3-style templates already think by default).
Applied to both the responses and chat.completions paths in the text and
image nodes, so reasoning is controlled consistently regardless of which
endpoint the server supports.
Fix the duplicate "Model" field: the web extension used to inject a second
combo widget and hide the backend string widget (unreliably), leaving two
model fields on screen. The model input is now a native io.Combo populated
in place by Refresh Models; `model` is listed in validate_inputs so ComfyUI
skips its "value not in list" check (the real options are only known at
runtime).
Rework the Connect node UI: essentials first (Server URL -> Model ->
Thinking), secondary settings moved to advanced, buttons show progress
labels and toast feedback instead of a blocking alert, and the status line
reports the resolved thinking mode.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
When a reasoning model returns a response consisting solely of
<think>...</think> content, the emptiness guard fired before
strip_think_content ran, so the node silently returned an empty
string rather than signalling an error. Add a post-strip check
in both LMStudioTextGen and LMStudioImageToText that raises
ValueError if strip_think_content reduces the text to nothing.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
