@wendellresearch-ai/benchwarmers 0.2.0 → 0.2.1

package/metrics/guard.py

@@ -1,592 +1,617 @@
-"""One request check for both servers, the dashboard (serve.py) and the gateway.
-
-    python metrics/guard.py      # self-check
-
-Both bind to 127.0.0.1 by default, which keeps other machines out but not
-other web pages: any site open in the browser can send requests to localhost.
-
-- **Host allow-list, on every request.** DNS rebinding points a site's own
-  name at 127.0.0.1, so the browser treats our replies as that site's and
-  lets its script read them (prompts and answers, endpoints, keys' names).
-  The Host header still carries the site's name, so only our own names, port
-  included, are answered. More (a Docker service name, a proxy) come from
-  BENCH_ALLOWED_HOSTS, comma-separated host:port, and BENCHWARMERS_PUBLIC_URL.
-- **On POST, JSON only and never cross-site.** A page can send a plain form
-  or text/plain to us without asking, but not application/json: that needs
-  a CORS preflight. A browser also stamps Sec-Fetch-Site and Origin, which a
-  page cannot remove; cross-site or a foreign Origin is refused. API clients
-  (SDKs, curl, agents, Open WebUI's backend) send neither and are
-  unaffected. BENCH_ALLOWED_ORIGINS adds a browser UI on another port.
-- **CORS on the gateway's routes only**, for a browser app calling the
-  gateway: a preflight is answered, and Access-Control-Allow-Origin echoed,
-  only for an allowed origin (BENCH_ALLOWED_ORIGINS, BENCHWARMERS_PUBLIC_URL),
-  which may then be cross-site. Never `*`, and never Allow-Credentials: a
-  gateway key travels in a header the page sets, not in a cookie. The
-  dashboard answers no preflight, so it stays same-origin.
-- **A body-size cap**, so one request cannot make a server read gigabytes.
-
-WHO MAY ASK (RELEASE-PLAN Phase 5, "Token login before Docker")
----------------------------------------------------------------
-Decided by ROUTE, never by what a request happens to carry, so an API call
-with a bearer never meets the login cookie and a browser's cookie never opens
-the gateway:
-
-- the gateway's routes (`/v1/...`, `/bench/...`, on either port) take a
-  GATEWAY KEY as `Authorization: Bearer <key>`: named, revocable, kept only
-  as a SHA-256 in `config/gateway-keys.json`, shown once when made. A key
-  never opens the dashboard, so a chat UI or agent given one cannot change
-  settings, read keys or spend through it.
-- every other route is the dashboard's and takes the LOGIN COOKIE, set by
-  posting BENCHWARMERS_ADMIN_TOKEN to /login: `<expiry>.<HMAC of it keyed by
-  the token>`, HttpOnly, SameSite=Strict, Secure when the public URL is https.
-  Changing the token signs every browser out. Open without either: /health,
-  the login page, and Connect's callback (the provider's redirect back is
-  cross-site, so a Strict cookie never rides on it; its single-use,
-  unguessable state is what protects it).
-- NO ADMIN TOKEN, NO AUTH: on 127.0.0.1 without BENCHWARMERS_ADMIN_TOKEN both
-  servers answer as they always did (no login, no key), and the checks above
-  are what keep web pages out. Binding anywhere else refuses to start without
-  a token (`startup`), so a server on a network always asks.
-  BENCHWARMERS_REQUIRE_KEYS asks for gateway keys on this machine too (the
-  dashboard stays open), so each local app has its own name, scope and cap.
-
-A key may be limited to some collections (`collections`, none = all: 403
-otherwise, and /v1/models lists only those) and to a number of chat
-completions a minute (`rpm`: 429 with Retry-After). Each records when it was
-last used, and the gateway writes its name on every call it makes.
-"""
-
-from __future__ import annotations
-
-import datetime as dt
-import hashlib
-import hmac
-import http.cookies
-import ipaddress
-import os
-import re
-import secrets
-import sys
-import threading
-import time
-import urllib.parse
-from collections import deque
-
-from paths import DATA  # noqa: E402  -- the install's data folder (metrics/paths.py)
-import locked  # noqa: E402
-
-KEYS = DATA / "config" / "gateway-keys.json"
-KEY_NAME = re.compile(r"^[a-z0-9][a-z0-9-]{1,31}$")
-PREFIX = "bwk_"                    # every gateway key starts so: a secret scanner, or a person, can tell it
-COOKIE = "bw_session"
-SESSION_S = 7 * 24 * 3600
-MIN_TOKEN = 16
-GATEWAY_ROUTES = ("/v1/", "/bench/")
-
-
-class Refused(Exception):
-    """The request is not served. `status` is the HTTP code to answer with,
-    `headers` any the answer must carry (WWW-Authenticate, Retry-After)."""
-
-    def __init__(self, status: int, message: str, headers: dict | None = None):
-        super().__init__(message)
-        self.status = status
-        self.headers = headers or {}
-
-
-def _env(name: str) -> set[str]:
-    return {v.strip().lower() for v in os.environ.get(name, "").split(",") if v.strip()}
-
-
-def bind() -> str:
-    """The address both servers listen on: BENCH_BIND, else this machine only."""
-    return os.environ.get("BENCH_BIND", "").strip() or "127.0.0.1"
-
-
-def public_url() -> str:
-    return os.environ.get("BENCHWARMERS_PUBLIC_URL", "").strip().rstrip("/")
-
-
-def _public() -> tuple[str, str]:
-    """(host, origin) of BENCHWARMERS_PUBLIC_URL, or ("", "") when unset or not
-    an http(s) URL. ponytail: the host only; serving under a path prefix behind
-    a proxy is not supported."""
-    u = urllib.parse.urlsplit(public_url())
-    if u.scheme not in ("http", "https") or not u.netloc or "@" in u.netloc:
-        return "", ""
-    return u.netloc.lower(), f"{u.scheme}://{u.netloc}".lower()
-
-
-def hosts(port: int) -> set[str]:
-    public = _public()[0]
-    return ({f"127.0.0.1:{port}", f"localhost:{port}", f"[::1]:{port}"} | _env("BENCH_ALLOWED_HOSTS")
-            | ({public} if public else set()))
-
-
-def origins(port: int) -> set[str]:
-    """Browser origins allowed to call: this server's own, BENCH_ALLOWED_ORIGINS
-    and the public URL's."""
-    return {f"http://{h}" for h in hosts(port)} | _env("BENCH_ALLOWED_ORIGINS") | ({_public()[1]} - {""})
-
-
-def cors_origin(headers, port: int) -> str:
-    """The request's Origin, as sent, when it is allowed; else "". What a
-    gateway reply echoes in Access-Control-Allow-Origin."""
-    origin = (headers.get("Origin") or "").strip() if headers else ""
-    return origin if origin.lower() in origins(port) else ""
-
-
-def check(headers, port: int, post: bool, max_body: int, cors: bool = False) -> int:
-    """Raise Refused, or return how many body bytes to read (0 for a GET).
-    `cors`: a gateway route, where an allowed origin may be cross-site."""
-    host = (headers.get("Host") or "").strip().lower()
-    if host not in hosts(port):
-        raise Refused(403, f"unknown Host {host or '(none)'!r}; set BENCHWARMERS_PUBLIC_URL"
-                           " or BENCH_ALLOWED_HOSTS to add one")
-    if not post:
-        return 0
-    origin = (headers.get("Origin") or "").strip().lower()
-    allowed = origins(port)
-    if (headers.get("Sec-Fetch-Site") or "").lower() == "cross-site" and not (cors and origin in allowed):
-        raise Refused(403, "cross-site requests are refused")
-    if origin and origin not in allowed:
-        raise Refused(403, f"Origin {origin!r} is not allowed; set BENCH_ALLOWED_ORIGINS to add one")
-    if not (headers.get("Content-Type") or "").lower().startswith("application/json"):
-        raise Refused(415, "send Content-Type: application/json")
-    try:
-        length = int(headers.get("Content-Length") or 0)
-    except ValueError:
-        raise Refused(400, "Content-Length is not a number") from None
-    if length < 0:
-        raise Refused(400, "Content-Length is negative")
-    if length > max_body:
-        raise Refused(413, f"body is {length:,} bytes; the limit is {max_body:,}")
-    return length
-
-
-# ---- auth: decided by route ------------------------------------------------
-
-def admin_token() -> str:
-    return os.environ.get("BENCHWARMERS_ADMIN_TOKEN", "")
-
-
-def auth_on() -> bool:
-    return bool(admin_token())
-
-
-def keys_required() -> bool:
-    """Gateway keys are asked for: always with an admin token, and without one
-    when BENCHWARMERS_REQUIRE_KEYS is set. Any value but 0/false/no/off counts,
-    so a misspelt "required" fails closed rather than silently open."""
-    return auth_on() or os.environ.get("BENCHWARMERS_REQUIRE_KEYS", "").strip().lower() \
-        not in ("", "0", "false", "no", "off")
-
-
-def _loopback(host: str) -> bool:
-    if host.lower() == "localhost":
-        return True
-    try:
-        return ipaddress.ip_address(host).is_loopback
-    except ValueError:
-        return False
-
-
-def startup(address: str) -> None:
-    """Refuse (SystemExit) a configuration that must not serve: a network
-    address without an admin token, a guessable token, a public URL that is
-    not one."""
-    token = admin_token()
-    if token and len(token) < MIN_TOKEN:
-        # ponytail: no login rate limit; this floor is what makes guessing
-        # hopeless. Add a limit if short tokens ever have to be allowed.
-        raise SystemExit(f"BENCHWARMERS_ADMIN_TOKEN is {len(token)} characters; use {MIN_TOKEN} or more,"
-                         ' e.g. python -c "import secrets; print(secrets.token_urlsafe(24))"')
-    if public_url() and not _public()[0]:
-        raise SystemExit(f"BENCHWARMERS_PUBLIC_URL must be http(s)://host[:port], not {public_url()!r}")
-    if not token and not _loopback(address):
-        raise SystemExit(f"refusing to listen on {address}: beyond this machine every request needs"
-                         " a login, so set BENCHWARMERS_ADMIN_TOKEN first (or BENCH_BIND=127.0.0.1)")
-
-
-def gateway_route(path: str) -> bool:
-    """A route the gateway answers, on its own port or the dashboard's."""
-    return path.startswith(GATEWAY_ROUTES)
-
-
-def _sign(expires: int) -> bytes:
-    return hmac.new(admin_token().encode(), f"session|{expires}".encode(), hashlib.sha256).hexdigest().encode()
-
-
-def login(given: str) -> bool:
-    """True when `given` is the admin token. Digests compared, so neither the
-    token's length nor its content leaks through timing."""
-    token = admin_token()
-    return bool(token) and hmac.compare_digest(hashlib.sha256(given.encode()).digest(),
-                                               hashlib.sha256(token.encode()).digest())
-
-
-def session_cookie() -> str:
-    """The Set-Cookie value for a good login."""
-    expires = int(time.time()) + SESSION_S
-    secure = "; Secure" if _public()[1].startswith("https://") else ""
-    return (f"{COOKIE}={expires}.{_sign(expires).decode()}; Path=/; Max-Age={SESSION_S};"
-            f" HttpOnly; SameSite=Strict{secure}")
-
-
-def _session_ok(headers) -> bool:
-    jar = http.cookies.SimpleCookie()
-    try:
-        jar.load(headers.get("Cookie") or "")
-    except http.cookies.CookieError:
-        return False
-    morsel = jar.get(COOKIE)
-    expires, _, sig = (morsel.value if morsel else "").partition(".")
-    if not (expires.isascii() and expires.isdigit()) or int(expires) < time.time():
-        return False
-    return hmac.compare_digest(sig.encode(), _sign(int(expires)))
-
-
-def admin_caller(headers) -> str:
-    """Who is asking for a dashboard route: "admin" with a good login cookie,
-    "local" when no admin token is set. Raises Refused(401)."""
-    if not auth_on():
-        return "local"
-    if not _session_ok(headers):
-        raise Refused(401, "sign in first: /login")
-    return "admin"
-
-
-def gateway_caller(headers) -> str:
-    """Who is asking for a gateway route: the gateway key's name, or "local"
-    when keys are not required (any bearer is ignored then: SDKs send a dummy
-    one). Raises Refused(401), saying which of missing, malformed or unknown
-    it was. The cookie is never looked at."""
-    if not keys_required():
-        return "local"
-    challenge = {"WWW-Authenticate": 'Bearer realm="benchwarmers"'}     # RFC 6750: every 401 says how
-    raw = headers.get("Authorization")
-    if raw is None:
-        raise Refused(401, "this route needs a gateway key (Authorization: Bearer <key>);"
-                           " make one on the dashboard's Settings page", challenge)
-    scheme, _, value = raw.strip().partition(" ")
-    value = value.strip()
-    if scheme.lower() != "bearer" or not value:
-        raise Refused(401, "send the gateway key as 'Authorization: Bearer <key>'", challenge)
-    name = key_name(value)
-    if not name:
-        raise Refused(401, "unknown or revoked gateway key" + ("" if value.startswith(PREFIX) else
-                           f" (gateway keys start {PREFIX}; this is some other key)"), challenge)
-    _touch(name)
-    return name
-
-
-def scope(name: str) -> list[str] | None:
-    """The collections key `name` may call: [] for all (and for "local", no
-    key), or None for none, when its entry is not a list (a hand edit gone
-    wrong fails closed)."""
-    entry = gateway_keys().get(name)
-    allowed = entry.get("collections") if isinstance(entry, dict) else None
-    if allowed is None or allowed == []:               # not `not allowed`: "" or 0 must fail closed
-        return []
-    return allowed if isinstance(allowed, list) and all(isinstance(c, str) for c in allowed) else None
-
-
-def may_call(name: str, collection: str) -> bool:
-    allowed = scope(name)
-    return allowed == [] or (allowed is not None and collection in allowed)
-
-
-_SPENT: dict[str, deque] = {}
-_SPENT_LOCK = threading.Lock()
-
-
-def spend(name: str) -> None:
-    """Count one chat completion against key `name`'s per-minute cap, if it has
-    one; Refused(429) with Retry-After over it. Each call fans out to every
-    seat, so a client stuck in a loop would otherwise empty free tiers in
-    seconds. ponytail: counted in this process, so the gateway's port and the
-    dashboard's count apart and a restart forgets; a shared count needs a
-    store both read."""
-    entry = gateway_keys().get(name)
-    cap = entry.get("rpm") if isinstance(entry, dict) else None
-    if not isinstance(cap, int) or isinstance(cap, bool) or cap < 1:
-        return
-    now = time.monotonic()
-    with _SPENT_LOCK:
-        recent = _SPENT.setdefault(name, deque())
-        while recent and recent[0] <= now - 60:
-            recent.popleft()
-        if len(recent) >= cap:
-            wait = int(recent[0] + 60 - now) + 1
-            raise Refused(429, f"the gateway key {name} may make {cap} requests a minute;"
-                               f" try again in {wait}s", {"Retry-After": str(wait)})
-        recent.append(now)
-
-
-# ---- gateway keys ------------------------------------------------------------
-
-def _digest(value: str) -> bytes:
-    return hashlib.sha256(value.encode()).hexdigest().encode()
-
-
-def gateway_keys() -> dict[str, dict]:
-    """name -> {"sha256", "created", "hint", "collections", "rpm", "last_used"}.
-    Never a key itself."""
-    return locked.read_json(KEYS)
-
-
-def make_key(name: str, collections: list[str] | None = None, rpm: int | None = None) -> str:
-    """A new key, returned this once: only its SHA-256 is kept, and its last
-    four characters, so a key found in a log or a repository can be told
-    apart (the bwk_ prefix says whose it is). A slow hash would add nothing,
-    the key being 256 random bits rather than a password. `collections`: the
-    ones it may call (none = all); `rpm`: chat completions a minute (none = no cap)."""
-    if not KEY_NAME.match(name):
-        raise ValueError("name: 2-32 lower-case letters, digits or '-', e.g. open-webui")
-    if name == "local":
-        raise ValueError("local is what a call without a key is recorded as; pick another name")
-    collections = sorted(set(collections or []))
-    if not all(isinstance(c, str) and KEY_NAME.match(c) for c in collections):
-        raise ValueError("collections: names such as core, as they follow bench/")
-    if rpm is not None and (not isinstance(rpm, int) or isinstance(rpm, bool) or not 1 <= rpm <= 10_000):
-        raise ValueError("requests a minute: a whole number from 1 to 10000, or empty for no cap")
-    value = PREFIX + secrets.token_urlsafe(32)
-
-    def add(data: dict) -> None:
-        if name in data:
-            raise ValueError(f"there is already a key named {name}: revoke it first, or pick another name")
-        data[name] = {"sha256": _digest(value).decode(), "hint": value[-4:],
-                      "created": dt.datetime.now().astimezone().isoformat(timespec="seconds"),
-                      **({"collections": collections} if collections else {}),
-                      **({"rpm": rpm} if rpm else {})}
-    locked.update(KEYS, add)
-    return value
-
-
-def _touch(name: str) -> None:
-    """Note when key `name` was last used, at most once a minute: chat UIs poll
-    /v1/models, and a write per poll would be a write per page load."""
-    now = dt.datetime.now().astimezone()
-
-    def stamp(data: dict) -> None:
-        entry = data.get(name)
-        if not isinstance(entry, dict):
-            return                                     # revoked meanwhile
-        try:
-            if (now - dt.datetime.fromisoformat(entry.get("last_used", ""))).total_seconds() < 60:
-                return
-        except (TypeError, ValueError):
-            pass
-        entry["last_used"] = now.isoformat(timespec="seconds")
-    try:
-        locked.update(KEYS, stamp)
-    except (OSError, TimeoutError):                    # a busy file must not refuse the call
-        pass
-
-
-def revoke_key(name: str) -> bool:
-    return locked.update(KEYS, lambda data: data.pop(name, None)) is not None
-
-
-def key_name(value: str) -> str | None:
-    """The name of the gateway key `value` is, or None. Every stored digest is
-    compared in constant time, and all of them, whatever matched."""
-    if not value:
-        return None
-    given, found = _digest(value), None
-    for name, entry in gateway_keys().items():
-        stored = str(entry.get("sha256", "")).encode() if isinstance(entry, dict) else b""
-        if hmac.compare_digest(given, stored):
-            found = name
-    return found
-
-
-def access(handler, code) -> None:
-    """One access-log line on stderr: method, route, status, who. Never the
-    query (a sign-in's code travels in one) or a Connect callback's id. A
-    request line too broken to parse has no path or command yet."""
-    route = re.sub(r"^(/connect/callback)/.*", r"\1/...", (getattr(handler, "path", "") or "").split("?", 1)[0])
-    sys.stderr.write(f"{dt.datetime.now():%Y-%m-%d %H:%M:%S} {getattr(handler, 'command', '') or '-'}"
-                     f" {ascii(route[:200])[1:-1]}"
-                     f" {getattr(code, 'value', code)} {getattr(handler, 'who', '-')}\n")
-
-
-def _selftest() -> None:
-    import pathlib
-    import tempfile
-    global KEYS
-    ok = {"Host": "127.0.0.1:8123", "Content-Type": "application/json", "Content-Length": "12"}
-
-    def status(port=8123, post=True, cap=100, cors=False, **h):
-        try:
-            check({**ok, **h}, port, post, cap, cors)
-            return 0
-        except Refused as r:
-            return r.status
-    assert status() == 0
-    assert status(Host="localhost:8123") == 0 and status(Host="LOCALHOST:8123") == 0
-    assert status(Host="evil.example:8123") == 403            # DNS rebinding
-    assert status(Host="127.0.0.1") == 403                    # port is part of the name
-    assert status(port=8122) == 403                           # the other server's port
-    assert status(post=False, Host="evil.example") == 403     # GETs too
-    assert status(post=False, **{"Content-Type": ""}) == 0    # a GET needs no JSON type
-    assert status(**{"Content-Type": "text/plain"}) == 415    # the form a page can send freely
-    assert status(Origin="http://evil.example") == 403
-    assert status(Origin="http://127.0.0.1:8123") == 0
-    assert status(**{"Sec-Fetch-Site": "cross-site"}) == 403
-    assert status(**{"Sec-Fetch-Site": "same-origin"}) == 0
-    assert status(**{"Content-Length": "101"}) == 413
-    assert status(**{"Content-Length": "x"}) == 400
-    assert status(**{"Content-Length": "-1"}) == 400
-    saved = {k: os.environ.pop(k, None) for k in ("BENCH_ALLOWED_HOSTS", "BENCH_ALLOWED_ORIGINS",
-                                                    "BENCHWARMERS_PUBLIC_URL", "BENCHWARMERS_ADMIN_TOKEN",
-                                                    "BENCHWARMERS_REQUIRE_KEYS")}
-    try:
-        os.environ["BENCH_ALLOWED_HOSTS"] = "bench:8123"
-        os.environ["BENCH_ALLOWED_ORIGINS"] = "http://localhost:3000"
-        assert status(Host="bench:8123") == 0
-        assert status(Origin="http://localhost:3000") == 0
-        # The public URL is a Host and an Origin, https included.
-        os.environ["BENCHWARMERS_PUBLIC_URL"] = "https://Bench.example.com/"
-        assert status(Host="bench.example.com", Origin="https://bench.example.com") == 0
-        assert status(Host="bench.example.com", Origin="https://evil.example") == 403
-        # CORS: on a gateway route an allowed origin may be cross-site; the dashboard's rule stands.
-        site = {"Sec-Fetch-Site": "cross-site"}
-        assert status(cors=True, Origin="http://localhost:3000", **site) == 0
-        assert status(Origin="http://localhost:3000", **site) == 403
-        assert status(cors=True, Origin="http://evil.example", **site) == 403
-        assert status(cors=True, **site) == 403                             # no Origin: not a CORS call
-        assert cors_origin({"Origin": "http://localhost:3000"}, 8123) == "http://localhost:3000"
-        assert cors_origin({"Origin": "http://evil.example"}, 8123) == "" and cors_origin({}, 8123) == ""
-
-        # Binding: this machine needs nothing; anywhere else needs a long enough token.
-        def refused(address: str) -> bool:
-            try:
-                startup(address)
-                return False
-            except SystemExit:
-                return True
-        assert not refused("127.0.0.1") and not refused("localhost") and not refused("::1")
-        assert refused("0.0.0.0") and refused("") and refused("192.0.2.7")
-        os.environ["BENCHWARMERS_ADMIN_TOKEN"] = "short"
-        assert refused("127.0.0.1")
-        token = os.environ["BENCHWARMERS_ADMIN_TOKEN"] = "t" * 24
-        assert not refused("0.0.0.0")
-        os.environ["BENCHWARMERS_PUBLIC_URL"] = "bench.example.com"       # no scheme
-        assert refused("0.0.0.0")
-        os.environ["BENCHWARMERS_PUBLIC_URL"] = "https://bench.example.com"
-
-        # The login cookie: only the token makes one; tampered, expired or foreign ones fail.
-        assert login(token) and not login("t" * 23) and not login("")
-        cookie = session_cookie()
-        assert "HttpOnly" in cookie and "SameSite=Strict" in cookie and "Secure" in cookie and token not in cookie
-        value = cookie.split(";", 1)[0]
-        assert admin_caller({"Cookie": value}) == "admin"
-        assert admin_caller({"Cookie": f"theme=x; {value}"}) == "admin"
-        for bad in ("", value + "0", value.replace(f"{COOKIE}=", f"{COOKIE}=1"), f"{COOKIE}=1.abc",
-                    f"{COOKIE}=²." + "0" * 64, f"{COOKIE}=0.{_sign(0).decode()}", 'x="'):
-            try:
-                admin_caller({"Cookie": bad})
-                raise AssertionError(f"accepted cookie {bad!r}")
-            except Refused as r:
-                assert r.status == 401
-        os.environ["BENCHWARMERS_ADMIN_TOKEN"] = "u" * 24                  # a new token signs everyone out
-        try:
-            admin_caller({"Cookie": value})
-            raise AssertionError("an old token's cookie still works")
-        except Refused:
-            pass
-
-        # Gateway keys: shown once, kept hashed, constant-time, revocable; never a cookie's job.
-        with tempfile.TemporaryDirectory() as tmp:
-            real, KEYS = KEYS, pathlib.Path(tmp) / "gateway-keys.json"
-            try:
-                key = make_key("chat-ui")
-                assert key.startswith("bwk_") and key not in KEYS.read_text(encoding="utf-8")
-                assert gateway_keys()["chat-ui"]["sha256"] == hashlib.sha256(key.encode()).hexdigest()
-                for bad in ("Bad Name", "x", "chat-ui"):                    # form, length, taken
-                    try:
-                        make_key(bad)
-                        raise AssertionError(f"made a key named {bad!r}")
-                    except ValueError:
-                        pass
-                other = make_key("agent")
-                assert key_name(key) == "chat-ui" and key_name(other) == "agent"
-                assert key_name("") is None and key_name(key + "x") is None and key_name("bwk_é") is None
-                assert gateway_caller({"Authorization": f"Bearer {key}"}) == "chat-ui"
-                assert gateway_caller({"Authorization": f"bearer  {key} "}) == "chat-ui"
-                entry = gateway_keys()["chat-ui"]
-                assert entry["hint"] == key[-4:] and entry["last_used"] and "rpm" not in entry, entry
-                # Each refusal says which it was, and how to authenticate (RFC 6750).
-                for headers, said in (({}, "needs a gateway key"), ({"Cookie": value}, "needs a gateway key"),
-                                      ({"Authorization": key}, "Bearer <key>"), ({"Authorization": "Bearer "}, "Bearer <key>"),
-                                      ({"Authorization": f"Basic {key}"}, "Bearer <key>"),
-                                      ({"Authorization": "Bearer sk-dummy"}, "some other key"),
-                                      ({"Authorization": f"Bearer {key}x"}, "unknown or revoked")):
-                    try:
-                        gateway_caller(headers)
-                        raise AssertionError(f"gateway let in {headers}")
-                    except Refused as r:
-                        assert r.status == 401 and said in str(r), (headers, str(r))
-                        assert r.headers["WWW-Authenticate"].startswith("Bearer"), r.headers
-                # Scope and cap, set when made: checked per collection, and per minute.
-                for name, kw in (("scoped", {"collections": ["Core"]}), ("scoped", {"rpm": 0}),
-                                 ("scoped", {"rpm": True}), ("scoped", {"rpm": "5"}), ("local", {})):
-                    try:
-                        make_key(name, **kw)
-                        raise AssertionError(f"made a key with {kw}")
-                    except ValueError:
-                        pass
-                make_key("scoped", ["vision", "core", "core"], 2)
-                assert gateway_keys()["scoped"]["collections"] == ["core", "vision"]
-                assert may_call("scoped", "vision") and not may_call("scoped", "reviewers")
-                assert may_call("chat-ui", "reviewers") and may_call("local", "reviewers")
-                spend("scoped")
-                spend("scoped")
-                try:
-                    spend("scoped")
-                    raise AssertionError("a third request in a minute on a cap of 2")
-                except Refused as r:
-                    assert r.status == 429 and 1 <= int(r.headers["Retry-After"]) <= 61, r.headers
-                for _ in range(5):
-                    spend("chat-ui")                                    # no cap
-                for bad in ("core", "", 0, False, [1]):                 # a bad edit fails closed
-                    locked.update(KEYS, lambda d: d["scoped"].__setitem__("collections", bad))
-                    assert scope("scoped") is None and not may_call("scoped", "core"), bad
-                try:                                                    # a key never opens the dashboard
-                    admin_caller({"Authorization": f"Bearer {key}"})
-                    raise AssertionError("a gateway key opened the dashboard")
-                except Refused:
-                    pass
-                assert revoke_key("chat-ui") and not revoke_key("chat-ui") and key_name(key) is None
-                assert key_name(other) == "agent"
-                # No admin token: as before keys existed, whatever is sent.
-                del os.environ["BENCHWARMERS_ADMIN_TOKEN"]
-                assert gateway_caller({}) == "local" and gateway_caller({"Authorization": "Bearer x"}) == "local"
-                assert admin_caller({}) == "local"
-                # ...unless keys are required on this machine too; a misspelt value still asks.
-                for v in ("1", "required", "TRUE"):
-                    os.environ["BENCHWARMERS_REQUIRE_KEYS"] = v
-                    assert gateway_caller({"Authorization": f"Bearer {other}"}) == "agent"
-                    try:
-                        gateway_caller({"Authorization": "Bearer x"})
-                        raise AssertionError(f"BENCHWARMERS_REQUIRE_KEYS={v} let a dummy bearer in")
-                    except Refused as r:
-                        assert r.status == 401
-                    assert admin_caller({}) == "local"                  # the dashboard stays open
-                os.environ["BENCHWARMERS_REQUIRE_KEYS"] = "off"
-                assert not keys_required() and gateway_caller({}) == "local"
-            finally:
-                KEYS = real
-        assert gateway_route("/v1/models") and gateway_route("/bench/verdict")
-        assert not gateway_route("/api/endpoints") and not gateway_route("/v1") and not gateway_route("//v1/x")
-    finally:
-        for k, v in saved.items():
-            os.environ.pop(k, None)
-            if v is not None:
-                os.environ[k] = v
-    print("ok")
-
-
-if __name__ == "__main__":
-    _selftest()
+"""One request check for both servers, the dashboard (serve.py) and the gateway.
+
+    python metrics/guard.py      # self-check
+
+Both bind to 127.0.0.1 by default, which keeps other machines out but not
+other web pages: any site open in the browser can send requests to localhost.
+
+- **Host allow-list, on every request.** DNS rebinding points a site's own
+  name at 127.0.0.1, so the browser treats our replies as that site's and
+  lets its script read them (prompts and answers, endpoints, keys' names).
+  The Host header still carries the site's name, so only our own names, port
+  included, are answered. More (a Docker service name, a proxy) come from
+  BENCH_ALLOWED_HOSTS, comma-separated host:port, and BENCHWARMERS_PUBLIC_URL.
+- **On POST, JSON only and never cross-site.** A page can send a plain form
+  or text/plain to us without asking, but not application/json: that needs
+  a CORS preflight. A browser also stamps Sec-Fetch-Site and Origin, which a
+  page cannot remove; cross-site or a foreign Origin is refused. API clients
+  (SDKs, curl, agents, Open WebUI's backend) send neither and are
+  unaffected. BENCH_ALLOWED_ORIGINS adds a browser UI on another port.
+- **CORS on the gateway's routes only**, for a browser app calling the
+  gateway: a preflight is answered, and Access-Control-Allow-Origin echoed,
+  only for an allowed origin (BENCH_ALLOWED_ORIGINS, BENCHWARMERS_PUBLIC_URL),
+  which may then be cross-site. Never `*`, and never Allow-Credentials: a
+  gateway key travels in a header the page sets, not in a cookie. The
+  dashboard answers no preflight, so it stays same-origin.
+- **A body-size cap**, so one request cannot make a server read gigabytes.
+
+WHO MAY ASK (RELEASE-PLAN Phase 5, "Token login before Docker")
+---------------------------------------------------------------
+Decided by ROUTE, never by what a request happens to carry, so an API call
+with a bearer never meets the login cookie and a browser's cookie never opens
+the gateway:
+
+- the gateway's routes (`/v1/...`, `/bench/...`, on either port) take a
+  GATEWAY KEY as `Authorization: Bearer <key>`: named, revocable, kept only
+  as a SHA-256 in `config/gateway-keys.json`, shown once when made. A key
+  never opens the dashboard, so a chat UI or agent given one cannot change
+  settings, read keys or spend through it.
+- every other route is the dashboard's and takes the LOGIN COOKIE, set by
+  posting BENCHWARMERS_ADMIN_TOKEN to /login: `<expiry>.<HMAC of it keyed by
+  the token>`, HttpOnly, SameSite=Strict, Secure when the public URL is https.
+  Changing the token signs every browser out. Open without either: /health,
+  the login page, and Connect's callback (the provider's redirect back is
+  cross-site, so a Strict cookie never rides on it; its single-use,
+  unguessable state is what protects it).
+- NO ADMIN TOKEN, NO AUTH: on 127.0.0.1 without BENCHWARMERS_ADMIN_TOKEN both
+  servers answer as they always did (no login, no key), and the checks above
+  are what keep web pages out. Binding anywhere else refuses to start without
+  a token (`startup`), so a server on a network always asks.
+  BENCHWARMERS_REQUIRE_KEYS asks for gateway keys on this machine too (the
+  dashboard stays open), so each local app has its own name, scope and cap.
+
+A key may be limited to some collections (`collections`, none = all: 403
+otherwise, and /v1/models lists only those) and to a number of chat
+completions a minute (`rpm`: 429 with Retry-After). Each records when it was
+last used, and the gateway writes its name on every call it makes.
+"""
+
+from __future__ import annotations
+
+import datetime as dt
+import hashlib
+import hmac
+import http.cookies
+import ipaddress
+import os
+import re
+import secrets
+import sys
+import threading
+import time
+import urllib.parse
+from collections import deque
+
+from paths import DATA  # noqa: E402  -- the install's data folder (metrics/paths.py)
+import locked  # noqa: E402
+
+KEYS = DATA / "config" / "gateway-keys.json"
+KEY_NAME = re.compile(r"^[a-z0-9][a-z0-9-]{1,31}$")
+PREFIX = "bwk_"                    # every gateway key starts so: a secret scanner, or a person, can tell it
+COOKIE = "bw_session"
+SESSION_S = 7 * 24 * 3600
+MIN_TOKEN = 16
+GATEWAY_ROUTES = ("/v1/", "/bench/")
+
+
+class Refused(Exception):
+    """The request is not served. `status` is the HTTP code to answer with,
+    `headers` any the answer must carry (WWW-Authenticate, Retry-After)."""
+
+    def __init__(self, status: int, message: str, headers: dict | None = None):
+        super().__init__(message)
+        self.status = status
+        self.headers = headers or {}
+
+
+def _env(name: str) -> set[str]:
+    return {v.strip().lower() for v in os.environ.get(name, "").split(",") if v.strip()}
+
+
+def bind() -> str:
+    """The address both servers listen on: BENCH_BIND, else this machine only."""
+    return os.environ.get("BENCH_BIND", "").strip() or "127.0.0.1"
+
+
+def public_url() -> str:
+    return os.environ.get("BENCHWARMERS_PUBLIC_URL", "").strip().rstrip("/")
+
+
+def _public() -> tuple[str, str]:
+    """(host, origin) of BENCHWARMERS_PUBLIC_URL, or ("", "") when unset or not
+    an http(s) URL. ponytail: the host only; serving under a path prefix behind
+    a proxy is not supported."""
+    u = urllib.parse.urlsplit(public_url())
+    if u.scheme not in ("http", "https") or not u.netloc or "@" in u.netloc:
+        return "", ""
+    return u.netloc.lower(), f"{u.scheme}://{u.netloc}".lower()
+
+
+def hosts(port: int) -> set[str]:
+    public = _public()[0]
+    return ({f"127.0.0.1:{port}", f"localhost:{port}", f"[::1]:{port}"} | _env("BENCH_ALLOWED_HOSTS")
+            | ({public} if public else set()))
+
+
+def origins(port: int) -> set[str]:
+    """Browser origins allowed to call: this server's own, BENCH_ALLOWED_ORIGINS
+    and the public URL's."""
+    return {f"http://{h}" for h in hosts(port)} | _env("BENCH_ALLOWED_ORIGINS") | ({_public()[1]} - {""})
+
+
+def cors_origin(headers, port: int) -> str:
+    """The request's Origin, as sent, when it is allowed; else "". What a
+    gateway reply echoes in Access-Control-Allow-Origin."""
+    origin = (headers.get("Origin") or "").strip() if headers else ""
+    return origin if origin.lower() in origins(port) else ""
+
+
+def check(headers, port: int, post: bool, max_body: int, cors: bool = False) -> int:
+    """Raise Refused, or return how many body bytes to read (0 for a GET).
+    `cors`: a gateway route, where an allowed origin may be cross-site."""
+    host = (headers.get("Host") or "").strip().lower()
+    if host not in hosts(port):
+        raise Refused(403, f"unknown Host {host or '(none)'!r}; set BENCHWARMERS_PUBLIC_URL"
+                           " or BENCH_ALLOWED_HOSTS to add one")
+    if not post:
+        return 0
+    origin = (headers.get("Origin") or "").strip().lower()
+    allowed = origins(port)
+    if (headers.get("Sec-Fetch-Site") or "").lower() == "cross-site" and not (cors and origin in allowed):
+        raise Refused(403, "cross-site requests are refused")
+    if origin and origin not in allowed:
+        raise Refused(403, f"Origin {origin!r} is not allowed; set BENCH_ALLOWED_ORIGINS to add one")
+    if not (headers.get("Content-Type") or "").lower().startswith("application/json"):
+        raise Refused(415, "send Content-Type: application/json")
+    try:
+        length = int(headers.get("Content-Length") or 0)
+    except ValueError:
+        raise Refused(400, "Content-Length is not a number") from None
+    if length < 0:
+        raise Refused(400, "Content-Length is negative")
+    if length > max_body:
+        raise Refused(413, f"body is {length:,} bytes; the limit is {max_body:,}")
+    return length
+
+
+# ---- auth: decided by route ------------------------------------------------
+
+def admin_token() -> str:
+    return os.environ.get("BENCHWARMERS_ADMIN_TOKEN", "")
+
+
+def auth_on() -> bool:
+    return bool(admin_token())
+
+
+def keys_required() -> bool:
+    """Gateway keys are asked for: always with an admin token, and without one
+    when BENCHWARMERS_REQUIRE_KEYS is set. Any value but 0/false/no/off counts,
+    so a misspelt "required" fails closed rather than silently open."""
+    return auth_on() or os.environ.get("BENCHWARMERS_REQUIRE_KEYS", "").strip().lower() \
+        not in ("", "0", "false", "no", "off")
+
+
+def _loopback(host: str) -> bool:
+    if host.lower() == "localhost":
+        return True
+    try:
+        return ipaddress.ip_address(host).is_loopback
+    except ValueError:
+        return False
+
+
+def startup(address: str) -> None:
+    """Refuse (SystemExit) a configuration that must not serve: a network
+    address without an admin token, a guessable token, a public URL that is
+    not one."""
+    token = admin_token()
+    if token and len(token) < MIN_TOKEN:
+        # ponytail: no login rate limit; this floor is what makes guessing
+        # hopeless. Add a limit if short tokens ever have to be allowed.
+        raise SystemExit(f"BENCHWARMERS_ADMIN_TOKEN is {len(token)} characters; use {MIN_TOKEN} or more,"
+                         ' e.g. python -c "import secrets; print(secrets.token_urlsafe(24))"')
+    if public_url() and not _public()[0]:
+        raise SystemExit(f"BENCHWARMERS_PUBLIC_URL must be http(s)://host[:port], not {public_url()!r}")
+    if not token and not _loopback(address):
+        raise SystemExit(f"refusing to listen on {address}: beyond this machine every request needs"
+                         " a login, so set BENCHWARMERS_ADMIN_TOKEN first (or BENCH_BIND=127.0.0.1)")
+
+
+def gateway_route(path: str) -> bool:
+    """A route the gateway answers, on its own port or the dashboard's."""
+    return path.startswith(GATEWAY_ROUTES)
+
+
+def _sign(expires: int) -> bytes:
+    return hmac.new(admin_token().encode(), f"session|{expires}".encode(), hashlib.sha256).hexdigest().encode()
+
+
+def login(given: str) -> bool:
+    """True when `given` is the admin token. Digests compared, so neither the
+    token's length nor its content leaks through timing."""
+    token = admin_token()
+    return bool(token) and hmac.compare_digest(hashlib.sha256(given.encode()).digest(),
+                                               hashlib.sha256(token.encode()).digest())
+
+
+def session_cookie() -> str:
+    """The Set-Cookie value for a good login."""
+    expires = int(time.time()) + SESSION_S
+    secure = "; Secure" if _public()[1].startswith("https://") else ""
+    return (f"{COOKIE}={expires}.{_sign(expires).decode()}; Path=/; Max-Age={SESSION_S};"
+            f" HttpOnly; SameSite=Strict{secure}")
+
+
+def _session_ok(headers) -> bool:
+    jar = http.cookies.SimpleCookie()
+    try:
+        jar.load(headers.get("Cookie") or "")
+    except http.cookies.CookieError:
+        return False
+    morsel = jar.get(COOKIE)
+    expires, _, sig = (morsel.value if morsel else "").partition(".")
+    if not (expires.isascii() and expires.isdigit()) or int(expires) < time.time():
+        return False
+    return hmac.compare_digest(sig.encode(), _sign(int(expires)))
+
+
+def admin_caller(headers) -> str:
+    """Who is asking for a dashboard route: "admin" with a good login cookie
+    or the admin token itself as `Authorization: Bearer` (the `bench` CLI),
+    "local" when no admin token is set. Raises Refused(401).
+
+    The bearer is safe beside the cookie: a page cannot make a browser send an
+    Authorization header cross-site without a preflight, which the dashboard
+    never answers. A gateway key is not the admin token, so it still never
+    opens a dashboard route."""
+    if not auth_on():
+        return "local"
+    scheme, _, given = (headers.get("Authorization") or "").strip().partition(" ")
+    if scheme.lower() == "bearer" and given.strip():
+        if login(given.strip()):
+            return "admin"
+        raise Refused(401, "that bearer is not the admin token (a gateway key opens /v1 only)")
+    if not _session_ok(headers):
+        raise Refused(401, "sign in first: /login, or send the admin token as Authorization: Bearer")
+    return "admin"
+
+
+def gateway_caller(headers) -> str:
+    """Who is asking for a gateway route: the gateway key's name, or "local"
+    when keys are not required (any bearer is ignored then: SDKs send a dummy
+    one). Raises Refused(401), saying which of missing, malformed or unknown
+    it was. The cookie is never looked at."""
+    if not keys_required():
+        return "local"
+    challenge = {"WWW-Authenticate": 'Bearer realm="benchwarmers"'}     # RFC 6750: every 401 says how
+    raw = headers.get("Authorization")
+    if raw is None:
+        raise Refused(401, "this route needs a gateway key (Authorization: Bearer <key>);"
+                           " make one on the dashboard's Settings page", challenge)
+    scheme, _, value = raw.strip().partition(" ")
+    value = value.strip()
+    if scheme.lower() != "bearer" or not value:
+        raise Refused(401, "send the gateway key as 'Authorization: Bearer <key>'", challenge)
+    name = key_name(value)
+    if not name:
+        raise Refused(401, "unknown or revoked gateway key" + ("" if value.startswith(PREFIX) else
+                           f" (gateway keys start {PREFIX}; this is some other key)"), challenge)
+    _touch(name)
+    return name
+
+
+def scope(name: str) -> list[str] | None:
+    """The collections key `name` may call: [] for all (and for "local", no
+    key), or None for none, when its entry is not a list (a hand edit gone
+    wrong fails closed)."""
+    entry = gateway_keys().get(name)
+    allowed = entry.get("collections") if isinstance(entry, dict) else None
+    if allowed is None or allowed == []:               # not `not allowed`: "" or 0 must fail closed
+        return []
+    return allowed if isinstance(allowed, list) and all(isinstance(c, str) for c in allowed) else None
+
+
+def may_call(name: str, collection: str) -> bool:
+    allowed = scope(name)
+    return allowed == [] or (allowed is not None and collection in allowed)
+
+
+_SPENT: dict[str, deque] = {}
+_SPENT_LOCK = threading.Lock()
+
+
+def spend(name: str) -> None:
+    """Count one chat completion against key `name`'s per-minute cap, if it has
+    one; Refused(429) with Retry-After over it. Each call fans out to every
+    seat, so a client stuck in a loop would otherwise empty free tiers in
+    seconds. ponytail: counted in this process, so the gateway's port and the
+    dashboard's count apart and a restart forgets; a shared count needs a
+    store both read."""
+    entry = gateway_keys().get(name)
+    cap = entry.get("rpm") if isinstance(entry, dict) else None
+    if not isinstance(cap, int) or isinstance(cap, bool) or cap < 1:
+        return
+    now = time.monotonic()
+    with _SPENT_LOCK:
+        recent = _SPENT.setdefault(name, deque())
+        while recent and recent[0] <= now - 60:
+            recent.popleft()
+        if len(recent) >= cap:
+            wait = int(recent[0] + 60 - now) + 1
+            raise Refused(429, f"the gateway key {name} may make {cap} requests a minute;"
+                               f" try again in {wait}s", {"Retry-After": str(wait)})
+        recent.append(now)
+
+
+# ---- gateway keys ------------------------------------------------------------
+
+def _digest(value: str) -> bytes:
+    return hashlib.sha256(value.encode()).hexdigest().encode()
+
+
+def gateway_keys() -> dict[str, dict]:
+    """name -> {"sha256", "created", "hint", "collections", "rpm", "last_used"}.
+    Never a key itself."""
+    return locked.read_json(KEYS)
+
+
+def make_key(name: str, collections: list[str] | None = None, rpm: int | None = None) -> str:
+    """A new key, returned this once: only its SHA-256 is kept, and its last
+    four characters, so a key found in a log or a repository can be told
+    apart (the bwk_ prefix says whose it is). A slow hash would add nothing,
+    the key being 256 random bits rather than a password. `collections`: the
+    ones it may call (none = all); `rpm`: chat completions a minute (none = no cap)."""
+    if not KEY_NAME.match(name):
+        raise ValueError("name: 2-32 lower-case letters, digits or '-', e.g. open-webui")
+    if name == "local":
+        raise ValueError("local is what a call without a key is recorded as; pick another name")
+    collections = sorted(set(collections or []))
+    if not all(isinstance(c, str) and KEY_NAME.match(c) for c in collections):
+        raise ValueError("collections: names such as core, as they follow bench/")
+    if rpm is not None and (not isinstance(rpm, int) or isinstance(rpm, bool) or not 1 <= rpm <= 10_000):
+        raise ValueError("requests a minute: a whole number from 1 to 10000, or empty for no cap")
+    value = PREFIX + secrets.token_urlsafe(32)
+
+    def add(data: dict) -> None:
+        if name in data:
+            raise ValueError(f"there is already a key named {name}: revoke it first, or pick another name")
+        data[name] = {"sha256": _digest(value).decode(), "hint": value[-4:],
+                      "created": dt.datetime.now().astimezone().isoformat(timespec="seconds"),
+                      **({"collections": collections} if collections else {}),
+                      **({"rpm": rpm} if rpm else {})}
+    locked.update(KEYS, add)
+    return value
+
+
+def _touch(name: str) -> None:
+    """Note when key `name` was last used, at most once a minute: chat UIs poll
+    /v1/models, and a write per poll would be a write per page load."""
+    now = dt.datetime.now().astimezone()
+
+    def stamp(data: dict) -> None:
+        entry = data.get(name)
+        if not isinstance(entry, dict):
+            return                                     # revoked meanwhile
+        try:
+            if (now - dt.datetime.fromisoformat(entry.get("last_used", ""))).total_seconds() < 60:
+                return
+        except (TypeError, ValueError):
+            pass
+        entry["last_used"] = now.isoformat(timespec="seconds")
+    try:
+        locked.update(KEYS, stamp)
+    except (OSError, TimeoutError):                    # a busy file must not refuse the call
+        pass
+
+
+def revoke_key(name: str) -> bool:
+    return locked.update(KEYS, lambda data: data.pop(name, None)) is not None
+
+
+def key_name(value: str) -> str | None:
+    """The name of the gateway key `value` is, or None. Every stored digest is
+    compared in constant time, and all of them, whatever matched."""
+    if not value:
+        return None
+    given, found = _digest(value), None
+    for name, entry in gateway_keys().items():
+        stored = str(entry.get("sha256", "")).encode() if isinstance(entry, dict) else b""
+        if hmac.compare_digest(given, stored):
+            found = name
+    return found
+
+
+def access(handler, code) -> None:
+    """One access-log line on stderr: method, route, status, who. Never the
+    query (a sign-in's code travels in one) or a Connect callback's id. A
+    request line too broken to parse has no path or command yet."""
+    route = re.sub(r"^(/connect/callback)/.*", r"\1/...", (getattr(handler, "path", "") or "").split("?", 1)[0])
+    sys.stderr.write(f"{dt.datetime.now():%Y-%m-%d %H:%M:%S} {getattr(handler, 'command', '') or '-'}"
+                     f" {ascii(route[:200])[1:-1]}"
+                     f" {getattr(code, 'value', code)} {getattr(handler, 'who', '-')}\n")
+
+
+def _selftest() -> None:
+    import pathlib
+    import tempfile
+    global KEYS
+    ok = {"Host": "127.0.0.1:8123", "Content-Type": "application/json", "Content-Length": "12"}
+
+    def status(port=8123, post=True, cap=100, cors=False, **h):
+        try:
+            check({**ok, **h}, port, post, cap, cors)
+            return 0
+        except Refused as r:
+            return r.status
+    assert status() == 0
+    assert status(Host="localhost:8123") == 0 and status(Host="LOCALHOST:8123") == 0
+    assert status(Host="evil.example:8123") == 403            # DNS rebinding
+    assert status(Host="127.0.0.1") == 403                    # port is part of the name
+    assert status(port=8122) == 403                           # the other server's port
+    assert status(post=False, Host="evil.example") == 403     # GETs too
+    assert status(post=False, **{"Content-Type": ""}) == 0    # a GET needs no JSON type
+    assert status(**{"Content-Type": "text/plain"}) == 415    # the form a page can send freely
+    assert status(Origin="http://evil.example") == 403
+    assert status(Origin="http://127.0.0.1:8123") == 0
+    assert status(**{"Sec-Fetch-Site": "cross-site"}) == 403
+    assert status(**{"Sec-Fetch-Site": "same-origin"}) == 0
+    assert status(**{"Content-Length": "101"}) == 413
+    assert status(**{"Content-Length": "x"}) == 400
+    assert status(**{"Content-Length": "-1"}) == 400
+    saved = {k: os.environ.pop(k, None) for k in ("BENCH_ALLOWED_HOSTS", "BENCH_ALLOWED_ORIGINS",
+                                                    "BENCHWARMERS_PUBLIC_URL", "BENCHWARMERS_ADMIN_TOKEN",
+                                                    "BENCHWARMERS_REQUIRE_KEYS")}
+    try:
+        os.environ["BENCH_ALLOWED_HOSTS"] = "bench:8123"
+        os.environ["BENCH_ALLOWED_ORIGINS"] = "http://localhost:3000"
+        assert status(Host="bench:8123") == 0
+        assert status(Origin="http://localhost:3000") == 0
+        # The public URL is a Host and an Origin, https included.
+        os.environ["BENCHWARMERS_PUBLIC_URL"] = "https://Bench.example.com/"
+        assert status(Host="bench.example.com", Origin="https://bench.example.com") == 0
+        assert status(Host="bench.example.com", Origin="https://evil.example") == 403
+        # CORS: on a gateway route an allowed origin may be cross-site; the dashboard's rule stands.
+        site = {"Sec-Fetch-Site": "cross-site"}
+        assert status(cors=True, Origin="http://localhost:3000", **site) == 0
+        assert status(Origin="http://localhost:3000", **site) == 403
+        assert status(cors=True, Origin="http://evil.example", **site) == 403
+        assert status(cors=True, **site) == 403                             # no Origin: not a CORS call
+        assert cors_origin({"Origin": "http://localhost:3000"}, 8123) == "http://localhost:3000"
+        assert cors_origin({"Origin": "http://evil.example"}, 8123) == "" and cors_origin({}, 8123) == ""
+
+        # Binding: this machine needs nothing; anywhere else needs a long enough token.
+        def refused(address: str) -> bool:
+            try:
+                startup(address)
+                return False
+            except SystemExit:
+                return True
+        assert not refused("127.0.0.1") and not refused("localhost") and not refused("::1")
+        assert refused("0.0.0.0") and refused("") and refused("192.0.2.7")
+        os.environ["BENCHWARMERS_ADMIN_TOKEN"] = "short"
+        assert refused("127.0.0.1")
+        token = os.environ["BENCHWARMERS_ADMIN_TOKEN"] = "t" * 24
+        assert not refused("0.0.0.0")
+        os.environ["BENCHWARMERS_PUBLIC_URL"] = "bench.example.com"       # no scheme
+        assert refused("0.0.0.0")
+        os.environ["BENCHWARMERS_PUBLIC_URL"] = "https://bench.example.com"
+
+        # The login cookie: only the token makes one; tampered, expired or foreign ones fail.
+        assert login(token) and not login("t" * 23) and not login("")
+        cookie = session_cookie()
+        assert "HttpOnly" in cookie and "SameSite=Strict" in cookie and "Secure" in cookie and token not in cookie
+        value = cookie.split(";", 1)[0]
+        assert admin_caller({"Cookie": value}) == "admin"
+        assert admin_caller({"Cookie": f"theme=x; {value}"}) == "admin"
+        for bad in ("", value + "0", value.replace(f"{COOKIE}=", f"{COOKIE}=1"), f"{COOKIE}=1.abc",
+                    f"{COOKIE}=²." + "0" * 64, f"{COOKIE}=0.{_sign(0).decode()}", 'x="'):
+            try:
+                admin_caller({"Cookie": bad})
+                raise AssertionError(f"accepted cookie {bad!r}")
+            except Refused as r:
+                assert r.status == 401
+        # The admin token as a bearer (the bench CLI): it, and only it.
+        assert admin_caller({"Authorization": f"Bearer {token}"}) == "admin"
+        assert admin_caller({"Authorization": f"bearer  {token} "}) == "admin"
+        for bad in (f"Bearer {token}x", "Bearer " + "t" * 23, f"Basic {token}"):
+            try:
+                admin_caller({"Authorization": bad})
+                raise AssertionError(f"accepted {bad!r}")
+            except Refused as r:
+                assert r.status == 401
+        try:                                                    # a wrong bearer is refused, not saved by a cookie
+            admin_caller({"Authorization": f"Bearer {token}x", "Cookie": value})
+            raise AssertionError("a wrong bearer fell back to the cookie")
+        except Refused:
+            pass
+        os.environ["BENCHWARMERS_ADMIN_TOKEN"] = "u" * 24                  # a new token signs everyone out
+        try:
+            admin_caller({"Cookie": value})
+            raise AssertionError("an old token's cookie still works")
+        except Refused:
+            pass
+
+        # Gateway keys: shown once, kept hashed, constant-time, revocable; never a cookie's job.
+        with tempfile.TemporaryDirectory() as tmp:
+            real, KEYS = KEYS, pathlib.Path(tmp) / "gateway-keys.json"
+            try:
+                key = make_key("chat-ui")
+                assert key.startswith("bwk_") and key not in KEYS.read_text(encoding="utf-8")
+                assert gateway_keys()["chat-ui"]["sha256"] == hashlib.sha256(key.encode()).hexdigest()
+                for bad in ("Bad Name", "x", "chat-ui"):                    # form, length, taken
+                    try:
+                        make_key(bad)
+                        raise AssertionError(f"made a key named {bad!r}")
+                    except ValueError:
+                        pass
+                other = make_key("agent")
+                assert key_name(key) == "chat-ui" and key_name(other) == "agent"
+                assert key_name("") is None and key_name(key + "x") is None and key_name("bwk_é") is None
+                assert gateway_caller({"Authorization": f"Bearer {key}"}) == "chat-ui"
+                assert gateway_caller({"Authorization": f"bearer  {key} "}) == "chat-ui"
+                entry = gateway_keys()["chat-ui"]
+                assert entry["hint"] == key[-4:] and entry["last_used"] and "rpm" not in entry, entry
+                # Each refusal says which it was, and how to authenticate (RFC 6750).
+                for headers, said in (({}, "needs a gateway key"), ({"Cookie": value}, "needs a gateway key"),
+                                      ({"Authorization": key}, "Bearer <key>"), ({"Authorization": "Bearer "}, "Bearer <key>"),
+                                      ({"Authorization": f"Basic {key}"}, "Bearer <key>"),
+                                      ({"Authorization": "Bearer sk-dummy"}, "some other key"),
+                                      ({"Authorization": f"Bearer {key}x"}, "unknown or revoked")):
+                    try:
+                        gateway_caller(headers)
+                        raise AssertionError(f"gateway let in {headers}")
+                    except Refused as r:
+                        assert r.status == 401 and said in str(r), (headers, str(r))
+                        assert r.headers["WWW-Authenticate"].startswith("Bearer"), r.headers
+                # Scope and cap, set when made: checked per collection, and per minute.
+                for name, kw in (("scoped", {"collections": ["Core"]}), ("scoped", {"rpm": 0}),
+                                 ("scoped", {"rpm": True}), ("scoped", {"rpm": "5"}), ("local", {})):
+                    try:
+                        make_key(name, **kw)
+                        raise AssertionError(f"made a key with {kw}")
+                    except ValueError:
+                        pass
+                make_key("scoped", ["vision", "core", "core"], 2)
+                assert gateway_keys()["scoped"]["collections"] == ["core", "vision"]
+                assert may_call("scoped", "vision") and not may_call("scoped", "reviewers")
+                assert may_call("chat-ui", "reviewers") and may_call("local", "reviewers")
+                spend("scoped")
+                spend("scoped")
+                try:
+                    spend("scoped")
+                    raise AssertionError("a third request in a minute on a cap of 2")
+                except Refused as r:
+                    assert r.status == 429 and 1 <= int(r.headers["Retry-After"]) <= 61, r.headers
+                for _ in range(5):
+                    spend("chat-ui")                                    # no cap
+                for bad in ("core", "", 0, False, [1]):                 # a bad edit fails closed
+                    locked.update(KEYS, lambda d: d["scoped"].__setitem__("collections", bad))
+                    assert scope("scoped") is None and not may_call("scoped", "core"), bad
+                try:                                                    # a key never opens the dashboard
+                    admin_caller({"Authorization": f"Bearer {key}"})
+                    raise AssertionError("a gateway key opened the dashboard")
+                except Refused:
+                    pass
+                assert revoke_key("chat-ui") and not revoke_key("chat-ui") and key_name(key) is None
+                assert key_name(other) == "agent"
+                # No admin token: as before keys existed, whatever is sent.
+                del os.environ["BENCHWARMERS_ADMIN_TOKEN"]
+                assert gateway_caller({}) == "local" and gateway_caller({"Authorization": "Bearer x"}) == "local"
+                assert admin_caller({}) == "local"
+                # ...unless keys are required on this machine too; a misspelt value still asks.
+                for v in ("1", "required", "TRUE"):
+                    os.environ["BENCHWARMERS_REQUIRE_KEYS"] = v
+                    assert gateway_caller({"Authorization": f"Bearer {other}"}) == "agent"
+                    try:
+                        gateway_caller({"Authorization": "Bearer x"})
+                        raise AssertionError(f"BENCHWARMERS_REQUIRE_KEYS={v} let a dummy bearer in")
+                    except Refused as r:
+                        assert r.status == 401
+                    assert admin_caller({}) == "local"                  # the dashboard stays open
+                os.environ["BENCHWARMERS_REQUIRE_KEYS"] = "off"
+                assert not keys_required() and gateway_caller({}) == "local"
+            finally:
+                KEYS = real
+        assert gateway_route("/v1/models") and gateway_route("/bench/verdict")
+        assert not gateway_route("/api/endpoints") and not gateway_route("/v1") and not gateway_route("//v1/x")
+    finally:
+        for k, v in saved.items():
+            os.environ.pop(k, None)
+            if v is not None:
+                os.environ[k] = v
+    print("ok")
+
+
+if __name__ == "__main__":
+    _selftest()