EmAySee Deep Reasoning Connector

This is a big one hiding in a kitchen-sink pack: a fully wired LLM connector that talks to any OpenAI-compatible /v1/chat/completions endpoint, with one feature you barely see elsewhere - it splits the model's chain-of-thought out of the answer and hands you both. Point it at a local oobabooga text-generation-webui or llama.cpp server running a reasoning model (Qwen3, DeepSeek-R1 distills, anything that emits <think> blocks) and you get a thinking output and an answer output. No API key required for local use, no account, no rate limits. Just a POST to your own machine.

The name is accurate in a way the pack's README isn't: this one targets reasoning models specifically. That's what the reasoning_effort knob and the think-tag parsing are about.

How it works

The node builds a standard chat payload - system_prompt + prompt, with enable_thinking: True and a pile of text-generation-webui sampler knobs (min_p, top_k, repetition_penalty_range, xtc_*, dry_*) - and POSTs it non-streaming. Two details make it genuinely useful:

• Auto-continue. If the response's finish_reason is length/maxlen (the model hit max_tokens_per_call mid-thought), it appends the partial reply as an assistant message and re-posts, up to max_continues times. That's how a 4096-token default gets a 32k-token response out of a local model without you babysitting it.
• Think-tag parsing. After the loop it regexes the full text for <think>/<thought> blocks. Everything inside becomes the thinking output; everything outside becomes answer. If the model opens a think tag and never closes it, you get [REASONING DID NOT FINISH] in the answer - a loud, honest failure mode instead of a silently truncated reply.

The inputs you actually set

Most of the 20+ inputs are sampler values you can leave alone. The ones that matter:

• api_url - the default is http://10.0.0.71:5000/v1/chat/completions, which is the author's own LAN box. You must change this to your text-generation-webui or llama.cpp server's address. This is the #1 footgun.
• api_key - for local textgen, any string works (it's ignored); for a real remote service, use the real key.
• model_name - "default" is fine for textgen; set it explicitly for llama.cpp/Ollama.
• auto_continue + max_continues - leave on unless you want short, fast replies.
• reasoning_effort - low/medium/high, passed straight through; low is the sensible default for prompt-tagging work.

Outputs: answer (clean reply), thinking (the chain-of-thought), full_raw (everything, for debugging). Wire answer into a text display or back into prompt-building; wire thinking to its own viewer so you can watch the model talk to itself.

Install

Part of ComfyUI_EmAySee_CustomNodes (Manager → search "EmAySee", or git clone https://github.com/EmAySee/ComfyUI_EmAySee_CustomNodes into ComfyUI/custom_nodes, then restart). It needs requests, which ComfyUI already bundles - no extra pip install, no model files.

Gotchas and the honest take

Non-streaming means a long reasoning run blocks the entire ComfyUI queue for minutes (there's a 600-second timeout, so it won't hang forever). The dry_*/xtc_* fields are textgen-specific sampler options; llama.cpp ignores unknown keys, textgen honors them. And remember the LLMVISION lesson from this ecosystem: custom nodes run unsandboxed, and this one sends your prompts to whatever URL you configure. A local endpoint keeps everything on your machine. Pointing it at a random public API means your prompts go there too - be deliberate. For a seedable, reproducible version of this exact node, grab EmAySee_DeepReasoningConnectorWithSeed.