Stream thinking+answer live; restyle UI with BottleCap AI branding

Backend (app.py):
- generate() is now a generator yielding {"text": ...} as tokens stream via
TextIteratorStreamer, then a final dict with parsed thinking/answer + token
count. The reasoning trace streams live, not just the answer.

Frontend (index.html):
- BottleCap AI dark monochrome palette (logo is pure #101010 on #f0f0f0).
- BottleCap AI logo image in header, assistant avatar, and empty state.
- Live streaming render: shows the trace as it grows, then finalizes
into a collapsible Reasoning panel + answer with token count.

Co-Authored-By: Claude <noreply@anthropic.com>

app.py

@@ -18,7 +18,8 @@ On Spaces:    select ZeroGPU hardware + xlarge GPU (96 GB) in Space settings
 import os
 import re
 import sys
-from typing import Optional
+import threading
+from typing import Iterator, Optional
 
 import spaces
 import torch
@@ -27,6 +28,7 @@ from fastapi.responses import HTMLResponse
 from gradio import Server
 from gradio.data_classes import FileData
 from transformers import AutoModelForImageTextToText, AutoProcessor
+from transformers import TextIteratorStreamer
 
 MODEL_ID = os.environ.get("MODEL_ID", "bottlecapai/ThinkingCap-Qwen3.6-27B")
 
@@ -36,13 +38,15 @@ DEFAULT_TEMPERATURE = 1.0
 DEFAULT_TOP_P = 0.95
 DEFAULT_TOP_K = 20
 
-# Qwen3 reasoning-trace tags. Built by concatenation so the literal markers
-# never appear together in source.
+# Qwen3 reasoning-trace tags + EOS markers. Built by concatenation so the
+# literal markers never appear together in source.
 _THINK_OPEN = "<" + "think" + ">"
 _THINK_CLOSE = "<" + "/think" + ">"
 _THINK_RE = re.compile(
     re.escape(_THINK_OPEN) + r"(.*?)" + re.escape(_THINK_CLOSE), re.DOTALL
 )
+_IM_END = "<" + "|im_end|" + ">"
+_EOT = "<" + "|endoftext|" + ">"
 
 # ---------------------------------------------------------------------------
 # Model load (root level so ZeroGPU's CUDA emulation is active during startup;
@@ -108,8 +112,8 @@ def _split_thinking(text: str) -> tuple[str, str]:
 
 
 def _tidy(text: str) -> str:
-    """Drop trailing partial tags / leftover special tokens for a clean answer."""
-    text = text.strip()
+    """Drop leftover special tokens (eos) and an unclosed think block."""
+    text = text.replace(_IM_END, "").replace(_EOT, "").strip()
     if _THINK_OPEN in text and _THINK_CLOSE not in text:
         text = text.split(_THINK_OPEN, 1)[0].strip()
     return text
@@ -126,12 +130,13 @@ def generate(
     top_p: float = DEFAULT_TOP_P,
     top_k: int = DEFAULT_TOP_K,
     enable_thinking: bool = True,
-) -> dict:
-    """Generate one response from ThinkingCap-Qwen3.6-27B.
+) -> Iterator[dict]:
+    """Stream a response from ThinkingCap-Qwen3.6-27B.
 
-    Parameters mirror the model card's recommended sampling. Returns the raw
-    text plus parsed thinking/answer and a thinking-token count, so the
-    frontend can render a collapsible reasoning panel.
+    Yields ``{"text": <full text so far>}`` dicts as tokens are produced
+    (so the frontend can render the reasoning trace live), then a final dict
+    with parsed ``thinking`` / ``answer`` and a thinking-token count.
+    Parameters mirror the model card's recommended sampling.
     """
     history = history or []
     image_pil = _load_image(image)
@@ -165,27 +170,40 @@ def generate(
         return_tensors="pt",
     ).to(model.device)
 
-    prompt_len = inputs["input_ids"].shape[-1]
+    streamer = TextIteratorStreamer(
+        processor.tokenizer,
+        skip_prompt=True,
+        skip_special_tokens=False,
+        timeout=60.0,
+    )
 
-    with torch.inference_mode():
-        out_ids = model.generate(
-            **inputs,
-            max_new_tokens=int(max_new_tokens),
-            do_sample=float(temperature) > 0,
-            temperature=max(float(temperature), 1e-5),
-            top_p=float(top_p),
-            top_k=int(top_k),
-        )
+    gen_kwargs = dict(
+        **inputs,
+        max_new_tokens=int(max_new_tokens),
+        do_sample=float(temperature) > 0,
+        temperature=max(float(temperature), 1e-5),
+        top_p=float(top_p),
+        top_k=int(top_k),
+        streamer=streamer,
+    )
+
+    thread = threading.Thread(target=lambda: model.generate(**gen_kwargs))
+    thread.start()
+
+    accumulated = ""
+    for piece in streamer:
+        accumulated += piece
+        # Echo the full text-so-far so the client can render the live trace.
+        yield {"text": accumulated}
 
-    generated = out_ids[0, prompt_len:]
-    full = processor.tokenizer.decode(generated, skip_special_tokens=False)
-    full = _tidy(full)
+    thread.join()
 
+    full = _tidy(accumulated)
     thinking, answer = _split_thinking(full)
     thinking_tokens = len(processor.tokenizer.encode(thinking)) if thinking else 0
 
-    return {
-        "raw": full,
+    yield {
+        "text": full,
         "thinking": thinking,
         "answer": answer,
         "thinking_tokens": thinking_tokens,

index.html

@@ -6,27 +6,34 @@
 <title>ThinkingCap · Qwen 3.6 27B</title>
 <style>
   :root {
-    --hf-yellow: #FFD21E;
-    --hf-yellow-dark: #FFB400;
-    --bg: #fafafa;
-    --panel: #ffffff;
-    --ink: #1f2937;
-    --muted: #6b7280;
-    --line: #e5e7eb;
-    --user-bubble: #fff7d6;
-    --user-border: #ffe78a;
-    --assistant-bubble: #ffffff;
-    --think-bg: #f3f4f6;
-    --think-ink: #4b5563;
-    --danger: #dc2626;
+    /* BottleCap AI identity: the logo is pure black (#101010) on near-white
+       (#f0f0f0). Dark, minimal, monochrome UI. */
+    --bc-mark: #101010;
+    --bc-bg: #0e0e0f;
+    --bc-surface: #18181b;
+    --bc-surface-2: #1f1f23;
+    --bc-line: #2a2a2e;
+    --bc-ink: #f0f0f0;
+    --bc-muted: #9a9aa0;
+    --bc-faint: #6a6a70;
+    --bc-accent: #2f2f33;      /* monochrome "accent" (borders/hover) */
+    --bc-accent-strong: #e8e8e8;
+    --bc-user-bubble: #202023;
+    --bc-user-border: #2f2f33;
+    --bc-assistant-bubble: #18181b;
+    --bc-think-bg: #131316;
+    --bc-think-ink: #9a9aa0;
+    --bc-danger: #f87171;
+    --bc-ok: #86efac;
     --radius: 14px;
+    --logo: url("https://cdn-avatars.huggingface.co/v1/production/uploads/6410186e06c3b5ca88451747/suQ_-Q9WpitObx-oBv0i_.png");
   }
   * { box-sizing: border-box; }
   html, body { height: 100%; margin: 0; }
   body {
     font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, Helvetica, Arial, sans-serif;
-    background: var(--bg);
-    color: var(--ink);
+    background: var(--bc-bg);
+    color: var(--bc-ink);
     display: flex;
     flex-direction: column;
     height: 100vh;
@@ -38,56 +45,61 @@
     align-items: center;
     gap: 12px;
     padding: 12px 18px;
-    background: var(--panel);
-    border-bottom: 1px solid var(--line);
-    box-shadow: 0 1px 2px rgba(0,0,0,0.03);
+    background: var(--bc-surface);
+    border-bottom: 1px solid var(--bc-line);
     z-index: 5;
   }
   .logo {
-    width: 38px; height: 38px;
-    background: var(--hf-yellow);
+    width: 40px; height: 40px;
+    background: #f0f0f0;
     border-radius: 10px;
-    display: grid; place-items: center;
-    font-size: 22px;
+    background-image: var(--logo);
+    background-size: cover;
+    background-position: center;
+    flex-shrink: 0;
   }
   .titles { line-height: 1.2; }
-  .titles h1 { margin: 0; font-size: 16px; font-weight: 700; }
-  .titles p { margin: 2px 0 0; font-size: 12px; color: var(--muted); }
+  .titles h1 { margin: 0; font-size: 16px; font-weight: 700; color: var(--bc-ink); letter-spacing: -0.2px; }
+  .titles p { margin: 2px 0 0; font-size: 12px; color: var(--bc-muted); }
   .spacer { flex: 1; }
   .pill {
-    font-size: 11px; color: var(--muted);
-    border: 1px solid var(--line); border-radius: 999px;
-    padding: 4px 10px; background: var(--panel);
+    font-size: 11px; color: var(--bc-faint);
+    border: 1px solid var(--bc-line); border-radius: 999px;
+    padding: 4px 10px; background: var(--bc-surface-2);
   }
   .icon-btn {
-    border: 1px solid var(--line); background: var(--panel);
+    border: 1px solid var(--bc-line); background: var(--bc-surface-2);
     border-radius: 10px; width: 38px; height: 38px;
-    cursor: pointer; font-size: 18px; color: var(--ink);
+    cursor: pointer; font-size: 18px; color: var(--bc-ink);
     display: grid; place-items: center;
   }
-  .icon-btn:hover { background: #f3f4f6; }
+  .icon-btn:hover { background: var(--bc-accent); }
 
   /* Settings drawer */
   #settings {
     display: none;
-    background: var(--panel);
-    border-bottom: 1px solid var(--line);
+    background: var(--bc-surface);
+    border-bottom: 1px solid var(--bc-line);
     padding: 14px 18px;
     grid-template-columns: repeat(auto-fit, minmax(220px, 1fr));
     gap: 14px 22px;
   }
   #settings.open { display: grid; }
   .field { display: flex; flex-direction: column; gap: 6px; }
-  .field label { font-size: 12px; font-weight: 600; color: var(--muted); }
+  .field > label { font-size: 12px; font-weight: 600; color: var(--bc-muted); }
   .field .row { display: flex; align-items: center; gap: 8px; }
-  .field input[type=range] { flex: 1; }
-  .field .val { font-size: 12px; min-width: 42px; text-align: right; color: var(--ink); font-variant-numeric: tabular-nums; }
-  .field input[type=number] { width: 90px; padding: 6px 8px; border: 1px solid var(--line); border-radius: 8px; font-size: 13px; }
+  .field input[type=range] { flex: 1; accent-color: var(--bc-accent-strong); }
+  .field .val { font-size: 12px; min-width: 42px; text-align: right; color: var(--bc-ink); font-variant-numeric: tabular-nums; }
+  .field input[type=number] {
+    width: 90px; padding: 6px 8px;
+    border: 1px solid var(--bc-line); border-radius: 8px;
+    background: var(--bc-surface-2); color: var(--bc-ink); font-size: 13px;
+  }
   .switch { position: relative; width: 40px; height: 22px; }
   .switch input { opacity: 0; width: 0; height: 0; }
-  .slider { position: absolute; inset: 0; background: #cbd5e1; border-radius: 999px; transition: .2s; cursor: pointer; }
-  .slider:before { content: ""; position: absolute; width: 16px; height: 16px; left: 3px; top: 3px; background: #fff; border-radius: 50%; transition: .2s; }
-  .switch input:checked + .slider { background: var(--hf-yellow-dark); }
+  .slider { position: absolute; inset: 0; background: var(--bc-accent); border-radius: 999px; transition: .2s; cursor: pointer; }
+  .slider:before { content: ""; position: absolute; width: 16px; height: 16px; left: 3px; top: 3px; background: #f0f0f0; border-radius: 50%; transition: .2s; }
+  .switch input:checked + .slider { background: var(--bc-accent-strong); }
   .switch input:checked + .slider:before { transform: translateX(18px); }
   .reset-btn { align-self: end; }
 
@@ -97,59 +109,73 @@
 
   .msg { display: flex; gap: 10px; max-width: 100%; }
   .msg.user { flex-direction: row-reverse; }
-  .avatar { width: 30px; height: 30px; border-radius: 50%; flex-shrink: 0; display: grid; place-items: center; font-size: 15px; }
-  .msg.user .avatar { background: var(--hf-yellow); }
-  .msg.assistant .avatar { background: #eef2ff; }
-  .bubble { padding: 12px 14px; border-radius: var(--radius); border: 1px solid var(--line); max-width: 78%; word-wrap: break-word; overflow-wrap: anywhere; line-height: 1.5; font-size: 14.5px; }
-  .msg.user .bubble { background: var(--user-bubble); border-color: var(--user-border); border-top-right-radius: 4px; }
-  .msg.assistant .bubble { background: var(--assistant-bubble); border-top-left-radius: 4px; }
+  .avatar { width: 30px; height: 30px; border-radius: 50%; flex-shrink: 0; display: grid; place-items: center; font-size: 15px; overflow: hidden; }
+  .msg.user .avatar { background: var(--bc-surface-2); color: var(--bc-ink); }
+  .msg.assistant .avatar {
+    background-image: var(--logo); background-size: cover; background-position: center;
+    background-color: #f0f0f0;
+  }
+  .bubble { padding: 12px 14px; border-radius: var(--radius); border: 1px solid var(--bc-line); max-width: 78%; word-wrap: break-word; overflow-wrap: anywhere; line-height: 1.5; font-size: 14.5px; }
+  .msg.user .bubble { background: var(--bc-user-bubble); border-color: var(--bc-user-border); border-top-right-radius: 4px; }
+  .msg.assistant .bubble { background: var(--bc-assistant-bubble); border-top-left-radius: 4px; }
   .bubble img.thumb { max-width: 220px; max-height: 220px; border-radius: 8px; margin-bottom: 8px; display: block; }
 
   /* Reasoning panel */
-  .thinking { margin-bottom: 8px; border: 1px solid var(--line); border-radius: 10px; background: var(--think-bg); }
-  .thinking summary { cursor: pointer; padding: 8px 12px; font-size: 12.5px; font-weight: 600; color: var(--think-ink); display: flex; align-items: center; gap: 8px; list-style: none; }
-  .thinking summary::-webkit-details-marker { display: none; }
-  .thinking summary .chev { transition: transform .15s; }
-  .thinking[open] summary .chev { transform: rotate(90deg); }
-  .thinking .body { padding: 0 12px 10px; font-size: 13px; color: var(--think-ink); white-space: pre-wrap; line-height: 1.55; }
-  .tok-count { margin-left: auto; font-weight: 500; color: var(--hf-yellow-dark); font-variant-numeric: tabular-nums; }
+  .thinking { margin-bottom: 8px; border: 1px solid var(--bc-line); border-radius: 10px; background: var(--bc-think-bg); }
+  .thinking > summary { cursor: pointer; padding: 8px 12px; font-size: 12.5px; font-weight: 600; color: var(--bc-think-ink); display: flex; align-items: center; gap: 8px; list-style: none; }
+  .thinking > summary::-webkit-details-marker { display: none; }
+  .thinking > summary .chev { transition: transform .15s; }
+  .thinking[open] > summary .chev { transform: rotate(90deg); }
+  .thinking .body { padding: 0 12px 10px; font-size: 13px; color: var(--bc-think-ink); white-space: pre-wrap; line-height: 1.55; }
+  .tok-count { margin-left: auto; font-weight: 500; color: var(--bc-accent-strong); font-variant-numeric: tabular-nums; }
   .answer { white-space: pre-wrap; }
 
-  /* Pending / error */
-  .pending .bubble { display: flex; align-items: center; gap: 10px; color: var(--muted); }
-  .spinner { width: 14px; height: 14px; border: 2px solid #d1d5db; border-top-color: var(--hf-yellow-dark); border-radius: 50%; animation: spin .8s linear infinite; }
+  /* Pending / streaming */
+  .pending .bubble { display: flex; align-items: center; gap: 10px; color: var(--bc-muted); }
+  .spinner { width: 14px; height: 14px; border: 2px solid var(--bc-accent); border-top-color: var(--bc-accent-strong); border-radius: 50%; animation: spin .8s linear infinite; flex-shrink: 0; }
   @keyframes spin { to { transform: rotate(360deg); } }
   .elapsed { font-size: 12px; font-variant-numeric: tabular-nums; }
-  .error .bubble { border-color: #fecaca; background: #fef2f2; color: var(--danger); }
+  .live-think { font-size: 13px; color: var(--bc-think-ink); white-space: pre-wrap; line-height: 1.55; }
+  .live-answer { white-space: pre-wrap; }
+  .caret::after { content: "▍"; color: var(--bc-accent-strong); animation: blink 1s steps(1) infinite; margin-left: 1px; }
+  @keyframes blink { 50% { opacity: 0; } }
+  .error .bubble { border-color: #5b2a2a; background: #2a1516; color: var(--bc-danger); }
 
   /* Empty state */
-  #empty { text-align: center; color: var(--muted); padding: 60px 20px; max-width: 540px; margin: 0 auto; }
-  #empty h2 { color: var(--ink); font-size: 19px; margin: 14px 0 6px; }
+  #empty { text-align: center; color: var(--bc-muted); padding: 60px 20px; max-width: 540px; margin: 0 auto; }
+  #empty .empty-logo { width: 64px; height: 64px; margin: 0 auto 14px; border-radius: 16px; background: #f0f0f0; background-image: var(--logo); background-size: cover; background-position: center; }
+  #empty h2 { color: var(--bc-ink); font-size: 19px; margin: 0 0 6px; }
   #empty p { font-size: 14px; line-height: 1.55; }
   .examples { display: flex; flex-wrap: wrap; gap: 8px; justify-content: center; margin-top: 18px; }
-  .examples button { font: inherit; font-size: 13px; padding: 8px 12px; border: 1px solid var(--line); background: var(--panel); border-radius: 999px; cursor: pointer; color: var(--ink); }
-  .examples button:hover { border-color: var(--hf-yellow-dark); background: #fffdf0; }
+  .examples button { font: inherit; font-size: 13px; padding: 8px 12px; border: 1px solid var(--bc-line); background: var(--bc-surface-2); border-radius: 999px; cursor: pointer; color: var(--bc-ink); }
+  .examples button:hover { border-color: var(--bc-accent-strong); background: var(--bc-accent); }
 
   /* Composer */
-  footer { border-top: 1px solid var(--line); background: var(--panel); padding: 12px 18px; }
+  footer { border-top: 1px solid var(--bc-line); background: var(--bc-surface); padding: 12px 18px; }
   .composer { max-width: 820px; margin: 0 auto; display: flex; gap: 10px; align-items: flex-end; }
   .composer .attach { position: relative; }
   .composer .attach input { display: none; }
-  .composer .btn { border: 1px solid var(--line); background: var(--panel); border-radius: 10px; width: 40px; height: 40px; cursor: pointer; font-size: 18px; color: var(--ink); display: grid; place-items: center; flex-shrink: 0; }
-  .composer .btn:hover { background: #f3f4f6; }
-  .composer .btn.primary { background: var(--hf-yellow); border-color: var(--hf-yellow-dark); color: #1f2937; font-weight: 700; }
-  .composer .btn.primary:disabled { opacity: .5; cursor: not-allowed; }
-  #prompt { flex: 1; resize: none; border: 1px solid var(--line); border-radius: 10px; padding: 10px 12px; font: inherit; font-size: 14.5px; line-height: 1.4; max-height: 180px; outline: none; }
-  #prompt:focus { border-color: var(--hf-yellow-dark); }
-  .img-chip { display: inline-flex; align-items: center; gap: 6px; background: #fff7d6; border: 1px solid var(--user-border); border-radius: 999px; padding: 4px 8px 4px 4px; font-size: 12px; margin-right: 8px; }
+  .composer .btn { border: 1px solid var(--bc-line); background: var(--bc-surface-2); border-radius: 10px; width: 40px; height: 40px; cursor: pointer; font-size: 18px; color: var(--bc-ink); display: grid; place-items: center; flex-shrink: 0; }
+  .composer .btn:hover { background: var(--bc-accent); }
+  .composer .btn.primary { background: var(--bc-mark); border-color: var(--bc-mark); color: #f0f0f0; font-weight: 700; }
+  .composer .btn.primary:disabled { opacity: .4; cursor: not-allowed; }
+  #prompt { flex: 1; resize: none; border: 1px solid var(--bc-line); border-radius: 10px; padding: 10px 12px; font: inherit; font-size: 14.5px; line-height: 1.4; max-height: 180px; outline: none; background: var(--bc-surface-2); color: var(--bc-ink); }
+  #prompt::placeholder { color: var(--bc-faint); }
+  #prompt:focus { border-color: var(--bc-accent-strong); }
+  .img-chip { display: inline-flex; align-items: center; gap: 6px; background: var(--bc-user-bubble); border: 1px solid var(--bc-line); border-radius: 999px; padding: 4px 8px 4px 4px; font-size: 12px; margin-right: 8px; }
   .img-chip img { width: 22px; height: 22px; border-radius: 50%; object-fit: cover; }
-  .img-chip .x { cursor: pointer; color: var(--muted); font-weight: 700; }
-  .hint { max-width: 820px; margin: 6px auto 0; font-size: 11px; color: var(--muted); text-align: center; }
+  .img-chip .x { cursor: pointer; color: var(--bc-faint); font-weight: 700; }
+  .hint { max-width: 820px; margin: 6px auto 0; font-size: 11px; color: var(--bc-faint); text-align: center; }
+
+  /* Scrollbar (dark) */
+  main::-webkit-scrollbar { width: 10px; }
+  main::-webkit-scrollbar-thumb { background: var(--bc-accent); border-radius: 6px; border: 2px solid var(--bc-bg); }
+  main::-webkit-scrollbar-track { background: transparent; }
 </style>
 </head>
 <body>
   <header>
-    <div class="logo">🤗</div>
+    <div class="logo" role="img" aria-label="BottleCap AI"></div>
     <div class="titles">
       <h1>ThinkingCap · Qwen 3.6 27B</h1>
       <p>Token-efficient reasoning — same answers, ~50% fewer thinking tokens</p>
@@ -164,7 +190,7 @@
       <label>Thinking mode</label>
       <div class="row">
         <label class="switch"><input type="checkbox" id="s-enable" checked /><span class="slider"></span></label>
-        <span style="font-size:12px;color:var(--muted)">emit &lt;think&gt; trace</span>
+        <span style="font-size:12px;color:var(--bc-muted)">stream the  trace</span>
       </div>
     </div>
     <div class="field">
@@ -184,16 +210,16 @@
       <div class="row"><input type="range" id="s-topk" min="0" max="100" step="1" value="20" /></div>
     </div>
     <div class="field reset-btn">
-      <button class="btn" id="s-reset" style="width:auto;height:34px;padding:0 14px;font-size:13px;">Reset defaults</button>
+      <button class="btn" id="s-reset" style="width:auto;height:34px;padding:0 14px;font-size:13px;border:1px solid var(--bc-line);background:var(--bc-surface-2);color:var(--bc-ink);border-radius:8px;cursor:pointer;">Reset defaults</button>
     </div>
   </div>
 
   <main>
     <div id="messages">
       <div id="empty">
-        <div style="font-size:40px">🧠</div>
+        <div class="empty-logo"></div>
         <h2>Chat with ThinkingCap-Qwen3.6-27B</h2>
-        <p>A finetune of Qwen 3.6 27B that keeps the answer quality while cutting thinking tokens by ~50% on average (and &gt;90% in best cases). Ask anything — reasoning, math, code, or upload an image.</p>
+        <p>A finetune of Qwen 3.6 27B that keeps answer quality while cutting thinking tokens by ~50% on average (and &gt;90% in best cases). Ask anything — reasoning, math, code, or upload an image.</p>
         <div class="examples">
           <button>Explain why 0.1 + 0.2 ≠ 0.3 in floating point</button>
           <button>Write a Python function to debounce calls</button>
@@ -227,11 +253,16 @@ const fileInput = $("#file-input");
 const statusEl = $("#status");
 const imgHint = $("#img-hint");
 
-// Conversation history sent to the backend (text-only; images are per-turn).
+// Reasoning-trace markers (built so the literal tags never sit together here).
+const THINK_OPEN  = "<" + "think" + ">";
+const THINK_CLOSE  = "<" + "/think" + ">";
+const IM_END      = "<" + "|im_end|" + ">";
+const EOT         = "<" + "|endoftext|" + ">";
+
 let history = [];
 let client = null;
 let busy = false;
-let pendingFile = null;       // File object selected for the next message
+let pendingFile = null;
 
 // --- Settings --------------------------------------------------------------
 const DEFAULTS = { enable: true, max: 32768, temp: 1.0, topp: 0.95, topk: 20 };
@@ -265,10 +296,10 @@ $("#gear").addEventListener("click", () => $("#settings").classList.toggle("open
   try {
     client = await Client.connect(window.location.origin);
     statusEl.textContent = "ready";
-    statusEl.style.color = "#16a34a";
+    statusEl.style.color = "var(--bc-ok)";
   } catch (e) {
     statusEl.textContent = "offline";
-    statusEl.style.color = "var(--danger)";
+    statusEl.style.color = "var(--bc-danger)";
     console.error(e);
   }
 })();
@@ -284,7 +315,6 @@ promptEl.addEventListener("keydown", (e) => {
 });
 sendBtn.addEventListener("click", send);
 
-// Image attach
 $("#attach-btn").addEventListener("click", () => fileInput.click());
 fileInput.addEventListener("change", () => {
   const f = fileInput.files[0];
@@ -300,7 +330,6 @@ function clearImage() {
   imgHint.innerHTML = "";
 }
 
-// Example chips
 document.querySelectorAll(".examples button").forEach(b =>
   b.addEventListener("click", () => { promptEl.value = b.textContent; autosize(); promptEl.focus(); })
 );
@@ -309,6 +338,29 @@ document.querySelectorAll(".examples button").forEach(b =>
 function escapeHtml(s) {
   return String(s).replace(/[&<>"']/g, c => ({ "&":"&amp;", "<":"&lt;", ">":"&gt;", '"':"&quot;", "'":"&#39;" }[c]));
 }
+function stripMarkers(s) {
+  return s.split(IM_END).join("").split(EOT).join("").replace(/</?think>/g, "");
+}
+// Split accumulated text into a live (thinking, answer) view.
+function splitLive(text) {
+  let think = "", answer = "", state = "answer";
+  const o = text.indexOf(THINK_OPEN);
+  const c = text.indexOf(THINK_CLOSE);
+  if (o === -1 && c === -1) {
+    // No thinking block at all so far → treat everything as answer.
+    return { thinking: "", answer: text };
+  }
+  if (o !== -1) {
+    state = "think";
+    think = text.slice(o + THINK_OPEN.length);
+    if (c !== -1 && c > o) {
+      think = text.slice(o + THINK_OPEN.length, c);
+      answer = text.slice(c + THINK_CLOSE.length);
+      state = "answer";
+    }
+  }
+  return { thinking: think, answer, state };
+}
 
 function addUserBubble(text, imgUrl) {
   if (emptyEl) emptyEl.remove();
@@ -320,18 +372,21 @@ function addUserBubble(text, imgUrl) {
   scrollDown();
 }
 
-function addPendingBubble() {
+function addStreamingBubble() {
   const el = document.createElement("div");
   el.className = "msg assistant pending";
-  el.id = "pending";
-  el.innerHTML = `<div class="avatar">🧠</div>
-    <div class="bubble"><span class="spinner"></span> thinking… <span class="elapsed" data-t0>0.0s</span></div>`;
+  el.innerHTML = `<div class="avatar"></div>
+    <div class="bubble">
+      <span class="spinner"></span> thinking…
+      <span class="elapsed" data-t0>0.0s</span>
+      <div class="live"></div>
+    </div>`;
   messagesEl.appendChild(el);
   scrollDown();
   const t0 = performance.now();
   const tick = () => {
     const node = el.querySelector("[data-t0]");
-    if (!node) return;
+    if (!node || el.dataset.done) return;
     node.textContent = ((performance.now() - t0) / 1000).toFixed(1) + "s";
     el._timer = requestAnimationFrame(tick);
   };
@@ -339,31 +394,51 @@ function addPendingBubble() {
   return el;
 }
 
-function finalizeBubble(pendingEl, data) {
-  cancelAnimationFrame(pendingEl._timer);
+function updateLive(el, text) {
+  const live = el.querySelector(".live");
+  const head = el.querySelector(".bubble");
+  const { thinking, answer, state } = splitLive(text);
+  if (thinking || state === "think") {
+    head.querySelector(".spinner").style.display = "";
+    head.querySelector(".elapsed").style.display = "";
+    live.innerHTML = `<div class="live-think caret">${escapeHtml(thinking)}</div>`;
+  } else if (answer) {
+    head.querySelector(".spinner").style.display = "none";
+    head.querySelector(".elapsed").style.display = "none";
+    live.innerHTML = `<div class="live-answer caret">${escapeHtml(answer)}</div>`;
+  }
+  scrollDown();
+}
+
+function finalizeBubble(el, data) {
+  el.dataset.done = "1";
+  cancelAnimationFrame(el._timer);
   const { thinking, answer, thinking_tokens } = data;
   const thinkHtml = thinking
-    ? `<details class="thinking" ${settings.enable ? "" : ""}>
+    ? `<details class="thinking" ${stateOpen(el) ? "open" : ""}>
          <summary><span class="chev">▸</span> Reasoning <span class="tok-count">${thinking_tokens} thinking tokens</span></summary>
          <div class="body">${escapeHtml(thinking)}</div>
        </details>`
     : "";
   const ansHtml = answer ? `<div class="answer">${escapeHtml(answer)}</div>` : "";
-  pendingEl.classList.remove("pending");
-  pendingEl.querySelector(".bubble").innerHTML = thinkHtml + ansHtml;
+  el.classList.remove("pending");
+  el.querySelector(".bubble").innerHTML = thinkHtml + ansHtml;
   scrollDown();
 }
-
-function showError(pendingEl, msg) {
-  cancelAnimationFrame(pendingEl._timer);
-  pendingEl.classList.remove("pending");
-  pendingEl.classList.add("error");
-  pendingEl.querySelector(".bubble").innerHTML = "⚠️ " + escapeHtml(msg);
+// Keep the reasoning panel open across finalize if the user expanded it while streaming.
+function stateOpen(el) { return el.querySelector(".live-think") && el.dataset.userOpen === "1"; }
+
+function showError(el, msg) {
+  el.dataset.done = "1";
+  cancelAnimationFrame(el._timer);
+  el.classList.remove("pending");
+  el.classList.add("error");
+  el.querySelector(".bubble").innerHTML = "⚠️ " + escapeHtml(msg);
 }
 
 function scrollDown() { messagesEl.parentElement.scrollTop = messagesEl.parentElement.scrollHeight; }
 
-// --- Send -----------------------------------------------------------------
+// --- Send (streaming via Gradio JS client) --------------------------------
 async function send() {
   const text = promptEl.value.trim();
   if ((!text && !pendingFile) || busy || !client) return;
@@ -384,23 +459,37 @@ async function send() {
     enable_thinking: settings.enable,
   };
 
-  // Record the user turn in history (text only).
   history.push({ role: "user", content: text || "(image)" });
-
   promptEl.value = "";
   clearImage();
   autosize();
 
-  const pendingEl = addPendingBubble();
+  const el = addStreamingBubble();
+  let accumulated = "";
+
   try {
-    const result = await client.predict("/generate", payload);
-    const data = result.data[0] || {};
-    finalizeBubble(pendingEl, data);
-    history.push({ role: "assistant", content: data.answer || data.raw || "" });
+    const stream = await client.predict("/generate", payload, { events: ["generating", "done"] });
+
+    for await (const msg of stream) {
+      if (msg.type === "generating" && msg.data) {
+        const chunk = msg.data[0];
+        if (chunk && typeof chunk.text === "string") {
+          accumulated = chunk.text;             // server echoes full text so far
+          updateLive(el, accumulated);
+        }
+      } else if (msg.type === "done") {
+        const data = (msg.data && msg.data[0]) || {};
+        finalizeBubble(el, {
+          thinking: data.thinking || "",
+          answer: data.answer || data.text || "",
+          thinking_tokens: data.thinking_tokens || 0,
+        });
+        history.push({ role: "assistant", content: data.answer || data.text || "" });
+      }
+    }
   } catch (e) {
     console.error(e);
-    showError(pendingEl, (e && e.message) ? e.message : "Generation failed.");
-    // Roll back the user turn we optimistically added.
+    showError(el, (e && e.message) ? e.message : "Generation failed.");
     history.pop();
   } finally {
     busy = false;