fix: custom VLM API compatibility with vLLM reasoning models

[lonrencn]

Problem

Custom VLM API (OpenAI-compatible) has three bugs that prevent it from working with vLLM-hosted reasoning models like Qwen3.6:

1. Empty response from reasoning models

vLLM with --reasoning-parser qwen3 returns message.content = null and puts the actual text in message.reasoning. The extractor only checked content, so the result was always empty — causing SuperPrompt to silently fall back to the original prompt.

Additionally, reasoning models spend all max_tokens on thinking and never produce content. Adding chat_template_kwargs: {"enable_thinking": false} to the request makes the model respond directly in content.

2. VLM chat panel sends wrong version

readSelectedVlmVersion() reads the Gradio Dropdown's display label (the custom model name, e.g. "Qwen3.6-35B-AWQ-4b") instead of its internal value ("Custom") when the component has no native <input> element. The backend receives an unrecognized version string and falls back to the default local model.

Fix: If the version string matches the custom model name, treat it as Custom mode.

3. VLM selection resets on page refresh

"Custom" was explicitly excluded from admin default persistence (persist_admin and version != VLM.CUSTOM_VERSION). On page load, init_nav_bars restores the admin default (a local model) and its .change() event cascades to override the user-facing dropdown — even though load_main_vlm_user_settings had correctly restored "Custom" from local_settings.json.

Fix: Remove the exclusion so Custom persists as the admin default.

Changes (3 files, +14/-3)

File	Change
enhanced/vlm.py	enable_thinking:false in request; reasoning/reasoning_content fallback in extractor
javascript/describe_vlm_chat.js	Model-name heuristic in readSelectedVlmVersion()
webui.py	Remove Custom exclusion from admin persistence

Tested with

• vLLM 0.x + Qwen3.6-35B-AWQ-4b (--reasoning-parser qwen3)
• SuperPrompt, VLM chat, and page refresh persistence all working