chore: add .codex to .gitignore

.gitignore

@@ -32,5 +32,12 @@ htmlcov/
 # IDE and editor files
 .idea/
 .vscode/
+.codex/
 *.swp
 *.swo
+
+# debug resources
+debug/
+
+# Node dependencies for browser automation tooling
+node_modules/

AGENTS.md

@@ -0,0 +1,32 @@
+# Repository Guidelines
+
+## Project Structure & Module Organization
+
+Core poker service code lives in `texas_holdem/`. Important modules include `engine.py` for Texas Hold'em rules, `service.py` for game management, `server.py` for the HTTP API, `agents.py` for local/HTTP agents, and `ai_client.py` / `human_client.py` for standalone agents. Prompt templates live in `texas_holdem/prompts/`.
+
+Replay UI code lives in `texas_holdem_replay/`, with static browser assets under `texas_holdem_replay/static/`. Tests are in `tests/`, named by feature area such as `test_engine.py`, `test_service.py`, and `test_replay_server.py`. Design notes belong in `docs/`.
+
+## Build, Test, and Development Commands
+
+- `python -m unittest discover -v` runs the full test suite.
+- `python -m compileall texas_holdem texas_holdem_replay tests` checks import and syntax validity.
+- `python -m texas_holdem.server --host 127.0.0.1 --port 8000` starts the game service.
+- `python -m texas_holdem.human_client --port 9001 --keep-history` starts an interactive human HTTP agent.
+- `python -m texas_holdem.ai_client --port 9101 --api-key "$OPENAI_API_KEY" --model gpt-4o-mini` starts an OpenAI-compatible AI agent.
+- `python -m texas_holdem_replay.server --port 8088` starts the replay viewer.
+
+## Coding Style & Naming Conventions
+
+Use Python 3.11+ standard-library APIs unless a dependency is intentionally added to `pyproject.toml`. Keep modules focused and prefer explicit dataclasses for wire/state models. Use 4-space indentation, type hints, `snake_case` for functions and variables, `PascalCase` for classes, and concise comments only where logic is non-obvious.
+
+## Testing Guidelines
+
+Use `unittest`. Add tests near the behavior changed: engine rules in `tests/test_engine.py`, HTTP/service behavior in `tests/test_service.py`, agent transport in `tests/test_agents.py`, and replay UI server helpers in `tests/test_replay_server.py`. New bug fixes should include a regression test. Avoid tests that require external network access or real LLM calls.
+
+## Commit & Pull Request Guidelines
+
+History uses short Conventional Commit-style subjects, for example `feat: add replay server and web client` and `fix: game service api block when a game is running`. Keep commits scoped to one behavior change. Pull requests should include a short summary, test commands run, linked issue or motivation, and screenshots only for replay UI or visible terminal-output changes.
+
+## Security & Configuration Tips
+
+Do not commit API keys. Pass LLM credentials through `OPENAI_API_KEY` or CLI flags in local shells only. HTTP Agent endpoints are exclusive per active game; preserve this invariant when changing service concurrency.

README.md

@@ -11,6 +11,7 @@
 - 支持盲注、四条街下注、弃牌、过牌、跟注、下注、加注、全下、边池和摊牌结算。
 - 支持本地 Agent 和 HTTP Agent。
 - 支持 Human Agent 和 OpenAI-compatible AI Agent 的终端过程输出。
+- 游戏运行中可以并发查询状态；查询返回上一手完成后的稳定快照。
 
 ## 运行服务
 
@@ -51,6 +52,12 @@ curl -X POST http://127.0.0.1:8000/games/demo/hands/run \
 curl http://127.0.0.1:8000/games/demo
 ```
 
+也可以使用单数别名：
+
+```bash
+curl http://127.0.0.1:8000/game/demo
+```
+
 ## HTTP Agent 协议
 
 玩家配置可以使用远程 HTTP Agent：
@@ -62,12 +69,19 @@ curl http://127.0.0.1:8000/games/demo
   "agent": {
     "type": "http",
     "endpoint": "http://127.0.0.1:9101",
-    "timeout_seconds": 10
+    "timeout_seconds": 10,
+    "game_update_timeout_seconds": 3,
+    "retries": 2,
+    "retry_backoff_seconds": 0.25
   }
 }
 ```
 
-服务会向 `endpoint` 发送当前行动玩家的观察 JSON。Agent 返回：
+服务会向 `endpoint + /game` 推送每手开始时的游戏快照，向 `endpoint + /act` 发送当前行动玩家的观察 JSON。`endpoint` 也可以传入历史形式的 `/act` 或 `/game` 后缀，服务会归一化为 base URL。
+
+同一个 HTTP Agent endpoint 不能同时被不同游戏占用；后创建的游戏会返回错误。服务会给 HTTP Agent 请求自动重试，`/act` 重试仍失败时，规则引擎会按 `check > call > fold` 选择默认动作，避免整桌卡死。
+
+Agent 返回：
 
 ```json
 {"action": "call"}

docs/replay_view_design.md

@@ -0,0 +1,81 @@
+# Texas Hold X 回放视图设计方案
+
+## 目标
+
+构建一个与核心游戏服务和 Agent 解耦的独立 Web 服务，用于读取游戏详情 JSON 并以动画方式回放 Texas Hold'em 对局。它可以部署在任意能运行 Python 标准库 HTTP 服务的环境中，不要求核心服务增加前端路由，也不改变 Human HTTP Agent / AI HTTP Agent 协议。
+
+## 架构
+
+新增 `texas_holdem_replay` 包：
+
+- `texas_holdem_replay.server`：标准库 HTTP 服务，负责托管静态前端文件，并提供 `/api/fetch-game` 抓取代理。
+- `texas_holdem_replay/static/index.html`：独立页面入口。
+- `texas_holdem_replay/static/styles.css`：像素风牌桌、卡牌、座位和响应式布局。
+- `texas_holdem_replay/static/app.js`：数据归一化、手牌时间轴生成、动画播放、上传 JSON、手动抓取和自动轮询。
+
+运行方式：
+
+```bash
+python -m texas_holdem_replay.server --host 127.0.0.1 --port 8088
+```
+
+也可以通过安装后的脚本启动：
+
+```bash
+texas-holdem-replay --host 127.0.0.1 --port 8088
+```
+
+## 数据输入
+
+视图支持三种输入方式：
+
+1. 填写核心游戏服务地址和 `game_id`，点击“获取”。
+   前端请求自身的 `/api/fetch-game?base_url=...&game_id=...`，由回放服务去访问核心服务的 `/games/{game_id}`，避免浏览器跨域限制。
+2. 上传静态 JSON 文件。
+   文件在浏览器本地解析，不依赖核心服务。
+3. 开启自动获取。
+   按指定秒数轮询同一个核心服务和 `game_id`，用于观察正在运行的游戏快照。新快照会先尝试和当前回放位置合并；如果当前手牌追加了 action、showdown 或 award，回放会接续到新增 frame，而不是重头播放。
+
+`/api/fetch-game` 也支持传入完整 `url`，便于未来接入网关或静态 JSON 服务。
+
+## 数据模型与归一化
+
+当前游戏详情返回结构包含：
+
+- `players`：玩家最终状态。
+- `hands`：历史手牌列表。
+- 每手 `actions`：行动记录，包含 `street`、`player_id`、`action`、`amount`、行动后 `street_bet` 和 `stack`。
+- `awards`：底池分配。
+- `showdown_hands`：摊牌玩家手牌。
+
+前端不会依赖核心服务内部对象，只读取 JSON 字段并做归一化：
+
+- 根据 `players` 和 `actions` 得到座位顺序。
+- 从 `starting_stack` 开始，按历史 `actions.stack` 和 `awards` 推演每手牌开始时的筹码。
+- 每手牌生成一组离散 frame：开局、跨街发公共牌、玩家行动、摊牌、结算。
+- 每个 frame 都有稳定 key。上传 JSON、手动获取和自动获取都会复用同一套合并逻辑：同一 `game_id` 且能找到当前手牌时，保留当前 frame；如果用户停在最新进度末尾，且当前手牌或后续手牌出现新增 frame，则从当前位置接续播放增量。用户正在查看历史手牌时，不会被新轮询强制跳走。
+- 非 showdown 玩家手牌显示卡背。
+- 已预留 `hand.hole_cards[player_id]` 和 `hand.private_hands[player_id]` 兼容点，后续核心服务返回非 showdown 手牌时可直接展示。
+
+## 动画与交互
+
+牌桌采用卡通像素风格：
+
+- 椭圆绿色牌桌、木质像素边框、像素筹码状态条。
+- 玩家围绕牌桌分布，当前行动玩家高亮并显示冒泡文字。
+- 公共牌按 flop / turn / river 分阶段发出。
+- 动作之间默认保留约 1.1-1.5 秒间隔，用户可用“节奏”滑杆调慢或调快。
+- 支持上一帧、下一帧、播放/暂停、重置、选择指定手牌。
+
+## 响应式设计
+
+桌面端为三栏布局：数据控制、牌桌、事件日志。中等屏幕下事件日志下移，手机和平板窄屏下改为单列，牌桌高度固定到可观看的移动端比例，座位尺寸通过 CSS clamp 控制，避免文字和控件溢出。
+
+## 解耦边界
+
+回放服务只消费核心服务的公开 HTTP JSON，不导入 `texas_holdem.engine`、`service` 或 Agent 代码，也不要求游戏服务开放 CORS。未来可以单独部署在 CDN + 轻量代理、容器或任意 Python 运行环境中。
+
+## 后续适配
+
+- 核心服务返回非 showdown 玩家手牌后，只需要让每手 JSON 包含 `hole_cards` 或 `private_hands` 映射，前端现有归一化会优先读取。
+- 如果服务端未来提供增量事件流，可以在 `app.js` 增加一个 source adapter，把事件流转成同样的 frame 列表，动画层无需重写。

pyproject.toml

@@ -9,6 +9,10 @@ dependencies = []
 texas-holdem-server = "texas_holdem.server:main"
 texas-holdem-human = "texas_holdem.human_client:main"
 texas-holdem-ai = "texas_holdem.ai_client:main"
+texas-holdem-replay = "texas_holdem_replay.server:main"
+
+[tool.setuptools.package-data]
+texas_holdem_replay = ["static/*"]
 
 [tool.pytest.ini_options]
 testpaths = ["tests"]

tests/test_agents.py

@@ -0,0 +1,60 @@
+import json
+import unittest
+from unittest.mock import patch
+from urllib.error import URLError
+
+from texas_holdem.agents import HttpAgent, normalise_http_agent_endpoint
+
+
+class FakeResponse:
+    def __init__(self, payload: dict[str, object]) -> None:
+        self.payload = payload
+
+    def read(self) -> bytes:
+        return json.dumps(self.payload).encode("utf-8")
+
+    def __enter__(self) -> "FakeResponse":
+        return self
+
+    def __exit__(self, *args: object) -> None:
+        return None
+
+
+class AgentTests(unittest.TestCase):
+    def test_normalise_http_agent_endpoint_accepts_action_or_game_paths(self) -> None:
+        self.assertEqual(
+            normalise_http_agent_endpoint("http://127.0.0.1:9101/act"),
+            "http://127.0.0.1:9101",
+        )
+        self.assertEqual(
+            normalise_http_agent_endpoint("http://127.0.0.1:9101/game/"),
+            "http://127.0.0.1:9101",
+        )
+
+    def test_http_agent_post_retries_and_sets_player_header(self) -> None:
+        calls = []
+
+        def fake_urlopen(request, timeout):  # type: ignore[no-untyped-def]
+            calls.append((request, timeout))
+            if len(calls) == 1:
+                raise URLError("temporary")
+            return FakeResponse({"ok": True})
+
+        agent = HttpAgent(
+            "http://agent.test/act",
+            player_id="p1",
+            retries=1,
+            retry_backoff_seconds=0,
+        )
+
+        with patch("texas_holdem.agents.urlopen", fake_urlopen):
+            payload = agent._post_json("/game", {"game_id": "g1"}, timeout_seconds=2)
+
+        self.assertEqual(payload, {"ok": True})
+        self.assertEqual(len(calls), 2)
+        self.assertEqual(calls[1][0].headers["X-player-id"], "p1")
+        self.assertEqual(calls[1][1], 2)
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_replay_server.py

@@ -0,0 +1,28 @@
+from __future__ import annotations
+
+import unittest
+
+from texas_holdem_replay.server import build_game_url
+
+
+class ReplayServerTests(unittest.TestCase):
+    def test_build_game_url_from_base_and_game_id(self) -> None:
+        self.assertEqual(
+            build_game_url({"base_url": ["http://127.0.0.1:8000/"], "game_id": ["game 1"]}),
+            "http://127.0.0.1:8000/games/game%201",
+        )
+
+    def test_build_game_url_accepts_full_url(self) -> None:
+        self.assertEqual(
+            build_game_url({"url": ["https://example.test/games/demo"]}),
+            "https://example.test/games/demo",
+        )
+
+    def test_build_game_url_rejects_non_http_url(self) -> None:
+        with self.assertRaises(ValueError):
+            build_game_url({"url": ["file:///tmp/game.json"]})
+
+
+if __name__ == "__main__":
+    unittest.main()
+

tests/test_service.py

@@ -1,8 +1,26 @@
 import unittest
+from threading import Event, Thread
 
+from texas_holdem.agents import PokerAgent
+from texas_holdem.models import Observation, PlayerAction
 from texas_holdem.service import GameManager
 
 
+class BlockingAgent(PokerAgent):
+    def __init__(self, entered: Event, release: Event) -> None:
+        self.entered = entered
+        self.release = release
+
+    def decide(self, observation: Observation) -> PlayerAction:
+        self.entered.set()
+        if not self.release.wait(timeout=5):
+            raise RuntimeError("test timed out waiting to release blocking agent")
+        for action in observation.legal_actions:
+            if action["action"] == "check":
+                return PlayerAction("check")
+        return PlayerAction("call")
+
+
 class ServiceTests(unittest.TestCase):
     def test_create_and_run_game(self) -> None:
         manager = GameManager()
@@ -23,7 +41,62 @@ class ServiceTests(unittest.TestCase):
         hands = manager.run_hands(game.game_id, count=1)
 
         self.assertEqual(len(hands), 1)
-        self.assertEqual(manager.get_game("demo").to_dict()["hand_number"], 1)
+        self.assertEqual(manager.get_game_state("demo")["hand_number"], 1)
+
+    def test_get_game_state_does_not_block_during_run(self) -> None:
+        manager = GameManager()
+        entered = Event()
+        release = Event()
+        game = manager.create_game(
+            {
+                "game_id": "blocking",
+                "seed": 13,
+                "starting_stack": 200,
+                "small_blind": 5,
+                "big_blind": 10,
+                "players": [
+                    {"id": "a", "type": "calling"},
+                    {"id": "b", "type": "calling"},
+                ],
+            }
+        )
+        manager.run_hands("blocking", count=1)
+        game.agents["a"] = BlockingAgent(entered, release)
+
+        thread = Thread(target=lambda: manager.run_hands("blocking", count=1))
+        thread.start()
+        self.assertTrue(entered.wait(timeout=2))
+
+        state = manager.get_game_state("blocking")
+
+        release.set()
+        thread.join(timeout=2)
+        self.assertFalse(thread.is_alive())
+        self.assertEqual(state["hand_number"], 1)
+        self.assertEqual(len(state["hands"]), 1)
+
+    def test_duplicate_http_agent_endpoint_is_rejected_across_active_games(self) -> None:
+        manager = GameManager()
+        payload = {
+            "starting_stack": 200,
+            "small_blind": 5,
+            "big_blind": 10,
+            "players": [
+                {
+                    "id": "ai",
+                    "agent": {
+                        "type": "http",
+                        "endpoint": "http://127.0.0.1:9101/act",
+                    },
+                },
+                {"id": "b", "type": "calling"},
+            ],
+        }
+
+        manager.create_game({"game_id": "g1", **payload})
+
+        with self.assertRaisesRegex(ValueError, "already belongs to game g1"):
+            manager.create_game({"game_id": "g2", **payload})
 
 
 if __name__ == "__main__":

texas_holdem/agents.py

@@ -2,10 +2,11 @@ from __future__ import annotations
 
 import json
 import sys
+import time
 from abc import ABC, abstractmethod
 from random import Random
 from typing import IO, Any
-from urllib.error import URLError
+from urllib.error import HTTPError, URLError
 from urllib.request import Request, urlopen
 
 from texas_holdem.human_io import clear_screen, prompt_action, render_observation
@@ -54,6 +55,27 @@ class CallingStationAgent(PokerAgent):
         return PlayerAction("fold")
 
 
+def normalise_http_agent_endpoint(raw: str) -> str:
+    """Return the canonical base URL for an HTTP agent endpoint."""
+    url = raw.rstrip("/")
+    if url.endswith("/act"):
+        url = url[: -len("/act")]
+    if url.endswith("/game"):
+        url = url[: -len("/game")]
+    return url
+
+
+def http_agent_endpoint_from_spec(spec: dict[str, Any]) -> str | None:
+    """Extract the canonical HTTP endpoint from an agent spec, if present."""
+    agent_type = str(spec.get("type", "calling")).lower()
+    if agent_type != "http":
+        return None
+    endpoint = spec.get("endpoint")
+    if not endpoint:
+        raise ValueError("http agent requires an endpoint")
+    return normalise_http_agent_endpoint(str(endpoint))
+
+
 class HttpAgent(PokerAgent):
     """Remote agent that talks to a base URL exposing ``/act`` and ``/game``.
 
@@ -66,28 +88,36 @@ class HttpAgent(PokerAgent):
     ACT_PATH = "/act"
     GAME_PATH = "/game"
 
-    def __init__(self, endpoint: str, timeout_seconds: float = 10.0) -> None:
+    def __init__(
-        self.base_url = self._normalise_base_url(endpoint)
+        self,
+        endpoint: str,
+        timeout_seconds: float = 10.0,
+        player_id: str | None = None,
+        game_update_timeout_seconds: float | None = None,
+        retries: int = 2,
+        retry_backoff_seconds: float = 0.25,
+    ) -> None:
+        self.base_url = normalise_http_agent_endpoint(endpoint)
         self.timeout_seconds = timeout_seconds
-
+        self.player_id = player_id
-    @staticmethod
+        self.game_update_timeout_seconds = (
-    def _normalise_base_url(raw: str) -> str:
+            float(game_update_timeout_seconds)
-        """Strip a trailing slash so URL joins do not produce double slashes.
+            if game_update_timeout_seconds is not None
-
+            else min(timeout_seconds, 3.0)
-        Centralising this also tolerates the legacy "endpoint already points
+        )
-        at /act" mistake by chopping off a redundant ``/act`` suffix.
+        self.retries = max(0, retries)
-        """
+        self.retry_backoff_seconds = max(0.0, retry_backoff_seconds)
-        url = raw.rstrip("/")
-        if url.endswith("/act"):
-            url = url[: -len("/act")]
-        return url
 
     def _url(self, path: str) -> str:
         """Compose a full URL by joining the base with a path component."""
         return f"{self.base_url}{path}"
 
     def decide(self, observation: Observation) -> PlayerAction:
-        payload = self._post_json(self.ACT_PATH, observation.to_dict())
+        payload = self._post_json(
+            self.ACT_PATH,
+            observation.to_dict(),
+            timeout_seconds=self.timeout_seconds,
+        )
         if not isinstance(payload, dict):
             raise RuntimeError("agent endpoint must return a JSON object")
         return PlayerAction.from_dict(payload)
@@ -100,30 +130,54 @@ class HttpAgent(PokerAgent):
         only by way of the raised exception bubbling to the engine guard.
         """
         try:
-            self._post_json(self.GAME_PATH, game_state)
+            self._post_json(
+                self.GAME_PATH,
+                game_state,
+                timeout_seconds=self.game_update_timeout_seconds,
+            )
         except RuntimeError:
             # ``/game`` is informational; treat any HTTP error as a benign
             # drop rather than reraising and aborting the hand loop.
             return None
 
-    def _post_json(self, path: str, payload: dict[str, Any]) -> Any:
+    def _post_json(
+        self,
+        path: str,
+        payload: dict[str, Any],
+        timeout_seconds: float,
+    ) -> Any:
         """POST ``payload`` as JSON to ``base_url + path`` and return parsed body.
 
         Extracted as a tiny helper so ``decide`` and ``on_game_update`` share
         identical transport semantics (timeout, error wrapping, content-type).
         """
         body = json.dumps(payload).encode("utf-8")
-        request = Request(
+        last_error: BaseException | None = None
-            self._url(path),
+        raw = ""
-            data=body,
+        for attempt in range(self.retries + 1):
-            headers={"Content-Type": "application/json"},
+            request = Request(
-            method="POST",
+                self._url(path),
-        )
+                data=body,
-        try:
+                headers=self._headers(),
-            with urlopen(request, timeout=self.timeout_seconds) as response:
+                method="POST",
-                raw = response.read().decode("utf-8")
+            )
-        except (OSError, URLError) as exc:
+            try:
-            raise RuntimeError(f"agent endpoint failed: {self._url(path)}") from exc
+                with urlopen(request, timeout=timeout_seconds) as response:
+                    raw = response.read().decode("utf-8")
+                break
+            except HTTPError as exc:
+                detail = exc.read().decode("utf-8", errors="replace")
+                last_error = RuntimeError(
+                    f"agent endpoint failed with HTTP {exc.code}: "
+                    f"{self._url(path)} {detail}"
+                )
+            except (OSError, URLError) as exc:
+                last_error = exc
+            if attempt < self.retries and self.retry_backoff_seconds > 0:
+                time.sleep(self.retry_backoff_seconds * (2**attempt))
+        else:
+            raise RuntimeError(f"agent endpoint failed: {self._url(path)}") from last_error
+
         if not raw:
             return None
         try:
@@ -133,6 +187,12 @@ class HttpAgent(PokerAgent):
                 f"agent endpoint returned invalid JSON: {self._url(path)}"
             ) from exc
 
+    def _headers(self) -> dict[str, str]:
+        headers = {"Content-Type": "application/json", "Connection": "close"}
+        if self.player_id:
+            headers["X-Player-Id"] = self.player_id
+        return headers
+
 
 class HumanAgent(PokerAgent):
     """Interactive CLI agent for debugging and manual play.
@@ -189,7 +249,11 @@ class HumanAgent(PokerAgent):
         return line.rstrip("\n")
 
 
-def build_agent(spec: dict[str, Any], rng: Random | None = None) -> PokerAgent:
+def build_agent(
+    spec: dict[str, Any],
+    rng: Random | None = None,
+    player_id: str | None = None,
+) -> PokerAgent:
     agent_type = str(spec.get("type", "calling")).lower()
     if agent_type == "random":
         return RandomAgent(rng)
@@ -199,7 +263,18 @@ def build_agent(spec: dict[str, Any], rng: Random | None = None) -> PokerAgent:
         endpoint = spec.get("endpoint")
         if not endpoint:
             raise ValueError("http agent requires an endpoint")
-        return HttpAgent(str(endpoint), float(spec.get("timeout_seconds", 10.0)))
+        return HttpAgent(
+            str(endpoint),
+            timeout_seconds=float(spec.get("timeout_seconds", 10.0)),
+            player_id=player_id,
+            game_update_timeout_seconds=(
+                float(spec["game_update_timeout_seconds"])
+                if "game_update_timeout_seconds" in spec
+                else None
+            ),
+            retries=int(spec.get("retries", 2)),
+            retry_backoff_seconds=float(spec.get("retry_backoff_seconds", 0.25)),
+        )
     if agent_type in {"human", "cli", "interactive"}:
         return HumanAgent()
     raise ValueError(f"unknown agent type: {agent_type}")

texas_holdem/ai_client.py

@@ -69,6 +69,109 @@ ANSI_RESET = "\x1b[0m"
 # ---------------------------------------------------------------------------
 
 
+class _ThinkingIndicator:
+    """Animated "thinking..." marquee for the AI agent console.
+
+    Design rationale:
+    - Encapsulated as its own class so the animation lifecycle (timer
+      thread, frame state, screen erase sequence) does not pollute the
+      surrounding console class.
+    - Runs in a daemon background thread driven by ``threading.Event`` so
+      ``stop`` returns promptly even if the current frame is mid-sleep.
+    - Uses ANSI ``\\r`` plus a clearing escape sequence to overwrite the
+      previous frame in place, avoiding scrollback noise. The frames
+      cycle through 0/1/2/3 dots every 0.5s as requested.
+    - ``start``/``stop`` are idempotent so the higher-level console can
+      call ``stop`` defensively (e.g. on the fallback path) without
+      tracking whether a marquee is actually running.
+    """
+
+    # Frame interval in seconds; matches the user-visible cadence.
+    _FRAME_INTERVAL = 0.5
+    # 0..3 dots, looping.
+    _FRAMES = ("thinking", "thinking.", "thinking..", "thinking...")
+    # ANSI escape that clears from the cursor to the end of the line; we
+    # combine it with a leading carriage return to redraw the frame in
+    # place.
+    _ERASE_LINE = "\r\x1b[K"
+
+    def __init__(
+        self,
+        write_fn: Callable[[str], None],
+        gray_fn: Callable[[str], str],
+    ) -> None:
+        self._write = write_fn
+        self._gray = gray_fn
+        self._stop_event = threading.Event()
+        self._thread: threading.Thread | None = None
+        # ``_active`` reflects whether a frame is currently visible on
+        # screen; ``stop`` uses it to decide whether to emit the final
+        # erase sequence.
+        self._active = False
+        # Guard against concurrent start/stop calls from different
+        # threads (e.g. content-delta handler vs. end_llm_stream).
+        self._lifecycle_lock = threading.Lock()
+
+    def start(self) -> None:
+        """Begin the marquee in a background thread.
+
+        Calling ``start`` while already running is a no-op.
+        """
+        with self._lifecycle_lock:
+            if self._thread is not None and self._thread.is_alive():
+                return
+            self._stop_event.clear()
+            self._active = True
+            thread = threading.Thread(
+                target=self._run,
+                name="ai-thinking-indicator",
+                daemon=True,
+            )
+            self._thread = thread
+            thread.start()
+
+    def stop(self) -> None:
+        """Stop the marquee and erase the current frame from the screen.
+
+        Safe to call when not running.
+        """
+        with self._lifecycle_lock:
+            thread = self._thread
+            if thread is None:
+                return
+            self._stop_event.set()
+            self._thread = None
+        # Wait for the worker outside the lifecycle lock so an in-flight
+        # ``_render_frame`` cannot deadlock against ``start`` from
+        # another thread.
+        thread.join()
+        if self._active:
+            # Wipe the last frame so the model's actual content begins on
+            # a clean line.
+            self._write(self._ERASE_LINE)
+            self._active = False
+
+    def _run(self) -> None:
+        """Background loop: redraw the next frame every ``_FRAME_INTERVAL``."""
+        index = 0
+        while not self._stop_event.is_set():
+            self._render_frame(self._FRAMES[index % len(self._FRAMES)])
+            index += 1
+            # ``Event.wait`` returns immediately when ``set`` is called,
+            # so ``stop`` is responsive even mid-frame.
+            if self._stop_event.wait(self._FRAME_INTERVAL):
+                return
+
+    def _render_frame(self, label: str) -> None:
+        """Emit one frame in place using carriage-return + erase-EOL."""
+        self._write(f"{self._ERASE_LINE}{self._gray(label)}")
+
+
+# ---------------------------------------------------------------------------
+# AI agent console
+# ---------------------------------------------------------------------------
+
+
 class AIAgentConsole:
     """Serialised terminal output for the standalone AI agent.
 
@@ -84,11 +187,30 @@ class AIAgentConsole:
         output_stream: IO[str] | None = None,
         keep_history: bool = False,
         use_color: bool = True,
+        show_reasoning: bool = True,
     ) -> None:
         self._output = output_stream if output_stream is not None else sys.stdout
         self._keep_history = keep_history
         self._use_color = use_color
+        # ``show_reasoning`` controls whether the LLM's chain-of-thought
+        # ("reasoning") deltas are printed to the terminal. The final
+        # answer ("content") is always printed so operators can still see
+        # the action being chosen.
+        self._show_reasoning = show_reasoning
+        # ``_lock`` serialises whole act/game render blocks (coarse grain).
+        # ``_io_lock`` is a finer-grained mutex protecting just the
+        # ``self._output.write`` calls so the thinking-indicator background
+        # thread can interleave safely with the main rendering thread
+        # without being blocked by the coarse lock.
         self._lock = threading.Lock()
+        self._io_lock = threading.Lock()
+        # Animated "thinking..." marquee shown while reasoning output is
+        # suppressed. Created up-front so callers can ``start``/``stop``
+        # idempotently regardless of the show_reasoning flag.
+        self._thinking = _ThinkingIndicator(
+            write_fn=self._write,
+            gray_fn=self._gray,
+        )
 
     @contextmanager
     def act_log(self, observation: dict[str, Any]) -> Iterator[None]:
@@ -106,13 +228,32 @@ class AIAgentConsole:
 
     def begin_llm_stream(self) -> None:
         self._write(self._gray("AI MODEL STREAM\n"))
+        # When reasoning output is hidden, immediately start the marquee
+        # so the user sees liveness while the model is "thinking" before
+        # any content delta arrives.
+        if not self._show_reasoning:
+            self._thinking.start()
 
     def write_llm_delta(self, kind: str, text: str) -> None:
         if not text:
             return
+        # Skip "reasoning" deltas entirely when reasoning output is hidden;
+        # this keeps the terminal focused on the final answer for users
+        # who do not care about chain-of-thought traces.
+        if kind == "reasoning" and not self._show_reasoning:
+            return
+        # First non-reasoning delta means the model has started speaking
+        # the actual answer; tear down the marquee before printing so the
+        # animation does not collide with the content stream.
+        if kind == "content" and not self._show_reasoning:
+            self._thinking.stop()
         self._write(self._gray(text))
 
     def end_llm_stream(self) -> None:
+        # Defensive stop in case the request finished without ever
+        # producing a content delta (e.g. fallback path / error).
+        if not self._show_reasoning:
+            self._thinking.stop()
         self._write(self._gray("\n"))
 
     def announce_action(
@@ -120,11 +261,17 @@ class AIAgentConsole:
         action: dict[str, Any],
         source: str = "model",
     ) -> None:
+        # Defensive stop: error / fallback paths bypass end_llm_stream, so
+        # we ensure the marquee never leaks into action / warning output.
+        self._thinking.stop()
         body = json.dumps(action, ensure_ascii=False)
         self._write(f"\nAI ACTION ({source}) -> {body}\n")
         self._write("~" * 60 + "\n\n")
 
     def announce_warning(self, message: str) -> None:
+        # Same defensive stop as ``announce_action`` - warnings can fire
+        # before the LLM stream closes (HTTP error, JSON parse error...).
+        self._thinking.stop()
         self._write(f"\nAI WARNING -> {message}\n")
 
     def _gray(self, text: str) -> str:
@@ -133,8 +280,12 @@ class AIAgentConsole:
         return f"{ANSI_GRAY}{text}{ANSI_RESET}"
 
     def _write(self, text: str) -> None:
-        self._output.write(text)
+        # The thinking-indicator background thread writes from a different
-        self._output.flush()
+        # thread than the main /act handler; the fine-grained ``_io_lock``
+        # avoids tearing of escape sequences and keeps stdout consistent.
+        with self._io_lock:
+            self._output.write(text)
+            self._output.flush()
 
 
 # ---------------------------------------------------------------------------
@@ -296,8 +447,11 @@ def _format_action_history(history: list[dict[str, Any]]) -> str:
         return "(no actions yet)"
     # The engine never produces unbounded history within a single hand, but
     # we cap defensively so a malformed payload cannot blow up token usage.
+    # The cap is sized to comfortably cover the worst realistic case (a
+    # 12-handed table running ~10 betting rounds within one hand) so the
+    # LLM never sees a silently truncated history at full ring tables.
     rows = []
-    for record in history[-32:]:
+    for record in history[-128:]:
         rows.append(
             f"- [{record.get('street')}] {record.get('player_id')} -> "
             f"{record.get('action')} amount={record.get('amount', 0)}"
@@ -927,6 +1081,16 @@ def main() -> None:
         action="store_true",
         help="Disable ANSI gray coloring for streamed LLM output.",
     )
+    parser.add_argument(
+        "--hide-reasoning",
+        action="store_true",
+        help=(
+            "Hide the LLM's reasoning/chain-of-thought stream from the "
+            "terminal. The model still performs reasoning; only its "
+            "terminal output is suppressed. The final answer (content) "
+            "is still printed so operators can see the chosen action."
+        ),
+    )
     args = parser.parse_args()
 
     if not args.api_key:
@@ -944,6 +1108,7 @@ def main() -> None:
     console = AIAgentConsole(
         keep_history=args.keep_history,
         use_color=not args.no_color,
+        show_reasoning=not args.hide_reasoning,
     )
     service = AIAgentService(LLMClient(config), prompts, console=console)
     server = create_server(args.host, args.port, service, default_player_id=args.player_id)
@@ -956,6 +1121,7 @@ def main() -> None:
         f"  base_url    : {config.base_url}\n"
         f"  player_id   : {args.player_id}\n"
         f"  stream      : {'on' if config.stream else 'off'}\n"
+        f"  reasoning   : {'hidden (output suppressed)' if args.hide_reasoning else 'visible'}\n"
         f"  clear-screen: {'off (keep history)' if args.keep_history else 'on'}",
         file=sys.stderr,
         flush=True,

texas_holdem/engine.py

@@ -1,6 +1,8 @@
 from __future__ import annotations
 
+from copy import deepcopy
 from random import Random
+from threading import RLock
 from time import time
 
 from texas_holdem.agents import PokerAgent
@@ -8,6 +10,7 @@ from texas_holdem.cards import Deck
 from texas_holdem.evaluator import evaluate
 from texas_holdem.models import (
     ActionRecord,
+    BlindLevel,
     HandSummary,
     Observation,
     PlayerAction,
@@ -53,21 +56,59 @@ class TableGame:
         self.small_blind = small_blind
         self.big_blind = big_blind
         self.rng = rng or Random()
+        self.lock = RLock()
         self.hand_number = 0
         self.button_index: int | None = None
         self.board = []
         self.action_history: list[ActionRecord] = []
         self.hand_summaries: list[HandSummary] = []
+        # ``blind_history`` is an append-only log of every blind level change
+        # (including the initial one). Each entry's ``hand_number`` is the
+        # first hand that played under those stakes, which makes it trivial
+        # to reconstruct the schedule from the outside.
+        self.blind_history: list[BlindLevel] = []
+        self._completed_snapshot: dict[str, object] = self._to_dict_unlocked()
 
     @property
     def is_complete(self) -> bool:
         return len([player for player in self.players if player.stack > 0]) < 2
 
-    def run_hand(self) -> HandSummary:
+    def run_hand(
+        self,
+        small_blind: int | None = None,
+        big_blind: int | None = None,
+    ) -> HandSummary:
+        """Play a single hand.
+
+        ``small_blind`` / ``big_blind`` allow callers to bump the stakes
+        between hands without rebuilding the table. Either both must be
+        provided or both omitted (in which case the previously configured
+        blinds carry over). The resolved blind level is appended to
+        :attr:`blind_history` whenever it changes (including the very first
+        hand) so external observers can replay the schedule.
+        """
+        with self.lock:
+            return self._run_hand_locked(small_blind=small_blind, big_blind=big_blind)
+
+    def _run_hand_locked(
+        self,
+        small_blind: int | None = None,
+        big_blind: int | None = None,
+    ) -> HandSummary:
         if self.is_complete:
             raise GameComplete("game is complete")
 
+        self._apply_blinds_for_hand(small_blind, big_blind)
+
         self.hand_number += 1
+        # Stamp the active blind level onto the upcoming summary so a hand
+        # remains self-describing even after the blinds change later on.
+        active_blinds = BlindLevel(
+            hand_number=self.hand_number,
+            small_blind=self.small_blind,
+            big_blind=self.big_blind,
+        )
+        self._record_blind_level_if_new(active_blinds)
         started_at = time()
         self.board = []
         self.action_history = []
@@ -116,26 +157,69 @@ class TableGame:
             board=list(self.board),
             actions=list(self.action_history),
             awards=awards,
+            blinds=active_blinds,
             showdown_hands=self._collect_showdown_hands(),
             started_at=started_at,
             finished_at=time(),
         )
         self.hand_summaries.append(summary)
+        self._completed_snapshot = deepcopy(self._to_dict_unlocked())
         return summary
 
-    def run_hands(self, max_hands: int, until_one_left: bool = False) -> list[HandSummary]:
+    def run_hands(
-        if max_hands <= 0:
+        self,
-            raise ValueError("max_hands must be positive")
+        max_hands: int,
-        summaries = []
+        until_one_left: bool = False,
-        for _ in range(max_hands):
+        small_blind: int | None = None,
-            if self.is_complete:
+        big_blind: int | None = None,
-                break
+    ) -> list[HandSummary]:
-            summaries.append(self.run_hand())
+        """Play up to ``max_hands`` hands using a single blind configuration.
-            if until_one_left and self.is_complete:
+
-                break
+        Passing ``small_blind`` / ``big_blind`` bumps the stakes starting
-        return summaries
+        with the first hand of this call; subsequent calls can raise them
+        again. Leaving them ``None`` keeps the current level unchanged.
+        """
+        with self.lock:
+            if max_hands <= 0:
+                raise ValueError("max_hands must be positive")
+            summaries = []
+            for _ in range(max_hands):
+                if self.is_complete:
+                    break
+                # Only the first hand of the batch needs to apply the blind
+                # override; after that the engine reuses the stored values.
+                summaries.append(
+                    self._run_hand_locked(
+                        small_blind=small_blind,
+                        big_blind=big_blind,
+                    )
+                )
+                small_blind = None
+                big_blind = None
+                if until_one_left and self.is_complete:
+                    break
+            return summaries
 
     def to_dict(self) -> dict[str, object]:
+        with self.lock:
+            return self._to_dict_unlocked()
+
+    def snapshot_completed(self) -> dict[str, object]:
+        """Return a stable snapshot from the latest completed hand boundary.
+
+        If a hand is currently running under ``self.lock``, this method does
+        not block. It returns the most recent completed hand summary and
+        stacks captured in memory, which is exactly what status endpoints
+        need while a long-running HTTP-agent decision is in progress.
+        """
+        if self.lock.acquire(blocking=False):
+            try:
+                return deepcopy(self._to_dict_unlocked())
+            finally:
+                self.lock.release()
+        return deepcopy(self._completed_snapshot)
+
+    def _to_dict_unlocked(self) -> dict[str, object]:
         return {
             "game_id": self.game_id,
             "status": "complete" if self.is_complete else "running",
@@ -143,8 +227,18 @@ class TableGame:
             "button_seat": None
             if self.button_index is None
             else self.players[self.button_index].seat,
+            # ``small_blind`` / ``big_blind`` mirror the *current* level so
+            # legacy callers keep working. New consumers should prefer the
+            # structured ``blinds`` block which carries the full schedule.
             "small_blind": self.small_blind,
             "big_blind": self.big_blind,
+            "blinds": {
+                "current": {
+                    "small_blind": self.small_blind,
+                    "big_blind": self.big_blind,
+                },
+                "history": [level.to_dict() for level in self.blind_history],
+            },
             "starting_stack": self.starting_stack,
             "players": [player.public_dict() for player in self.players],
             # ``hands`` exposes every finished hand (each entry is the same
@@ -153,6 +247,47 @@ class TableGame:
             "hands": [summary.to_dict() for summary in self.hand_summaries],
         }
 
+    def _apply_blinds_for_hand(
+        self,
+        small_blind: int | None,
+        big_blind: int | None,
+    ) -> None:
+        """Validate and apply optional per-hand blind overrides.
+
+        Splitting this out keeps :meth:`run_hand` focused on the table flow
+        while letting us reuse the validation rules originally enforced by
+        ``__init__``. We require both values to be supplied together so the
+        configuration cannot drift into an inconsistent half-update.
+        """
+        if small_blind is None and big_blind is None:
+            return
+        if small_blind is None or big_blind is None:
+            raise ValueError(
+                "small_blind and big_blind must be provided together"
+            )
+        if small_blind <= 0 or big_blind <= 0 or small_blind > big_blind:
+            raise ValueError("blinds must satisfy 0 < small_blind <= big_blind")
+        self.small_blind = int(small_blind)
+        self.big_blind = int(big_blind)
+
+    def _record_blind_level_if_new(self, level: BlindLevel) -> None:
+        """Append ``level`` to :attr:`blind_history` when it differs.
+
+        Comparing against the latest entry (rather than blindly appending)
+        keeps the log compact: stretches of unchanged stakes only contribute
+        a single record. The very first hand always seeds an entry because
+        the history starts empty.
+        """
+        if not self.blind_history:
+            self.blind_history.append(level)
+            return
+        latest = self.blind_history[-1]
+        if (
+            latest.small_blind != level.small_blind
+            or latest.big_blind != level.big_blind
+        ):
+            self.blind_history.append(level)
+
     def _advance_button(self) -> None:
         if self.button_index is None:
             self.button_index = self._next_index(0, lambda index: self.players[index].stack > 0)
@@ -350,9 +485,18 @@ class TableGame:
         try:
             requested = agent.decide(observation)
         except Exception:
-            requested = PlayerAction("fold")
+            requested = self._default_action(observation.legal_actions)
         return self._coerce_action(requested, observation.legal_actions)
 
+    def _default_action(self, legal_actions: list[dict[str, object]]) -> PlayerAction:
+        by_action = {str(action["action"]): action for action in legal_actions}
+        for action_type in ("check", "call", "fold"):
+            if action_type in by_action:
+                legal = by_action[action_type]
+                return PlayerAction(action_type, int(legal.get("amount") or 0))
+        legal = legal_actions[0]
+        return PlayerAction(str(legal["action"]), int(legal.get("amount") or 0))
+
     def _coerce_action(
         self,
         requested: PlayerAction,
@@ -482,7 +626,7 @@ class TableGame:
         swallow individual exceptions so a flaky remote endpoint cannot
         break the table flow.
         """
-        snapshot = self.to_dict()
+        snapshot = self._to_dict_unlocked()
         for agent in self.agents.values():
             try:
                 agent.on_game_update(snapshot)

texas_holdem/models.py

@@ -130,6 +130,28 @@ class Observation:
         }
 
 
+@dataclass(slots=True)
+class BlindLevel:
+    """A snapshot of the blind configuration that took effect at a given hand.
+
+    The structure is intentionally append-only: every time the blinds change
+    (or the very first hand seeds the initial values) we push a new
+    ``BlindLevel`` so callers can reconstruct how the stakes evolved over the
+    course of the game without losing any prior state.
+    """
+
+    hand_number: int
+    small_blind: int
+    big_blind: int
+
+    def to_dict(self) -> dict[str, object]:
+        return {
+            "hand_number": self.hand_number,
+            "small_blind": self.small_blind,
+            "big_blind": self.big_blind,
+        }
+
+
 @dataclass(slots=True)
 class PotAward:
     amount: int
@@ -152,6 +174,10 @@ class HandSummary:
     board: list[Card]
     actions: list[ActionRecord]
     awards: list[PotAward]
+    # ``blinds`` records the exact blind level used by this hand. Storing it
+    # on the summary (rather than only on the game) guarantees historical
+    # hands remain self-describing even after the blinds are raised later.
+    blinds: BlindLevel | None = None
     showdown_hands: dict[str, list[Card]] = field(default_factory=dict)
     started_at: float = field(default_factory=time)
     finished_at: float = field(default_factory=time)
@@ -161,6 +187,7 @@ class HandSummary:
             "game_id": self.game_id,
             "hand_number": self.hand_number,
             "button_seat": self.button_seat,
+            "blinds": self.blinds.to_dict() if self.blinds else None,
             "board": [str(card) for card in self.board],
             "actions": [record.to_dict() for record in self.actions],
             "awards": [award.to_dict() for award in self.awards],

texas_holdem/server.py

@@ -25,8 +25,8 @@ class PokerRequestHandler(BaseHTTPRequestHandler):
             if path == ["games"]:
                 self._json({"games": MANAGER.list_games()})
                 return
-            if len(path) == 2 and path[0] == "games":
+            if len(path) == 2 and path[0] in {"game", "games"}:
-                self._json(MANAGER.get_game(path[1]).to_dict())
+                self._json(MANAGER.get_game_state(path[1]))
                 return
             self._json({"error": "not found"}, HTTPStatus.NOT_FOUND)
         except KeyError as exc:
@@ -35,23 +35,37 @@ class PokerRequestHandler(BaseHTTPRequestHandler):
     def do_POST(self) -> None:
         path = self._path_parts()
         try:
-            if path == ["games"]:
+            if path in (["game"], ["games"]):
                 game = MANAGER.create_game(self._read_json())
-                self._json(game.to_dict(), HTTPStatus.CREATED)
+                self._json(game.snapshot_completed(), HTTPStatus.CREATED)
                 return
-            if len(path) == 3 and path[0] == "games" and path[2] == "hands":
+            if len(path) == 3 and path[0] in {"game", "games"} and path[2] == "hands":
                 body = self._read_json()
                 count = int(body.get("count", 1))
                 until_one_left = bool(body.get("until_one_left", False))
-                summaries = MANAGER.run_hands(path[1], count, until_one_left)
+                small_blind, big_blind = self._extract_blinds(body)
-                self._json({"hands": summaries, "game": MANAGER.get_game(path[1]).to_dict()})
+                summaries = MANAGER.run_hands(
+                    path[1],
+                    count,
+                    until_one_left,
+                    small_blind=small_blind,
+                    big_blind=big_blind,
+                )
+                self._json({"hands": summaries, "game": MANAGER.get_game_state(path[1])})
                 return
-            if len(path) == 4 and path[0] == "games" and path[2] == "hands" and path[3] == "run":
+            if len(path) == 4 and path[0] in {"game", "games"} and path[2] == "hands" and path[3] == "run":
                 body = self._read_json()
                 count = int(body.get("count", 1))
                 until_one_left = bool(body.get("until_one_left", False))
-                summaries = MANAGER.run_hands(path[1], count, until_one_left)
+                small_blind, big_blind = self._extract_blinds(body)
-                self._json({"hands": summaries, "game": MANAGER.get_game(path[1]).to_dict()})
+                summaries = MANAGER.run_hands(
+                    path[1],
+                    count,
+                    until_one_left,
+                    small_blind=small_blind,
+                    big_blind=big_blind,
+                )
+                self._json({"hands": summaries, "game": MANAGER.get_game_state(path[1])})
                 return
             self._json({"error": "not found"}, HTTPStatus.NOT_FOUND)
         except KeyError as exc:
@@ -78,6 +92,25 @@ class PokerRequestHandler(BaseHTTPRequestHandler):
             raise ValueError("request body must be a JSON object")
         return payload
 
+    @staticmethod
+    def _extract_blinds(body: dict[str, Any]) -> tuple[int | None, int | None]:
+        """Parse optional blind overrides from a /hands POST body.
+
+        Callers may omit both keys (keep current level), or supply both to
+        raise the blinds for the upcoming batch. Providing only one is
+        treated as a client error and surfaced via ``ValueError`` so the
+        handler can reply with 400.
+        """
+        raw_small = body.get("small_blind")
+        raw_big = body.get("big_blind")
+        if raw_small is None and raw_big is None:
+            return None, None
+        if raw_small is None or raw_big is None:
+            raise ValueError(
+                "small_blind and big_blind must be provided together"
+            )
+        return int(raw_small), int(raw_big)
+
     def _json(self, payload: dict[str, Any], status: HTTPStatus = HTTPStatus.OK) -> None:
         body = json.dumps(payload, ensure_ascii=True).encode("utf-8")
         self.send_response(status)

texas_holdem/service.py

@@ -5,13 +5,14 @@ from threading import RLock
 from typing import Any
 from uuid import uuid4
 
-from texas_holdem.agents import build_agent
+from texas_holdem.agents import build_agent, http_agent_endpoint_from_spec
 from texas_holdem.engine import TableGame
 
 
 class GameManager:
     def __init__(self) -> None:
         self._games: dict[str, TableGame] = {}
+        self._http_endpoint_owners: dict[str, str] = {}
         self._lock = RLock()
 
     def create_game(self, payload: dict[str, Any]) -> TableGame:
@@ -29,12 +30,19 @@ class GameManager:
         big_blind = int(payload.get("big_blind", 10))
 
         specs = []
+        http_endpoints: set[str] = set()
         for seat, raw_spec in enumerate(players):
             if not isinstance(raw_spec, dict):
                 raise ValueError("each player must be an object")
             player_id = str(raw_spec.get("id") or raw_spec.get("player_id") or f"p{seat + 1}")
             name = str(raw_spec.get("name") or player_id)
-            agent = build_agent(raw_spec.get("agent", raw_spec), rng)
+            agent_spec = raw_spec.get("agent", raw_spec)
+            if not isinstance(agent_spec, dict):
+                raise ValueError("agent spec must be an object")
+            endpoint = http_agent_endpoint_from_spec(agent_spec)
+            if endpoint is not None:
+                http_endpoints.add(endpoint)
+            agent = build_agent(agent_spec, rng, player_id=player_id)
             specs.append((player_id, name, agent))
 
         game = TableGame(
@@ -46,9 +54,18 @@ class GameManager:
             rng=rng,
         )
         with self._lock:
+            self._release_completed_http_endpoints_locked()
             if game_id in self._games:
                 raise ValueError(f"game already exists: {game_id}")
+            for endpoint in http_endpoints:
+                owner = self._http_endpoint_owners.get(endpoint)
+                if owner is not None and owner != game_id:
+                    raise ValueError(
+                        f"http agent endpoint already belongs to game {owner}: {endpoint}"
+                    )
             self._games[game_id] = game
+            for endpoint in http_endpoints:
+                self._http_endpoint_owners[endpoint] = game_id
         return game
 
     def get_game(self, game_id: str) -> TableGame:
@@ -58,14 +75,54 @@ class GameManager:
             except KeyError as exc:
                 raise KeyError(f"game not found: {game_id}") from exc
 
+    def get_game_state(self, game_id: str) -> dict[str, object]:
+        return self.get_game(game_id).snapshot_completed()
+
     def list_games(self) -> list[dict[str, object]]:
         with self._lock:
-            return [game.to_dict() for game in self._games.values()]
+            games = list(self._games.values())
+        return [game.snapshot_completed() for game in games]
 
-    def run_hands(self, game_id: str, count: int = 1, until_one_left: bool = False) -> list[dict[str, object]]:
+    def run_hands(
+        self,
+        game_id: str,
+        count: int = 1,
+        until_one_left: bool = False,
+        small_blind: int | None = None,
+        big_blind: int | None = None,
+    ) -> list[dict[str, object]]:
+        """Run ``count`` hands, optionally raising the blinds first.
+
+        ``small_blind`` / ``big_blind`` are forwarded to the engine so the
+        blinds can change between batches. Leaving them as ``None`` keeps
+        the previously configured level, which preserves the original
+        no-argument behaviour.
+        """
         game = self.get_game(game_id)
-        with self._lock:
+        summaries = [
-            return [
+            summary.to_dict()
-                summary.to_dict()
+            for summary in game.run_hands(
-                for summary in game.run_hands(count, until_one_left=until_one_left)
+                count,
-            ]
+                until_one_left=until_one_left,
+                small_blind=small_blind,
+                big_blind=big_blind,
+            )
+        ]
+        if game.is_complete:
+            with self._lock:
+                self._release_http_endpoints_for_game_locked(game_id)
+        return summaries
+
+    def _release_completed_http_endpoints_locked(self) -> None:
+        for game_id, game in list(self._games.items()):
+            if game.lock.acquire(blocking=False):
+                try:
+                    if game.is_complete:
+                        self._release_http_endpoints_for_game_locked(game_id)
+                finally:
+                    game.lock.release()
+
+    def _release_http_endpoints_for_game_locked(self, game_id: str) -> None:
+        for endpoint, owner in list(self._http_endpoint_owners.items()):
+            if owner == game_id:
+                del self._http_endpoint_owners[endpoint]

texas_holdem_replay/__init__.py

@@ -0,0 +1,2 @@
+"""Standalone web replay viewer for Texas Hold X game snapshots."""
+

texas_holdem_replay/server.py

@@ -0,0 +1,153 @@
+from __future__ import annotations
+
+import argparse
+import json
+import mimetypes
+from http import HTTPStatus
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+from pathlib import Path
+from typing import Any
+from urllib.error import HTTPError, URLError
+from urllib.parse import quote, unquote, urlparse, parse_qs
+from urllib.request import Request, urlopen
+
+
+STATIC_DIR = Path(__file__).with_name("static")
+
+
+def build_game_url(query: dict[str, list[str]]) -> str:
+    raw_url = _first(query, "url")
+    if raw_url:
+        return _validate_http_url(raw_url)
+
+    base_url = _first(query, "base_url")
+    game_id = _first(query, "game_id")
+    if not base_url or not game_id:
+        raise ValueError("provide either url or both base_url and game_id")
+
+    base = _validate_http_url(base_url).rstrip("/")
+    safe_game_id = quote(game_id.strip("/"), safe="")
+    return f"{base}/games/{safe_game_id}"
+
+
+class ReplayRequestHandler(BaseHTTPRequestHandler):
+    server_version = "TexasHoldemReplay/0.1"
+
+    def do_GET(self) -> None:
+        parsed = urlparse(self.path)
+        if parsed.path == "/api/health":
+            self._json({"ok": True, "service": "texas-holdem-replay"})
+            return
+        if parsed.path == "/api/fetch-game":
+            self._handle_fetch_game(parsed.query)
+            return
+        self._serve_static(parsed.path)
+
+    def do_OPTIONS(self) -> None:
+        self.send_response(HTTPStatus.NO_CONTENT)
+        self._cors_headers()
+        self.end_headers()
+
+    def log_message(self, format: str, *args: Any) -> None:
+        return
+
+    def _handle_fetch_game(self, raw_query: str) -> None:
+        query = parse_qs(raw_query)
+        try:
+            target = build_game_url(query)
+            timeout = float(_first(query, "timeout") or 8)
+            request = Request(target, headers={"Accept": "application/json"})
+            with urlopen(request, timeout=max(1, min(timeout, 30))) as response:
+                payload = response.read()
+                content_type = response.headers.get("Content-Type", "")
+            try:
+                data = json.loads(payload.decode("utf-8"))
+            except json.JSONDecodeError as exc:
+                raise ValueError("target did not return valid JSON") from exc
+            if not isinstance(data, dict):
+                raise ValueError("target JSON must be an object")
+            self._json({"source": target, "content_type": content_type, "game": data})
+        except HTTPError as exc:
+            body = exc.read().decode("utf-8", errors="replace")[:600]
+            self._json(
+                {"error": f"upstream returned HTTP {exc.code}", "detail": body},
+                HTTPStatus.BAD_GATEWAY,
+            )
+        except (URLError, TimeoutError) as exc:
+            self._json({"error": "failed to reach upstream", "detail": str(exc)}, HTTPStatus.BAD_GATEWAY)
+        except ValueError as exc:
+            self._json({"error": str(exc)}, HTTPStatus.BAD_REQUEST)
+
+    def _serve_static(self, raw_path: str) -> None:
+        relative = unquote(raw_path.lstrip("/")) or "index.html"
+        if relative.endswith("/"):
+            relative += "index.html"
+        candidate = (STATIC_DIR / relative).resolve()
+        root = STATIC_DIR.resolve()
+        if root not in candidate.parents and candidate != root:
+            self._json({"error": "not found"}, HTTPStatus.NOT_FOUND)
+            return
+        if not candidate.is_file():
+            candidate = STATIC_DIR / "index.html"
+        body = candidate.read_bytes()
+        content_type = mimetypes.guess_type(candidate.name)[0] or "application/octet-stream"
+        self.send_response(HTTPStatus.OK)
+        self._cors_headers()
+        self.send_header("Content-Type", content_type)
+        self.send_header("Cache-Control", "no-store" if candidate.name == "index.html" else "public, max-age=60")
+        self.send_header("Content-Length", str(len(body)))
+        self.end_headers()
+        self.wfile.write(body)
+
+    def _json(self, payload: dict[str, Any], status: HTTPStatus = HTTPStatus.OK) -> None:
+        body = json.dumps(payload, ensure_ascii=False).encode("utf-8")
+        self.send_response(status)
+        self._cors_headers()
+        self.send_header("Content-Type", "application/json; charset=utf-8")
+        self.send_header("Content-Length", str(len(body)))
+        self.end_headers()
+        self.wfile.write(body)
+
+    def _cors_headers(self) -> None:
+        self.send_header("Access-Control-Allow-Origin", "*")
+        self.send_header("Access-Control-Allow-Methods", "GET, OPTIONS")
+        self.send_header("Access-Control-Allow-Headers", "Content-Type")
+
+
+def create_server(host: str, port: int) -> ThreadingHTTPServer:
+    return ThreadingHTTPServer((host, port), ReplayRequestHandler)
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="Run the standalone Texas Hold X replay web viewer.")
+    parser.add_argument("--host", default="127.0.0.1")
+    parser.add_argument("--port", default=8088, type=int)
+    args = parser.parse_args()
+
+    server = create_server(args.host, args.port)
+    print(f"Texas Hold X replay viewer listening on http://{args.host}:{args.port}")
+    try:
+        server.serve_forever()
+    except KeyboardInterrupt:
+        pass
+    finally:
+        server.server_close()
+
+
+def _first(query: dict[str, list[str]], key: str) -> str | None:
+    values = query.get(key)
+    if not values:
+        return None
+    value = values[0].strip()
+    return value or None
+
+
+def _validate_http_url(value: str) -> str:
+    parsed = urlparse(value.strip())
+    if parsed.scheme not in {"http", "https"} or not parsed.netloc:
+        raise ValueError("url must be an absolute http(s) URL")
+    return value.strip()
+
+
+if __name__ == "__main__":
+    main()

texas_holdem_replay/static/app.js

@@ -0,0 +1,832 @@
+// =============================================================================
+// Texas Hold X Replay — viewer logic
+// -----------------------------------------------------------------------------
+// Architecture overview:
+//   * `state`   — single mutable runtime store; never read directly by render
+//                 helpers, instead passed via the active frame snapshot.
+//   * `el`      — cached DOM references (set once on load, not per render).
+//   * Frames    — pre-computed, immutable per-hand snapshots produced by
+//                 buildFrames(). Each frame carries the full visual state
+//                 (board cards, pot, players, active speaker, action kind),
+//                 making renderFrame() a pure function of state.frameIndex.
+//   * Avatars   — pixel-art sprites styled in CSS via 8x8 box-shadow grids.
+//                 Per-player palette is derived deterministically from
+//                 player_id so the same player always looks identical.
+//   * Animation — seat sprites get a transient `act-${kind}` class that maps
+//                 to an avatar @keyframes animation in styles.css.
+// =============================================================================
+
+const state = {
+  game: null,
+  hands: [],
+  currentHandIndex: 0,
+  frames: [],
+  frameIndex: 0,
+  playing: false,
+  timer: null,
+  pollTimer: null,
+  source: "",
+};
+
+const el = {
+  subtitle: document.querySelector("#subtitle"),
+  sourceBadge: document.querySelector("#sourceBadge"),
+  pollBadge: document.querySelector("#pollBadge"),
+  serverUrl: document.querySelector("#serverUrl"),
+  gameId: document.querySelector("#gameId"),
+  fetchBtn: document.querySelector("#fetchBtn"),
+  fileInput: document.querySelector("#fileInput"),
+  autoPoll: document.querySelector("#autoPoll"),
+  pollSeconds: document.querySelector("#pollSeconds"),
+  handSelect: document.querySelector("#handSelect"),
+  pace: document.querySelector("#pace"),
+  prevBtn: document.querySelector("#prevBtn"),
+  playBtn: document.querySelector("#playBtn"),
+  nextBtn: document.querySelector("#nextBtn"),
+  resetBtn: document.querySelector("#resetBtn"),
+  progressBar: document.querySelector("#progressBar"),
+  gameStatus: document.querySelector("#gameStatus"),
+  playerCount: document.querySelector("#playerCount"),
+  handCount: document.querySelector("#handCount"),
+  blindLevel: document.querySelector("#blindLevel"),
+  handBadge: document.querySelector("#handBadge"),
+  streetLabel: document.querySelector("#streetLabel"),
+  potLabel: document.querySelector("#potLabel"),
+  boardCards: document.querySelector("#boardCards"),
+  tableMessage: document.querySelector("#tableMessage"),
+  seatLayer: document.querySelector("#seatLayer"),
+  eventLog: document.querySelector("#eventLog"),
+};
+
+// ---------------------------------------------------------------------------
+// Static label tables. Kept module-scope so render functions are fully pure.
+// ---------------------------------------------------------------------------
+const STREET_LABELS = {
+  setup: "准备",
+  preflop: "翻牌前",
+  flop: "翻牌",
+  turn: "转牌",
+  river: "河牌",
+  showdown: "摊牌",
+  awards: "结算",
+};
+
+const ACTION_LABELS = {
+  small_blind: "小盲",
+  big_blind: "大盲",
+  fold: "弃牌",
+  check: "过牌",
+  call: "跟注",
+  bet: "下注",
+  raise: "加注",
+  all_in: "全下",
+  award: "赢得",
+  showdown: "亮牌",
+  deal: "发牌",
+};
+
+// Bubble color category — bucket many actions into a few visual classes.
+// (The .speech.kind-* CSS classes paint distinct backgrounds for fold/
+// call/check/bet/raise/all_in/award. Blinds are rendered as bet-style.)
+
+// ---------------------------------------------------------------------------
+// Pixel avatar palette — picked from a small 8-bit-friendly set so every
+// player is visually distinct. Derivation is deterministic via FNV-1a over
+// player_id, so a given player always renders with the same colors across
+// frames and hands.
+// ---------------------------------------------------------------------------
+const AVATAR_PALETTE = {
+  skin:  ["#f7c98a", "#e2a96b", "#c98c5a", "#8b5a3c", "#f2d4ad"],
+  hair:  ["#2a1e16", "#5b3a23", "#a85d2a", "#d8a13a", "#7a3b8e", "#3a4d8a", "#b9362f"],
+  shirt: ["#c44c4c", "#3a7fbf", "#3a9b62", "#b87a1f", "#7a3b8e", "#2a3a5a", "#d8a13a"],
+};
+
+/**
+ * Hash a string into a non-negative 32-bit integer using FNV-1a.
+ * Used to deterministically pick avatar palette entries from player_id.
+ */
+function fnv1a(value) {
+  let hash = 0x811c9dc5;
+  const text = String(value);
+  for (let i = 0; i < text.length; i += 1) {
+    hash ^= text.charCodeAt(i);
+    hash = Math.imul(hash, 0x01000193) >>> 0;
+  }
+  return hash >>> 0;
+}
+
+/**
+ * Pick a deterministic avatar palette for the given player_id.
+ * Returns an object with { skin, hair, shirt } CSS colors.
+ */
+function avatarPalette(playerId) {
+  const hash = fnv1a(playerId);
+  const pickFrom = (list, salt) => list[(hash >>> salt) % list.length];
+  return {
+    skin:  pickFrom(AVATAR_PALETTE.skin, 0),
+    hair:  pickFrom(AVATAR_PALETTE.hair, 5),
+    shirt: pickFrom(AVATAR_PALETTE.shirt, 11),
+  };
+}
+
+// 8x8 sprite map for the player portrait.
+//   . = transparent / dark frame backdrop
+//   H = hair    S = skin    E = eye    M = mouth (red)    T = shirt body
+// Two side hair pixels on rows 2-3 give a cohesive helmet shape; row 5 mouth
+// adds personality. Row 7 is full shirt body so the avatar reads as a bust.
+const AVATAR_SPRITE = [
+  ".HHHHHH.",
+  "HHHHHHHH",
+  "HSSSSSSH",
+  ".SESSESS",
+  ".SSSSSS.",
+  ".SSMMSS.",
+  ".TTTTTT.",
+  "TTTTTTTT",
+];
+
+/**
+ * Build an inline SVG string for a pixel-art avatar using the player's
+ * deterministic palette. Each filled cell becomes a 1x1 <rect> in an 8x8
+ * viewBox; SVG `shape-rendering="crispEdges"` keeps the squares sharp.
+ */
+function avatarSvg(playerId) {
+  const palette = avatarPalette(playerId);
+  const colors = {
+    H: palette.hair,
+    S: palette.skin,
+    E: "#1c0e08",
+    M: "#b9362f",
+    T: palette.shirt,
+  };
+  let rects = "";
+  for (let y = 0; y < AVATAR_SPRITE.length; y += 1) {
+    const row = AVATAR_SPRITE[y];
+    for (let x = 0; x < row.length; x += 1) {
+      const ch = row[x];
+      const fill = colors[ch];
+      if (!fill) continue;
+      rects += `<rect x="${x}" y="${y}" width="1" height="1" fill="${fill}"/>`;
+    }
+  }
+  return `<svg viewBox="0 0 8 8" preserveAspectRatio="xMidYMid meet" xmlns="http://www.w3.org/2000/svg" shape-rendering="crispEdges">${rects}</svg>`;
+}
+
+// ---------------------------------------------------------------------------
+// Card rendering helpers.
+// ---------------------------------------------------------------------------
+function cardParts(raw) {
+  if (!raw || raw === "back") return { rank: "", suit: "", red: false, back: true };
+  const suitCode = raw.slice(-1).toLowerCase();
+  const rank = raw.slice(0, -1).toUpperCase();
+  const suits = { h: "♥", d: "♦", c: "♣", s: "♠" };
+  return {
+    rank,
+    suit: suits[suitCode] || suitCode,
+    red: suitCode === "h" || suitCode === "d",
+    back: false,
+  };
+}
+
+function renderCard(raw) {
+  const parts = cardParts(raw);
+  const card = document.createElement("div");
+  card.className = `card${parts.red ? " red" : ""}${parts.back ? " back" : ""}`;
+  if (!parts.back) {
+    card.innerHTML = `<span class="rank">${parts.rank}</span><span class="suit">${parts.suit}</span><span class="rank bottom">${parts.rank}</span>`;
+  }
+  return card;
+}
+
+function money(value) {
+  return Number(value || 0).toLocaleString("en-US");
+}
+
+// ---------------------------------------------------------------------------
+// Game normalization: hydrate the raw JSON into a uniform shape with
+// player list and per-hand starting stacks.
+// ---------------------------------------------------------------------------
+function uniquePlayers(game) {
+  const byId = new Map();
+  for (const player of game.players || []) {
+    byId.set(player.player_id, {
+      player_id: player.player_id,
+      name: player.name || player.player_id,
+      seat: Number(player.seat || 0),
+      stack: Number(player.stack || 0),
+    });
+  }
+  for (const hand of game.hands || []) {
+    for (const action of hand.actions || []) {
+      if (!byId.has(action.player_id)) {
+        byId.set(action.player_id, {
+          player_id: action.player_id,
+          name: action.player_id,
+          seat: byId.size,
+          stack: Number(game.starting_stack || 0),
+        });
+      }
+    }
+  }
+  return Array.from(byId.values()).sort((a, b) => a.seat - b.seat);
+}
+
+function inferHandStarts(game) {
+  const players = uniquePlayers(game);
+  const stacks = new Map(players.map((player) => [player.player_id, Number(game.starting_stack || player.stack || 0)]));
+  const starts = new Map();
+  for (const hand of game.hands || []) {
+    starts.set(hand.hand_number, new Map(stacks));
+    for (const action of hand.actions || []) {
+      if (stacks.has(action.player_id)) stacks.set(action.player_id, Number(action.stack || 0));
+    }
+    for (const award of hand.awards || []) {
+      const winners = award.winners || [];
+      const share = winners.length ? Math.floor(Number(award.amount || 0) / winners.length) : 0;
+      for (const winner of winners) stacks.set(winner, Number(stacks.get(winner) || 0) + share);
+    }
+  }
+  return starts;
+}
+
+function normalizeGame(raw) {
+  const game = raw.game && raw.game.hands ? raw.game : raw;
+  if (!game || !Array.isArray(game.hands)) throw new Error("JSON 中未找到 hands 数组");
+  const players = uniquePlayers(game);
+  const starts = inferHandStarts(game);
+  const hands = game.hands.map((hand) => ({ ...hand, startingStacks: starts.get(hand.hand_number) || new Map() }));
+  return { ...game, players, hands };
+}
+
+function clonePlayersForHand(game, hand) {
+  return game.players.map((player) => ({
+    ...player,
+    stack: Number(hand.startingStacks.get(player.player_id) ?? game.starting_stack ?? player.stack ?? 0),
+    folded: false,
+    all_in: false,
+    in_hand: Number(hand.startingStacks.get(player.player_id) ?? game.starting_stack ?? player.stack ?? 0) > 0,
+    street_bet: 0,
+    total_bet: 0,
+    hole_cards: [],
+  }));
+}
+
+// ---------------------------------------------------------------------------
+// Frame builder — converts a hand's chronological action list into a list of
+// fully-resolved frame snapshots. Each frame is independent and renderable.
+// ---------------------------------------------------------------------------
+function buildFrames(game, hand) {
+  const players = clonePlayersForHand(game, hand);
+  const byId = new Map(players.map((player) => [player.player_id, player]));
+  const frames = [];
+  const pot = { value: 0 };
+  let street = "preflop";
+  let boardCount = 0;
+
+  // `actionKind` is what drives the seat sprite animation classes. It maps
+  // the raw action verb (fold/call/raise/...) onto a CSS animation token.
+  const snapshot = (type, message, activePlayerId = null, key = `${type}:${frames.length}`, extras = {}) => ({
+    key: `hand:${hand.hand_number}:${key}`,
+    type,
+    message,
+    activePlayerId,
+    street,
+    pot: pot.value,
+    board: (hand.board || []).slice(0, boardCount),
+    players: players.map((player) => ({ ...player, hole_cards: [...player.hole_cards] })),
+    ...extras,
+  });
+
+  frames.push(snapshot("setup", `第 ${hand.hand_number} 手牌开始`, null, "setup"));
+
+  const revealStreet = (nextStreet) => {
+    const counts = { flop: 3, turn: 4, river: 5 };
+    const nextCount = counts[nextStreet] || boardCount;
+    for (const player of players) player.street_bet = 0;
+    if (nextCount > boardCount) {
+      street = nextStreet;
+      boardCount = Math.min(nextCount, (hand.board || []).length);
+      frames.push(snapshot("deal", `${STREET_LABELS[nextStreet]}发出`, null, `deal:${nextStreet}:${boardCount}`));
+    } else {
+      street = nextStreet;
+    }
+  };
+
+  for (const [actionIndex, action] of (hand.actions || []).entries()) {
+    if (action.street !== street && action.street !== "preflop") revealStreet(action.street);
+    const player = byId.get(action.player_id);
+    if (!player) continue;
+    const previousTotal = Number(player.total_bet || 0);
+    player.street_bet = Number(action.street_bet || 0);
+    player.stack = Number(action.stack || 0);
+    if (action.action === "fold") player.folded = true;
+    if (action.action === "all_in" || player.stack === 0) player.all_in = true;
+    player.total_bet = previousTotal + Math.max(0, Number(action.amount || 0));
+    pot.value += Math.max(0, Number(action.amount || 0));
+    frames.push(snapshot(
+      "action",
+      actionText(action),
+      action.player_id,
+      actionKey(action, actionIndex),
+      { actionKind: action.action, bubbleText: bubbleTextFor(action) },
+    ));
+  }
+
+  if ((hand.board || []).length > boardCount) {
+    for (const nextStreet of ["flop", "turn", "river"]) {
+      if (({ flop: 3, turn: 4, river: 5 }[nextStreet] || 0) > boardCount) revealStreet(nextStreet);
+    }
+  }
+
+  const shown = hand.showdown_hands || {};
+  for (const [playerId, cards] of Object.entries(shown)) {
+    const player = byId.get(playerId);
+    if (player) player.hole_cards = cards;
+  }
+  if (Object.keys(shown).length) {
+    street = "showdown";
+    frames.push(snapshot("showdown", "摊牌亮牌", null, `showdown:${Object.keys(shown).sort().join(",")}`));
+  }
+
+  street = "awards";
+  for (const [awardIndex, award] of (hand.awards || []).entries()) {
+    const winners = award.winners || [];
+    const share = winners.length ? Math.floor(Number(award.amount || 0) / winners.length) : 0;
+    const remainder = winners.length ? Number(award.amount || 0) % winners.length : 0;
+    winners.forEach((winnerId, index) => {
+      const player = byId.get(winnerId);
+      if (player) player.stack += share + (index < remainder ? 1 : 0);
+    });
+    pot.value = Math.max(0, pot.value - Number(award.amount || 0));
+    frames.push(snapshot(
+      "award",
+      awardText(award),
+      winners[0] || null,
+      awardKey(award, awardIndex),
+      { actionKind: "award", bubbleText: `赢得 ${money(award.amount)}` },
+    ));
+  }
+
+  if (frames.length === 1) frames.push(snapshot("empty", "这手牌没有可回放动作", null, "empty"));
+  return frames;
+}
+
+function actionKey(action, index) {
+  return [
+    "action", index, action.street, action.player_id, action.action,
+    Number(action.amount || 0), Number(action.street_bet || 0), Number(action.stack || 0),
+  ].join(":");
+}
+
+function awardKey(award, index) {
+  return [
+    "award", index, Number(award.amount || 0),
+    (award.winners || []).join(","),
+    award.hand_value?.name || "",
+  ].join(":");
+}
+
+function actionText(action) {
+  const label = ACTION_LABELS[action.action] || action.action;
+  if (["check", "fold"].includes(action.action)) return `${action.player_id} ${label}`;
+  return `${action.player_id} ${label} ${money(action.amount)}`;
+}
+
+/**
+ * Build a SHORT speech-bubble label (player_id is implicit since the bubble
+ * already points at the seat). Keeps text readable inside narrow bubbles.
+ */
+function bubbleTextFor(action) {
+  const label = ACTION_LABELS[action.action] || action.action;
+  if (["check", "fold"].includes(action.action)) return label;
+  return `${label} ${money(action.amount)}`;
+}
+
+function awardText(award) {
+  const winners = (award.winners || []).join(", ");
+  const handName = award.hand_value?.name ? ` · ${award.hand_value.name}` : "";
+  return `${winners} 赢得 ${money(award.amount)}${handName}`;
+}
+
+// ---------------------------------------------------------------------------
+// Seat layout — six radial presets, with mobile fallback.
+// Returned (x, y) values are percentages relative to the poker-table box
+// (NOT the felt-shell). The poker-table reserves vertical padding so the
+// top/bottom seats and their bubbles never overlap the table headers.
+// ---------------------------------------------------------------------------
+function seatPosition(index, count) {
+  const mobile = window.matchMedia("(max-width: 760px)").matches;
+  const presets = mobile ? mobileSeatPreset(count) : desktopSeatPreset(count);
+  if (presets[index]) return presets[index];
+  const radiusX = mobile ? 36 : 39;
+  const radiusY = mobile ? 41 : 39;
+  const start = -90;
+  const angle = (start + index * (360 / Math.max(count, 1))) * Math.PI / 180;
+  return {
+    x: 50 + Math.cos(angle) * radiusX,
+    y: 50 + Math.sin(angle) * radiusY,
+  };
+}
+
+// Coordinates expressed in percent of the poker-table box (which already
+// includes vertical padding around the felt). `y < ~25` lands above the
+// felt (top rail), `y > ~75` lands below the felt (bottom rail).
+function desktopSeatPreset(count) {
+  const presets = {
+    2: [{ x: 50, y: 18 }, { x: 50, y: 82 }],
+    3: [{ x: 50, y: 18 }, { x: 80, y: 72 }, { x: 20, y: 72 }],
+    4: [{ x: 50, y: 18 }, { x: 84, y: 60 }, { x: 50, y: 84 }, { x: 16, y: 60 }],
+    5: [{ x: 50, y: 18 }, { x: 84, y: 44 }, { x: 72, y: 84 }, { x: 28, y: 84 }, { x: 16, y: 44 }],
+    6: [{ x: 50, y: 18 }, { x: 82, y: 33 }, { x: 82, y: 70 }, { x: 50, y: 86 }, { x: 18, y: 70 }, { x: 18, y: 33 }],
+  };
+  return presets[count] || [];
+}
+
+function mobileSeatPreset(count) {
+  const presets = {
+    2: [{ x: 50, y: 16 }, { x: 50, y: 84 }],
+    3: [{ x: 50, y: 15 }, { x: 80, y: 78 }, { x: 20, y: 78 }],
+    4: [{ x: 50, y: 14 }, { x: 83, y: 60 }, { x: 50, y: 86 }, { x: 17, y: 60 }],
+  };
+  return presets[count] || [];
+}
+
+/**
+ * Map a seat coordinate to a `seat-top|bottom|left|right|mid` zone class.
+ * The CSS uses this to flip the speech bubble below the seat for top-rail
+ * seats so it never extends outside the visible table area.
+ */
+function seatZone(pos) {
+  if (pos.y < 32) return "seat-top";
+  if (pos.y > 70) return "seat-bottom";
+  if (pos.x < 34) return "seat-left";
+  if (pos.x > 66) return "seat-right";
+  return "seat-mid";
+}
+
+// ---------------------------------------------------------------------------
+// DOM rendering: assemble a single seat element from a frame's player snapshot.
+// Extracted into its own function so renderFrame stays declarative.
+// ---------------------------------------------------------------------------
+function renderSeat(player, position, frame, hand) {
+  const zone = seatZone(position);
+  const isActive = player.player_id === frame.activePlayerId;
+  const isDealer = player.seat === hand.button_seat;
+  const isCurrentActor = isActive && (frame.type === "action" || frame.type === "award");
+
+  const seat = document.createElement("div");
+  seat.className = [
+    "seat",
+    zone,
+    isActive ? "active" : "",
+    player.folded ? "folded" : "",
+    isDealer ? "dealer-seat" : "",
+    // The transient act-* class is what triggers the avatar's reaction
+    // animation (shake/nod/fold/cheer). Only apply on the actor's frame.
+    isCurrentActor && frame.actionKind ? `act-${frame.actionKind}` : "",
+  ].filter(Boolean).join(" ");
+  seat.style.setProperty("--x", `${position.x}%`);
+  seat.style.setProperty("--y", `${position.y}%`);
+
+  // Speech bubble — shown on the active player's action/award frames using
+  // a SHORT label (e.g. "加注 50") so it fits the small bubble width.
+  const bubbleHtml = (isCurrentActor && frame.bubbleText)
+    ? `<div class="speech kind-${frame.actionKind || "info"}">${escapeHtml(frame.bubbleText)}</div>`
+    : "";
+
+  // Avatar — inline SVG pixel-art sprite, deterministic per player_id so
+  // the same player keeps a stable look across hands.
+  const avatarMarkup = `<div class="avatar" aria-hidden="true">${avatarSvg(player.player_id)}</div>`;
+
+  seat.innerHTML = `
+    ${bubbleHtml}
+    <div class="player-box">
+      <div class="player-head">
+        ${avatarMarkup}
+        <span class="player-name">${escapeHtml(player.name || player.player_id)}</span>
+        <span class="dealer">D</span>
+      </div>
+      <div class="player-meta">
+        <span class="stack">Stack ${money(player.stack)}</span>
+        <span class="bet">${player.folded ? "Folded" : player.all_in ? "All In" : player.street_bet ? `Bet ${money(player.street_bet)}` : ""}</span>
+      </div>
+    </div>
+  `;
+
+  // Hole cards live inside the player-box so they share its layout flow.
+  const hole = document.createElement("div");
+  hole.className = "card-row hole-cards";
+  for (const card of knownCardsForPlayer(player, hand, frame)) hole.appendChild(renderCard(card));
+  seat.querySelector(".player-box").appendChild(hole);
+
+  return seat;
+}
+
+function renderFrame() {
+  const hand = state.hands[state.currentHandIndex];
+  const frame = state.frames[state.frameIndex];
+  const game = state.game;
+  const hasData = Boolean(game && hand && frame);
+
+  el.handBadge.textContent = hasData ? `Hand ${hand.hand_number}` : "Hand -";
+  el.streetLabel.textContent = hasData ? (STREET_LABELS[frame.street] || frame.street) : "未加载";
+  el.potLabel.textContent = hasData ? `Pot ${money(frame.pot)}` : "Pot 0";
+  el.tableMessage.textContent = hasData ? frame.message : "上传 JSON 或从游戏服务获取快照";
+  el.tableMessage.style.display = hasData ? "" : "block";
+  el.progressBar.style.width = hasData ? `${((state.frameIndex + 1) / state.frames.length) * 100}%` : "0";
+
+  el.boardCards.replaceChildren();
+  for (const card of hasData ? frame.board : []) el.boardCards.appendChild(renderCard(card));
+
+  el.seatLayer.replaceChildren();
+  if (hasData) {
+    frame.players.forEach((player, index) => {
+      const position = seatPosition(index, frame.players.length);
+      el.seatLayer.appendChild(renderSeat(player, position, frame, hand));
+    });
+  }
+
+  renderLog();
+  syncControls();
+}
+
+function knownCardsForPlayer(player, hand, frame) {
+  const futureHoleCards = hand.hole_cards?.[player.player_id] || hand.private_hands?.[player.player_id];
+  const canRevealShowdown = ["showdown", "awards"].includes(frame.street) || ["showdown", "award"].includes(frame.type);
+  const showdownCards = canRevealShowdown ? hand.showdown_hands?.[player.player_id] : null;
+  const shown = futureHoleCards || showdownCards || player.hole_cards;
+  if (shown && shown.length) return shown;
+  if (!player.in_hand && !player.folded) return [];
+  return ["back", "back"];
+}
+
+// ---------------------------------------------------------------------------
+// Event log — one line per visible event, with auto-scroll keeping the
+// current item in view. Items past the current frame index are dimmed via
+// the `.past` class.
+// ---------------------------------------------------------------------------
+function renderLog() {
+  const hand = state.hands[state.currentHandIndex];
+  el.eventLog.replaceChildren();
+  if (!hand) return;
+  const events = [
+    { text: `第 ${hand.hand_number} 手牌开始`, kind: "setup" },
+    ...(hand.actions || []).map((action) => ({ text: actionText(action), kind: "action" })),
+    ...(hand.awards || []).map((award) => ({ text: awardText(award), kind: "award" })),
+  ];
+  let currentLi = null;
+  events.forEach((event, index) => {
+    const li = document.createElement("li");
+    li.textContent = `${index + 1}. ${event.text}`;
+    if (index < state.frameIndex) li.classList.add("past");
+    if (index === state.frameIndex) {
+      li.classList.add("current");
+      currentLi = li;
+    }
+    el.eventLog.appendChild(li);
+  });
+  // Keep the focused event visible without yanking the page when an event
+  // is already in view.
+  if (currentLi) {
+    currentLi.scrollIntoView({ block: "nearest", behavior: "smooth" });
+  }
+}
+
+function syncControls() {
+  const loaded = Boolean(state.game);
+  const hasPreviousFrame = state.frameIndex > 0 || state.currentHandIndex > 0;
+  const hasNextFrame = state.frameIndex < state.frames.length - 1 || state.currentHandIndex < state.hands.length - 1;
+  el.playBtn.disabled = !loaded;
+  el.prevBtn.disabled = !loaded || !hasPreviousFrame;
+  el.nextBtn.disabled = !loaded || !hasNextFrame;
+  el.resetBtn.disabled = !loaded;
+  el.playBtn.textContent = state.playing ? "Ⅱ" : "▶";
+}
+
+// ---------------------------------------------------------------------------
+// Game lifecycle: load / merge / select hand / step / play.
+// ---------------------------------------------------------------------------
+function loadGame(raw, source, options = {}) {
+  const nextGame = normalizeGame(raw);
+  const wasPlaying = state.playing;
+  const mergeResult = options.allowMerge !== false ? mergeGame(nextGame) : { merged: false, advanced: false };
+  if (!mergeResult.merged) {
+    pause();
+    state.game = nextGame;
+    state.hands = state.game.hands;
+    state.currentHandIndex = Math.max(0, state.hands.length - 1);
+    state.frames = buildFrames(state.game, state.hands[state.currentHandIndex]);
+    state.frameIndex = 0;
+  }
+  state.source = source;
+  el.sourceBadge.textContent = source;
+  el.subtitle.textContent = `${state.game.game_id || "未命名游戏"} · ${state.hands.length} hands`;
+  renderSummary();
+  populateHands();
+  el.handSelect.value = String(state.currentHandIndex);
+  renderFrame();
+  if (mergeResult.merged && (wasPlaying || mergeResult.advanced)) play();
+}
+
+function mergeGame(nextGame) {
+  if (!state.game || state.game.game_id !== nextGame.game_id) return { merged: false, advanced: false };
+  const currentHand = state.hands[state.currentHandIndex];
+  const nextHandIndex = nextGame.hands.findIndex((hand) => hand.hand_number === currentHand?.hand_number);
+  if (nextHandIndex < 0) return { merged: false, advanced: false };
+
+  const oldFrame = state.frames[state.frameIndex];
+  const nextFrames = buildFrames(nextGame, nextGame.hands[nextHandIndex]);
+  const oldKeyIndex = oldFrame ? nextFrames.findIndex((frame) => frame.key === oldFrame.key) : -1;
+  const atKnownFrame = oldKeyIndex >= 0;
+  const wasAtTail = state.frameIndex >= state.frames.length - 1;
+  const hadNewFramesOnCurrentHand = nextFrames.length > state.frames.length;
+  const currentWasLatestHand = state.currentHandIndex >= state.hands.length - 1;
+  const hasNewLaterHandFromCurrent = currentWasLatestHand && nextGame.hands.length > state.hands.length;
+  let advanced = false;
+
+  state.game = nextGame;
+  state.hands = nextGame.hands;
+  state.currentHandIndex = nextHandIndex;
+  state.frames = nextFrames;
+
+  if (atKnownFrame) {
+    advanced = shouldStepIntoIncrement(wasAtTail, hadNewFramesOnCurrentHand, hasNewLaterHandFromCurrent);
+    state.frameIndex = advanced
+      ? Math.min(oldKeyIndex + 1, nextFrames.length - 1)
+      : oldKeyIndex;
+  } else {
+    state.frameIndex = Math.min(state.frameIndex, nextFrames.length - 1);
+  }
+
+  if (state.frameIndex >= state.frames.length - 1 && hasNewLaterHandFromCurrent) {
+    state.currentHandIndex += 1;
+    state.frames = buildFrames(state.game, state.hands[state.currentHandIndex]);
+    state.frameIndex = 0;
+    advanced = true;
+  }
+
+  return { merged: true, advanced };
+}
+
+function shouldStepIntoIncrement(wasAtTail, hadNewFramesOnCurrentHand, hasLaterHands) {
+  return wasAtTail && (hadNewFramesOnCurrentHand || hasLaterHands);
+}
+
+function renderSummary() {
+  const game = state.game;
+  const hand = state.hands[state.currentHandIndex];
+  const blinds = hand?.blinds || game;
+  el.gameStatus.textContent = game?.status || "-";
+  el.playerCount.textContent = game?.players?.length ?? "-";
+  el.handCount.textContent = game?.hands?.length ?? "-";
+  el.blindLevel.textContent = game ? `${money(blinds.small_blind)} / ${money(blinds.big_blind)}` : "-";
+}
+
+function populateHands() {
+  el.handSelect.replaceChildren();
+  state.hands.forEach((hand, index) => {
+    const option = document.createElement("option");
+    option.value = String(index);
+    option.textContent = `Hand ${hand.hand_number} · ${hand.actions?.length || 0} actions`;
+    el.handSelect.appendChild(option);
+  });
+}
+
+function selectHand(index) {
+  pause();
+  state.currentHandIndex = Math.max(0, Math.min(index, state.hands.length - 1));
+  state.frames = buildFrames(state.game, state.hands[state.currentHandIndex]);
+  state.frameIndex = 0;
+  el.handSelect.value = String(state.currentHandIndex);
+  renderSummary();
+  renderFrame();
+}
+
+function nextFrame() {
+  if (!state.frames.length) return;
+  if (state.frameIndex < state.frames.length - 1) {
+    state.frameIndex += 1;
+    renderFrame();
+    return;
+  }
+  if (state.currentHandIndex < state.hands.length - 1) {
+    state.currentHandIndex += 1;
+    state.frames = buildFrames(state.game, state.hands[state.currentHandIndex]);
+    state.frameIndex = 0;
+    el.handSelect.value = String(state.currentHandIndex);
+    renderSummary();
+    renderFrame();
+    return;
+  }
+  pause();
+}
+
+function prevFrame() {
+  pause();
+  if (state.frameIndex === 0 && state.currentHandIndex > 0) {
+    state.currentHandIndex -= 1;
+    state.frames = buildFrames(state.game, state.hands[state.currentHandIndex]);
+    state.frameIndex = state.frames.length - 1;
+    el.handSelect.value = String(state.currentHandIndex);
+    renderSummary();
+    renderFrame();
+    return;
+  }
+  state.frameIndex = Math.max(0, state.frameIndex - 1);
+  renderFrame();
+}
+
+function play() {
+  if (!state.game || state.playing) return;
+  state.playing = true;
+  tick();
+  renderFrame();
+}
+
+function pause() {
+  state.playing = false;
+  if (state.timer) window.clearTimeout(state.timer);
+  state.timer = null;
+  syncControls();
+}
+
+function tick() {
+  if (!state.playing) return;
+  const frame = state.frames[state.frameIndex];
+  const baseDelay = frame?.type === "deal" || frame?.type === "award" ? 1500 : 1150;
+  const delay = baseDelay * Number(el.pace.value || 1);
+  state.timer = window.setTimeout(() => {
+    nextFrame();
+    tick();
+  }, delay);
+}
+
+// ---------------------------------------------------------------------------
+// Network / file I/O.
+// ---------------------------------------------------------------------------
+async function fetchFromServer() {
+  const base = el.serverUrl.value.trim();
+  const gameId = el.gameId.value.trim();
+  if (!base || !gameId) throw new Error("请填写游戏服务地址和 Game ID");
+  const url = `/api/fetch-game?${new URLSearchParams({ base_url: base, game_id: gameId })}`;
+  const response = await fetch(url);
+  const payload = await response.json();
+  if (!response.ok) throw new Error(payload.error || "获取失败");
+  loadGame(payload.game, `Server · ${gameId}`);
+}
+
+function setAutoPoll(enabled) {
+  if (state.pollTimer) window.clearInterval(state.pollTimer);
+  state.pollTimer = null;
+  el.pollBadge.textContent = enabled ? `Auto ${el.pollSeconds.value}s` : "Auto Off";
+  if (!enabled) return;
+  const interval = Math.max(5, Number(el.pollSeconds.value || 12)) * 1000;
+  state.pollTimer = window.setInterval(() => {
+    fetchFromServer().catch((error) => showMessage(error.message));
+  }, interval);
+}
+
+function showMessage(message) {
+  el.tableMessage.textContent = message;
+  el.sourceBadge.textContent = "Error";
+}
+
+function escapeHtml(value) {
+  return String(value)
+    .replaceAll("&", "&amp;")
+    .replaceAll("<", "&lt;")
+    .replaceAll(">", "&gt;")
+    .replaceAll('"', "&quot;")
+    .replaceAll("'", "&#039;");
+}
+
+// ---------------------------------------------------------------------------
+// Wiring.
+// ---------------------------------------------------------------------------
+el.fetchBtn.addEventListener("click", () => {
+  fetchFromServer().catch((error) => showMessage(error.message));
+});
+
+el.fileInput.addEventListener("change", async (event) => {
+  const file = event.target.files?.[0];
+  if (!file) return;
+  try {
+    const raw = JSON.parse(await file.text());
+    loadGame(raw, `File · ${file.name}`);
+  } catch (error) {
+    showMessage(error.message);
+  } finally {
+    event.target.value = "";
+  }
+});
+
+el.autoPoll.addEventListener("change", () => setAutoPoll(el.autoPoll.checked));
+el.pollSeconds.addEventListener("change", () => setAutoPoll(el.autoPoll.checked));
+el.handSelect.addEventListener("change", () => selectHand(Number(el.handSelect.value)));
+el.playBtn.addEventListener("click", () => state.playing ? pause() : play());
+el.nextBtn.addEventListener("click", () => nextFrame());
+el.prevBtn.addEventListener("click", () => prevFrame());
+el.resetBtn.addEventListener("click", () => selectHand(state.currentHandIndex));
+window.addEventListener("resize", () => renderFrame());
+
+syncControls();

texas_holdem_replay/static/index.html

@@ -0,0 +1,132 @@
+<!doctype html>
+<html lang="zh-CN">
+  <head>
+    <meta charset="utf-8" />
+    <meta name="viewport" content="width=device-width, initial-scale=1" />
+    <title>Texas Hold X Replay</title>
+    <link rel="stylesheet" href="/styles.css" />
+  </head>
+  <body>
+    <div class="app-shell">
+      <header class="topbar">
+        <div class="brand-lockup">
+          <div class="chip-mark" aria-hidden="true">TX</div>
+          <div class="brand-meta">
+            <h1>Texas Hold X Replay</h1>
+            <p id="subtitle">等待加载游戏数据</p>
+          </div>
+        </div>
+        <div class="status-strip" aria-live="polite">
+          <span id="sourceBadge" class="badge badge-gold">No Data</span>
+          <span id="pollBadge" class="badge badge-blue">Auto Off</span>
+        </div>
+      </header>
+
+      <main class="layout-grid">
+        <!-- Stage zone: pure visualization (table, seats, animations).
+             Placed first in DOM so mobile/tablet layouts keep it on top. -->
+        <section class="stage-zone" aria-label="牌桌动画回放">
+          <div class="stage-head">
+            <div class="stage-head-left">
+              <span id="handBadge" class="badge badge-gold">Hand -</span>
+              <strong id="streetLabel">未加载</strong>
+            </div>
+            <div class="stage-head-right">
+              <span id="potLabel" class="badge badge-gold">Pot 0</span>
+            </div>
+          </div>
+          <div id="table" class="poker-table">
+            <!-- felt-shell encapsulates the rounded green felt with overflow:hidden,
+                 so player speech bubbles drawn in seat-layer can overflow freely
+                 above and below the table without being clipped. -->
+            <div class="felt-shell" aria-hidden="true">
+              <div class="felt-rail"></div>
+              <div class="felt-surface">
+                <div class="felt-grid"></div>
+                <div class="felt-glow"></div>
+                <div class="felt-mark">TX</div>
+              </div>
+            </div>
+            <div class="community-area">
+              <div id="boardCards" class="card-row board"></div>
+              <div id="tableMessage" class="table-message">上传 JSON 或从游戏服务获取快照</div>
+            </div>
+            <div id="seatLayer" class="seat-layer"></div>
+          </div>
+        </section>
+
+        <!-- Interaction zone: data source + replay controls + summary. -->
+        <section class="control-panel" aria-label="数据与播放控制">
+          <div class="panel-section">
+            <h2>数据源</h2>
+            <label>
+              <span>游戏服务</span>
+              <input id="serverUrl" type="url" value="http://127.0.0.1:8000" placeholder="http://127.0.0.1:8000" />
+            </label>
+            <label>
+              <span>Game ID</span>
+              <input id="gameId" type="text" value="game1" placeholder="game1" />
+            </label>
+            <div class="button-row">
+              <button id="fetchBtn" class="primary-btn" type="button">获取</button>
+              <label class="file-btn">
+                上传 JSON
+                <input id="fileInput" type="file" accept="application/json,.json" />
+              </label>
+            </div>
+            <div class="auto-grid">
+              <label class="toggle-line">
+                <input id="autoPoll" type="checkbox" />
+                <span>自动获取</span>
+              </label>
+              <label>
+                <span>间隔秒</span>
+                <input id="pollSeconds" type="number" min="5" max="300" value="12" />
+              </label>
+            </div>
+          </div>
+
+          <div class="panel-section">
+            <h2>回放</h2>
+            <label>
+              <span>手牌</span>
+              <select id="handSelect"></select>
+            </label>
+            <label>
+              <span>节奏</span>
+              <input id="pace" type="range" min="0.75" max="1.8" step="0.05" value="1" />
+            </label>
+            <div class="transport-row">
+              <button id="prevBtn" type="button" title="上一帧">◀</button>
+              <button id="playBtn" class="primary-btn" type="button" title="播放/暂停">▶</button>
+              <button id="nextBtn" type="button" title="下一帧">▶</button>
+              <button id="resetBtn" type="button" title="重置">↺</button>
+            </div>
+            <div class="progress-wrap">
+              <div id="progressBar"></div>
+            </div>
+          </div>
+
+          <div class="panel-section dense">
+            <h2>牌局摘要</h2>
+            <dl class="stat-list">
+              <div><dt>状态</dt><dd id="gameStatus">-</dd></div>
+              <div><dt>玩家</dt><dd id="playerCount">-</dd></div>
+              <div><dt>总手数</dt><dd id="handCount">-</dd></div>
+              <div><dt>盲注</dt><dd id="blindLevel">-</dd></div>
+            </dl>
+          </div>
+        </section>
+
+        <aside class="event-panel" aria-label="事件日志">
+          <div class="panel-section">
+            <h2>事件</h2>
+            <ol id="eventLog" class="event-log"></ol>
+          </div>
+        </aside>
+      </main>
+    </div>
+
+    <script src="/app.js" type="module"></script>
+  </body>
+</html>

texas_holdem_replay/static/styles.css

@@ -0,0 +1,735 @@
+/* =========================================================================
+   Texas Hold X Replay — Pixel-art skin & responsive layout
+   -------------------------------------------------------------------------
+   Design goals:
+   1. Stage zone (table + seats + animations) is visually isolated from the
+      interaction zone (controls + event log). Each zone has independent
+      overflow rules so speech bubbles never get clipped.
+   2. Pixel-art aesthetic: hard edges, stepped shadows (no blur), 8-bit
+      palette, monospace typography.
+   3. Three responsive breakpoints (desktop 3-col → tablet 2-col → mobile
+      stacked) — see media queries at the bottom of this file.
+   ========================================================================= */
+
+:root {
+  color-scheme: dark;
+
+  /* Palette — keep limited and high-contrast for that 8-bit look. */
+  --ink: #f7efd2;
+  --ink-dim: #c8b98a;
+  --muted: #8e8466;
+  --panel: #221a17;
+  --panel-2: #2d231f;
+  --panel-3: #3a2c25;
+  --line: #5d4638;
+  --line-soft: #3b2c25;
+
+  /* Felt greens. */
+  --felt: #1d8a5f;
+  --felt-dark: #0c4a37;
+  --felt-light: #4ec089;
+  --felt-rail: #6c3a20;
+  --felt-rail-dark: #3a1d10;
+
+  /* Accents. */
+  --gold: #f0b64d;
+  --gold-dark: #b87a1f;
+  --red: #e24b4b;
+  --blue: #53a6de;
+  --green: #63cb73;
+  --purple: #b773d3;
+
+  --shadow: rgba(8, 6, 5, 0.55);
+  --pixel: 3px;
+
+  /* Seat sizing scales with the stage width via container query fallback
+     (clamp on viewport). */
+  --seat-size: clamp(96px, 13vw, 138px);
+  --avatar-size: 44px;
+}
+
+* { box-sizing: border-box; }
+
+body {
+  min-height: 100vh;
+  margin: 0;
+  color: var(--ink);
+  /* Pixel-art monospace stack — keeps numerals crisp & blocky. */
+  font-family: "Courier New", "Lucida Console", "Press Start 2P", monospace;
+  font-size: 13px;
+  letter-spacing: 0.2px;
+  background:
+    /* Tiny noise made from offset diagonals for that CRT pixel-grid feel. */
+    linear-gradient(45deg, rgba(255,255,255,0.025) 25%, transparent 25%) 0 0 / 8px 8px,
+    linear-gradient(135deg, rgba(255,255,255,0.025) 25%, transparent 25%) 0 0 / 8px 8px,
+    radial-gradient(ellipse at 50% 0%, #2a1f1a 0%, #14100e 60%, #0d0a08 100%);
+  image-rendering: pixelated;
+  -webkit-font-smoothing: none;
+  -moz-osx-font-smoothing: grayscale;
+}
+
+button, input, select { font: inherit; }
+
+/* ---------- Buttons (chunky pixel "press" feel) ---------- */
+button, .file-btn {
+  min-height: 42px;
+  border: var(--pixel) solid #17100d;
+  color: var(--ink);
+  background: #3a2b24;
+  box-shadow:
+    inset -3px -3px 0 rgba(0,0,0,0.35),
+    inset  3px  3px 0 rgba(255,255,255,0.08),
+    0 4px 0 #120d0b;
+  cursor: pointer;
+  transition: transform 90ms steps(2), filter 90ms steps(2);
+}
+button:hover, .file-btn:hover { filter: brightness(1.12); }
+button:active, .file-btn:active {
+  transform: translateY(3px);
+  box-shadow:
+    inset -3px -3px 0 rgba(0,0,0,0.35),
+    inset  3px  3px 0 rgba(255,255,255,0.08),
+    0 1px 0 #120d0b;
+}
+button:disabled { opacity: 0.4; cursor: not-allowed; }
+
+input, select {
+  width: 100%;
+  min-height: 40px;
+  border: 2px solid var(--line);
+  border-radius: 0;
+  color: var(--ink);
+  background: #15110f;
+  padding: 9px 10px;
+  outline: none;
+}
+input:focus, select:focus { border-color: var(--gold); }
+
+/* ---------- Shell / Topbar ---------- */
+.app-shell {
+  width: min(1640px, 100%);
+  margin: 0 auto;
+  padding: 18px;
+}
+
+.topbar {
+  display: flex;
+  justify-content: space-between;
+  align-items: center;
+  gap: 16px;
+  padding: 14px 16px;
+  border: var(--pixel) solid #49382e;
+  background:
+    linear-gradient(180deg, #322620 0%, #1f1715 100%);
+  box-shadow: 0 8px 0 #0d0908, 0 18px 34px var(--shadow);
+}
+
+.brand-lockup { display: flex; gap: 14px; align-items: center; min-width: 0; }
+.brand-meta { min-width: 0; }
+
+.chip-mark {
+  display: grid;
+  place-items: center;
+  width: 58px;
+  height: 58px;
+  flex: 0 0 auto;
+  border: 4px dashed #fff4bc;
+  border-radius: 50%;
+  color: #20130b;
+  background: radial-gradient(circle, #ffe28a 0 42%, #c73f3d 43% 100%);
+  font-weight: 900;
+  text-shadow: 1px 1px 0 rgba(255,255,255,0.45);
+  /* Slow 8-bit chip rotation when idle for a touch of life. */
+  animation: chipSpin 6s linear infinite;
+}
+
+h1, h2, p { margin: 0; }
+h1 { font-size: clamp(18px, 2.3vw, 28px); }
+h2 { margin-bottom: 12px; color: var(--gold); font-size: 15px; }
+
+#subtitle { margin-top: 5px; color: var(--ink-dim); font-size: 12px; }
+
+.status-strip {
+  display: flex;
+  flex-wrap: wrap;
+  justify-content: flex-end;
+  gap: 8px;
+}
+
+/* Generic pixel "tag" badge — shared by status, hand, pot, etc. */
+.badge {
+  display: inline-block;
+  border: 2px solid #16100d;
+  padding: 7px 10px;
+  color: #1b120c;
+  font-weight: 700;
+  white-space: nowrap;
+  box-shadow: inset -2px -2px 0 rgba(0,0,0,0.18);
+}
+.badge-gold { background: var(--gold); }
+.badge-blue { background: var(--blue); color: #0f1c2e; }
+
+/* ---------- Layout Grid ---------- */
+/* Desktop: stage in middle, controls left, events right. The stage column
+   is the dominant area (1.4fr) so the table breathes. */
+.layout-grid {
+  display: grid;
+  grid-template-columns: minmax(260px, 320px) minmax(0, 1.4fr) minmax(280px, 340px);
+  grid-template-areas: "controls stage events";
+  gap: 16px;
+  margin-top: 18px;
+  align-items: start;
+}
+.control-panel { grid-area: controls; display: grid; gap: 14px; }
+.stage-zone    { grid-area: stage;    min-width: 0; }
+.event-panel   { grid-area: events;   display: grid; gap: 14px; }
+
+/* ---------- Panel sections (the chunky bordered cards) ---------- */
+.panel-section {
+  border: var(--pixel) solid #49382e;
+  background: linear-gradient(180deg, var(--panel-2), var(--panel));
+  padding: 14px;
+  box-shadow: 0 7px 0 #0e0a08, 0 14px 26px var(--shadow);
+}
+.panel-section label {
+  display: grid;
+  gap: 6px;
+  margin-bottom: 10px;
+  color: var(--ink-dim);
+  font-size: 12px;
+}
+
+.button-row, .transport-row, .auto-grid { display: grid; gap: 9px; }
+.button-row    { grid-template-columns: 1fr 1fr; }
+.transport-row { grid-template-columns: repeat(4, minmax(42px, 1fr)); }
+.auto-grid     { grid-template-columns: 1fr 86px; align-items: end; }
+
+.toggle-line { display: flex !important; flex-direction: row; align-items: center; min-height: 40px; }
+.toggle-line input { width: auto; min-height: auto; accent-color: var(--gold); }
+
+.primary-btn { color: #1b120c; background: var(--gold); }
+
+.file-btn {
+  display: grid; place-items: center; text-align: center;
+  position: relative; overflow: hidden;
+}
+.file-btn input { position: absolute; inset: 0; opacity: 0; cursor: pointer; }
+
+.progress-wrap {
+  height: 12px; margin-top: 14px;
+  border: 2px solid #120d0b; background: #14100e;
+}
+#progressBar {
+  width: 0; height: 100%;
+  background: linear-gradient(90deg, var(--red), var(--gold), var(--green));
+  transition: width 220ms linear;
+}
+
+.stat-list { display: grid; gap: 8px; margin: 0; }
+.stat-list div {
+  display: flex; justify-content: space-between; gap: 12px;
+  border-bottom: 1px solid rgba(255,255,255,0.08);
+  padding-bottom: 7px;
+}
+.stat-list dt { color: var(--ink-dim); }
+.stat-list dd { margin: 0; text-align: right; }
+
+/* ---------- Stage zone ---------- */
+.stage-head {
+  display: flex; justify-content: space-between; align-items: center;
+  gap: 10px; margin-bottom: 12px;
+}
+.stage-head-left { display: flex; gap: 10px; align-items: center; min-width: 0; }
+#streetLabel { font-size: clamp(16px, 2vw, 22px); color: var(--ink); }
+
+/* The poker-table is the positioning context for seats. Its overflow MUST
+   stay visible so seats placed near the table edges (and their speech
+   bubbles) can extend slightly outside the green felt without being clipped.
+   The visual clipping is handled by .felt-shell instead. */
+.poker-table {
+  position: relative;
+  /* Reserve vertical breathing room above/below the felt for seats that
+     visually sit on the rail and for speech bubbles. */
+  padding: 70px 12px 80px;
+  /* Explicit responsive height — required because seats and felt are
+     absolutely positioned and the inner content (community area) is
+     center-aligned. clamp() keeps it usable from 640px to ~820px. */
+  height: clamp(620px, 64vw, 820px);
+  overflow: visible;
+}
+
+/* Felt shell — actual visible green table. Absolutely positioned and
+   inset within poker-table so seats/bubbles can spill outside. */
+.felt-shell {
+  position: absolute;
+  inset: 60px 0 70px;
+  border: 6px solid var(--felt-rail);
+  border-radius: 46% / 36%;
+  background: var(--felt-rail-dark);
+  box-shadow:
+    inset 0 0 0 10px var(--felt-rail),
+    inset 0 0 0 16px #2c1b12,
+    0 10px 0 #130c09,
+    0 24px 44px var(--shadow);
+  overflow: hidden;
+  pointer-events: none;
+}
+.felt-rail {
+  /* Decorative pixel "studs" running along the rail. */
+  position: absolute; inset: -2px;
+  background:
+    repeating-linear-gradient(90deg,
+      transparent 0 22px, rgba(255,220,160,0.18) 22px 24px),
+    repeating-linear-gradient(0deg,
+      transparent 0 22px, rgba(255,220,160,0.18) 22px 24px);
+  mix-blend-mode: screen;
+  opacity: 0.4;
+}
+.felt-surface {
+  position: absolute;
+  inset: 12px;
+  border-radius: 44% / 32%;
+  background:
+    radial-gradient(ellipse at 50% 50%, var(--felt-light) 0%, var(--felt) 38%, var(--felt-dark) 100%);
+  overflow: hidden;
+}
+.felt-grid {
+  position: absolute; inset: 0;
+  opacity: 0.18;
+  /* 1px-wide pixel grid for the felt — gives a chess-board-like 8-bit feel. */
+  background:
+    linear-gradient(90deg, transparent calc(100% - 2px), rgba(255,255,255,0.4) 0) 0 0 / 28px 28px,
+    linear-gradient(180deg, transparent calc(100% - 2px), rgba(255,255,255,0.28) 0) 0 0 / 28px 28px;
+  image-rendering: pixelated;
+}
+.felt-glow {
+  position: absolute; inset: 0;
+  background: radial-gradient(ellipse at 50% 45%, rgba(255,255,255,0.18), transparent 55%);
+  pointer-events: none;
+}
+.felt-mark {
+  position: absolute;
+  left: 50%; top: 50%;
+  transform: translate(-50%, -50%);
+  font-size: clamp(48px, 7vw, 96px);
+  font-weight: 900;
+  color: rgba(0,0,0,0.18);
+  letter-spacing: 6px;
+  text-shadow: 2px 2px 0 rgba(255,255,255,0.06);
+  user-select: none;
+}
+
+/* ---------- Community area (board + message) ---------- */
+.community-area {
+  position: absolute;
+  left: 50%;
+  top: 50%;
+  transform: translate(-50%, -50%);
+  width: min(60%, 560px);
+  display: grid;
+  justify-items: center;
+  gap: 14px;
+  z-index: 3;
+  pointer-events: none;
+}
+
+.card-row {
+  display: flex;
+  justify-content: center;
+  gap: clamp(6px, 1vw, 12px);
+  min-height: 76px;
+}
+.card-row.board { min-height: 92px; }
+
+/* ---------- Cards ---------- */
+.card {
+  display: grid;
+  grid-template-rows: auto 1fr auto;
+  width: clamp(48px, 5.4vw, 70px);
+  height: clamp(66px, 7.4vw, 96px);
+  border: 3px solid #1c1411;
+  background: #fff7df;
+  color: #17100d;
+  padding: 4px 6px;
+  box-shadow:
+    inset -2px -2px 0 rgba(0,0,0,0.12),
+    inset  2px  2px 0 rgba(255,255,255,0.6),
+    0 5px 0 rgba(0,0,0,0.42);
+  transform-origin: center;
+  animation: cardDeal 520ms cubic-bezier(.2,.9,.2,1);
+  font-family: "Courier New", monospace;
+}
+.card.red { color: #b92732; }
+.card.back {
+  background:
+    /* 8-bit checker pattern for card back. */
+    linear-gradient(45deg,  #2d579a 25%, transparent 25%) 0 0 / 8px 8px,
+    linear-gradient(-45deg, #2d579a 25%, transparent 25%) 0 0 / 8px 8px,
+    linear-gradient(45deg,  transparent 75%, #2d579a 75%) 0 0 / 8px 8px,
+    linear-gradient(-45deg, transparent 75%, #2d579a 75%) 0 0 / 8px 8px,
+    #173a72;
+}
+.card .rank { font-size: clamp(12px, 1.2vw, 16px); font-weight: 900; line-height: 1; }
+.card .suit { display: grid; place-items: center; font-size: clamp(18px, 2vw, 28px); }
+.card .rank.bottom { transform: rotate(180deg); justify-self: end; }
+
+.table-message {
+  min-height: 32px;
+  max-width: min(540px, 82%);
+  border: 3px solid #14100d;
+  padding: 8px 12px;
+  color: #1b120c;
+  background: #ffe28a;
+  text-align: center;
+  box-shadow: 0 4px 0 rgba(0,0,0,0.38);
+}
+
+/* ---------- Seats ---------- */
+.seat-layer {
+  position: absolute;
+  inset: 0;
+  z-index: 4;
+  pointer-events: none;
+}
+
+.seat {
+  --x: 50%;
+  --y: 50%;
+  position: absolute;
+  left: var(--x);
+  top: var(--y);
+  width: var(--seat-size);
+  transform: translate(-50%, -50%);
+  transition: transform 220ms steps(4), filter 220ms ease;
+  pointer-events: auto;
+}
+
+/* Active seat — slight lift + glow + sprite "hop" animation. */
+.seat.active {
+  filter: drop-shadow(0 0 14px rgba(240,182,77,0.85));
+  z-index: 6;
+}
+.seat.active .avatar { animation: avatarHop 520ms ease; }
+
+.seat.folded { opacity: 0.55; filter: grayscale(0.6); }
+.seat.folded .avatar { transform: rotate(-8deg); }
+
+/* Action-driven sprite reactions (added by JS as transient classes). */
+.seat.act-fold .avatar    { animation: avatarFold 600ms ease forwards; }
+.seat.act-bet .avatar,
+.seat.act-raise .avatar,
+.seat.act-all_in .avatar  { animation: avatarShake 480ms ease; }
+.seat.act-call .avatar,
+.seat.act-check .avatar   { animation: avatarNod 480ms ease; }
+.seat.act-award .avatar   { animation: avatarCheer 900ms ease; }
+
+/* Speech bubble — placed above the seat by default; below for top seats so
+   it does not punch out of the table viewport. */
+.speech {
+  position: absolute;
+  left: 50%;
+  bottom: calc(100% + 8px);
+  min-width: 72px;
+  max-width: 160px;
+  transform: translateX(-50%);
+  border: 3px solid #160f0c;
+  padding: 6px 10px;
+  color: #1a110b;
+  background: #fff2b7;
+  text-align: center;
+  font-weight: 700;
+  font-size: 12px;
+  box-shadow: 0 4px 0 rgba(0,0,0,0.38);
+  animation: bubblePop 380ms ease;
+  z-index: 8;
+  white-space: nowrap;
+  overflow: hidden;
+  text-overflow: ellipsis;
+}
+/* Bubble tail — pixel-art triangle made from a rotated solid square. */
+.speech::after {
+  content: "";
+  position: absolute;
+  left: 50%; top: 100%;
+  transform: translate(-50%, -3px) rotate(45deg);
+  width: 12px; height: 12px;
+  background: #fff2b7;
+  border-right: 3px solid #160f0c;
+  border-bottom: 3px solid #160f0c;
+}
+/* For top-row seats, flip the bubble below so it does not get clipped. */
+.seat.seat-top .speech {
+  bottom: auto;
+  top: calc(100% + 8px);
+}
+.seat.seat-top .speech::after {
+  top: auto; bottom: 100%;
+  transform: translate(-50%, 3px) rotate(225deg);
+}
+/* Color the bubble by action category. */
+.speech.kind-fold   { background: #d8d2bc; }
+.speech.kind-call,
+.speech.kind-check  { background: #c2e6c8; }
+.speech.kind-bet,
+.speech.kind-raise,
+.speech.kind-all_in { background: #ffc7a8; color: #5a1f0a; }
+.speech.kind-award  { background: #ffe28a; }
+
+/* Player name+avatar+stack box. */
+.player-box {
+  border: 3px solid #15100d;
+  background: linear-gradient(180deg, #423128, #261c18);
+  padding: 8px;
+  box-shadow:
+    inset -3px -3px 0 rgba(0,0,0,0.35),
+    inset  3px  3px 0 rgba(255,255,255,0.08),
+    0 6px 0 #120d0b;
+  position: relative;
+}
+
+.player-head {
+  display: grid;
+  grid-template-columns: var(--avatar-size) 1fr auto;
+  gap: 8px;
+  align-items: center;
+  min-width: 0;
+}
+
+/* Avatar — wraps an inline SVG pixel-art portrait generated by JS. The SVG
+   is an 8x8 grid of <rect> elements; the host element only provides the
+   square frame, border, and animation hook. */
+.avatar {
+  width: var(--avatar-size);
+  height: var(--avatar-size);
+  flex: 0 0 auto;
+  position: relative;
+  border: 2px solid #15100d;
+  background: #211814;
+  image-rendering: pixelated;
+  display: grid;
+  place-items: stretch;
+  overflow: hidden;
+  box-shadow: inset -2px -2px 0 rgba(0,0,0,0.35);
+}
+.avatar svg {
+  width: 100%;
+  height: 100%;
+  display: block;
+  shape-rendering: crispEdges;
+}
+
+.player-name {
+  min-width: 0;
+  overflow: hidden;
+  text-overflow: ellipsis;
+  white-space: nowrap;
+  font-weight: 900;
+  font-size: 12px;
+  color: var(--ink);
+}
+
+.dealer {
+  display: none;
+  flex: 0 0 auto;
+  border: 2px solid #130e0b;
+  padding: 2px 6px;
+  color: #17100d;
+  background: var(--gold);
+  font-weight: 900;
+  font-size: 11px;
+}
+.seat.dealer-seat .dealer { display: inline-block; }
+
+.player-meta {
+  display: flex;
+  justify-content: space-between;
+  gap: 8px;
+  margin-top: 6px;
+  font-size: 11px;
+}
+.stack { color: var(--green); }
+.bet   { min-height: 14px; color: var(--gold); }
+
+.hole-cards {
+  justify-content: flex-start;
+  min-height: 38px;
+  margin-top: 6px;
+  gap: 4px;
+}
+.hole-cards .card {
+  width: clamp(26px, 2.8vw, 36px);
+  height: clamp(36px, 4vw, 50px);
+  padding: 2px 3px;
+  border-width: 2px;
+  box-shadow:
+    inset -1px -1px 0 rgba(0,0,0,0.12),
+    inset  1px  1px 0 rgba(255,255,255,0.6),
+    0 3px 0 rgba(0,0,0,0.42);
+}
+.hole-cards .card .rank { font-size: 10px; }
+.hole-cards .card .suit { font-size: 14px; }
+
+/* Chip stack indicator drawn near the player's bet (gives action visual
+   weight even before the bubble). */
+.chip-pile {
+  position: absolute;
+  left: 50%;
+  top: -14px;
+  transform: translateX(-50%);
+  display: flex;
+  gap: 2px;
+  pointer-events: none;
+  opacity: 0;
+  transition: opacity 200ms steps(2);
+}
+.chip-pile.visible { opacity: 1; }
+.chip-pile .chip {
+  width: 12px; height: 12px;
+  border-radius: 50%;
+  border: 2px solid #160f0c;
+  background: var(--gold);
+  box-shadow: 0 2px 0 rgba(0,0,0,0.5);
+}
+.chip-pile .chip:nth-child(2) { background: var(--red); }
+.chip-pile .chip:nth-child(3) { background: var(--blue); }
+
+/* ---------- Event log ---------- */
+.event-panel .panel-section { display: flex; flex-direction: column; }
+.event-log {
+  display: grid;
+  gap: 8px;
+  /* Use viewport-relative max-height for desktop, with a fallback minimum
+     so the log never collapses to nothing. The actual scrollable height is
+     plenty for 20+ events while still aligning with the table's height. */
+  max-height: clamp(420px, 70vh, 760px);
+  overflow: auto;
+  margin: 0;
+  padding: 0 4px 0 26px;
+  scrollbar-width: thin;
+  scrollbar-color: var(--gold-dark) #1a1310;
+}
+.event-log::-webkit-scrollbar { width: 10px; }
+.event-log::-webkit-scrollbar-track { background: #1a1310; }
+.event-log::-webkit-scrollbar-thumb {
+  background: var(--gold-dark);
+  border: 2px solid #1a1310;
+}
+.event-log li {
+  border-left: 4px solid var(--line);
+  padding: 7px 8px;
+  color: var(--ink-dim);
+  background: rgba(0,0,0,0.18);
+  word-break: break-word;
+  white-space: normal;
+  line-height: 1.4;
+}
+.event-log li.current {
+  border-color: var(--gold);
+  color: var(--ink);
+  background: rgba(240,182,77,0.14);
+  box-shadow: inset 2px 0 0 var(--gold);
+}
+.event-log li.past { opacity: 0.85; }
+
+/* ---------- Animations ---------- */
+@keyframes cardDeal {
+  from { opacity: 0; transform: translateY(-18px) rotate(-6deg) scale(0.86); }
+  to   { opacity: 1; transform: translateY(0)     rotate(0)     scale(1); }
+}
+@keyframes bubblePop {
+  from { opacity: 0; transform: translate(-50%, 8px) scale(0.8); }
+  to   { opacity: 1; transform: translate(-50%, 0)   scale(1); }
+}
+@keyframes chipSpin {
+  0%, 90%, 100% { transform: rotate(0); }
+  95%           { transform: rotate(8deg); }
+}
+@keyframes avatarHop {
+  0%, 100% { transform: translateY(0); }
+  40%      { transform: translateY(-6px); }
+  70%      { transform: translateY(-2px); }
+}
+@keyframes avatarShake {
+  0%, 100% { transform: translateX(0); }
+  20%      { transform: translateX(-3px) rotate(-3deg); }
+  40%      { transform: translateX(3px)  rotate(3deg); }
+  60%      { transform: translateX(-2px) rotate(-2deg); }
+  80%      { transform: translateX(2px)  rotate(2deg); }
+}
+@keyframes avatarNod {
+  0%, 100% { transform: translateY(0); }
+  50%      { transform: translateY(3px); }
+}
+@keyframes avatarFold {
+  0%   { transform: rotate(0)     translateY(0); }
+  100% { transform: rotate(-12deg) translateY(2px); }
+}
+@keyframes avatarCheer {
+  0%, 100% { transform: translateY(0)  rotate(0); }
+  20%      { transform: translateY(-8px) rotate(-6deg); }
+  50%      { transform: translateY(-4px) rotate(6deg); }
+  80%      { transform: translateY(-6px) rotate(-3deg); }
+}
+
+/* =========================================================================
+   Responsive breakpoints
+   - Tablet (<=1180px): drop to 2 columns. Stage on top spans full width;
+     controls + events sit side-by-side underneath.
+   - Mobile (<=760px): single column; stage first, then controls, then log.
+   ========================================================================= */
+
+@media (max-width: 1180px) {
+  .layout-grid {
+    grid-template-columns: minmax(260px, 1fr) minmax(260px, 1fr);
+    grid-template-areas:
+      "stage    stage"
+      "controls events";
+  }
+  .event-log { max-height: 360px; }
+  :root { --seat-size: clamp(96px, 16vw, 130px); }
+}
+
+@media (max-width: 760px) {
+  .app-shell { padding: 10px; }
+  .topbar {
+    display: grid;
+    grid-template-columns: 1fr;
+  }
+  .status-strip { justify-content: flex-start; }
+
+  .layout-grid {
+    grid-template-columns: 1fr;
+    grid-template-areas:
+      "stage"
+      "controls"
+      "events";
+  }
+
+  .stage-head {
+    flex-wrap: wrap;
+  }
+
+  .poker-table {
+    padding: 56px 4px 64px;
+    height: 600px;
+  }
+  .felt-shell {
+    inset: 48px 0 56px;
+    border-radius: 28px;
+  }
+  .felt-surface { border-radius: 22px; }
+
+  .community-area { width: min(76%, 420px); }
+
+  .event-log { max-height: 240px; }
+
+  :root {
+    --seat-size: clamp(86px, 30vw, 120px);
+    --avatar-size: 36px;
+  }
+}
+
+/* Reduce-motion users get static sprites (no shake/hop). */
+@media (prefers-reduced-motion: reduce) {
+  *, *::before, *::after { animation: none !important; transition: none !important; }
+}