diff --git a/.gitignore b/.gitignore
index 47a6807f7..5d9605dae 100644
--- a/.gitignore
+++ b/.gitignore
@@ -150,3 +150,4 @@ frontends/conductor_im_plugins/*
 
 # 本地 bug / 局限报告草稿（交给维护者参考，不进仓库）
 /BUG_*.md
+uv.lock
diff --git a/assets/configure_mykey.py b/assets/configure_mykey.py
index b9ed19d7e..de685e6cc 100644
--- a/assets/configure_mykey.py
+++ b/assets/configure_mykey.py
@@ -90,6 +90,7 @@
             'name': 'deepseek', 'apikey': 'sk-<your-deepseek-key>',
             'apibase': 'https://api.deepseek.com', 'model': 'deepseek-v4-pro',
             'api_mode': 'chat_completions', 'reasoning_effort': 'high',
+            'context_win': 1000000, 'max_tokens': 384000,
         },
         'key_hint': '在 https://platform.deepseek.com/api_keys 获取',
         'model_choices': ['deepseek-v4-pro', 'deepseek-v4-flash'],
@@ -266,6 +267,7 @@
             'apibase': 'https://api.xiaomimimo.com/v1',
             'model': 'mimo-v2.5-pro',
             'api_mode': 'chat_completions',
+            'context_win': 1000000, 'max_tokens': 128000,
         },
         'key_hint': '在 https://x.xiaomi.com/ 获取 API Key',
         'model_choices': ['mimo-v2.5-pro', 'mimo-v2-flash'],
diff --git a/frontends/cost_tracker.py b/frontends/cost_tracker.py
index 744258982..d322f6533 100644
--- a/frontends/cost_tracker.py
+++ b/frontends/cost_tracker.py
@@ -39,15 +39,15 @@ def elapsed_seconds(self) -> float:
         return max(0.0, time.time() - self.started_at)
 
 
-# GA's real context budget lives on `BaseSession.context_win` (chars). The
-# trim trigger is `context_win * 3` (see llmcore.trim_messages_history), so
-# `/cost` compares actual-history chars against that cap for consistent units.
+# GA's real context budget lives on `BaseSession.context_win` (chars).
+# `/cost` displays this value directly so users see the actual context window.
 def context_window_chars(backend) -> int:
-    """`context_win * 3` — the char cap before `trim_messages_history` kicks
-    in. Reads dynamically so a `mykey.py` override propagates. Returns 0 on
-    bad/missing backend so the caller can hide the row."""
+    """Returns the actual context_win (no multiplier). The internal trim
+    trigger is `context_win * 3` (see llmcore.trim_messages_history), but the
+    display should show the real model capacity. Returns 0 on bad/missing
+    backend so the caller can hide the row."""
     try:
-        return int(getattr(backend, 'context_win', 0)) * 3
+        return int(getattr(backend, 'context_win', 0))
     except (TypeError, ValueError):
         return 0
 
diff --git a/frontends/tui_v3.py b/frontends/tui_v3.py
index 0684d538d..1d1e458e6 100644
--- a/frontends/tui_v3.py
+++ b/frontends/tui_v3.py
@@ -90,7 +90,7 @@
         "Tip: /rewind [n] rewinds the last n turns; /stop aborts the current task.",
         "Tip: /continue lists past sessions — arrow keys to pick, Enter to restore.",
         "Tip: Ctrl+J / Shift+Enter inserts a newline in multi-line input; Enter sends.",
-        "Tip: put [multi-select] in an ask_user prompt to switch to a multi-pick picker.",
+        "Tip: in a multi-choice picker, Space toggles options and Enter confirms your selection.",
         "Tip: /cost shows token usage; /llm views / switches the model.",
         "Tip: /new [name] starts a fresh session; /language switches the interface language.",
         "Tip: /export clip copies the last reply to your system clipboard; /export all prints the log path.",
@@ -113,7 +113,7 @@
         "Tip: /rewind [n] 回退最近 n 轮对话；/stop 中止当前任务。",
         "Tip: /continue 列出历史会话 —— 方向键选择，Enter 恢复。",
         "Tip: 多行输入用 Ctrl+J / Shift+Enter 换行；Enter 直接发送。",
-        "Tip: ask_user 题目里写 [多选] 会自动切到多选 picker。",
+        "Tip: 多选选择器中，空格切换选项，Enter 确认选择。",
         "Tip: /cost 查看 token 用量；/llm 查看 / 切换模型。",
         "Tip: /new [name] 新建会话；/language 切换界面语言。",
         "Tip: /export clip 把最后回复复制到系统剪贴板；/export all 打印日志路径。",
@@ -1349,14 +1349,6 @@ def _extract_ask_user(ctx: dict | None) -> AskUserEvent | None:
         return None
     data = payload.get('data') or {}
     candidates = data.get('candidates') or []
-    # v2 parity: skip the ask card when the agent didn't supply candidates —
-    # the 'Waiting for your answer ...' marker already lands in scrollback as
-    # part of the assistant stream, and the user replies via the normal input
-    # box.  Pushing an empty-candidate event onto the queue would route us
-    # through _enter_ask → free-text ask card, which freezes the live region
-    # in some terminals.
-    if not candidates:
-        return None
     return AskUserEvent(
         question=data.get('question', ''),
         candidates=candidates,
@@ -2333,7 +2325,7 @@ def _indent_rows(rows: list[str], width: int) -> list[str]:
 
 
 def _cost_str(agent) -> str:
-    """Context-window usage view (cc/v2 style): used / cap of context_win*3."""
+    """Context-window usage view (cc/v2 style): used / cap of context_win."""
     try:
         from frontends import cost_tracker
         be = agent.llmclient.backend
@@ -3180,6 +3172,7 @@ def row(text: str, style: str = '') -> list[str]:
             return r
 
         rows = [top]
+        multi = self._picker_mode == 'multi'
         for ln in (ae.question or _t('ask.default_q')).strip().split('\n'):
             rows.extend(row(ln, _BOLD))
         if ae.candidates:
@@ -3188,7 +3181,6 @@ def row(text: str, style: str = '') -> list[str]:
             # current row and (in multi) a [ ]/[x] marker; in free-text mode
             # candidates show only as numbered hints — typing the number
             # still picks that row at submit time.
-            multi = self._picker_mode == 'multi'
             for i, c in enumerate(ae.candidates):
                 if multi:
                     mark = '[x]' if i in self._picker_checked else '[ ]'
@@ -4478,8 +4470,8 @@ def _cmd(self, raw: str) -> None:
             try:
                 from frontends import cost_tracker as _ct
                 cap = _ct.context_window_chars(be) if be is not None else 0
-                used = _ct.context_chars_used(be) if be is not None else 0
-                ctx_use = _t('status.ctx.fmt', used=used, cap=cap * 3) if cap else _t('status.ctx.unknown')
+                used = _ct.current_input_chars(be) if be is not None else 0
+                ctx_use = _t('status.ctx.fmt', used=used, cap=cap) if cap else _t('status.ctx.unknown')
             except Exception:
                 ctx_use = _t('status.ctx.unknown')
             cwd = os.getcwd().replace(os.path.expanduser('~'), '~')
@@ -5970,6 +5962,12 @@ def _pre_run() -> None:
             _w('\x1b[>4;1m')    # ask supporting terminals to distinguish Shift+Enter
             self.commit(Block('banner', ''))
 
+        # Some terminals (IDE consoles, certain SSH configs, etc.) do not
+        # respond to CPR (cursor position report) requests, causing a loud
+        # warning.  Suppress just the warning — keep CPR probing enabled so
+        # that _min_available_height stays accurate for dynamic-height regions
+        # like the ask_user picker card.
+        _ptk_output = create_output(stdout=so)
         app = Application(
             layout=layout,
             key_bindings=kb,
@@ -5982,8 +5980,14 @@ def _pre_run() -> None:
             terminal_size_polling_interval=0.2,
             mouse_support=False,
             before_render=_before_render,
-            output=create_output(stdout=so),
+            output=_ptk_output,
         )
+        # Silence the CPR warning.  The Application's callback is checked by
+        # prompt_toolkit's Application.run() path, but the Renderer captures its
+        # own copy during __init__ and that is the one that actually fires the
+        # warning.  Both must be no-ops to prevent the brief flash on startup.
+        app.cpr_not_supported_callback = lambda: None
+        app.renderer.cpr_not_supported_callback = lambda: None
         self._ptk_app = app
 
         try:
diff --git a/llmcore.py b/llmcore.py
index 5eaa016d4..9fb57791d 100644
--- a/llmcore.py
+++ b/llmcore.py
@@ -530,7 +530,7 @@ def __init__(self, cfg):
         self.model = cfg.get('model', '')
         default_context_win = 30000
         if 'deepseek' in self.model.lower():
-            default_context_win = 70000; self.cut_msg_interval = 25; self.trim_keep_rate = 0.3
+            default_context_win = 1000000; self.cut_msg_interval = 25; self.trim_keep_rate = 0.3
         self.context_win = cfg.get('context_win', default_context_win)
         self.history = []; self.lock = threading.Lock(); self.system = ""
         self.name = cfg.get('name', self.model)
diff --git a/pyproject.toml b/pyproject.toml
index cf243449c..dfbdb7d4a 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -14,6 +14,8 @@ dependencies = [
     "bottle>=0.12",
     "simple-websocket-server>=0.4",
     "aiohttp>=3.9",
+    "rich>=15.0.0",
+    "prompt-toolkit>=3.0.52",
 ]
 
 [project.optional-dependencies]
@@ -48,4 +50,4 @@ py-modules = []
 packages = ["ga_cli"]
 
 [project.scripts]
-ga = "ga_cli.cli:main"
\ No newline at end of file
+ga = "ga_cli.cli:main"