alphabell-research commited on
Commit
d5d5f09
·
verified ·
1 Parent(s): 3c48ae4

Launch the alphaBell Flint Arcade gateway

Browse files
Files changed (3) hide show
  1. Dockerfile +9 -0
  2. README.md +32 -5
  3. proxy.py +229 -0
Dockerfile ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ FROM python:3.11.14-slim
2
+ ENV PYTHONDONTWRITEBYTECODE=1 PYTHONUNBUFFERED=1
3
+ RUN pip install --no-cache-dir fastapi==0.159.1 httpx==0.28.1 uvicorn==0.54.0 \
4
+ && useradd --create-home --uid 1000 arcade
5
+ WORKDIR /app
6
+ COPY --chown=arcade:arcade proxy.py /app/proxy.py
7
+ USER 1000
8
+ EXPOSE 7860
9
+ CMD ["uvicorn", "proxy:create_app", "--factory", "--host", "0.0.0.0", "--port", "7860", "--workers", "1", "--no-access-log", "--no-proxy-headers"]
README.md CHANGED
@@ -1,10 +1,37 @@
1
  ---
2
  title: Flint Arcade
3
- emoji: 🏢
4
- colorFrom: gray
5
- colorTo: gray
6
  sdk: docker
7
- pinned: false
 
 
 
 
 
 
 
 
 
 
 
 
 
8
  ---
9
 
10
- Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
  title: Flint Arcade
3
+ emoji: 🕹️
4
+ colorFrom: yellow
5
+ colorTo: red
6
  sdk: docker
7
+ app_port: 7860
8
+ license: apache-2.0
9
+ short_description: Describe a strange game. Play it. Make it stranger.
10
+ models:
11
+ - RecursiveSI/flint
12
+ datasets:
13
+ - alphabell/Flint-Arcade-Trajectories
14
+ tags:
15
+ - flint
16
+ - code-generation
17
+ - browser-games
18
+ - game-generation
19
+ - agents
20
+ - alphaBell
21
  ---
22
 
23
+ # Flint Arcade by alphaBell
24
+
25
+ Describe a tiny world. Flint writes the game. Play it, share the result and remix it with another sentence.
26
+
27
+ This Space presents the live [alphaBell Arcade](https://alphabell.com/arcade/) through a bounded server-side gateway. Model inference and durable game storage run on the alphaBell application backend. The initial release uses stock [Recursive Flint](https://huggingface.co/RecursiveSI/flint) Q8_0 with browser feedback and repair. Each build checks the selected inference model's metadata and requires 8-bit Q8_0.
28
+
29
+ New games adapt a small tested alphaBell foundation for input, rendering and game state. Flint returns a complete standalone HTML game, which the browser checker exercises before publication. Remixes use the actual parent game's code. There is no Arcade-trained adapter in this release.
30
+
31
+ Try [Moon Garden](https://alphabell.com/arcade/g/moon-garden), then give it one more twist. The [generation and repair dataset](https://huggingface.co/datasets/alphabell/Flint-Arcade-Trajectories) records alphaBell-authored briefs, actual candidate code, measured feedback and generation budgets, including failures. It excludes raw reasoning and infrastructure identifiers.
32
+
33
+ Published games are public. Keep private information out of your brief. Training-data contribution is optional and disabled by default. Browser smoke checks do not establish that every game is complete or winnable.
34
+
35
+ The gateway has no Hugging Face token. Its `ARCADE_EDGE_SECRET` Space secret signs a timestamp and a pseudonymous client identifier for backend quota enforcement. It never forwards raw IP addresses or exposes the signing secret to the browser. The primary backend enforces a durable global limit of 100 accepted builds per rolling 24 hours and a client limit of three per hour. The Space adds transient limits of three forwarded requests per client per hour and thirty total per rolling day; its counters reset on a Space restart.
36
+
37
+ [Source and evaluation method](https://github.com/alphabell-research/flint-arcade) · [alphaBell](https://alphabell.com)
proxy.py ADDED
@@ -0,0 +1,229 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Small, bounded same-origin gateway for the alphaBell Arcade Space."""
2
+ import asyncio
3
+ from collections import defaultdict, deque
4
+ from contextlib import asynccontextmanager
5
+ from dataclasses import dataclass
6
+ import hashlib
7
+ import hmac
8
+ from http.cookies import SimpleCookie
9
+ import ipaddress
10
+ import os
11
+ import re
12
+ import secrets
13
+ import time
14
+ from urllib.parse import urlsplit
15
+
16
+ from fastapi import FastAPI, Request
17
+ from fastapi.responses import JSONResponse, Response
18
+ import httpx
19
+
20
+ MAX_DOCUMENT = 1_500_000
21
+ MAX_ASSET = 500_000
22
+ EDGE_COOKIE = "arcade_edge"
23
+ ASSETS = {"/app.js", "/styles.css", "/favicon.svg"}
24
+ ASSET_ROUTES = re.compile(r"^/(?:preview/[A-Za-z0-9_-]{1,80}\.png|fonts/[A-Za-z0-9][A-Za-z0-9_-]{0,100}\.(?:woff2|txt))$")
25
+ ROUTES = re.compile(r"^/(?:g/[A-Za-z0-9_-]{1,80}|play/[A-Za-z0-9_-]{1,80}|download/[A-Za-z0-9_-]{1,80}|api/(?:config|games(?:/[A-Za-z0-9_-]{1,80})?|jobs(?:/[A-Za-z0-9_-]{1,80})?)|health)?$")
26
+
27
+
28
+ @dataclass(frozen=True)
29
+ class Settings:
30
+ origin: str = "https://alphabell.com/arcade"
31
+ edge_secret: str = ""
32
+ space_host: str = ""
33
+ per_hour: int = 3
34
+ daily_cap: int = 30
35
+
36
+ def __post_init__(self):
37
+ parsed = urlsplit(self.origin)
38
+ if parsed.scheme != "https" or parsed.netloc != "alphabell.com" or parsed.path.rstrip("/") != "/arcade" or parsed.query or parsed.fragment:
39
+ raise ValueError("The gateway target must be the alphaBell Arcade application")
40
+ if self.space_host and not re.fullmatch(r"[a-z0-9-]+\.hf\.space", self.space_host):
41
+ raise ValueError("Invalid Space host")
42
+
43
+
44
+ class Limits:
45
+ """A single-worker admission limit, in addition to durable backend quotas."""
46
+ def __init__(self):
47
+ self.hourly = defaultdict(deque)
48
+ self.daily = deque()
49
+ self.lock = asyncio.Lock()
50
+
51
+ async def admit(self, cookie, client):
52
+ now = time.time()
53
+ async with self.lock:
54
+ for key in list(self.hourly):
55
+ values = self.hourly[key]
56
+ while values and values[0] <= now - 3600:
57
+ values.popleft()
58
+ if not values:
59
+ del self.hourly[key]
60
+ while self.daily and self.daily[0] <= now - 86400:
61
+ self.daily.popleft()
62
+ if len(self.daily) >= self.daily_cap:
63
+ return False
64
+ keys = ("cookie:" + cookie, "client:" + client)
65
+ if any(len(self.hourly[key]) >= self.per_hour for key in keys):
66
+ return False
67
+ for key in keys:
68
+ self.hourly[key].append(now)
69
+ self.daily.append(now)
70
+ return True
71
+
72
+
73
+ def _signature(secret, value):
74
+ return hmac.new(secret.encode(), value.encode(), hashlib.sha256).hexdigest()
75
+
76
+
77
+ def _edge_cookie(request, secret):
78
+ value = request.cookies.get(EDGE_COOKIE, "")
79
+ identity, _, signature = value.partition(".")
80
+ if not re.fullmatch(r"[A-Za-z0-9_-]{32}", identity) or not hmac.compare_digest(signature, _signature(secret, "cookie\n" + identity)):
81
+ identity = secrets.token_urlsafe(24)
82
+ return identity, identity + "." + _signature(secret, "cookie\n" + identity)
83
+
84
+
85
+ def _client(request, settings):
86
+ # HF's gateway appends the address it observes. Never trust the leftmost
87
+ # element, which a caller can supply. Outside Spaces use the direct peer.
88
+ address = request.client.host if request.client else "unknown"
89
+ forwarded = request.headers.get("x-forwarded-for", "")
90
+ if settings.space_host and forwarded:
91
+ address = forwarded.split(",")[-1].strip()
92
+ try:
93
+ address = str(ipaddress.ip_address(address))
94
+ except ValueError:
95
+ address = "unknown"
96
+ return _signature(settings.edge_secret, "client-ip\n" + address)
97
+
98
+
99
+ def _upstream_cookie(raw):
100
+ cookies = SimpleCookie()
101
+ try:
102
+ cookies.load(raw)
103
+ except Exception:
104
+ return ""
105
+ item = cookies.get("arcade_owner")
106
+ if item and re.fullmatch(r"[A-Za-z0-9_-]{32}\.[0-9a-f]{64}", item.value):
107
+ return "arcade_owner=" + item.value
108
+ return ""
109
+
110
+
111
+ def _rewrite_cookie(raw):
112
+ if not raw.lower().startswith("arcade_owner="):
113
+ return None
114
+ # Drop Domain and origin-specific Path; keep only this gateway's cookie.
115
+ parts = [part.strip() for part in raw.split(";")]
116
+ kept = [parts[0]]
117
+ for part in parts[1:]:
118
+ name = part.split("=", 1)[0].lower()
119
+ if name in {"max-age", "expires"}:
120
+ kept.append(part)
121
+ return "; ".join(kept + ["Path=/", "HttpOnly", "Secure", "SameSite=None", "Partitioned"])
122
+
123
+
124
+ def _rewrite_location(location, origin):
125
+ parsed = urlsplit(location)
126
+ if parsed.scheme or parsed.netloc:
127
+ if parsed.scheme != "https" or parsed.netloc != "alphabell.com":
128
+ return None
129
+ path = parsed.path
130
+ else:
131
+ path = parsed.path
132
+ prefix = urlsplit(origin).path.rstrip("/")
133
+ if not (path == prefix or path.startswith(prefix + "/")):
134
+ return None
135
+ return (path[len(prefix):] or "/") + ("?" + parsed.query if parsed.query else "")
136
+
137
+
138
+ def create_app(settings=None, transport=None):
139
+ settings = settings or Settings(origin=os.environ.get("ARCADE_ORIGIN", "https://alphabell.com/arcade").rstrip("/"), edge_secret=os.environ.get("ARCADE_EDGE_SECRET", ""), space_host=os.environ.get("SPACE_HOST", ""))
140
+ client = httpx.AsyncClient(transport=transport, trust_env=False, follow_redirects=False, verify=True, limits=httpx.Limits(max_connections=12, max_keepalive_connections=4))
141
+ limits = Limits()
142
+ limits.per_hour, limits.daily_cap = settings.per_hour, settings.daily_cap
143
+
144
+ @asynccontextmanager
145
+ async def lifespan(app):
146
+ yield
147
+ await client.aclose()
148
+
149
+ app = FastAPI(docs_url=None, redoc_url=None, openapi_url=None, lifespan=lifespan)
150
+
151
+ @app.api_route("/{path:path}", methods=["GET", "HEAD", "POST"])
152
+ async def proxy(path, request: Request):
153
+ public_path = "/" + path
154
+ if settings.space_host and request.url.hostname != settings.space_host:
155
+ return JSONResponse({"detail": "Not found."}, status_code=404)
156
+ is_asset = public_path in ASSETS or bool(ASSET_ROUTES.fullmatch(public_path))
157
+ if b"%" in request.scope.get("raw_path", b"") or request.url.query or (not is_asset and not ROUTES.fullmatch(public_path)):
158
+ return JSONResponse({"detail": "Not found."}, status_code=404)
159
+ is_post = request.method == "POST"
160
+ if is_post and public_path != "/api/jobs":
161
+ return JSONResponse({"detail": "Method not allowed."}, status_code=405)
162
+ if not is_post and public_path == "/api/jobs":
163
+ return JSONResponse({"detail": "Method not allowed."}, status_code=405)
164
+ cookie = _upstream_cookie(request.headers.get("cookie", ""))
165
+ # An explicit (possibly empty) Cookie header prevents the shared HTTP
166
+ # client's upstream cookie jar from crossing browser ownership boundaries.
167
+ outgoing = {"Accept": "*/*", "Accept-Encoding": "identity", "Cookie": cookie}
168
+ body = b""
169
+ edge_cookie = None
170
+ if is_post:
171
+ if not settings.edge_secret:
172
+ return JSONResponse({"detail": "Generation is temporarily unavailable."}, status_code=503)
173
+ incoming = request.headers.get("origin")
174
+ expected_host = settings.space_host or request.url.netloc
175
+ if incoming:
176
+ parsed = urlsplit(incoming)
177
+ if parsed.scheme not in {"https", "http"} or parsed.netloc != expected_host or parsed.path or parsed.query or parsed.fragment or (settings.space_host and parsed.scheme != "https"):
178
+ return JSONResponse({"detail": "Cross-origin requests are not allowed."}, status_code=403)
179
+ if request.headers.get("content-type", "").split(";", 1)[0].strip().lower() != "application/json":
180
+ return JSONResponse({"detail": "A JSON request is required."}, status_code=415)
181
+ chunks, size = [], 0
182
+ async for chunk in request.stream():
183
+ size += len(chunk)
184
+ if size > 16_384:
185
+ return JSONResponse({"detail": "Request is too large."}, status_code=413)
186
+ chunks.append(chunk)
187
+ body = b"".join(chunks)
188
+ identity, edge_cookie = _edge_cookie(request, settings.edge_secret)
189
+ pseudonym = _client(request, settings)
190
+ if not await limits.admit(identity, pseudonym):
191
+ return JSONResponse({"detail": "The arcade build limit has been reached. Please try again later."}, status_code=429)
192
+ timestamp = str(int(time.time()))
193
+ outgoing.update({"Content-Type": "application/json", "Origin": "https://alphabell.com", "X-Arcade-Edge-Client": pseudonym, "X-Arcade-Edge-Timestamp": timestamp, "X-Arcade-Edge-Signature": _signature(settings.edge_secret, timestamp + "\n" + pseudonym)})
194
+ target = settings.origin.rstrip("/") + public_path
195
+ timeout = httpx.Timeout(20 if is_post else 35, connect=10)
196
+ try:
197
+ async with asyncio.timeout(20 if is_post else 35), client.stream(request.method, target, headers=outgoing, content=body, timeout=timeout) as upstream:
198
+ if upstream.status_code >= 500:
199
+ return JSONResponse({"detail": "The arcade is temporarily unavailable. Please try again shortly."}, status_code=502)
200
+ cap = MAX_ASSET if is_asset else MAX_DOCUMENT
201
+ payload = bytearray()
202
+ async for chunk in upstream.aiter_bytes():
203
+ if len(payload) + len(chunk) > cap:
204
+ return JSONResponse({"detail": "The arcade response is unavailable."}, status_code=502)
205
+ payload.extend(chunk)
206
+ headers = {}
207
+ for name in ("content-type", "content-disposition", "content-security-policy", "permissions-policy", "referrer-policy", "x-content-type-options"):
208
+ if name in upstream.headers:
209
+ headers[name] = upstream.headers[name]
210
+ headers["Cache-Control"] = "no-store" if public_path.startswith("/api/jobs") or is_post else "public, max-age=60"
211
+ if "location" in upstream.headers:
212
+ location = _rewrite_location(upstream.headers["location"], settings.origin)
213
+ if location is None:
214
+ return JSONResponse({"detail": "The arcade response is unavailable."}, status_code=502)
215
+ headers["Location"] = location
216
+ if public_path == "/" or public_path.startswith("/g/"):
217
+ payload = bytes(payload).replace(b'<base href="/arcade/">', b'<base href="/">')
218
+ response = Response(bytes(payload), status_code=upstream.status_code, headers=headers)
219
+ for raw in upstream.headers.get_list("set-cookie"):
220
+ rewritten = _rewrite_cookie(raw)
221
+ if rewritten:
222
+ response.headers.append("set-cookie", rewritten)
223
+ if edge_cookie:
224
+ response.headers.append("set-cookie", f"{EDGE_COOKIE}={edge_cookie}; Path=/; Max-Age=86400; HttpOnly; Secure; SameSite=None; Partitioned")
225
+ return response
226
+ except (httpx.HTTPError, ValueError, TimeoutError):
227
+ return JSONResponse({"detail": "The arcade is temporarily unavailable. Please try again shortly."}, status_code=502)
228
+
229
+ return app