fix: scroll ghost height reflow via overflow toggle; extend cache markers to last 2 user msgs; simplify cursor & merge JS fixes

frontends/stapp.py

@@ -27,6 +27,7 @@ agent = init()
 
 st.title("🖥️ Cowork")
 
+
 if 'autonomous_enabled' not in st.session_state: st.session_state.autonomous_enabled = False
 
 @st.fragment
@@ -97,7 +98,7 @@ def render_segments(segments, suffix=''):
         if seg['type'] == 'fold':
             with st.expander(seg['title'], expanded=False): st.markdown(seg['content'])
         else:
-            st.markdown(seg['content'] + suffix, unsafe_allow_html=not not suffix)
+            st.markdown(seg['content'] + suffix)
 
 def agent_backend_stream(prompt):
     display_queue = agent.put_task(prompt, source="user")
@@ -123,10 +124,28 @@ for msg in st.session_state.messages:
             if msg["role"] == "assistant": render_segments(fold_turns(msg["content"]))
             else: st.markdown(msg["content"])
 
-# IME composition fix (macOS only) - prevents Enter from submitting during CJK input
+# Scroll-height ghost fix: during streaming, expander open/close mid-animation can leave
-if os.name != 'nt':
+# phantom height → scrollbar long but can't scroll to bottom. Periodically detect & reflow.
 import streamlit.components.v1 as components
-    components.html('<script>!function(){if(window.parent.__imeFix)return;window.parent.__imeFix=1;var d=window.parent.document,c=0;d.addEventListener("compositionstart",()=>c=1,!0);d.addEventListener("compositionend",()=>c=0,!0);function f(){d.querySelectorAll("textarea[data-testid=stChatInputTextArea]").forEach(t=>{t.__imeFix||(t.__imeFix=1,t.addEventListener("keydown",e=>{e.key==="Enter"&&!e.shiftKey&&(e.isComposing||c||e.keyCode===229)&&(e.stopImmediatePropagation(),e.preventDefault())},!0))})}f();new MutationObserver(f).observe(d.body,{childList:1,subtree:1})}()</script>', height=0)
+_js_scroll_fix = ("!function(){var p=window.parent;if(p.__sfx)return;p.__sfx=1;"
+    "var d=p.document;setInterval(function(){"
+    "var m=d.querySelector('section.main');if(!m)return;"
+    "var b=m.querySelector('.block-container');if(!b)return;"
+    "if(m.scrollHeight>b.scrollHeight+150){"
+    "m.style.overflow='hidden';void m.offsetHeight;m.style.overflow=''}"
+    "},3000)}()")
+# IME composition fix (macOS only) - prevents Enter from submitting during CJK input
+_js_ime_fix = ("" if os.name == 'nt' else
+    "!function(){if(window.parent.__imeFix)return;window.parent.__imeFix=1;"
+    "var d=window.parent.document,c=0;"
+    "d.addEventListener('compositionstart',()=>c=1,!0);"
+    "d.addEventListener('compositionend',()=>c=0,!0);"
+    "function f(){d.querySelectorAll('textarea[data-testid=stChatInputTextArea]')"
+    ".forEach(t=>{t.__imeFix||(t.__imeFix=1,t.addEventListener('keydown',e=>{"
+    "e.key==='Enter'&&!e.shiftKey&&(e.isComposing||c||e.keyCode===229)&&"
+    "(e.stopImmediatePropagation(),e.preventDefault())},!0))})}"
+    "f();new MutationObserver(f).observe(d.body,{childList:1,subtree:1})}()")
+components.html(f'<script>{_js_scroll_fix};{_js_ime_fix}</script>', height=0)
 
 if prompt := st.chat_input("请输入指令"):
     st.session_state.messages.append({"role": "user", "content": prompt})
@@ -134,7 +153,7 @@ if prompt := st.chat_input("请输入指令"):
 
     with st.chat_message("assistant"):
         slot = st.empty(); response = ''
-        CURSOR = '<span style="animation: blink 1s step-start infinite; color: #0066cc;">▌</span><style>@keyframes blink { 50% { opacity: 0; } }</style>'
+        CURSOR = ' ▌'
         for response in agent_backend_stream(prompt):
             # 每轮整块重画（含 heartbeat 空转）：segments 不变时 Streamlit diff 零变更 → 不闪烁；
             # 而 slot.container() 调用本身保证 Streamlit 能抛 StopException（abort 生效）

llmcore.py

@@ -267,6 +267,19 @@ def _parse_openai_sse(resp_lines, api_mode="chat_completions"):
             blocks.append({"type": "tool_use", "id": tc["id"], "name": tc["name"], "input": inp})
         return blocks
 
+def _stamp_oai_cache_markers(messages, model):
+    """Add cache_control to last 2 user messages for Anthropic models via OAI-compatible relay."""
+    ml = model.lower()
+    if not any(k in ml for k in ('claude', 'anthropic')): return
+    user_idxs = [i for i, m in enumerate(messages) if m.get('role') == 'user']
+    for idx in user_idxs[-2:]:
+        c = messages[idx].get('content')
+        if isinstance(c, str):
+            messages[idx] = {**messages[idx], 'content': [{'type': 'text', 'text': c, 'cache_control': {'type': 'ephemeral'}}]}
+        elif isinstance(c, list) and c:
+            c = list(c); c[-1] = dict(c[-1], cache_control={'type': 'ephemeral'})
+            messages[idx] = {**messages[idx], 'content': c}
+
 def _openai_stream(api_base, api_key, messages, model, api_mode='chat_completions', *,
                    temperature=0.5, max_tokens=None, tools=None, reasoning_effort=None,
                    max_retries=0, connect_timeout=10, read_timeout=300, proxies=None):
@@ -281,6 +294,7 @@ def _openai_stream(api_base, api_key, messages, model, api_mode='chat_completion
         if reasoning_effort: payload["reasoning"] = {"effort": reasoning_effort}
     else:
         url = auto_make_url(api_base, "chat/completions")
+        _stamp_oai_cache_markers(messages, model)
         payload = {"model": model, "messages": messages, "temperature": temperature, "stream": True, "stream_options": {"include_usage": True}}
         if max_tokens: payload["max_tokens"] = max_tokens
         if reasoning_effort: payload["reasoning_effort"] = reasoning_effort
@@ -483,8 +497,9 @@ class ClaudeSession(BaseSession):
             return [{"type": "text", "text": err}]
     def make_messages(self, raw_list):
         msgs = [{"role": m['role'], "content": list(m['content'])} for m in raw_list]
-        c = msgs[-1]["content"]
+        user_idxs = [i for i, m in enumerate(msgs) if m['role'] == 'user']
-        c[-1] = dict(c[-1], cache_control={"type": "ephemeral"})
+        for idx in user_idxs[-2:]:
+            msgs[idx]["content"][-1] = dict(msgs[idx]["content"][-1], cache_control={"type": "ephemeral"})
         return msgs
 
 class LLMSession(BaseSession):
@@ -544,8 +559,10 @@ class NativeClaudeSession(BaseSession):
         if self.system:
             if self.fake_cc_system_prompt: messages[0]["content"].insert(0, {"type": "text", "text": self.system})
             else: payload["system"] = [{"type": "text", "text": self.system}]
-        messages[-1] = {**messages[-1], "content": list(messages[-1]["content"])}
+        user_idxs = [i for i, m in enumerate(messages) if m['role'] == 'user']
-        messages[-1]["content"][-1] = dict(messages[-1]["content"][-1], cache_control={"type": "ephemeral"})
+        for idx in user_idxs[-2:]:
+            messages[idx] = {**messages[idx], "content": list(messages[idx]["content"])}
+            messages[idx]["content"][-1] = dict(messages[idx]["content"][-1], cache_control={"type": "ephemeral"})
         try:
             resp = requests.post(auto_make_url(self.api_base, "messages")+'?beta=true', headers=headers, json=payload, stream=True, timeout=(self.connect_timeout, self.read_timeout))
             if resp.status_code != 200: raise Exception(f"HTTP {resp.status_code} {resp.content.decode('utf-8', errors='replace')[:500]}")
@@ -819,6 +836,7 @@ class MixinSession:
         for attempt in range(self._retries + 1):
             idx = (base + attempt) % n
             gen = self._orig_raw_asks[idx](*args, **kwargs)
+            print(f'[MixinSession] Using session ({self._sessions[idx].name})')
             last_chunk, return_val, yielded = None, [], False
             try:
                 while True: