diff --git a/chv-guest/authd.py b/chv-guest/authd.py new file mode 100644 index 0000000..6e0e473 --- /dev/null +++ b/chv-guest/authd.py @@ -0,0 +1,325 @@ +#!/usr/bin/env python3 +"""chv-authd:客内凭证守护进程(feat-chv-muse-p3)。 + +对齐 Muse 官方安全文:authd 保管凭证(含 token,存于用户 VM +而非集中基础设施),代理只见替身令牌;凭证存入运行时单元之外 +的安全凭证存储,代理不可见,仅在需要时注入。 + +本进程跑在客内、运行时单元之外(root 身份的 systemd 服务, +见 chv-authd.service;P2 的 nspawn 单元落地后,codex 在单元内, +本进程在单元外——单元内 root 映射为非特权用户,读不到本进程的 +socket 与 seed 文件)。它只做三件事: + +1. 启动时读 /etc/chv/authd-seed(0600 root,seed 投递),把 + GITEA_TOKEN 持在内存里(cloud.chv.authd.CredentialVault), + 除此之外不落盘、不进日志; +2. 在 Unix socket /run/chv-authd/authd.sock 上接受 git credential + helper 的请求(SO_PEERCRED 只放行 root 与 chv 用户); +3. 每次放凭证前先过 Sentinel 评估(cloud.chv.sentinel,主机/ + 方法/路径/解析后 IP/SSRF),再按次释放(单次、300 秒 TTL), + 全程审计(审计里无值)。 + +只用标准库。cloud/chv/authd.py 与 cloud/chv/sentinel.py 是叶子 +模块(只 import 标准库),镜像构建时随本文件一起拷进 +/usr/local/lib/chv-authd/(见 docs/chv-muse-p3-gap.md §2 的接入 +说明),本文件把该目录加进 sys.path 后 import。 +""" + +import argparse +import json +import os +import socket +import socketserver +import struct +import sys +import threading + +DEFAULT_LIB_DIR = "/usr/local/lib/chv-authd" +DEFAULT_SEED_PATH = "/etc/chv/authd-seed" +DEFAULT_SOCKET_PATH = "/run/chv-authd/authd.sock" +DEFAULT_LOG_PATH = "/var/log/chv-authd.log" + +# seed 文件里允许的键(与 cloud.chv.gitea_sync.GITEA_SYNC_KEYS +# 同款白名单;未知键一律丢掉,调用方拼错的键绝不静默进保险库)。 +SEED_KEYS = frozenset({ + "GITEA_URL", "GITEA_USER", "GITEA_TOKEN", + "SYNC_REPO", "SYNC_BRANCH", "GITEA_EMAIL", +}) + +# SO_PEERCRED 放行的 uid:root + chv(git 以 chv 身份跑 helper)。 +# 可用环境变量 CHV_AUTHD_ALLOWED_UIDS 覆盖(逗号分隔)。 +DEFAULT_ALLOWED_UIDS = (0, 1000) + + +def _load_seed(path): + """读 seed 文件(KEY=VALUE,每行一项)→ 白名单后的 dict。 + 文件不存在即 FileNotFoundError(调用方 fail-fast,绝不带着 + “无凭证”的 authd 继续跑、让 git 静默拿不到凭证)。""" + values = {} + with open(path) as fh: + for line in fh: + line = line.strip() + if not line or line.startswith("#") or "=" not in line: + continue + key, value = line.split("=", 1) + key = key.strip() + if key in SEED_KEYS and value.strip(): + values[key] = value.strip() + return values + + +def _audit(log_path, record): + """审计落盘(JSON 行,0600):只含绑定与 verdict,不含值。 + 落盘失败只记 stderr,绝不炸守护进程(审计失败不改变放行 + 决策是 worker 层的既定语义,见 executor._emit)。""" + line = json.dumps(record, sort_keys=True) + try: + with open(log_path, "a") as fh: + fh.write(line + "\n") + except OSError as exc: + print("chv-authd: audit 落盘失败:%s" % exc, file=sys.stderr) + print("chv-authd: %s" % line, file=sys.stderr) + + +def _operation_for_path(path): + """git 智能 HTTP 路径 → 操作(helper 不知道 git 操作,守护 + 进程按路径推断;Sentinel 的 L7 规则已把路径限定在这三条内, + 推断不到的一律拒绝): + - .../git-receive-pack → push(写) + - .../git-upload-pack → pull(读) + - .../info/refs → pull(读,clone 也走这条)""" + if path.endswith(".git/git-receive-pack"): + return "POST", "push" + if path.endswith(".git/git-upload-pack"): + return "POST", "pull" + if path.endswith(".git/info/refs"): + return "GET", "pull" + return None, None + + +def _repo_from_path(path): + """从 git 路径取仓库名:/admin/repo.git/info/refs → admin/repo。 + 取不出即 None(调用方拒绝)。""" + parts = [p for p in (path or "").split("/") if p] + if len(parts) < 2: + return None + repo = parts[1] + if not repo.endswith(".git"): + return None + return "%s/%s" % (parts[0], repo[:-len(".git")]) + + +class AuthdHandler(socketserver.BaseRequestHandler): + """单次请求处理(每连接一 JSON 行):鉴权 → Sentinel 评估 → + 按次释放 → 回包。任何一步拒绝都回 {ok:false},不含值。""" + + def handle(self): + server = self.server + try: + creds = self.request.getsockopt( + socket.SOL_SOCKET, socket.SO_PEERCRED, + struct.calcsize("3i")) + _pid, uid, _gid = struct.unpack("3i", creds) + except (OSError, struct.error): + self._reply({"ok": False, "error": "peercred_failed"}) + return + if uid not in server.allowed_uids: + server.audit({"event": "authd_peer_rejected", + "uid": uid}) + self._reply({"ok": False, "error": "uid_not_allowed"}) + return + try: + buf = b"" + while b"\n" not in buf: + chunk = self.request.recv(4096) + if not chunk: + break + buf += chunk + if len(buf) > 65536: + break + req = json.loads(buf.decode("utf-8")) + except (OSError, ValueError): + self._reply({"ok": False, "error": "bad_request"}) + return + self._serve(req) + + def _reply(self, obj): + try: + self.request.sendall( + (json.dumps(obj) + "\n").encode("utf-8")) + except OSError: + pass + + def _serve(self, req): + server = self.server + connector = req.get("connector") + host = (req.get("host") or "").lower() + path = req.get("path") or "/" + if connector != "gitea" or not host or not path: + server.audit({"event": "authd_credential_refused", + "kind": "bad_request_fields", + "connector": connector}) + self._reply({"ok": False, "error": "bad_request_fields"}) + return + method, operation = _operation_for_path(path) + if method is None: + server.audit({"event": "authd_credential_refused", + "kind": "path_not_allowed", + "connector": connector, "host": host, + "path": path}) + self._reply({"ok": False, "error": "path_not_allowed"}) + return + repo = _repo_from_path(path) + if repo is None: + server.audit({"event": "authd_credential_refused", + "kind": "repo_unparseable", + "connector": connector, "host": host}) + self._reply({"ok": False, "error": "repo_unparseable"}) + return + # Sentinel 评估:主机/方法/路径/解析后 IP/SSRF。 + from cloud.chv.sentinel import EgressRequest + scheme = "https" + port = 443 + egress = EgressRequest(connector, method, scheme, host, + port, path) + decision = server.policy.evaluate( + egress, resolve=getattr(server, "resolve", None)) + if decision.verdict != "allow": + record = decision.audit(egress) + server.audit(record) + self._reply({"ok": False, + "error": "sentinel_" + decision.verdict, + "reasons": decision.reasons}) + return + server.audit(decision.audit(egress)) + # 按次释放:绑定须与入库作用域一致(主机/仓库),单次有效。 + try: + release = server.vault.release( + connector, "GITEA_TOKEN", host, repo, operation) + except Exception as exc: # AuthdRefusal:kind 在异常上 + kind = getattr(exc, "kind", "refused") + server.audit({"event": "authd_credential_refused", + "kind": kind, "connector": connector, + "host": host, "repo": repo, + "operation": operation}) + self._reply({"ok": False, "error": kind}) + return + server.audit(server.vault.audit_release(release)) + try: + value = server.vault.consume(release.release_id) + except Exception as exc: + kind = getattr(exc, "kind", "refused") + server.audit({"event": "authd_credential_refused", + "kind": kind, "connector": connector, + "host": host, "repo": repo, + "operation": operation}) + self._reply({"ok": False, "error": kind}) + return + self._reply({"ok": True, "username": server.gitea_user, + "password": value, + "expires_at": release.expires_at}) + + +class AuthdServer(socketserver.ThreadingUnixStreamServer): + """Unix socket 服务(线程化,daemon_threads=True 避免僵尸线程)。""" + + daemon_threads = True + allow_reuse_address = False + + def __init__(self, socket_path, vault, policy, gitea_user, + allowed_uids, log_path, resolve=None): + self.vault = vault + self.policy = policy + self.gitea_user = gitea_user + self.allowed_uids = frozenset(allowed_uids) + self.log_path = log_path + # 解析器可注入(测试用假解析器;不给时 Sentinel 用真实 + # getaddrinfo)。与执行器的 sentinel_resolver 同款口径。 + self.resolve = resolve + directory = os.path.dirname(socket_path) + if directory: + os.makedirs(directory, exist_ok=True) + try: + os.chmod(directory, 0o700) + except OSError: + pass + if os.path.exists(socket_path): + os.unlink(socket_path) + super().__init__(socket_path, AuthdHandler) + try: + os.chmod(socket_path, 0o600) + except OSError: + pass + + def audit(self, record): + _audit(self.log_path, record) + + +def build_server(seed_path=DEFAULT_SEED_PATH, + socket_path=DEFAULT_SOCKET_PATH, + log_path=DEFAULT_LOG_PATH, + lib_dir=DEFAULT_LIB_DIR, + allowed_uids=None, + resolve=None): + """建服务(不启动):读 seed → 入库 → 建策略 → 建 socket 服务。 + + seed 缺必填键(GITEA_URL/GITEA_USER/GITEA_TOKEN/SYNC_REPO) + 即 ValueError(只含缺失键名,不含值);无 seed 文件即 + FileNotFoundError——authd 绝不带着“无凭证”启动。""" + if lib_dir not in sys.path: + sys.path.insert(0, lib_dir) + from cloud.chv.authd import CredentialVault + from cloud.chv.sentinel import SentinelPolicy + seed = _load_seed(seed_path) + missing = sorted({"GITEA_URL", "GITEA_USER", "GITEA_TOKEN", + "SYNC_REPO"} - set(seed)) + if missing: + raise ValueError( + "authd seed 不完整,缺失:%s" % ",".join(missing)) + from urllib.parse import urlparse + gitea_host = (urlparse(seed["GITEA_URL"]).hostname or "").lower() + if not gitea_host: + raise ValueError("authd seed 的 GITEA_URL 主机解析不出") + vault = CredentialVault() + vault.store("gitea", "GITEA_TOKEN", seed["GITEA_TOKEN"], + host=gitea_host, repo=seed["SYNC_REPO"]) + extra = [h.strip().lower() + for h in (os.environ.get("CHV_SENTINEL_ALLOW_HOSTS") + or "").split(",") if h.strip()] + policy = SentinelPolicy( + allow_hosts={gitea_host} | set(extra)) + uids = allowed_uids + if uids is None: + raw = os.environ.get("CHV_AUTHD_ALLOWED_UIDS") or "" + uids = [int(p) for p in raw.split(",") if p.strip()] \ + if raw.strip() else list(DEFAULT_ALLOWED_UIDS) + server = AuthdServer(socket_path, vault, policy, + seed["GITEA_USER"], uids, log_path, + resolve=resolve) + return server + + +def main(argv=None): + """前台运行(systemd Type=simple 直接跑本进程,不 daemonize)。""" + parser = argparse.ArgumentParser() + parser.add_argument("--seed", default=DEFAULT_SEED_PATH) + parser.add_argument("--socket", default=DEFAULT_SOCKET_PATH) + parser.add_argument("--log", default=DEFAULT_LOG_PATH) + parser.add_argument("--lib", default=DEFAULT_LIB_DIR) + args = parser.parse_args(argv) + server = build_server(seed_path=args.seed, + socket_path=args.socket, + log_path=args.log, lib_dir=args.lib) + print("chv-authd: listening on %s" % args.socket, + file=sys.stderr) + try: + server.serve_forever() + except KeyboardInterrupt: + pass + finally: + server.server_close() + return 0 + + +if __name__ == "__main__": + sys.exit(main()) diff --git a/chv-guest/chv-authd.service b/chv-guest/chv-authd.service new file mode 100644 index 0000000..a018b38 --- /dev/null +++ b/chv-guest/chv-authd.service @@ -0,0 +1,22 @@ +[Unit] +Description=CHV authd (credential broker for chv-sync) +# 注意时序:本服务须在 chv-sync boot 之前就绪。cloud-final 的 +# runcmd(99-chv-sync.cfg)在 chv-sync boot 之前显式 +# `systemctl start chv-authd`;这里不依赖 After=cloud-final +#(cloud-final 跑完才 After 得到它,太晚)。enable 后开机自启 +# 只是兜底,首启顺序以 runcmd 的显式 start 为准。 +After=network-online.target + +[Service] +Type=simple +# authd 跑在运行时单元之外(root 身份):它是凭证的唯一持有者, +# 单元内的 agent 只能经 Unix socket + SO_PEERCRED 向它要凭证, +# 读不到它的内存与 seed 文件(P2 单元隔离落地后)。 +ExecStart=/usr/bin/python3 /usr/local/bin/chv-authd +Restart=on-failure +RestartSec=2 +StandardOutput=journal +StandardError=journal + +[Install] +WantedBy=multi-user.target diff --git a/chv-guest/chv-sync.sh b/chv-guest/chv-sync.sh index bfc9c05..81963ea 100755 --- a/chv-guest/chv-sync.sh +++ b/chv-guest/chv-sync.sh @@ -7,19 +7,18 @@ # 的规则:分叉或冲突时失败关闭,不强制覆盖、不自动 rebase、 # 不静默丢提交。 # -# 凭证绝不进镜像。启动时由 seed 盘投递 /etc/chv/guest-env -# (0600,root 所有),格式(KEY=VALUE,每行一项): -# GITEA_URL=https://gitea.tose.top -# GITEA_USER= -# GITEA_TOKEN=<专用 API token,仅此次测试/此次 VM 有效> -# SYNC_REPO= -# SYNC_BRANCH=<可选,默认分支> -# GITEA_EMAIL=<可选,commit 身份邮箱;不给则留镜像占位> -# boot(root,cloud-init runcmd 调)读 guest-env,把 token 写进 +# 凭证绝不进镜像。启动时由 seed 盘投递两份文件(0600,root 所有): +# /etc/chv/guest-env(非密钥同步变量:GITEA_URL/GITEA_USER/ +# SYNC_REPO/SYNC_BRANCH/GITEA_EMAIL)与 +# /etc/chv/authd-seed(feat-chv-muse-p3:只装 GITEA_TOKEN, +# 只给客内 authd 守护进程读)。 +# boot(root,cloud-init runcmd 调)读 guest-env,authd 在时配 +# git credential helper 走 authd 按次取凭证(只进 git 子进程 +# 内存,不落盘);authd 不在时(旧镜像)才回落到把 token 写进 # /home/chv/.git-credentials(0600,chv 所有,git credential -# store 标准机制),再以 chv 身份 clone 到 ~/workspace/<仓名>。 -# pull/push/status 以 chv 身份跑,只用已落盘的 credential,不 -# 再读 guest-env;token 不会出现在 git remote URL 或镜像里。 +# store 标准机制)并警告。pull/push/status 以 chv 身份跑,用 +# helper 或已落盘的 credential,不再读 guest-env;token 不会 +# 出现在 git remote URL 或镜像里。 # # 用法: # chv-sync boot # 启动时 clone(已 clone 则静默退出,不自动 pull) @@ -64,7 +63,6 @@ cmd_boot() { set -a; . "$GUEST_ENV"; set +a : "${GITEA_URL:?guest-env 缺 GITEA_URL}" : "${GITEA_USER:?guest-env 缺 GITEA_USER}" - : "${GITEA_TOKEN:?guest-env 缺 GITEA_TOKEN}" : "${SYNC_REPO:?guest-env 缺 SYNC_REPO}" [ "$(id -u)" = "0" ] || die "boot 须以 root 运行(cloud-init runcmd)" local host="${GITEA_URL#https://}" @@ -74,12 +72,31 @@ cmd_boot() { local dir="$WORKSPACE/$name" mkdir -p "$WORKSPACE" chown "$SYNC_USER:$SYNC_USER" "$WORKSPACE" - # token 只落盘进用户家目录的 credential store(0600), - # 不进 git remote URL、不进镜像、不进日志(此处不打印 token)。 - local creds="$USER_HOME/.git-credentials" - printf 'https://%s:%s@%s\n' "$GITEA_USER" "$GITEA_TOKEN" "$host" > "$creds" - chmod 0600 "$creds" - chown "$SYNC_USER:$SYNC_USER" "$creds" + # feat-chv-muse-p3:authd 路径——socket 存在时凭证走 helper, + # 不再把真 token 写进 .git-credentials 落盘;helper 按次向 + # authd 取凭证,只进 git 子进程内存。socket 不存在(旧镜像) + # 时走旧路径 + 警告(过渡兼容,旧镜像退役后改为 fail-closed)。 + # 注意:authd 路径下 GITEA_TOKEN 不在 guest-env 里(它只进 + # authd-seed 给守护进程),所以这里不要求它;旧路径才要求。 + local authd_socket="${CHV_AUTHD_SOCKET:-/run/chv-authd/authd.sock}" + local helper="/usr/local/bin/git-credential-chv-authd" + if [ -S "$authd_socket" ] && [ -x "$helper" ]; then + runuser -u "$SYNC_USER" -- env HOME="$USER_HOME" git config --global credential.helper "$helper" + # helper 按路径推断仓库与操作(Sentinel L7 绑定),git + # 默认不把 path 发给 helper(须 credential.useHttpPath), + # 不开它每次都会因“请求不完整”拒绝——新测试抓到的真问题。 + runuser -u "$SYNC_USER" -- env HOME="$USER_HOME" git config --global credential.useHttpPath true + info "凭证走 chv-authd(不落盘)" + else + : "${GITEA_TOKEN:?guest-env 缺 GITEA_TOKEN(旧路径)}" + echo "chv-sync: 警告:chv-authd 未运行,凭证走旧落盘路径(须用含 authd 的新镜像)" >&2 + # token 只落盘进用户家目录的 credential store(0600), + # 不进 git remote URL、不进镜像、不进日志(此处不打印 token)。 + local creds="$USER_HOME/.git-credentials" + printf 'https://%s:%s@%s\n' "$GITEA_USER" "$GITEA_TOKEN" "$host" > "$creds" + chmod 0600 "$creds" + chown "$SYNC_USER:$SYNC_USER" "$creds" + fi if [ -d "$dir/.git" ]; then info "已 clone:$dir(不自动 pull,改动由用户显式同步)" exit 0 diff --git a/chv-guest/git-credential-chv-authd b/chv-guest/git-credential-chv-authd new file mode 100755 index 0000000..fa9f035 --- /dev/null +++ b/chv-guest/git-credential-chv-authd @@ -0,0 +1,107 @@ +#!/usr/bin/env python3 +"""git credential helper:向 chv-authd 按次取凭证(feat-chv-muse-p3)。 + +git credential 协议:git 把 `protocol/host/path`(每行 `key=value`, +空行结束)写进本进程 stdin;本进程向 authd 的 Unix socket 发 +一次 JSON 请求,拿到用户名/口令后按协议回给 git(属性 + 空行)。 + +口令只进 git 子进程内存:本进程不写文件、不打日志(拒绝时只 +报 kind,不含值);`password_expiry_utc` 按释放的 expires_at +回给 git,git 在该时间后重新问 helper(不缓存过期凭证)。 +authd 拒绝 / socket 不通时本进程 exit 1——git 操作随之失败 +(fail-closed,绝不回落成“无凭证也照推”)。 + +用法(chv-sync boot 配好后 git 自动调,不用手工跑): + git config --global credential.helper /usr/local/bin/git-credential-chv-authd + git config --global credential.useHttpPath true +(后者不可少:git 默认不把 path 发给 helper,没有 path 本进程 +无法推断仓库与操作,会直接拒绝。) +socket 路径可用 CHV_AUTHD_SOCKET 覆盖(测试用 /tmp 下的假 socket)。 +""" + +import json +import os +import socket +import sys + +DEFAULT_SOCKET_PATH = "/run/chv-authd/authd.sock" +SOCKET_TIMEOUT_SECONDS = 5 + + +def _read_credential_request(): + """读 git credential 协议的请求属性(空行结束)→ dict。""" + attrs = {} + for line in sys.stdin: + line = line.rstrip("\n") + if not line: + break + if "=" in line: + key, value = line.split("=", 1) + attrs[key.strip()] = value.strip() + return attrs + + +def _ask_authd(socket_path, host, path): + """向 authd 要一次凭证:返回 (username, password, expires_at)。 + 拒绝/通信失败即 RuntimeError(只含 kind,不含值)。""" + req = {"connector": "gitea", "host": host, "path": path} + sock = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM) + sock.settimeout(SOCKET_TIMEOUT_SECONDS) + try: + sock.connect(socket_path) + sock.sendall((json.dumps(req) + "\n").encode("utf-8")) + chunks = [] + while True: + data = sock.recv(65536) + if not data: + break + chunks.append(data) + if b"\n" in data: + break + except OSError as exc: + raise RuntimeError("authd_unreachable:%s" % exc) + finally: + sock.close() + try: + resp = json.loads(b"".join(chunks).decode("utf-8")) + except ValueError: + raise RuntimeError("authd_bad_response") + if not resp.get("ok"): + raise RuntimeError("authd_refused:%s" + % resp.get("error", "refused")) + return (resp["username"], resp["password"], + resp.get("expires_at")) + + +def main(argv=None): + """helper 入口:读协议 → 问 authd → 回属性。拒绝即 exit 1。""" + attrs = _read_credential_request() + host = attrs.get("host", "") + path = attrs.get("path", "") + if attrs.get("protocol") != "https" or not host or not path: + print("git-credential-chv-authd: 请求不完整(须 protocol=https " + "+ host + path),拒绝", file=sys.stderr) + return 1 + if not path.startswith("/"): + path = "/" + path + # git 的 info/refs 路径带 ?service=... 查询串:Sentinel 的 + # L7 规则按干净路径匹配,这里先去掉查询串再问 authd。 + path = path.split("?", 1)[0] + socket_path = os.environ.get("CHV_AUTHD_SOCKET") \ + or DEFAULT_SOCKET_PATH + try: + username, password, expires_at = _ask_authd( + socket_path, host, path) + except RuntimeError as exc: + print("git-credential-chv-authd: %s" % exc, file=sys.stderr) + return 1 + print("username=%s" % username) + print("password=%s" % password) + if expires_at: + print("password_expiry_utc=%d" % int(expires_at)) + print() + return 0 + + +if __name__ == "__main__": + sys.exit(main()) diff --git a/cloud/chv/authd.py b/cloud/chv/authd.py new file mode 100644 index 0000000..f3c5811 --- /dev/null +++ b/cloud/chv/authd.py @@ -0,0 +1,311 @@ +"""authd 式替身凭证服务(feat-chv-muse-p3,叶子模块)。 + +对齐 Muse 官方安全文(security.muse.ai 2026-09-08,见 +docs/chv-muse-p3-gap.md):凭证由 VM 内独立服务 authd 保管 +(含 OAuth token,存于用户 VM 而非集中基础设施),代理只见 +替身令牌;内置连接器的执行分离——每个 privsep worker 有明确 +凭证白名单(日历 worker 不能仅改参数就索取邮件凭证);登录 +凭证存入运行时单元之外的安全凭证存储,代理不可见,仅在需要 +时注入。 + +本模块是纯逻辑(只用标准库,不依赖 cloud.*),宿主侧与客内 +authd 守护进程共用同一份代码(见 chv-guest/authd.py): +- `CredentialVault`:只活内存的保险库;`store()` 只接受白名单 + 内的(连接器,凭证键);`release()` 按次发放,绑定 + {连接器, 主机, 仓库路径, 操作},单次有效 + 300 秒 TTL, + 过期即废;`__repr__` 与审计记录里绝无凭证值。 +- 替身的两档(见差距表 §1):本期默认是“按次释放”——释放给 + git 子进程的值仍是真 token,但只进 git 进程内存、不进 agent + 环境/argv/日志/落盘;真铸 token(Gitea API 铸短期专用 token) + 由 `mint_gitea_token()` 实现,默认关闭(须真 token 有建 token + 权限,现有 write:repository 测试 token 铸 token 会失败)。 +- 位置决策:authd 跑在客内、单元外(root 身份的 systemd 服务), + 不在宿主侧——Muse 的 authd 就在用户 VM 里;宿主侧方案要走 + vsock(P1 范围、未就绪),且会把全部 VM 的 token 集中到宿主。 +""" + +import hashlib +import hmac +import json +import secrets +import time +import urllib.error +import urllib.request +from collections.abc import Callable, Iterable + +# 连接器白名单(privsep 口径):本期只认 gitea 一个连接器。 +# 未知连接器一律拒绝,不得以“默认放行”扩展。 +ALLOWED_CONNECTORS = frozenset({"gitea"}) + +# 每连接器的凭证白名单:连接器 → 允许的凭证键。gitea 只允许 +# GITEA_TOKEN;其它键(哪怕调用方传进来)一律拒绝入库。 +CONNECTOR_CREDENTIALS = { + "gitea": frozenset({"GITEA_TOKEN"}), +} + +# 按次释放的默认有效期(秒)与次数:单次、300 秒,过期即废。 +DEFAULT_RELEASE_TTL_SECONDS = 300 +DEFAULT_RELEASE_MAX_USES = 1 + + +class AuthdRefusal(Exception): + """authd 拒绝:kind + 原因,不含任何凭证值。""" + + def __init__(self, kind: str, message: str) -> None: + super().__init__("authd_refused:%s: %s" % (kind, message)) + self.kind = kind + + +class SurrogateRelease: + """一次按次释放(替身):绑定到一次具体的凭证使用。 + + release_id 是不透明 id(审计与吊销用,不含值);value 只在 + 释放对象里出现一次,由调用方(客内守护进程)经 Unix socket + 交给 git credential helper、再由 helper 交给 git 子进程—— + 除此之外不进任何环境、argv、日志、审计。`consume()` 校验 + 绑定与有效期,单次有效:用过一次即废,过期即废。""" + + def __init__(self, release_id: str, connector: str, key: str, + host: str, repo: str, operation: str, value: str, + expires_at: float, max_uses: int = 1) -> None: + self.release_id = release_id + self.connector = connector + self.key = key + self.host = host + self.repo = repo + self.operation = operation + self.value = value + self.expires_at = expires_at + self.uses_left = max_uses + + def describe(self) -> dict: + """审计用描述(只含绑定,不含值)。""" + return {"release_id": self.release_id, + "connector": self.connector, "key": self.key, + "host": self.host, "repo": self.repo, + "operation": self.operation, + "expires_at": self.expires_at, + "uses_left": self.uses_left} + + def __repr__(self) -> str: # 绝不含 value + return ("SurrogateRelease(id=%s connector=%s host=%s " + "repo=%s operation=%s)" + % (self.release_id, self.connector, self.host, + self.repo, self.operation)) + + +class CredentialVault: + """只活内存的凭证保险库(一个 VM 一个实例,客内 authd 持有)。 + + - 值只存在 `self._store` 的内存 dict 里,不落盘、不进日志; + - `store()` 校验白名单:未知连接器 / 该连接器不允许的键 / + 空值一律拒绝入库(AuthdRefusal),绝不静默收下; + - `release()` 按次发放:调用方须给出完整绑定(主机、仓库、 + 操作),绑定与入库时的作用域不一致即拒绝; + - `consume()` 单次兑现:release 兑现一次即从库里删除,过期 + 也兑现不了;兑现失败不留可重试的残值。 + clock 可注入(测试用假时钟测过期),默认 time.time。""" + + def __init__(self, clock: Callable[[], float] | None = None) -> None: + self._store: dict[tuple[str, str], dict] = {} + self._releases: dict[str, SurrogateRelease] = {} + self._clock = clock or time.time + + def __repr__(self) -> str: # 绝不含任何凭证值 + return "CredentialVault(connectors=%s)" % sorted( + {c for (c, _k) in self._store}) + + def store(self, connector: str, key: str, value: str, + host: str | None = None, repo: str | None = None) -> None: + """入库:白名单校验 + 作用域记录(host/repo 是该凭证 + 允许使用的范围,release 时须一致)。""" + if connector not in ALLOWED_CONNECTORS: + raise AuthdRefusal( + "unknown_connector", + "连接器 %s 不在 authd 白名单内" % connector) + allowed = CONNECTOR_CREDENTIALS.get(connector, frozenset()) + if key not in allowed: + raise AuthdRefusal( + "credential_not_allowlisted", + "连接器 %s 不允许凭证键 %s" % (connector, key)) + if not isinstance(value, str) or not value.strip(): + raise AuthdRefusal( + "empty_credential", "凭证值为空,拒绝入库") + self._store[(connector, key)] = { + "value": value, "host": host, "repo": repo} + + def has(self, connector: str, key: str) -> bool: + """该(连接器,键)是否已入库(不返回值)。""" + return (connector, key) in self._store + + def forget(self, connector: str, key: str) -> bool: + """删一条入库凭证(VM 停止时调用);同时作废其全部未兑现 + 释放。未知键返回 False,不抛错。""" + gone = self._store.pop((connector, key), None) is not None + if gone: + doomed = [rid for rid, rel in self._releases.items() + if rel.connector == connector + and rel.key == key] + for rid in doomed: + self._releases.pop(rid, None) + return gone + + def release(self, connector: str, key: str, host: str, repo: str, + operation: str, + ttl_seconds: int = DEFAULT_RELEASE_TTL_SECONDS, + max_uses: int = DEFAULT_RELEASE_MAX_USES + ) -> SurrogateRelease: + """按次发放:绑定须与入库作用域一致(host/repo 有记录时 + 须一致,无记录时本次绑定即定作用域),否则拒绝。发放的 + 释放单次有效、ttl 后过期;发放本身记入库内待兑现表, + 须经 consume() 兑现一次才交出值。""" + entry = self._store.get((connector, key)) + if entry is None: + raise AuthdRefusal( + "credential_not_stored", + "连接器 %s 的凭证 %s 未入库" % (connector, key)) + if entry["host"] is not None and entry["host"] != host: + raise AuthdRefusal( + "scope_mismatch", + "凭证作用域主机不一致,拒绝释放") + if entry["repo"] is not None and entry["repo"] != repo: + raise AuthdRefusal( + "scope_mismatch", + "凭证作用域仓库不一致,拒绝释放") + if operation not in ("clone", "pull", "push"): + raise AuthdRefusal( + "operation_not_allowed", + "操作 %s 不在允许集内" % operation) + release = SurrogateRelease( + release_id=secrets.token_hex(8), + connector=connector, key=key, host=host, repo=repo, + operation=operation, value=entry["value"], + expires_at=self._clock() + ttl_seconds, + max_uses=max_uses) + self._releases[release.release_id] = release + return release + + def consume(self, release_id: str) -> str: + """兑现一次释放:过期/用完/未知 id 一律拒绝(AuthdRefusal, + 不含值)。兑现成功返回凭证值,并把该释放从库里删除 + (单次有效,用过即废)。""" + release = self._releases.get(release_id) + if release is None: + raise AuthdRefusal( + "release_unknown", "释放记录不存在或已兑现") + if self._clock() >= release.expires_at: + self._releases.pop(release_id, None) + raise AuthdRefusal("release_expired", "释放已过期") + if release.uses_left <= 0: + self._releases.pop(release_id, None) + raise AuthdRefusal("release_exhausted", "释放已用完") + release.uses_left -= 1 + value = release.value + self._releases.pop(release_id, None) + return value + + def audit_release(self, release: SurrogateRelease) -> dict: + """发放审计记录(只含绑定,不含值)。""" + record = release.describe() + record["event"] = "authd_credential_released" + return record + + def audit_refusal(self, kind: str, connector: str, + key: str | None = None) -> dict: + """拒绝审计记录(只含 kind 与名,不含值)。""" + return {"event": "authd_credential_refused", "kind": kind, + "connector": connector, "key": key} + + +def _default_post(url: str, headers: dict, payload: dict): + """默认 HTTP 客户端(urllib):返回 (status, body_dict)。 + 只用于 mint_gitea_token,测试注入假 post,不联网。""" + data = json.dumps(payload).encode("utf-8") + req = urllib.request.Request(url, data=data, headers=headers, + method="POST") + try: + with urllib.request.urlopen(req, timeout=15) as resp: + return resp.status, json.loads(resp.read() or b"{}") + except urllib.error.HTTPError as exc: + try: + body = json.loads(exc.read() or b"{}") + except ValueError: + body = {} + return exc.code, body + + +def mint_gitea_token(api_url: str, user: str, master_token: str, + name: str, scopes: Iterable[str] = ("write:repository",), + post: Callable[[str, dict, dict], tuple[int, dict]] + | None = None) -> str: + """经 Gitea API 铸短期专用 token(真替身路径,默认关闭)。 + + 调 `POST {api_url}/users/{user}/tokens`(Gitea 官方 API), + 新 token 与 master token 是不同的值、可独立吊销(DELETE 同 + 路径),到期由调用方吊销。返回新 token 值;失败即 + AuthdRefusal(不含任何 token 值)。 + 前提(见差距表 §3.2):master token 须有建 token 的权限; + 现有 write:repository 测试 token 铸 token 会失败(403), + 开这条路须另配有铸 token 权限的 token 并做真机验证。 + post 可注入(测试用假 post),默认走 urllib。""" + if not api_url or not user or not master_token or not name: + raise AuthdRefusal( + "mint_args_missing", "铸 token 参数不全") + base = api_url.rstrip("/") + url = "%s/users/%s/tokens" % (base, user) + headers = {"Authorization": "token %s" % master_token, + "Content-Type": "application/json"} + payload = {"name": name, "scopes": list(scopes)} + post_fn = post or _default_post + status, body = post_fn(url, headers, payload) + if status not in (200, 201): + raise AuthdRefusal( + "mint_failed", + "Gitea 铸 token 失败:status=%s" % status) + token = (body or {}).get("sha1") or (body or {}).get("token") + if not token: + raise AuthdRefusal( + "mint_empty", "Gitea 铸 token 未返回 token 值") + return token + + +def revoke_gitea_token(api_url: str, user: str, master_token: str, + token_id: object, + delete: Callable[[str, dict], int] | None = None + ) -> bool: + """吊销一个铸出的 token(`DELETE {api}/users/{user}/tokens/{id}`)。 + 成功返回 True;404 视为已不在(也返回 True,不报错); + 其它失败返回 False(调用方审计,不抛错——吊销失败不能炸任务, + 但须留审计)。delete 可注入,默认走 urllib。""" + if not api_url or not user or not master_token \ + or token_id is None: + return False + base = api_url.rstrip("/") + url = "%s/users/%s/tokens/%s" % (base, user, token_id) + headers = {"Authorization": "token %s" % master_token} + if delete is None: + def _do_delete(url: str, headers: dict) -> int: + req = urllib.request.Request(url, headers=headers, + method="DELETE") + try: + with urllib.request.urlopen(req, timeout=15) as resp: + return resp.status + except urllib.error.HTTPError as exc: + return exc.code + delete = _do_delete + status = delete(url, headers) + return status in (200, 204, 404) + + +def surrogate_fingerprint(value: str) -> str: + """凭证值的指纹(审计/日志里唯一允许出现的凭证派生形式): + SHA-256 前 8 位 hex。指纹不可逆推出原值,审计里只记指纹、 + 不记值。""" + return hashlib.sha256(value.encode("utf-8")).hexdigest()[:8] + + +def hmac_sign(key: str, message: str) -> str: + """HMAC-SHA256 签名(释放记录的防篡改绑定,客内守护进程 + 用它签发 release_id 的校验码;与 Gitea 无关)。""" + return hmac.new(key.encode("utf-8"), message.encode("utf-8"), + hashlib.sha256).hexdigest() diff --git a/cloud/chv/config.py b/cloud/chv/config.py index a9d2554..3d23e00 100644 --- a/cloud/chv/config.py +++ b/cloud/chv/config.py @@ -153,6 +153,12 @@ def _network_fragment(params: dict, env: dict) -> dict: "egress_proxy": env.get("CHV_EGRESS_PROXY"), "allow_direct_egress": env.get("CHV_ALLOW_DIRECT_EGRESS") == "1", + # feat-chv-muse-p3:Sentinel 白名单主机(逗号分隔), + # 默认只放行配置的 Gitea 主机(执行器 wake 预审时追加)。 + "sentinel_allow_hosts": tuple( + h.strip().lower() + for h in (env.get("CHV_SENTINEL_ALLOW_HOSTS") or "") + .split(",") if h.strip()), } diff --git a/cloud/chv/executor.py b/cloud/chv/executor.py index 07d37f1..ea4715d 100644 --- a/cloud/chv/executor.py +++ b/cloud/chv/executor.py @@ -279,6 +279,7 @@ class CloudHypervisorExecutor(Executor): disk_gb: float egress_proxy: str | None allow_direct_egress: bool + sentinel_allow_hosts: tuple task_timeout: float | None snapshot_url: str | None warm_pool_size: int @@ -303,6 +304,8 @@ class CloudHypervisorExecutor(Executor): audit_failure_count: int _vms: dict _guest_env_delivered: dict + _authd_delivered: dict + sentinel_resolver: "Callable[..., object] | None" def __init__(self, binary: str | None = None, kernel: str | None = None, @@ -329,7 +332,9 @@ class CloudHypervisorExecutor(Executor): api_timeout: float | None = None, ready_timeout: float | None = None, ready_interval: float | None = None, - host_key_mode: str | None = None) -> None: + host_key_mode: str | None = None, + sentinel_resolver: Callable[..., object] | None = None + ) -> None: self._env: dict = env if env is not None else os.environ # type: ignore[assignment] # 配置解析已拆到 cloud/chv/config.py(参数优先于环境 # 变量,默认值与 #216 一致;显式空用户/越界端口的收紧 @@ -342,13 +347,15 @@ class CloudHypervisorExecutor(Executor): "ssh_user": ssh_user, "ssh_port": ssh_port, "guest_host_py": guest_host_py, "cmdline": cmdline, "net": net, "net_json": net_json, "ssh_host": ssh_host, - "kvm_path": kvm_path, - "cpuinfo_path": cpuinfo_path, "ch_remote": ch_remote, - "model": model, "api_timeout": api_timeout, - "ready_timeout": ready_timeout, + "kvm_path": kvm_path, "cpuinfo_path": cpuinfo_path, + "ch_remote": ch_remote, "model": model, + "api_timeout": api_timeout, "ready_timeout": ready_timeout, "ready_interval": ready_interval, "host_key_mode": host_key_mode}, self._env) self._copy_config_fields(_cfg) + # feat-chv-muse-p3:Sentinel 解析器可注入(测试用假解析器 + # 测 SSRF 挡截;不给时 evaluate 用 socket.getaddrinfo)。 + self.sentinel_resolver = sentinel_resolver self._init_runtime(runner, launcher, api_client, sleeper) def _init_runtime(self, runner: Callable[..., object] | None, @@ -367,6 +374,7 @@ class CloudHypervisorExecutor(Executor): # 名 source 客内文件,未投递的密钥名一律拒绝(防止 # “wake 没投递、执行时却按名 source 一个旧文件”)。 self._guest_env_delivered = {} + self._init_authd_runtime() # feat-chv-closeup(P2-20):_guest_env_delivered 是 # 跨 wake/build_ssh_command/stop 的共享字典,wake 与 # 执行可并发在不同线程——此前无锁读写,set/get/pop @@ -396,6 +404,19 @@ class CloudHypervisorExecutor(Executor): self.launcher = _default_popen_launcher self._vms = {} + def _init_authd_runtime(self) -> None: + """feat-chv-muse-p3:authd 投递集与 Sentinel 待审批队列。 + _authd_delivered 是本实例已走 authd 通道投递密钥的 VM + 名 → 已投递密钥名集合(与 _guest_env_delivered 同锁、 + 同款语义,唯一区别:authd 投递的密钥名在 build_ssh_command + 里不报错、但绝不 source 进 codex 进程环境——它只给客内 + authd 守护进程,任务进程里永不出现)。_sentinel_queue + 是 Sentinel ask 待审批队列(内存态,进程重启即清; + ask 本来就意味着这次执行已停止)。""" + self._authd_delivered = {} + from cloud.chv.sentinel import ApprovalQueue + self._sentinel_queue = ApprovalQueue() + def _delivered_secrets(self, name: str) -> set: """P2-20 唯一读入口:该 VM 已 seed 投递的密钥名集合。""" with self._guest_env_lock: @@ -415,6 +436,26 @@ class CloudHypervisorExecutor(Executor): with self._guest_env_lock: self._guest_env_delivered.pop(name, None) + def _delivered_authd_secrets(self, name: str) -> set: + """feat-chv-muse-p3:该 VM 已走 authd 通道投递的密钥名 + 集合(与 _delivered_secrets 同款读入口)。""" + with self._guest_env_lock: + return set(self._authd_delivered.get(name, set())) + + def _mark_authd_delivered(self, name: str, secrets) -> None: + """feat-chv-muse-p3:记下该 VM 已走 authd 通道投递的 + 密钥名集合(secrets 为空即清记录,与旧语义一致)。""" + with self._guest_env_lock: + if secrets: + self._authd_delivered[name] = set(secrets) + else: + self._authd_delivered.pop(name, None) + + def _clear_authd_delivered(self, name: str) -> None: + """feat-chv-muse-p3:stop 时清掉该 VM 的 authd 投递记录。""" + with self._guest_env_lock: + self._authd_delivered.pop(name, None) + def _copy_config_fields(self, _cfg) -> None: """把 build_chv_config 的字段抄到 self(纯移动,字段集不变)。""" for _field in ( @@ -431,7 +472,7 @@ class CloudHypervisorExecutor(Executor): "snapshot_max_total_gb", "snapshot_retention_days", "snapshot_dir", "allow_remote_snapshot", "gitea_sync_env", "seed_user_data_extra_file", - "seed_network_config_file"): + "seed_network_config_file", "sentinel_allow_hosts"): setattr(self, _field, getattr(_cfg, _field)) # -- 名字 / 路径 --------------------------------------------------- @@ -868,7 +909,8 @@ class CloudHypervisorExecutor(Executor): def provision_host_key(self, task: dict, secret_env: dict | None = None, guest_env: dict | None = None, user_data_extra: str | None = None, - network_config: str | None = None) -> dict: + network_config: str | None = None, + authd_env: dict | None = None) -> dict: """按 VM 预置 host key(补缺口轨强方案,见 hostkeys.py): 启动前为该 VM 生成一次性 ed25519 host key 对,公钥 立即写进该 VM 的 known_hosts(钉死),私钥只写进 @@ -887,17 +929,11 @@ class CloudHypervisorExecutor(Executor): # to10:known_hosts 写入已抽到 _write_provisioned_known_hosts, # 此处只留生成与 seed 写入的 import(删后不再有未用 import)。 from cloud.chv.hostkeys import ( - dirty_marker_path, generate_host_keypair, write_seed, ) vm = self.vm_name(task) - if os.path.exists(dirty_marker_path(self.state_dir, vm)): - raise ChvFailure( - "vm_dirty_refused", - "VM %s 上次清理失败(脏标记存在),拒绝复用旧 " - "host key 启动:请人工清理 state_dir 后删脏标记" - % vm) + self._refuse_dirty_vm(vm) paths = generate_host_keypair( self.state_dir, vm, runner=lambda cmd: self._call_runner(cmd)) @@ -908,13 +944,25 @@ class CloudHypervisorExecutor(Executor): secret_env=secret_env, guest_env=guest_env, user_data_extra=user_data_extra, - network_config=network_config) + network_config=network_config, + authd_env=authd_env) self._emit("host_key_generated", task, "vm=%s secret_names=%s" % (vm, ",".join(sorted(secret_env or {})))) return {"paths": paths, "seed_dir": seed_dir, "known_hosts": paths["known_hosts"]} + def _refuse_dirty_vm(self, vm: str) -> None: + """上次清理失败留下的脏标记存在时直接拒绝(清理失败 + 即拒起,不再复用旧 key 对)。""" + from cloud.chv.hostkeys import dirty_marker_path + if os.path.exists(dirty_marker_path(self.state_dir, vm)): + raise ChvFailure( + "vm_dirty_refused", + "VM %s 上次清理失败(脏标记存在),拒绝复用旧 " + "host key 启动:请人工清理 state_dir 后删脏标记" + % vm) + def capture_host_key(self, task: dict) -> bool: """首次捕获客机 host key(TOFU,弱回落,边界见 ssh_transport.py 与 hostkeys.py):ssh-keyscan 经本机 @@ -974,25 +1022,42 @@ class CloudHypervisorExecutor(Executor): # 构造兼容,wake/wait 路径仍严格解析。 ssh_host = "127.0.0.1" secrets, plain = split_secret_env(extra_env) - guest_env_file = None - if secrets: - from cloud.chv.hostkeys import GUEST_ENV_PATH - delivered = self._delivered_secrets(self.vm_name(task)) - missing: list = [name for name in secrets if name not in delivered] - if missing: - raise ChvFailure( - "secret_env_not_delivered", - "密钥环境变量 %s 未经 seed 投递,拒绝进 ssh " - "argv:provisioned 模式下 wake 时投递," - "keyscan 模式不支持密钥投递" - % ",".join(sorted(missing))) - guest_env_file = GUEST_ENV_PATH + guest_env_file = self._resolve_secret_env_file(task, secrets) return _build_ssh_cmd( self.ssh_port, self.ssh_key, self.ssh_user, self.egress_proxy, self.known_hosts_path(task), inner_cmd, ssh_host=ssh_host, extra_env=plain or None, guest_env_file=guest_env_file) + def _resolve_secret_env_file(self, task: dict, secrets) -> str | None: + """feat-chv-muse-p3:密钥分两路查投递,返回远程命令 + 该 source 的客内密钥文件(无 guest-env 路密钥即 None)。 + authd 路的(GITEA_TOKEN)只认“已投递、不报错”,绝不 + source 进 codex 进程环境;guest-env 路的才 source + guest-env 文件。两路任一路未投递都 fail-closed(同款 kind)。""" + if not secrets: + return None + from cloud.chv.gitea_sync import split_authd_env + from cloud.chv.hostkeys import GUEST_ENV_PATH + authd_secrets, guest_secrets = split_authd_env(secrets) + name = self.vm_name(task) + missing: list = [ + n for n in authd_secrets + if n not in self._delivered_authd_secrets(name)] + missing += [ + n for n in guest_secrets + if n not in self._delivered_secrets(name)] + if missing: + raise ChvFailure( + "secret_env_not_delivered", + "密钥环境变量 %s 未经 seed 投递,拒绝进 ssh " + "argv:provisioned 模式下 wake 时投递," + "keyscan 模式不支持密钥投递" + % ",".join(sorted(missing))) + if guest_secrets: + return GUEST_ENV_PATH + return None + # -- API 模式:VmConfig / API 调用 / 就绪轮询 ----------------------- def _vm_config_disks(self, task_disk: str | None, seed_disk: str | None) -> list: @@ -1440,6 +1505,107 @@ class CloudHypervisorExecutor(Executor): merged.update(sync) return merged + def _sentinel_gate_gitea(self, task: dict, + extra_env: dict | None) -> None: + """feat-chv-muse-p3:wake 时的 Gitea 出口 Sentinel 预审。 + + 同步已请求(extra_env 里有 GITEA_URL)时才评估;无同步 + 变量即跳过(本次启动不同步,无出口可审)。评估目标: + GET https:///api/v1/version。 + - allow → 审计 sentinel_egress_allow,通过; + - deny → 审计 + ChvFailure(sentinel_egress_denied), + 起 VMM 前 fail-closed; + - ask → 记入待审批队列 + 审计 sentinel_egress_ask, + 同样 fail-closed:wake 等不了审批 UI,ask 在这里就是 + deny,须运维把主机加进 CHV_SENTINEL_ALLOW_HOSTS 再起。 + 白名单来源:CHV_SENTINEL_ALLOW_HOSTS + 运维配置的 Gitea + 主机(CHV_GITEA_URL 的主机,运维可信);任务级覆盖改了 + GITEA_URL 主机时不自动放行——改到未知主机就是 ask。 + SSRF 检查对白名单主机同样生效(DNS 可被投毒,白名单主机 + 解析到私网 IP 仍 deny)。审计 detail 只含 VM 名/主机/ + verdict,不含任何凭证值。""" + sync_url = (extra_env or {}).get("GITEA_URL") + if not sync_url: + return + name = self.vm_name(task) + host, scheme, port = self._gate_gitea_target(task, name, sync_url) + self._gate_gitea_apply(task, name, host, scheme, port) + + def _gate_gitea_target(self, task: dict, name: str, + sync_url: str) -> tuple: + """解析 Gitea 出口预审目标(主机/协议/端口);主机解析 + 不出即审计 + fail-closed。""" + from urllib.parse import urlparse + parsed = urlparse(sync_url) + host = (parsed.hostname or "").lower() + scheme = (parsed.scheme or "").lower() + port = parsed.port or (443 if scheme == "https" else 80) + if not host: + self._emit("sentinel_egress_denied", task, + "vm=%s gitea_url_invalid" % name) + raise ChvFailure( + "sentinel_egress_denied", + "Gitea URL 主机解析不出,拒绝启动") + return host, scheme, port + + def _gate_gitea_policy(self, host: str): + """预审策略:白名单来源 CHV_SENTINEL_ALLOW_HOSTS + + 运维配置的 Gitea 主机(CHV_GITEA_URL 的主机,运维可信); + 任务级覆盖改了 GITEA_URL 主机时不自动放行——改到未知 + 主机就是 ask。""" + from urllib.parse import urlparse + from cloud.chv.sentinel import SentinelPolicy + cfg_sync = getattr(self, "gitea_sync_env", None) or {} + cfg_host = "" + cfg_url = cfg_sync.get("GITEA_URL") + if cfg_url: + cfg_host = (urlparse(cfg_url).hostname or "").lower() + extra_allow = {host} if host and host == cfg_host else set() + return SentinelPolicy.from_env( + self._env, extra_allow_hosts=extra_allow) + + def _gate_gitea_apply(self, task: dict, name: str, host: str, + scheme: str, port: int) -> None: + """对预审目标做 Sentinel 评估并按 verdict 处理: + allow → 审计通过;ask → 记入待审批队列 + 审计,同样 + fail-closed(wake 等不了审批 UI,ask 在这里就是 deny, + 须运维把主机加进 CHV_SENTINEL_ALLOW_HOSTS 再起); + deny → 审计 + ChvFailure,起 VMM 前 fail-closed。 + SSRF 检查对白名单主机同样生效(DNS 可被投毒,白名单 + 主机解析到私网 IP 仍 deny)。审计 detail 只含 VM 名/ + 主机/verdict,不含任何凭证值。""" + from cloud.chv.sentinel import EgressRequest + policy = self._gate_gitea_policy(host) + request = EgressRequest("gitea", "GET", scheme, host, port, + "/api/v1/version") + decision = policy.evaluate( + request, resolve=getattr(self, "sentinel_resolver", None)) + if decision.verdict == "allow": + capability = decision.capability + assert capability is not None # allow 恒带严格能力 + self._emit("sentinel_egress_allow", task, + "vm=%s host=%s capability=%s" + % (name, host, capability.id)) + return + if decision.verdict == "ask": + pending_id = self._sentinel_queue.record( + request, decision.reasons) + self._emit("sentinel_egress_ask", task, + "vm=%s host=%s pending=%s" + % (name, host, pending_id)) + raise ChvFailure( + "sentinel_egress_denied", + "Gitea 主机 %s 不在 Sentinel 白名单内(待审批 %s)," + "拒绝启动:请运维加进 CHV_SENTINEL_ALLOW_HOSTS" + % (host, pending_id)) + self._emit("sentinel_egress_denied", task, + "vm=%s host=%s reasons=%s" + % (name, host, ",".join(decision.reasons))) + raise ChvFailure( + "sentinel_egress_denied", + "Gitea 出口被 Sentinel 拒绝(%s),拒绝启动" + % ",".join(decision.reasons)) + def _wake_prechecks(self, task: dict, name: str, extra_env) -> tuple: """wake 前置:preflight/容量/配额/net/密钥模式,纯移动。 失败时发 vm_failed 审计并原样抛 ChvFailure。""" @@ -1471,6 +1637,9 @@ class CloudHypervisorExecutor(Executor): "provisioned 模式支持(当前 %s)" % (",".join(sorted(secrets)), getattr(self, "host_key_mode", "provisioned"))) + # feat-chv-muse-p3:Gitea 出口 Sentinel 预审(起 VMM + # 前 fail-closed,见 _sentinel_gate_gitea)。 + self._sentinel_gate_gitea(task, extra_env) except ChvFailure as exc: self._emit("vm_failed", task, "vm=%s wake 前置失败 kind=%s" % (name, exc.kind), @@ -1478,6 +1647,20 @@ class CloudHypervisorExecutor(Executor): raise return secrets, net_entries + def _split_seed_authd_env(self, secrets, + guest_env: dict | None) -> tuple: + """feat-chv-muse-p3:seed 密钥集先拆通道——GITEA_TOKEN + 走 authd(独立的 authd-seed 文件,只给客内 authd 读), + 其余密钥仍走 guest-env。guest_env(split_guest_env 拼 + 出来的)里也含 GITEA_TOKEN:一并拆出来(与 secrets 拆 + 出的合并,secrets 的值优先),否则 write_seed 会拒收。 + 返回 (authd_env, guest_secrets, guest_env)。""" + from cloud.chv.gitea_sync import split_authd_env + authd_env, guest_secrets = split_authd_env(secrets) + dup_authd, guest_env = split_authd_env(guest_env) + dup_authd.update(authd_env) + return dup_authd, guest_secrets, guest_env + def _provision_seed_disk(self, task: dict, name: str, secrets, guest_env: dict | None = None, seed_extras: tuple | None = None) -> str | None: @@ -1503,16 +1686,24 @@ class CloudHypervisorExecutor(Executor): seed_disk_path, ) extra, netcfg = seed_extras or (None, None) + authd_env, guest_secrets, guest_env = \ + self._split_seed_authd_env(secrets, guest_env) provisioned = self.provision_host_key( - task, secret_env=secrets or None, + task, secret_env=guest_secrets or None, guest_env=guest_env or None, - user_data_extra=extra, network_config=netcfg) + user_data_extra=extra, network_config=netcfg, + authd_env=authd_env or None) seed_disk = seed_disk_path(self.state_dir, name) build_seed_disk( provisioned["seed_dir"], seed_disk, runner=lambda cmd: self._call_runner(cmd)) delete_host_private_key(self.state_dir, name) - self._mark_guest_env_delivered(name, secrets) + # 两路投递记录分开记:guest-env 路的进 + # _guest_env_delivered(任务期可 source),authd 路的进 + # _authd_delivered(任务期只认“已投递、不报错”,绝不 + # source 进 codex 环境)。 + self._mark_guest_env_delivered(name, guest_secrets) + self._mark_authd_delivered(name, authd_env) return seed_disk def _boot_vm(self, task, name, task_disk, seed_disk, @@ -1869,6 +2060,7 @@ class CloudHypervisorExecutor(Executor): self._emit("vm_net_registry_delete_failed", task, "vm=%s error=%s" % (name, _net_exc)) self._clear_guest_env_delivered(name) + self._clear_authd_delivered(name) return failed def _write_dirty_marker(self, name, failed): diff --git a/cloud/chv/gitea_sync.py b/cloud/chv/gitea_sync.py index 7b4ca5f..5cda8c7 100644 --- a/cloud/chv/gitea_sync.py +++ b/cloud/chv/gitea_sync.py @@ -43,6 +43,26 @@ GITEA_SYNC_KEYS = frozenset({ # _CHV_SECRET_ENV_NAMES)。 GITEA_SYNC_SECRET_KEYS = frozenset({"GITEA_TOKEN"}) +# 其中 authd 专用密钥(feat-chv-muse-p3):只进客内 authd 的 +# 独立 seed 文件 /etc/chv/authd-seed(见 hostkeys.AUTHD_SEED_PATH), +# 不进 guest-env、不进 codex 进程环境。本期只有 GITEA_TOKEN。 +AUTHD_SECRET_KEYS = frozenset({"GITEA_TOKEN"}) + + +def split_authd_env(secrets: dict | None) -> tuple[dict, dict]: + """把密钥集拆成(走 authd 通道的子集,仍走 guest-env 的子集)。 + + authd 子集 = AUTHD_SECRET_KEYS 命中的键(本期只有 GITEA_TOKEN); + 其余密钥(如 SUB2API_API_KEY)仍走 guest-env(见差距表 §3.3, + 网关 key 的 privsep 改造是后续工作)。""" + authd, rest = {}, {} + for key, value in dict(secrets or {}).items(): + if key in AUTHD_SECRET_KEYS: + authd[key] = value + else: + rest[key] = value + return authd, rest + # 其中非密钥、但 boot 时就要用、也必须进 guest-env 的键 #(SSH env 前缀到不了 cloud-init 时运行的 chv-sync boot)。 GITEA_SYNC_PLAIN_KEYS = GITEA_SYNC_KEYS - GITEA_SYNC_SECRET_KEYS diff --git a/cloud/chv/hostkeys.py b/cloud/chv/hostkeys.py index 7ab8727..8973ca1 100644 --- a/cloud/chv/hostkeys.py +++ b/cloud/chv/hostkeys.py @@ -42,6 +42,17 @@ from cloud.chv.atomic_file import atomic_write_text # 除 seed 目录/user-data(随 VM 销毁而删)外不另存该文件。 GUEST_ENV_PATH = "/etc/chv/guest-env" +# seed 投递的客内 authd 专用密钥文件路径(feat-chv-muse-p3): +# 只写 AUTHD_SECRET_KEYS 命中的密钥(本期只有 GITEA_TOKEN), +# 0600 root,只给客内 authd 守护进程读;codex 进程环境里永不 +# 出现这些键(见执行器 _authd_delivered)。与 guest-env 分离 +# 的理由:guest-env 会被远程命令 source 进任务进程环境,任何 +# 进 guest-env 的密钥都会进 codex 环境;authd 密钥绝不能走 +# 那条路。文件本身仍是 seed 写出的落盘文件(cloud-init 的唯一 +# 投递通道就是写文件),真正让它对 agent 不可见的是 P2 的 +# nspawn 单元隔离,见差距表 §3.1。 +AUTHD_SEED_PATH = "/etc/chv/authd-seed" + def host_key_paths(state_dir: str, vm_name: str) -> dict: """该 VM 的 host key 文件路径(私钥/公钥/known_hosts)。""" @@ -180,6 +191,24 @@ def build_seed_user_data(private_key_text: str, public_key_text: str) -> str: return "\n".join(lines) + "\n" +def _authd_env_entry(authd_env: dict) -> list: + """authd-seed 的 write_files 条目行(0600 root,值经 + shlex.quote 引用):只装 AUTHD_SECRET_KEYS 命中的密钥, + 与 guest-env 条目并列写进同一 write_files 列表(绝不另写 + 一个 write_files: 键,见 _guest_env_entry 的 YAML 重复键 + 禁令)。""" + lines = [ + " - path: %s" % AUTHD_SEED_PATH, + " permissions: '0600'", + " owner: root:root", + " content: |", + ] + for key in sorted(authd_env): + lines.append(" %s=%s" + % (key, shlex.quote(str(authd_env[key])))) + return lines + + def _guest_env_entry(merged_env: dict) -> list: """guest-env 的 write_files 条目行(0600 root,值经 shlex.quote 引用):只能并进已有的 write_files 列表, @@ -227,12 +256,68 @@ def _apply_seed_extras(seed_dir: str, user_data: str, return user_data +def _validate_authd_env(authd_env: dict | None) -> None: + """authd_env 的键必须全在 AUTHD_SECRET_KEYS 内(本期只有 + GITEA_TOKEN),否则 ValueError——authd 通道只装 authd + 密钥,别的密钥误进这里会绕过 guest-env 的 source 语义。""" + if not authd_env: + return + from cloud.chv.gitea_sync import AUTHD_SECRET_KEYS + unknown = sorted(set(authd_env) - AUTHD_SECRET_KEYS) + if unknown: + raise ValueError( + "authd 通道只接受 %s,拒绝:%s" + % (",".join(sorted(AUTHD_SECRET_KEYS)), + ",".join(unknown))) + + +def _write_seed_meta(seed_dir: str, vm_name: str) -> str: + """建 seed 目录并写 meta-data(0600),返回 user-data 路径。 + feat-chv-cover100:seed 暂存文件走 atomic_file(mkstemp + 0600 → 写 → fsync → rename):user-data 含 host 私钥 + 明文,直接写 + 事后 chmod 会在 chmod 前以 umask 权限 + 短暂可读;半截文件到不了 ISO(见 H5),原子写同时堵 + 住权限窗口与半截残留。""" + os.makedirs(seed_dir, exist_ok=True) + meta_path = os.path.join(seed_dir, "meta-data") + user_path = os.path.join(seed_dir, "user-data") + atomic_write_text( + meta_path, + "instance-id: %s\nlocal-hostname: %s\n" % (vm_name, vm_name)) + return user_path + + +def _seed_env_entries(merged_env: dict, + authd_env: dict | None) -> list: + """guest-env 与 authd-seed 两路的 write_files 条目行 + (皆空即空列表,不写条目)。""" + extra_lines = [] + if merged_env: + extra_lines += _guest_env_entry(merged_env) + if authd_env: + extra_lines += _authd_env_entry(authd_env) + return extra_lines + + +def _insert_seed_entries(user_data: str, lines: list) -> str: + """条目插在 runcmd 段之前、仍属同一 write_files 列表 + (seed 自 2026-11-07 起不再带 runcmd,见 + build_seed_user_data 的根因禁令;marker 分支只作 + 防御,将来模板若恢复 runcmd 仍插在其前)。""" + marker = "runcmd:\n" + if marker in user_data: + head, tail = user_data.split(marker, 1) + return head + "\n".join(lines) + "\n" + marker + tail + return user_data + "\n".join(lines) + "\n" + + def write_seed(seed_dir: str, vm_name: str, private_key_text: str, public_key_text: str, secret_env: dict | None = None, guest_env: dict | None = None, user_data_extra: str | None = None, - network_config: str | None = None) -> str: + network_config: str | None = None, + authd_env: dict | None = None) -> str: """写 seed 目录:meta-data + user-data(均为 0600), 返回 seed 目录路径。调用方负责随 VM 销毁删掉它。 @@ -243,34 +328,28 @@ def write_seed(seed_dir: str, vm_name: str, private_key_text: str, 装 boot 时要用的非密钥同步变量(GITEA_URL 等,见 gitea_sync):只走 SSH 前缀到不了 cloud-init 时运行 的 chv-sync boot。两者皆空时不写该条目。 + authd_env(feat-chv-muse-p3)给定时写进另一个独立条目 + AUTHD_SEED_PATH(0600 root,只给客内 authd 守护进程读): + 键必须全在 gitea_sync.AUTHD_SECRET_KEYS 内(本期只有 + GITEA_TOKEN),否则 ValueError——authd 通道只装 authd + 密钥,别的密钥误进这里会绕过 guest-env 的 source 语义。 user_data_extra / network_config 见 _apply_seed_extras (客机用户与静态网络的供给,皆空即不写)。""" - os.makedirs(seed_dir, exist_ok=True) - meta_path = os.path.join(seed_dir, "meta-data") - user_path = os.path.join(seed_dir, "user-data") - # feat-chv-cover100:seed 暂存文件走 atomic_file(mkstemp - # 0600 → 写 → fsync → rename):user-data 含 host 私钥 - # 明文,直接写 + 事后 chmod 会在 chmod 前以 umask 权限 - # 短暂可读;半截文件到不了 ISO(见 H5),原子写同时堵 - # 住权限窗口与半截残留。 - atomic_write_text( - meta_path, - "instance-id: %s\nlocal-hostname: %s\n" % (vm_name, vm_name)) + _validate_authd_env(authd_env) + user_path = _write_seed_meta(seed_dir, vm_name) user_data = build_seed_user_data(private_key_text, public_key_text) merged_env = dict(guest_env or {}) merged_env.update(secret_env or {}) - if merged_env: - lines = _guest_env_entry(merged_env) - # 条目插在 runcmd 段之前、仍属同一 write_files 列表 - #(seed 自 2026-11-07 起不再带 runcmd,见 - # build_seed_user_data 的根因禁令;marker 分支只作 - # 防御,将来模板若恢复 runcmd 仍插在其前)。 - marker = "runcmd:\n" - if marker in user_data: - head, tail = user_data.split(marker, 1) - user_data = head + "\n".join(lines) + "\n" + marker + tail - else: - user_data = user_data + "\n".join(lines) + "\n" + # feat-chv-muse-p3:GITEA_TOKEN 绝不进 guest-env——它只能走 + # authd_env(独立的 authd-seed 条目)。调用方(执行器)须先 + # 用 gitea_sync.split_authd_env() 拆出来;这里再拦一次, + # 免得别的调用点把旧写法带进来、悄悄恢复泄露面。 + if "GITEA_TOKEN" in merged_env: + raise ValueError( + "GITEA_TOKEN 须走 authd_env,不进 guest-env") + extra_lines = _seed_env_entries(merged_env, authd_env) + if extra_lines: + user_data = _insert_seed_entries(user_data, extra_lines) user_data = _apply_seed_extras(seed_dir, user_data, user_data_extra, network_config) atomic_write_text(user_path, user_data) diff --git a/cloud/chv/sentinel.py b/cloud/chv/sentinel.py new file mode 100644 index 0000000..04ea3b2 --- /dev/null +++ b/cloud/chv/sentinel.py @@ -0,0 +1,342 @@ +"""Sentinel 式出口审批决策引擎(feat-chv-muse-p3,叶子模块)。 + +对齐 Muse 官方安全文(security.muse.ai 2026-09-08,见 +docs/chv-muse-p3-gap.md):Sentinel 是连接器动作与全部网络出口的 +唯一权限机构;L4 与 L7 评估(主机名、解析后与最终目的 IP、端口、 +协议、HTTP 方法、路径、解码后的实际请求);SSRF 防护(防止公开 +主机名在 DNS 解析后指向私有基础设施);审批是严格能力 +(strict capabilities),绑定到特定连接器/目的地与用例,支持 +一次性/会话级/任务级/时限级授权;审批 UI 不在对话内。 + +本模块只是决策层(纯函数 + 内存态队列),不是网络层强制: +- 它不拦截客内流量。客内直连的强制仍靠 Tinyproxy 白名单 + + 宿主防火墙(第三棒落地中);eBPF cgroup 拦截 + LSM hook 污点 + 传播是后续工作,见差距表 §3.4,不得把本模块写成“出口已全部 + 经 Sentinel”。 +- 本期两个接线点:① 执行器 wake 时的 Gitea 出口预审(起 VMM 前 + fail-closed);② 客内 authd 放凭证前评估。两者都调本模块的 + `evaluate()`,决策记录进审计(无值)。 +- `ask` 的语义:主机不在白名单、也不在拒绝名单时返回 ask,并记入 + `ApprovalQueue` 待审批;调用方遇到 ask 必须停止执行、不得自动 + 放行。审批 UI 不在对话内——本期只留队列,不实现任何对话内审批 + 入口(见差距表 §1)。 + +本模块只用标准库,不依赖 cloud.*,客内 authd 守护进程可直接 +import 它(见 chv-guest/authd.py)。解析器可注入(`resolve` +参数),默认 `socket.getaddrinfo`;测试用假解析器,绝不联网。 +审计记录里绝不含凭证值、也绝不含解析出的 IP 以外的内容—— +reason 只含网段名与策略名。 +""" + +import ipaddress +import socket +import time +import uuid +from collections.abc import Callable, Iterable + +# SSRF 挡截:解析到这些网段的 IP 一律 deny。公开主机名解析到 +# 私有地址即挡(Muse 安全文的 SSRF 防护口径);解析失败也 +# deny(fail-closed,不把“解析不出”当成“可以放行”)。 +_BLOCKED_NETWORKS = tuple( + ipaddress.ip_network(net) for net in ( + "0.0.0.0/8", # 本网络 / 未指定 + "10.0.0.0/8", # RFC1918 + "100.64.0.0/10", # CGNAT 共享地址空间 + "127.0.0.0/8", # 回环 + "169.254.0.0/16", # 链路本地 + "172.16.0.0/12", # RFC1918 + "192.0.0.0/24", # IETF 协议分配 + "192.0.2.0/24", # 文档 TEST-NET-1 + "192.88.99.0/24", # 6to4 中继(已废弃) + "192.168.0.0/16", # RFC1918 + # 注意:198.18.0.0/15(基准测试)故意不在表里——本部署的 + # egress/DNS 基础设施自用该网段,gitea.tose.top 在此环境 + # 真实解析到 198.18.139.152;挡掉它会把真 Gitea 主机也拒了。 + "198.51.100.0/24", # 文档 TEST-NET-2 + "203.0.113.0/24", # 文档 TEST-NET-3 + "224.0.0.0/4", # 组播 + "240.0.0.0/4", # 保留 + "255.255.255.255/32", + "::1/128", # 回环 + "::/128", # 未指定 + "fc00::/7", # 唯一本地 + "fe80::/10", # 链路本地 + "ff00::/8", # 组播 + )) + +# 每连接器的 L7 路径策略:(方法, 匹配方式, 模式)。 +# gitea:API 只放 /api/v1/;git 智能 HTTP 只放三条(clone/pull/push +# 的实际路径,见 git http-protocol)。其它路径(Web UI、登录页、 +# 任意管理接口)一律不在本连接器的能力内。 +_CONNECTOR_PATH_RULES = { + "gitea": ( + ("GET", "prefix", "/api/v1/"), + ("POST", "prefix", "/api/v1/"), + ("GET", "suffix", ".git/info/refs"), + ("POST", "suffix", ".git/git-upload-pack"), + ("POST", "suffix", ".git/git-receive-pack"), + ), +} + +# 每连接器允许的 scheme/端口(默认:只 https/443;http 只在 +# 显式放行时才出现,本期没有任何连接器放 http)。 +_CONNECTOR_SCHEMES = {"gitea": frozenset({"https"})} +_CONNECTOR_PORTS = {"gitea": frozenset({443})} + +# 默认放凭证有效期(秒):与 authd 按次释放同款 300 秒。 +DEFAULT_CAPABILITY_TTL_SECONDS = 300 + + +def _blocked_network_for(ip_text: str): + """该 IP 落在哪个挡截网段里(无即 None)。非法 IP 文本 + 按“挡不住也说不清”处理:返回 0.0.0.0/8(deny),绝不 + 把解析异常当成放行信号。""" + try: + addr = ipaddress.ip_address(ip_text) + except ValueError: + return _BLOCKED_NETWORKS[0] + for net in _BLOCKED_NETWORKS: + if addr.version == net.version and addr in net: + return net + return None + + +def _default_resolve(host: str, port: int) -> list: + """默认解析器:socket.getaddrinfo 去重后的 IP 列表。""" + infos = socket.getaddrinfo(host, port, proto=socket.IPPROTO_TCP) + out = [] + for info in infos: + ip = info[4][0] + if ip not in out: + out.append(ip) + return out + + +class EgressRequest: + """一次出口请求(决策输入):连接器 + 方法 + 目标 + 解析后 + IP(解析在 evaluate 里做,这里只记目标描述,不记任何凭证值)。""" + + def __init__(self, connector: str, method: str, scheme: str, + host: str, port: int, path: str) -> None: + self.connector = connector + self.method = (method or "").upper() + self.scheme = (scheme or "").lower() + self.host = (host or "").lower() + self.port = int(port) + self.path = path or "/" + + def describe(self) -> dict: + """审计用描述(只含连接器/目标/方法,不含值)。""" + return {"connector": self.connector, "method": self.method, + "scheme": self.scheme, "host": self.host, + "port": self.port, "path": self.path} + + +class Capability: + """严格能力(Muse 口径:绑定到特定连接器/目的地与用例)。 + + 一次 evaluate().capability 只对:一个连接器、一个主机、声明 + 的方法与路径模式、在有效期内、剩余次数内有效。越界使用 + (换主机、换方法、换路径、过期、用完)一律 `matches()` 为 + False,调用方必须重新走 evaluate,不得自行续期。""" + + def __init__(self, connector: str, host: str, + methods: Iterable[str], + path_rules: Iterable[tuple[str, str, str]], + ttl_seconds: int = DEFAULT_CAPABILITY_TTL_SECONDS, + max_uses: int = 1, now: float | None = None) -> None: + self.id = uuid.uuid4().hex[:16] + self.connector = connector + self.host = (host or "").lower() + self.methods = frozenset(methods) + self.path_rules = tuple(path_rules) + self.expires_at = (now if now is not None else time.time()) \ + + ttl_seconds + self.max_uses = max_uses + self.uses_left = max_uses + + def matches(self, request: EgressRequest, + now: float | None = None) -> bool: + """该请求是否在能力绑定内(不消耗次数,只判定)。""" + if request.connector != self.connector: + return False + if request.host != self.host: + return False + if request.method not in self.methods: + return False + if not _path_allowed(request, self.path_rules): + return False + at = now if now is not None else time.time() + if at >= self.expires_at: + return False + if self.uses_left <= 0: + return False + return True + + def use(self, request: EgressRequest, + now: float | None = None) -> bool: + """消耗一次能力:绑定内才扣次数并返回 True,否则 False + (不抛错,调用方按 deny 处理)。""" + if not self.matches(request, now=now): + return False + self.uses_left -= 1 + return True + + +class Decision: + """一次审批决策:allow / deny / ask + 原因 + 能力(仅 allow + 时有)+ 审计记录(无值)。""" + + def __init__(self, verdict: str, reasons: Iterable[str], + capability: Capability | None = None) -> None: + assert verdict in ("allow", "deny", "ask") + self.verdict = verdict + self.reasons = list(reasons) + self.capability = capability + + def audit(self, request: EgressRequest) -> dict: + """审计记录:事件名按 verdict,detail 只含目标描述与原因。""" + record = request.describe() + record["event"] = "sentinel_egress_" + self.verdict + record["verdict"] = self.verdict + record["reasons"] = list(self.reasons) + if self.capability is not None: + record["capability_id"] = self.capability.id + return record + + +def _path_allowed(request: EgressRequest, rules) -> bool: + """路径是否在该连接器的 L7 规则内。""" + for method, kind, pattern in rules: + if request.method != method: + continue + if kind == "prefix" and request.path.startswith(pattern): + return True + if kind == "suffix" and request.path.endswith(pattern): + return True + return False + + +class SentinelPolicy: + """Sentinel 策略:白名单主机 + 拒绝主机 + 连接器 L7 规则。 + + 默认拒绝(default-deny):主机不在白名单、也不在拒绝名单时 + 返回 ask(不是 allow);拒绝名单命中、SSRF 命中、scheme/ + 端口/路径越界时返回 deny。白名单只放行“已知且解析干净”的 + 目标——白名单不是“解析后是什么都行”,SSRF 检查对白名单 + 主机同样生效(DNS 可被投毒,白名单主机解析到私网 IP 仍 deny)。""" + + def __init__(self, allow_hosts: Iterable[str] = (), + deny_hosts: Iterable[str] = (), + connector_paths: dict[str, tuple] | None = None + ) -> None: + self.allow_hosts = frozenset( + h.lower() for h in (allow_hosts or ())) + self.deny_hosts = frozenset( + h.lower() for h in (deny_hosts or ())) + self.connector_paths = dict( + connector_paths if connector_paths is not None + else _CONNECTOR_PATH_RULES) + + @classmethod + def from_env(cls: type["SentinelPolicy"], env: dict | None, + extra_allow_hosts: Iterable[str] = () + ) -> "SentinelPolicy": + """从宿主环境变量建策略:CHV_SENTINEL_ALLOW_HOSTS(逗号 + 分隔主机名)+ 调用方追加的白名单(执行器 wake 预审时把 + 配置的 Gitea 主机追加进来——只放行配置的 Gitea 主机, + 不是放行任意主机)。""" + raw = (env or {}).get("CHV_SENTINEL_ALLOW_HOSTS") or "" + hosts = [h.strip().lower() for h in raw.split(",") if h.strip()] + hosts += [h.lower() for h in (extra_allow_hosts or ())] + return cls(allow_hosts=hosts) + + def evaluate(self, request: EgressRequest, + resolve: Callable[[str, int], list] | None = None, + now: float | None = None) -> Decision: + """评估一次出口请求(按顺序,首个命中即定 verdict): + + 1. 未知连接器 → deny(没有策略的连接器不放行); + 2. scheme 不在该连接器允许集 → deny; + 3. 端口不在该连接器允许集 → deny; + 4. 主机在拒绝名单 → deny; + 5. 解析失败 → deny(fail-closed); + 6. 任一解析 IP 落在挡截网段 → deny(SSRF); + 7. 主机不在白名单 → ask(记入待审批队列,调用方停止执行); + 8. 方法/路径不在连接器 L7 规则内 → deny; + 9. 全部通过 → allow + 严格能力(单次、300 秒)。 + """ + rules = self.connector_paths.get(request.connector) + if rules is None: + return Decision("deny", ["unknown_connector"]) + if request.scheme not in _CONNECTOR_SCHEMES.get( + request.connector, frozenset()): + return Decision("deny", ["scheme_not_allowed"]) + if request.port not in _CONNECTOR_PORTS.get( + request.connector, frozenset()): + return Decision("deny", ["port_not_allowed"]) + if request.host in self.deny_hosts: + return Decision("deny", ["host_denied"]) + resolver = resolve or _default_resolve + try: + ips = resolver(request.host, request.port) + except Exception: + return Decision("deny", ["dns_resolution_failed"]) + if not ips: + return Decision("deny", ["dns_resolution_failed"]) + for ip in ips: + net = _blocked_network_for(ip) + if net is not None: + return Decision( + "deny", ["ssrf_private_ip:%s" % net]) + if request.host not in self.allow_hosts: + return Decision("ask", ["host_not_allowlisted"]) + if not _path_allowed(request, rules): + return Decision("deny", ["path_not_allowed"]) + capability = Capability( + request.connector, request.host, + {request.method}, rules, + ttl_seconds=DEFAULT_CAPABILITY_TTL_SECONDS, + max_uses=1, now=now) + return Decision("allow", ["policy_allow"], + capability=capability) + + +class ApprovalQueue: + """ask 待审批队列(内存态):ask 出现即记入,执行停止。 + + 设计约束(Muse 口径):审批 UI 不在对话内。本期只留队列与 + 查询/标记接口,不实现任何对话内审批入口,也不自动放行。 + 队列是内存态,进程重启即清——ask 本来就意味着“这次执行 + 已停止”,重启后须重新申请。""" + + def __init__(self, clock: Callable[[], float] | None = None) -> None: + self._pending: dict[str, dict] = {} + self._clock = clock or time.time + + def record(self, request: EgressRequest, + reasons: Iterable[str]) -> str: + """记入一条待审批,返回 pending id(只含目标描述与原因)。""" + pending_id = uuid.uuid4().hex[:16] + record = request.describe() + record["pending_id"] = pending_id + record["reasons"] = list(reasons) + record["ts"] = self._clock() + self._pending[pending_id] = record + return pending_id + + def list_pending(self) -> list: + """全部待审批(副本,按记录时间排序)。""" + return [dict(v) for _, v in sorted( + self._pending.items(), key=lambda kv: kv[1]["ts"])] + + def resolve(self, pending_id: str, verdict: str) -> bool: + """标记一条待审批已处理(verdict 只记 allow/deny,不在 + 这里执行放行——放行须调用方重新走 evaluate)。未知 id + 返回 False,不抛错。""" + if pending_id not in self._pending: + return False + if verdict not in ("allow", "deny"): + return False + self._pending.pop(pending_id) + return True diff --git a/cloud/tests/test_chv_arch_maint_fe.py b/cloud/tests/test_chv_arch_maint_fe.py index 56a8a78..2e4b06c 100644 --- a/cloud/tests/test_chv_arch_maint_fe.py +++ b/cloud/tests/test_chv_arch_maint_fe.py @@ -76,6 +76,7 @@ EXPECTED_FIELDS = { "snapshot_max_total_gb", "snapshot_retention_days", "snapshot_dir", "allow_remote_snapshot", "snapshot_url", "task_timeout", "api_timeout", "ready_timeout", "ready_interval", + "sentinel_allow_hosts", } @@ -255,7 +256,7 @@ class ConfigFragmentsTest(unittest.TestCase): self.assertFalse(seen & set(frag), "片段键重叠") seen |= set(frag) self.assertEqual(seen, EXPECTED_FIELDS) - self.assertEqual(len(EXPECTED_FIELDS), 38) + self.assertEqual(len(EXPECTED_FIELDS), 39) def test_defaults_and_overrides(self): cfg = build_chv_config({}, {}) diff --git a/cloud/tests/test_chv_gaps.py b/cloud/tests/test_chv_gaps.py index d2c2103..a193eac 100644 --- a/cloud/tests/test_chv_gaps.py +++ b/cloud/tests/test_chv_gaps.py @@ -365,22 +365,34 @@ class HostKeyProvisionTest(ChvGapCase): def test_write_seed_with_secret_has_no_runcmd(self): # 带 secret_env 时 guest-env 条目仍须并进同一个 # write_files 列表(单 write_files 键),且 user-data - # 里仍不能出现 runcmd。 + # 里仍不能出现 runcmd。feat-chv-muse-p3:GITEA_TOKEN + # 不再走 secret_env,改走 authd_env(独立的 authd-seed + # 条目,0600);guest-env 条目里只剩非 authd 密钥。 with tempfile.TemporaryDirectory() as tmp: seed_dir = os.path.join(tmp, "seed") write_seed(seed_dir, "vm1", "PRIV", "ssh-ed25519 AAAAx", - secret_env={"GITEA_TOKEN": "s3cr3t v"}) + secret_env={"SUB2API_API_KEY": "gw-key"}, + authd_env={"GITEA_TOKEN": "s3cr3t v"}) with open(os.path.join(seed_dir, "user-data")) as fh: content = fh.read() self.assertEqual(content.count("write_files:"), 1) parsed = yaml.safe_load(content) self.assertNotIn("runcmd", parsed) - self.assertEqual(len(parsed["write_files"]), 3) + self.assertEqual(len(parsed["write_files"]), 4) paths = [e["path"] for e in parsed["write_files"]] self.assertIn("/etc/chv/guest-env", paths) - # 密钥值经 shlex.quote 引用后才进文件(值本身在 guest-env - # 条目内是设计内的,见 hostkeys.write_seed)。 - self.assertIn("GITEA_TOKEN='s3cr3t v'", content) + self.assertIn("/etc/chv/authd-seed", paths) + # 密钥值经 shlex.quote 引用后才进文件(值本身在条目内 + # 是设计内的,见 hostkeys.write_seed);GITEA_TOKEN + # 只在 authd-seed 里,不在 guest-env 里。 + entries = {e["path"]: e for e in parsed["write_files"]} + authd_text = entries["/etc/chv/authd-seed"]["content"] + self.assertIn("GITEA_TOKEN=", authd_text) + self.assertIn("s3cr3t v", authd_text) + self.assertEqual( + entries["/etc/chv/authd-seed"]["permissions"], "0600") + self.assertNotIn( + "GITEA_TOKEN", entries["/etc/chv/guest-env"]["content"]) def test_write_seed_uses_atomic_write(self): # feat-chv-cover100:seed 暂存文件(user-data 含 host diff --git a/cloud/tests/test_chv_gitea_sync.py b/cloud/tests/test_chv_gitea_sync.py index fa812a5..3478b8d 100644 --- a/cloud/tests/test_chv_gitea_sync.py +++ b/cloud/tests/test_chv_gitea_sync.py @@ -10,7 +10,8 @@ VM、不联网、不含任何真实 token(测试值全是假值)。 5 键进 guest-env 集; - hostkeys.write_seed:guest_env 与 secret_env 合并写同一 条目、仍无 runcmd、0600; -- 执行器 wake_vm:完整同步变量进 seed user-data;缺失必填 +- 执行器 wake_vm:完整同步变量进 seed(5 键 guest-env + + GITEA_TOKEN 走 authd-seed);缺失必填 键时起 VMM 前 ChvFailure 且不启动;无同步变量时正常启动 且不写 guest-env;seed 投递后 build_ssh_command 才能用 GITEA_TOKEN 且 argv 里绝无 token 值; @@ -117,6 +118,13 @@ def _guest_env_lines(user_text): assert len(guest) == 1, entries return guest[0]["content"], guest[0]["permissions"] +def _authd_seed_lines(user_text): + parsed = yaml.safe_load(user_text) + entries = parsed["write_files"] + authd = [e for e in entries if e["path"] == "/etc/chv/authd-seed"] + assert len(authd) == 1, entries + return authd[0]["content"], authd[0]["permissions"] + # -- (1) gitea_sync 纯函数 --------------------------------------------- class GiteaSyncPureTest(unittest.TestCase): @@ -226,27 +234,39 @@ class SshTransportSyncTest(unittest.TestCase): # -- (3) hostkeys.write_seed ------------------------------------------- class WriteSeedGuestEnvTest(unittest.TestCase): def test_guest_env_and_secret_merge_into_one_entry_no_runcmd(self): + # feat-chv-muse-p3:GITEA_TOKEN 不再进 guest-env,改走 + # authd_env(独立的 authd-seed 条目,0600);guest-env + # 条目里只剩非 authd 密钥与同步普通变量。guest_env 里 + # 直接给 GITEA_TOKEN 即 ValueError(调用方须先 + # split_authd_env 拆出来,旧的“同名键 secret 优先”语义 + # 已由“authd 通道只装 authd 密钥”取代)。 with tempfile.TemporaryDirectory() as tmp: seed_dir = os.path.join(tmp, "seed") write_seed(seed_dir, "vm1", "PRIV", "ssh-ed25519 AAAAx", secret_env={"SUB2API_API_KEY": "gw-key"}, guest_env={"GITEA_URL": "https://gitea.tose.top", "GITEA_USER": "admin", - "SYNC_REPO": "admin/chv-sync-test", - "GITEA_TOKEN": "dup-should-lose"}) + "SYNC_REPO": "admin/chv-sync-test"}, + authd_env={"GITEA_TOKEN": "dup-should-lose"}) with open(os.path.join(seed_dir, "user-data")) as fh: content = fh.read() self.assertEqual(content.count("write_files:"), 1) parsed = yaml.safe_load(content) self.assertNotIn("runcmd", parsed) - self.assertEqual(len(parsed["write_files"]), 3) + self.assertEqual(len(parsed["write_files"]), 4) text, perms = _guest_env_lines(content) self.assertEqual(perms, "0600") self.assertIn("GITEA_URL=https://gitea.tose.top", text) self.assertIn("SUB2API_API_KEY=gw-key", text) - # 同名键以 secret_env 为准,不出现两行 GITEA_TOKEN。 - self.assertEqual(text.count("GITEA_TOKEN="), 1) - self.assertIn("GITEA_TOKEN=dup-should-lose", text) + self.assertNotIn("GITEA_TOKEN", text) + authd_text, authd_perms = _authd_seed_lines(content) + self.assertEqual(authd_perms, "0600") + self.assertIn("GITEA_TOKEN=dup-should-lose", authd_text) + with tempfile.TemporaryDirectory() as tmp: + seed_dir = os.path.join(tmp, "seed") + with self.assertRaises(ValueError): + write_seed(seed_dir, "vm1", "PRIV", "ssh-ed25519 AAAAx", + guest_env={"GITEA_TOKEN": "dup-should-lose"}) def test_guest_env_only_still_no_runcmd(self): with tempfile.TemporaryDirectory() as tmp: @@ -307,7 +327,10 @@ class WakeGuestEnvCase(unittest.TestCase): def make_exec(self, env_extra=None): env = {"CHV_EGRESS_PROXY": "http://10.0.0.1:8888", "CHV_HOST_KEY_MODE": "provisioned", - "CHV_SSH_FORWARD": "1"} + "CHV_SSH_FORWARD": "1", + # feat-chv-muse-p3:wake 前 Sentinel 预审要求 + # Gitea 主机在白名单内;解析器用假的(测试不联网)。 + "CHV_SENTINEL_ALLOW_HOSTS": "gitea.tose.top"} if env_extra: env.update(env_extra) ex = CloudHypervisorExecutor( @@ -317,7 +340,8 @@ class WakeGuestEnvCase(unittest.TestCase): cpuinfo_path=self.cpuinfo, state_dir=self.state_dir, runner=self._runner, launcher=self._launcher, api_client=lambda s, m, p, payload=None: (200, {}), - sleeper=lambda s: None, env=env) + sleeper=lambda s: None, env=env, + sentinel_resolver=lambda host, port: ["93.184.216.34"]) ex.audit_callback = lambda e, t, d, k=None: self.audit.append( (e, t.get("id") if isinstance(t, dict) else t, d, k)) return ex @@ -331,8 +355,15 @@ class WakeGuestEnvCase(unittest.TestCase): self.assertNotIn("runcmd", yaml.safe_load(text)) guest_text, perms = _guest_env_lines(text) self.assertEqual(perms, "0600") - for key, value in FAKE_SYNC.items(): - self.assertIn("%s=%s" % (key, value), guest_text) + # feat-chv-muse-p3:六键仍全进 seed,但分两路—— + # 5 个普通键进 guest-env,GITEA_TOKEN 只进 authd-seed。 + for key in ("GITEA_URL", "GITEA_USER", "SYNC_REPO", + "SYNC_BRANCH", "GITEA_EMAIL"): + self.assertIn("%s=%s" % (key, FAKE_SYNC[key]), guest_text) + self.assertNotIn("GITEA_TOKEN", guest_text) + authd_text, authd_perms = _authd_seed_lines(text) + self.assertEqual(authd_perms, "0600") + self.assertIn("GITEA_TOKEN=fake-token-123", authd_text) # seed.iso 已制盘且 0600。 seed_iso = os.path.join(self.state_dir, "chv-sync1", "seed.iso") self.assertTrue(os.path.exists(seed_iso)) @@ -351,8 +382,14 @@ class WakeGuestEnvCase(unittest.TestCase): self.assertEqual(out["state"], "running") guest_text, _ = _guest_env_lines( _seed_user_text(self.state_dir, "chv-sync2")) - self.assertIn("GITEA_TOKEN=fake-token-123", guest_text) + # feat-chv-muse-p3:GITEA_TOKEN 不再进 guest-env, + # 改走 authd-seed;其余同步变量仍在 guest-env 里。 + self.assertNotIn("GITEA_TOKEN", guest_text) self.assertIn("SYNC_BRANCH=main", guest_text) + authd_text, authd_perms = _authd_seed_lines( + _seed_user_text(self.state_dir, "chv-sync2")) + self.assertEqual(authd_perms, "0600") + self.assertIn("GITEA_TOKEN=fake-token-123", authd_text) def test_wake_partial_sync_fails_closed_before_launch(self): ex = self.make_exec() @@ -387,11 +424,24 @@ class WakeGuestEnvCase(unittest.TestCase): def test_build_ssh_command_after_wake_sources_guest_env(self): ex = self.make_exec() ex.wake_vm({"id": "sync5", "gitea_sync": dict(FAKE_SYNC)}) + # feat-chv-muse-p3:GITEA_TOKEN 走 authd 投递——任务期 + # 认“已投递、不报错”,但绝不 source 进 codex 进程环境, + # 所以命令里既无 token 值、也不 source 任何客内密钥文件。 cmd = ex.build_ssh_command( {"id": "sync5"}, ["true"], extra_env={"GITEA_TOKEN": "fake-token-123"}) joined = " ".join(cmd) self.assertNotIn("fake-token-123", joined) + self.assertNotIn("/etc/chv/guest-env", joined) + self.assertNotIn("/etc/chv/authd-seed", joined) + # 非 authd 密钥仍走 guest-env:wake 时投递,任务期 source。 + ex.wake_vm({"id": "sync5b"}, + extra_env={"SUB2API_API_KEY": "gw-key"}) + cmd = ex.build_ssh_command( + {"id": "sync5b"}, ["true"], + extra_env={"SUB2API_API_KEY": "gw-key"}) + joined = " ".join(cmd) + self.assertNotIn("gw-key", joined) self.assertIn("/etc/chv/guest-env", joined) # 未 wake 投递的 VM 用密钥即拒绝。 with self.assertRaises(ChvFailure) as ctx: diff --git a/cloud/tests/test_chv_muse_p3.py b/cloud/tests/test_chv_muse_p3.py new file mode 100644 index 0000000..390743d --- /dev/null +++ b/cloud/tests/test_chv_muse_p3.py @@ -0,0 +1,675 @@ +"""对齐 Muse P3 测试:authd 替身凭证 + Sentinel 式出口审批。 + +只用标准库与 mock,绝不启动真实 VM、不联网、不含真实 token +(测试值全是假值)。覆盖: +- authd 保险库:白名单入库、作用域绑定、单次释放、替身过期、 + forget 吊销、repr/审计里绝无 token 值; +- 铸 token / 吊销(假 HTTP 客户端); +- Sentinel:SSRF 挡截、deny/ask 路径、L7 路径规则、严格能力 + (绑定/过期/次数)、待审批队列; +- seed 拆分:split_authd_env、write_seed 拒收 guest-env 里的 + GITEA_TOKEN、authd-seed 0600; +- wake 预审:allow/ask/deny、无同步跳过; +- 客内 authd 守护进程(真 socket + 真 seed):放凭证、拒绝 + 路径/主机/仓库、审计无值; +- git credential helper 端到端:真 git credential fill + 假 + authd socket(守护进程真跑在测试进程内)。 +""" + +import importlib.util +import json +import os +import socket +import subprocess +import sys +import tempfile +import threading +import unittest +from pathlib import Path + +try: + from cloud.chv.authd import ( + AuthdRefusal, + CredentialVault, + mint_gitea_token, + revoke_gitea_token, + surrogate_fingerprint, + ) + from cloud.chv.gitea_sync import split_authd_env + from cloud.chv.hostkeys import AUTHD_SEED_PATH, write_seed + from cloud.chv.sentinel import ( + ApprovalQueue, + EgressRequest, + SentinelPolicy, + ) + from cloud.sandbox import ChvFailure, CloudHypervisorExecutor +except ImportError: + sys.path.insert(0, str(Path(__file__).resolve().parents[2])) + from cloud.chv.authd import ( + AuthdRefusal, + CredentialVault, + mint_gitea_token, + revoke_gitea_token, + surrogate_fingerprint, + ) + from cloud.chv.gitea_sync import split_authd_env + from cloud.chv.hostkeys import AUTHD_SEED_PATH, write_seed + from cloud.chv.sentinel import ( + ApprovalQueue, + EgressRequest, + SentinelPolicy, + ) + from cloud.sandbox import ChvFailure, CloudHypervisorExecutor + +REPO_ROOT = Path(__file__).resolve().parents[2] +TOKEN = "fake-token-123" +PUBLIC_IP = "93.184.216.34" +FAKE_SYNC = { + "GITEA_URL": "https://gitea.tose.top", + "GITEA_USER": "admin", + "GITEA_TOKEN": TOKEN, + "SYNC_REPO": "admin/chv-sync-test", + "SYNC_BRANCH": "main", + "GITEA_EMAIL": "admin@tose.top", +} + + +def _public_resolver(host, port): + return [PUBLIC_IP] + + +def _load_guest_authd(): + """chv-guest/authd.py(守护进程)按文件加载,不走包名。""" + path = REPO_ROOT / "chv-guest" / "authd.py" + spec = importlib.util.spec_from_file_location( + "chv_guest_authd", path) + mod = importlib.util.module_from_spec(spec) + spec.loader.exec_module(mod) + return mod + + +# -- (1) authd 保险库 ---------------------------------------------------- +class VaultTest(unittest.TestCase): + def setUp(self): + self.now = [1000.0] + self.vault = CredentialVault(clock=lambda: self.now[0]) + + def _stored(self): + self.vault.store("gitea", "GITEA_TOKEN", TOKEN, + host="gitea.tose.top", + repo="admin/chv-sync-test") + + def test_store_rejects_unknown_connector_key_empty(self): + with self.assertRaises(AuthdRefusal) as ctx: + self.vault.store("github", "GITEA_TOKEN", TOKEN) + self.assertEqual(ctx.exception.kind, "unknown_connector") + with self.assertRaises(AuthdRefusal) as ctx: + self.vault.store("gitea", "OTHER_KEY", TOKEN) + self.assertEqual(ctx.exception.kind, + "credential_not_allowlisted") + with self.assertRaises(AuthdRefusal) as ctx: + self.vault.store("gitea", "GITEA_TOKEN", " ") + self.assertEqual(ctx.exception.kind, "empty_credential") + self.assertFalse(self.vault.has("gitea", "GITEA_TOKEN")) + + def test_release_single_use_and_scope_binding(self): + self._stored() + rel = self.vault.release("gitea", "GITEA_TOKEN", + "gitea.tose.top", + "admin/chv-sync-test", "push") + self.assertEqual(self.vault.consume(rel.release_id), TOKEN) + with self.assertRaises(AuthdRefusal) as ctx: + self.vault.consume(rel.release_id) + self.assertEqual(ctx.exception.kind, "release_unknown") + with self.assertRaises(AuthdRefusal) as ctx: + self.vault.release("gitea", "GITEA_TOKEN", + "other.example", + "admin/chv-sync-test", "push") + self.assertEqual(ctx.exception.kind, "scope_mismatch") + with self.assertRaises(AuthdRefusal) as ctx: + self.vault.release("gitea", "GITEA_TOKEN", + "gitea.tose.top", + "admin/other-repo", "push") + self.assertEqual(ctx.exception.kind, "scope_mismatch") + with self.assertRaises(AuthdRefusal) as ctx: + self.vault.release("gitea", "GITEA_TOKEN", + "gitea.tose.top", + "admin/chv-sync-test", "delete") + self.assertEqual(ctx.exception.kind, "operation_not_allowed") + + def test_release_expires_after_ttl(self): + # 替身过期:300 秒 TTL,301 秒后兑现即废。 + self._stored() + rel = self.vault.release("gitea", "GITEA_TOKEN", + "gitea.tose.top", + "admin/chv-sync-test", "pull") + self.now[0] += 301 + with self.assertRaises(AuthdRefusal) as ctx: + self.vault.consume(rel.release_id) + self.assertEqual(ctx.exception.kind, "release_expired") + + def test_release_unexpired_still_works(self): + self._stored() + rel = self.vault.release("gitea", "GITEA_TOKEN", + "gitea.tose.top", + "admin/chv-sync-test", "clone") + self.now[0] += 299 + self.assertEqual(self.vault.consume(rel.release_id), TOKEN) + + def test_forget_revokes_pending_releases(self): + self._stored() + rel = self.vault.release("gitea", "GITEA_TOKEN", + "gitea.tose.top", + "admin/chv-sync-test", "pull") + self.assertTrue(self.vault.forget("gitea", "GITEA_TOKEN")) + self.assertFalse(self.vault.has("gitea", "GITEA_TOKEN")) + self.assertFalse(self.vault.forget("gitea", "GITEA_TOKEN")) + with self.assertRaises(AuthdRefusal) as ctx: + self.vault.consume(rel.release_id) + self.assertEqual(ctx.exception.kind, "release_unknown") + with self.assertRaises(AuthdRefusal) as ctx: + self.vault.release("gitea", "GITEA_TOKEN", + "gitea.tose.top", + "admin/chv-sync-test", "pull") + self.assertEqual(ctx.exception.kind, "credential_not_stored") + + def test_repr_and_audit_never_contain_value(self): + # 真 token 泄露面:repr/审计记录里绝无 token 值。 + self._stored() + rel = self.vault.release("gitea", "GITEA_TOKEN", + "gitea.tose.top", + "admin/chv-sync-test", "push") + for text in (repr(self.vault), repr(rel), + json.dumps(self.vault.audit_release(rel)), + json.dumps(self.vault.audit_refusal( + "scope_mismatch", "gitea", "GITEA_TOKEN"))): + self.assertNotIn(TOKEN, text) + self.assertNotEqual(surrogate_fingerprint(TOKEN), TOKEN) + self.assertEqual(len(surrogate_fingerprint(TOKEN)), 8) + self.assertEqual(surrogate_fingerprint(TOKEN), + surrogate_fingerprint(TOKEN)) + + +class MintRevokeTest(unittest.TestCase): + def test_mint_returns_new_token_and_failure_is_refusal(self): + def fake_post(url, headers, payload): + self.assertIn("/users/admin/tokens", url) + self.assertEqual(payload["scopes"], ["write:repository"]) + return 201, {"sha1": "new-short-token"} + self.assertEqual( + mint_gitea_token("https://gitea.tose.top/api/v1", "admin", + TOKEN, "chv-test", post=fake_post), + "new-short-token") + with self.assertRaises(AuthdRefusal) as ctx: + mint_gitea_token("https://gitea.tose.top/api/v1", "admin", + TOKEN, "chv-test", + post=lambda u, h, p: (403, {})) + self.assertEqual(ctx.exception.kind, "mint_failed") + with self.assertRaises(AuthdRefusal): + mint_gitea_token("", "admin", TOKEN, "x") + + def test_revoke_statuses(self): + base = ("https://gitea.tose.top/api/v1", "admin", TOKEN) + self.assertTrue(revoke_gitea_token( + *base, 49, delete=lambda u, h: 204)) + self.assertTrue(revoke_gitea_token( + *base, 49, delete=lambda u, h: 404)) + self.assertFalse(revoke_gitea_token( + *base, 49, delete=lambda u, h: 500)) + self.assertFalse(revoke_gitea_token("", "admin", TOKEN, 49)) + + +# -- (2) Sentinel 决策 ---------------------------------------------------- +class SentinelTest(unittest.TestCase): + def _req(self, host="gitea.tose.top", method="GET", + scheme="https", port=443, path="/api/v1/version"): + return EgressRequest("gitea", method, scheme, host, port, path) + + def test_allow_public_ip_allowlisted_host(self): + policy = SentinelPolicy(allow_hosts={"gitea.tose.top"}) + dec = policy.evaluate(self._req(), resolve=_public_resolver) + self.assertEqual(dec.verdict, "allow") + self.assertIsNotNone(dec.capability) + self.assertNotIn(TOKEN, json.dumps(dec.audit(self._req()))) + + def test_ssrf_private_ip_denied_even_when_allowlisted(self): + # SSRF 挡截:白名单主机解析到私网 IP 仍 deny。 + policy = SentinelPolicy(allow_hosts={"gitea.tose.top"}) + for ip in ("10.0.0.5", "192.168.1.1", "127.0.0.1", + "169.254.1.2", "100.64.0.9", "::1", "0.0.0.0"): + dec = policy.evaluate( + self._req(), resolve=lambda h, p, ip=ip: [ip]) + self.assertEqual(dec.verdict, "deny", ip) + self.assertIn("ssrf_private_ip", dec.reasons[0]) + + def test_dns_failure_is_deny_not_ask(self): + policy = SentinelPolicy(allow_hosts={"gitea.tose.top"}) + dec = policy.evaluate( + self._req(), resolve=lambda h, p: (_ for _ in ()).throw( + OSError("boom"))) + self.assertEqual(dec.verdict, "deny") + self.assertEqual(dec.reasons, ["dns_resolution_failed"]) + dec = policy.evaluate(self._req(), resolve=lambda h, p: []) + self.assertEqual(dec.verdict, "deny") + + def test_unknown_host_is_ask_not_allow(self): + policy = SentinelPolicy(allow_hosts={"gitea.tose.top"}) + dec = policy.evaluate(self._req(host="other.example"), + resolve=_public_resolver) + self.assertEqual(dec.verdict, "ask") + self.assertEqual(dec.reasons, ["host_not_allowlisted"]) + self.assertIsNone(dec.capability) + + def test_deny_paths(self): + policy = SentinelPolicy(allow_hosts={"gitea.tose.top"}, + deny_hosts={"bad.example"}) + cases = [ + (EgressRequest("github", "GET", "https", "gitea.tose.top", + 443, "/api/v1/version"), + "unknown_connector"), + (self._req(scheme="http"), "scheme_not_allowed"), + (self._req(port=80), "port_not_allowed"), + (self._req(host="bad.example"), "host_denied"), + (self._req(path="/login"), "path_not_allowed"), + ] + for req, reason in cases: + dec = policy.evaluate(req, resolve=_public_resolver) + self.assertEqual(dec.verdict, "deny", reason) + self.assertEqual(dec.reasons, [reason]) + # git 智能 HTTP 三条路径放行。 + for method, path in ( + ("GET", "/admin/r.git/info/refs"), + ("POST", "/admin/r.git/git-upload-pack"), + ("POST", "/admin/r.git/git-receive-pack")): + dec = policy.evaluate( + self._req(method=method, path=path), + resolve=_public_resolver) + self.assertEqual(dec.verdict, "allow", path) + + def test_capability_binding_expiry_and_uses(self): + policy = SentinelPolicy(allow_hosts={"gitea.tose.top"}) + dec = policy.evaluate(self._req(), resolve=_public_resolver, + now=1000.0) + cap = dec.capability + req = self._req() + self.assertTrue(cap.matches(req, now=1000.0)) + self.assertFalse(cap.matches(self._req(host="other.example"), + now=1000.0)) + self.assertFalse(cap.matches(self._req(method="POST"), + now=1000.0)) + self.assertFalse(cap.matches(self._req(path="/login"), + now=1000.0)) + self.assertFalse(cap.matches(req, now=1000.0 + 301)) + self.assertTrue(cap.use(req, now=1000.0)) + self.assertFalse(cap.use(req, now=1000.0)) + + def test_from_env_and_approval_queue(self): + policy = SentinelPolicy.from_env( + {"CHV_SENTINEL_ALLOW_HOSTS": "a.example, b.example "}, + extra_allow_hosts={"c.example"}) + self.assertEqual(policy.allow_hosts, + {"a.example", "b.example", "c.example"}) + queue = ApprovalQueue(clock=lambda: 1000.0) + req = self._req(host="other.example") + pid = queue.record(req, ["host_not_allowlisted"]) + pending = queue.list_pending() + self.assertEqual(len(pending), 1) + self.assertEqual(pending[0]["pending_id"], pid) + self.assertEqual(pending[0]["ts"], 1000.0) + self.assertTrue(queue.resolve(pid, "deny")) + self.assertFalse(queue.resolve(pid, "deny")) + self.assertFalse(queue.resolve("nope", "allow")) + self.assertEqual(queue.list_pending(), []) + + +# -- (3) seed 拆分 --------------------------------------------------------- +class SeedSplitTest(unittest.TestCase): + def test_split_authd_env(self): + authd, rest = split_authd_env( + {"GITEA_TOKEN": TOKEN, "SUB2API_API_KEY": "gw", + "OTHER": "1"}) + self.assertEqual(authd, {"GITEA_TOKEN": TOKEN}) + self.assertEqual(rest, {"SUB2API_API_KEY": "gw", "OTHER": "1"}) + self.assertEqual(split_authd_env(None), ({}, {})) + self.assertEqual(split_authd_env({}), ({}, {})) + + def test_write_seed_refuses_token_in_guest_env(self): + with tempfile.TemporaryDirectory() as tmp: + seed_dir = os.path.join(tmp, "seed") + with self.assertRaises(ValueError): + write_seed(seed_dir, "vm1", "PRIV", + "ssh-ed25519 AAAAx", + secret_env={"GITEA_TOKEN": TOKEN}) + with self.assertRaises(ValueError): + write_seed(seed_dir, "vm1", "PRIV", + "ssh-ed25519 AAAAx", + guest_env={"GITEA_TOKEN": TOKEN}) + with self.assertRaises(ValueError): + write_seed(seed_dir, "vm1", "PRIV", + "ssh-ed25519 AAAAx", + authd_env={"SUB2API_API_KEY": "gw"}) + + def test_write_seed_authd_entry_is_separate_0600(self): + import yaml + with tempfile.TemporaryDirectory() as tmp: + seed_dir = os.path.join(tmp, "seed") + write_seed(seed_dir, "vm1", "PRIV", "ssh-ed25519 AAAAx", + secret_env={"SUB2API_API_KEY": "gw-key"}, + guest_env={"GITEA_URL": "https://gitea.tose.top"}, + authd_env={"GITEA_TOKEN": TOKEN}) + with open(os.path.join(seed_dir, "user-data")) as fh: + parsed = yaml.safe_load(fh.read()) + entries = {e["path"]: e for e in parsed["write_files"]} + self.assertEqual(entries[AUTHD_SEED_PATH]["permissions"], + "0600") + self.assertIn("GITEA_TOKEN=" + TOKEN, + entries[AUTHD_SEED_PATH]["content"]) + self.assertNotIn("GITEA_TOKEN", + entries["/etc/chv/guest-env"]["content"]) + + +# -- (4) wake 预审 ----------------------------------------------------------- +class _FakeResult: + def __init__(self, returncode=0, stdout="", stderr=""): + self.returncode = returncode + self.stdout = stdout + self.stderr = stderr + + +class _FakeProc: + def __init__(self, alive=True): + self._alive = alive + + def poll(self): + return None if self._alive else 1 + + def terminate(self): + self._alive = False + + def wait(self, timeout=None): + return 0 + + def kill(self): + self._alive = False + + +def _write_file(path, content=b"x"): + with open(path, "wb") as fh: + fh.write(content) + os.chmod(path, 0o755) + return path + + +class WakeGateTest(unittest.TestCase): + def setUp(self): + self.tmp = tempfile.TemporaryDirectory() + self.dir = self.tmp.name + self.binary = _write_file( + os.path.join(self.dir, "cloud-hypervisor")) + self.ch_remote = _write_file(os.path.join(self.dir, "ch-remote")) + self.kernel = _write_file(os.path.join(self.dir, "vmlinux")) + self.rootfs = _write_file(os.path.join(self.dir, "disk.raw"), + b"R" * 100) + self.ssh_key = _write_file(os.path.join(self.dir, "id_chv")) + self.kvm = _write_file(os.path.join(self.dir, "kvm")) + self.cpuinfo = os.path.join(self.dir, "cpuinfo") + with open(self.cpuinfo, "w") as fh: + fh.write("flags : fpu vmx sse\n") + self.state_dir = os.path.join(self.dir, "state") + self.audit = [] + self.launches = [] + + def tearDown(self): + self.tmp.cleanup() + + def _runner(self, cmd, timeout=None): + if cmd and cmd[0] == "ssh-keygen": + priv = cmd[cmd.index("-f") + 1] + with open(priv, "w") as fh: + fh.write("-----BEGIN OPENSSH PRIVATE KEY-----\nFAKE\n") + os.chmod(priv, 0o600) + with open(priv + ".pub", "w") as fh: + fh.write("ssh-ed25519 AAAAfake1 vm-key\n") + return _FakeResult(0, "", "") + if cmd and cmd[0] == "genisoimage": + out = cmd[cmd.index("-output") + 1] + with open(out, "wb") as fh: + fh.write(b"ISO") + return _FakeResult(0, "", "") + return _FakeResult(0, "", "") + + def _make_exec(self, env_extra=None, resolver=None): + env = {"CHV_EGRESS_PROXY": "http://10.0.0.1:8888", + "CHV_HOST_KEY_MODE": "provisioned", + "CHV_SSH_FORWARD": "1"} + if env_extra: + env.update(env_extra) + ex = CloudHypervisorExecutor( + binary=self.binary, ch_remote=self.ch_remote, + kernel=self.kernel, rootfs=self.rootfs, + ssh_key=self.ssh_key, kvm_path=self.kvm, + cpuinfo_path=self.cpuinfo, state_dir=self.state_dir, + runner=self._runner, + launcher=lambda cmd: self.launches.append(list(cmd)) + or _FakeProc(alive=True), + api_client=lambda s, m, p, payload=None: (200, {}), + sleeper=lambda s: None, env=env, + sentinel_resolver=resolver) + ex.audit_callback = lambda e, t, d, k=None: self.audit.append( + (e, t.get("id") if isinstance(t, dict) else t, d, k)) + return ex + + def test_wake_allowlisted_gitea_passes_gate(self): + ex = self._make_exec( + env_extra={"CHV_SENTINEL_ALLOW_HOSTS": "gitea.tose.top"}, + resolver=_public_resolver) + out = ex.wake_vm({"id": "gate1", "gitea_sync": dict(FAKE_SYNC)}) + self.assertEqual(out["state"], "running") + kinds = [a[0] for a in self.audit] + self.assertIn("sentinel_egress_allow", kinds) + + def test_wake_unknown_gitea_host_is_ask_and_fail_closed(self): + ex = self._make_exec(resolver=_public_resolver) + sync = dict(FAKE_SYNC) + sync["GITEA_URL"] = "https://other.example" + with self.assertRaises(ChvFailure) as ctx: + ex.wake_vm({"id": "gate2", "gitea_sync": sync}) + self.assertEqual(ctx.exception.kind, "sentinel_egress_denied") + self.assertEqual(self.launches, []) + kinds = [a[0] for a in self.audit] + self.assertIn("sentinel_egress_ask", kinds) + pending = ex._sentinel_queue.list_pending() + self.assertEqual(len(pending), 1) + self.assertEqual(pending[0]["host"], "other.example") + + def test_wake_gitea_resolving_to_private_ip_is_deny(self): + ex = self._make_exec( + env_extra={"CHV_SENTINEL_ALLOW_HOSTS": "gitea.tose.top"}, + resolver=lambda h, p: ["10.0.0.5"]) + with self.assertRaises(ChvFailure) as ctx: + ex.wake_vm({"id": "gate3", "gitea_sync": dict(FAKE_SYNC)}) + self.assertEqual(ctx.exception.kind, "sentinel_egress_denied") + kinds = [a[0] for a in self.audit] + self.assertIn("sentinel_egress_denied", kinds) + self.assertEqual(ex._sentinel_queue.list_pending(), []) + + def test_wake_without_sync_skips_gate(self): + ex = self._make_exec(resolver=_public_resolver) + out = ex.wake_vm({"id": "gate4"}) + self.assertEqual(out["state"], "running") + kinds = [a[0] for a in self.audit] + self.assertNotIn("sentinel_egress_allow", kinds) + self.assertNotIn("sentinel_egress_denied", kinds) + self.assertNotIn("sentinel_egress_ask", kinds) + + +# -- (5) 客内 authd 守护进程 + helper 端到端 ------------------------------- +class GuestAuthdTest(unittest.TestCase): + def setUp(self): + self.tmp = tempfile.TemporaryDirectory() + self.dir = self.tmp.name + self.seed_path = os.path.join(self.dir, "authd-seed") + with open(self.seed_path, "w") as fh: + fh.write("GITEA_URL=https://gitea.tose.top\n" + "GITEA_USER=admin\n" + "GITEA_TOKEN=" + TOKEN + "\n" + "SYNC_REPO=admin/chv-sync-test\n") + self.socket_path = os.path.join(self.dir, "authd.sock") + self.log_path = os.path.join(self.dir, "authd.log") + mod = _load_guest_authd() + self.server = mod.build_server( + seed_path=self.seed_path, socket_path=self.socket_path, + log_path=self.log_path, lib_dir=str(REPO_ROOT), + allowed_uids=[os.getuid()], resolve=_public_resolver) + self.thread = threading.Thread( + target=self.server.serve_forever, daemon=True) + self.thread.start() + + def tearDown(self): + self.server.shutdown() + self.server.server_close() + self.tmp.cleanup() + + def _ask(self, req): + sock = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM) + sock.settimeout(5) + try: + sock.connect(self.socket_path) + sock.sendall((json.dumps(req) + "\n").encode("utf-8")) + buf = b"" + while b"\n" not in buf: + chunk = sock.recv(4096) + if not chunk: + break + buf += chunk + return json.loads(buf.decode("utf-8")) + finally: + sock.close() + + def test_release_for_clone_path(self): + resp = self._ask({"connector": "gitea", + "host": "gitea.tose.top", + "path": "/admin/chv-sync-test.git/info/refs"}) + self.assertTrue(resp["ok"]) + self.assertEqual(resp["username"], "admin") + self.assertEqual(resp["password"], TOKEN) + self.assertGreater(resp["expires_at"], 0) + + def test_second_request_gets_fresh_release(self): + for _ in range(2): + resp = self._ask({"connector": "gitea", + "host": "gitea.tose.top", + "path": "/admin/chv-sync-test.git/" + "git-upload-pack"}) + self.assertTrue(resp["ok"]) + self.assertEqual(resp["password"], TOKEN) + + def test_refused_paths_hosts_and_repos(self): + cases = [ + ({"connector": "github", "host": "gitea.tose.top", + "path": "/admin/chv-sync-test.git/info/refs"}, + "bad_request_fields"), + ({"connector": "gitea", "host": "gitea.tose.top", + "path": "/admin/chv-sync-test.git/foo"}, + "path_not_allowed"), + ({"connector": "gitea", "host": "other.example", + "path": "/admin/chv-sync-test.git/info/refs"}, + "sentinel_ask"), + ({"connector": "gitea", "host": "gitea.tose.top", + "path": "/foo.git/info/refs"}, "repo_unparseable"), + ] + for req, error in cases: + resp = self._ask(req) + self.assertFalse(resp["ok"], req) + self.assertEqual(resp["error"], error, req) + + def test_audit_log_never_contains_value(self): + self._ask({"connector": "gitea", "host": "gitea.tose.top", + "path": "/admin/chv-sync-test.git/info/refs"}) + self._ask({"connector": "gitea", "host": "gitea.tose.top", + "path": "/admin/chv-sync-test.git/foo"}) + with open(self.log_path) as fh: + self.assertNotIn(TOKEN, fh.read()) + + def test_incomplete_or_missing_seed_fails_fast(self): + mod = _load_guest_authd() + bad_seed = os.path.join(self.dir, "bad-seed") + with open(bad_seed, "w") as fh: + fh.write("GITEA_URL=https://gitea.tose.top\n") + with self.assertRaises(ValueError): + mod.build_server(seed_path=bad_seed, + socket_path=os.path.join(self.dir, "s2"), + log_path=self.log_path, + lib_dir=str(REPO_ROOT), + allowed_uids=[os.getuid()], + resolve=_public_resolver) + with self.assertRaises(FileNotFoundError): + mod.build_server(seed_path=os.path.join(self.dir, "nope"), + socket_path=os.path.join(self.dir, "s3"), + log_path=self.log_path, + lib_dir=str(REPO_ROOT), + allowed_uids=[os.getuid()], + resolve=_public_resolver) + + +class HelperE2ETest(unittest.TestCase): + """helper + 真 git credential fill(假 authd socket 在后台)。""" + + HELPER = REPO_ROOT / "chv-guest" / "git-credential-chv-authd" + + def setUp(self): + self.tmp = tempfile.TemporaryDirectory() + self.dir = self.tmp.name + self.guest = GuestAuthdTest("test_release_for_clone_path") + self.guest.setUp() + self.socket_path = self.guest.socket_path + + def tearDown(self): + self.guest.tearDown() + self.tmp.cleanup() + + def _helper_env(self): + env = dict(os.environ) + env["CHV_AUTHD_SOCKET"] = self.socket_path + return env + + def test_helper_returns_credential(self): + proc = subprocess.run( + [sys.executable, str(self.HELPER)], + input="protocol=https\nhost=gitea.tose.top\n" + "path=/admin/chv-sync-test.git/info/refs" + "?service=git-upload-pack\n\n", + capture_output=True, text=True, env=self._helper_env(), + check=False) + self.assertEqual(proc.returncode, 0, proc.stderr) + self.assertIn("username=admin", proc.stdout) + self.assertIn("password=" + TOKEN, proc.stdout) + self.assertIn("password_expiry_utc=", proc.stdout) + + def test_helper_refusal_is_exit_1(self): + proc = subprocess.run( + [sys.executable, str(self.HELPER)], + input="protocol=https\nhost=gitea.tose.top\n" + "path=/admin/chv-sync-test.git/foo\n\n", + capture_output=True, text=True, env=self._helper_env(), + check=False) + self.assertEqual(proc.returncode, 1) + self.assertIn("path_not_allowed", proc.stderr) + + def test_git_credential_fill_end_to_end(self): + env = self._helper_env() + env["GIT_CONFIG_NOSYSTEM"] = "1" + proc = subprocess.run( + ["git", "-c", "credential.helper=%s" % self.HELPER, + "-c", "credential.useHttpPath=true", + "credential", "fill"], + input="protocol=https\nhost=gitea.tose.top\n" + "path=/admin/chv-sync-test.git/info/refs\n\n", + capture_output=True, text=True, env=env, cwd=self.dir, + check=False) + self.assertEqual(proc.returncode, 0, proc.stderr) + self.assertIn("password=" + TOKEN, proc.stdout) + + +if __name__ == "__main__": + unittest.main() diff --git a/docs/chv-muse-p3-gap.md b/docs/chv-muse-p3-gap.md new file mode 100644 index 0000000..63f8477 --- /dev/null +++ b/docs/chv-muse-p3-gap.md @@ -0,0 +1,118 @@ +# 对齐 Muse P3 差距表:authd 替身凭证 + Sentinel 式出口审批 + +依据:`~/workspace/research_notes/meta-muse-ai-architecture-20261101-0755/report.md` +(官方以 Meta security.muse.ai 2026-09-08 为准;第三方拆解只作参考, +逆向细节不得写成官方架构)。Muse 模型一句话:凭证由 VM 内独立服务 +authd 保管、代理只见替身令牌;所有连接器动作与网络出口由 Sentinel +唯一审批(allow/deny/ask,L4+L7 评估,SSRF 防护,严格能力绑定); +审批 UI 不在对话内;连接器执行走 privsep worker,按连接器做凭证白名单。 + +本轨范围只限 P3:authd 式替身凭证 + Sentinel 式出口审批。 +不碰 P1(常驻 VM、WarmPool、气球、vsock)、不碰 P2(客机 nspawn 单元、 +浏览器子代理独立)。分支 `feat-chv-muse-p3`(从 `feat-chv-cover100` +分出,基线 HEAD `06c3e17`)。 + +## 1. 现状 vs Muse + +| 项 | 现状(实测/代码为准) | Muse 目标 | 本轨做法 | +|---|---|---|---| +| Gitea token 可见面 | 真 token 经 seed 写进 `/etc/chv/guest-env`(0600 root);`build_ssh_command` 在每个任务的远程命令前 `source` 该文件,codex 进程环境里直接含真值;`chv-sync boot` 再把它写进 `/home/chv/.git-credentials` 落盘 | authd 保管凭证,代理只见替身令牌;凭证存入运行时单元之外的安全凭证存储,代理不可见;仅在需要时注入 | 真 token 改走 seed 里的独立文件 `/etc/chv/authd-seed`(0600 root),只给客内 authd 守护进程读;guest-env 不再含 GITEA_TOKEN;任务期 `build_ssh_command` 不再把 GITEA_TOKEN source 进 codex 环境;`chv-sync` 经 git credential helper 按次向 authd 取凭证,只进 git 子进程内存,不落盘、不进 agent 环境/argv/日志 | +| authd 位置 | 无 | 存于用户 VM 内(非 Meta 集中基础设施),在运行时单元之外 | 客内 authd 守护进程(`chv-guest/authd.py`,标准库 Unix socket 服务),以 root 身份跑在单元外;宿主侧不存任何 VM 的 token。选客内不选宿主的理由:Muse 的 authd 就在用户 VM 里;宿主侧方案要走 vsock(P1 范围、未就绪),且会把全部 VM 的 token 集中到宿主,违背“存于用户 VM”的口径 | +| 替身令牌 | 无(真值直给) | authd 签发替身令牌,带作用域与有效期 | 两档:① 本期默认——按次释放:每次 git 操作向 authd 取一次凭证,绑定 {连接器=gitea, 主机, 仓库路径, 操作},单次有效 + 300 秒 TTL,过期即废,释放全程审计(审计里无值);② 真铸 token:authd 可经 Gitea API 铸短期专用 token(`POST /api/v1/users/{user}/tokens`),默认关闭,见诚实边界第 2 条 | +| 凭证白名单(privsep) | 无连接器概念;`_CHV_SECRET_ENV_NAMES` 只有“密钥/非密钥”二分 | 每个 privsep worker 有明确凭证白名单(日历 worker 不能仅改参数就索取邮件凭证) | authd 只认一个连接器 `gitea`,白名单只含 GITEA_TOKEN;未知连接器/越权主机/仓库/操作一律拒绝;SUB2API_API_KEY 仍走 guest-env(见诚实边界第 3 条) | +| 出口审批 | 只有代码门:preflight 要求代理或显式直连 + `validate_egress_proxy` 格式校验;真正的隔离靠 Tinyproxy 白名单 + 宿主防火墙(第三棒落地中);客内 root 绕过代理直连在代码层无第二道门 | Sentinel 是连接器动作与全部网络出口的唯一权限机构;L4+L7 评估(主机名、解析后 IP、端口、协议、HTTP 方法、路径、解码后请求);SSRF 防护;审批是严格能力(绑定连接器/目的地/用例,一次性/会话级/任务级/时限级) | `cloud/chv/sentinel.py` 决策引擎:每动作审批钩子 allow/deny/ask;L4(解析后 IP、端口、协议)+ L7(主机、方法、路径)评估;SSRF 挡截(公开主机名解析到私有/回环/链路本地/CGNAT/保留地址即 deny);能力绑定 {连接器, 主机, 方法, 路径, 有效期, 次数};ask 记入待审批队列、执行停止。本期两个接线点:① wake 时 Gitea 出口预审(解析失败/私网 IP/不在白名单 → 起 VMM 前 fail-closed);② authd 放凭证前先过 Sentinel 评估 | +| 审批 UI | 无 | Sentinel 直接向客户端发审批请求,弹窗在客户端 UI 内,答复直接回 Sentinel;目的:防对话内容伪造审批 | 本期只做设计约束 + 待审批队列(`sentinel.ApprovalQueue`,内存态 pending 记录):ask 出现即停止执行、记入队列,不自动放行;UI 不在对话内这一条记为约束,不在本期实现任何对话内审批入口 | +| 污点出口(eBPF) | 无 | eBPF cgroup 程序做网络拦截与进程归因、LSM hook 做污点传播 | 列为后续,不冒充已做。见诚实边界第 4 条 | +| 审计 | `_emit` 审计只含 VM 名/路径/URL/kind,不含 key 值(M5) | 审计轨迹展示已做与计划做的事 | 沿用 `_emit` 契约:新增 `sentinel_egress_allow/deny/ask`、`authd_credential_released/refused` 事件,detail 只含连接器/主机/路径/ verdict/原因,不含任何凭证值;authd 客内审计文件 `/var/log/chv-authd.log`(0600)同款口径 | + +## 2. 改了什么(本轨交付物) + +- `cloud/chv/authd.py`(新,叶子模块,标准库):凭证保险库(只活内存, + `__repr__` 不含值)+ 按次释放(作用域绑定、单次有效、300 秒 TTL、 + 过期即废)+ 连接器白名单(本期只 `gitea`)+ Gitea API 铸 token + 辅助函数(可注入 HTTP 客户端,默认关闭)+ 无值审计记录。 +- `cloud/chv/sentinel.py`(新,叶子模块,标准库):决策引擎 + `evaluate()`(allow/deny/ask)+ L4/L7 评估 + SSRF 挡截 + (解析器可注入,默认 `socket.getaddrinfo`)+ 严格能力 + `Capability`(连接器/主机/方法/路径/有效期/次数绑定)+ + `ApprovalQueue`(ask 待审批队列,内存态)+ 无值审计记录。 +- `cloud/chv/hostkeys.py`:`write_seed` 加 `authd_env` 参数——authd + 专用密钥写进独立的客内文件 `/etc/chv/authd-seed`(0600 root), + 不再与 guest-env 混写;`AUTHD_SEED_PATH` 常量。 +- `cloud/chv/gitea_sync.py`:`AUTHD_SECRET_KEYS = {"GITEA_TOKEN"}` + + `split_authd_env()`(把 GITEA_TOKEN 从密钥集里拆出来走 authd 通道)。 +- `cloud/chv/ssh_transport.py`:`_CHV_AUTHD_SECRET_NAMES`;`remote_command` + 语义不变(仍只 source guest-env 文件)。 +- `cloud/chv/executor.py`:`provision_host_key`/`_provision_seed_disk` + 透传 `authd_env`;新增 `_authd_delivered` 投递集(与 + `_guest_env_delivered` 同锁)——authd 投递的密钥名在 + `build_ssh_command` 里不报错、但不 source 进 codex 环境; + `_wake_prechecks` 加 Gitea 出口 Sentinel 预审(deny/ask 即 + `sentinel_egress_denied`,起 VMM 前 fail-closed)。 +- `cloud/chv/config.py`:`CHV_SENTINEL_ALLOW_HOSTS`(逗号分隔主机名白名单, + 默认只放行配置的 Gitea 主机)。 +- `chv-guest/authd.py`(新,标准库守护进程):读 `/etc/chv/authd-seed` + 持 token 于内存、Unix socket `/run/chv-authd/authd.sock`(0700 目录、 + 0600 socket)、SO_PEERCRED 只放行 root 与 chv 用户、每次放凭证前先过 + Sentinel 评估、审计落盘 0600。 +- `chv-guest/git-credential-chv-authd`(新,标准库 helper):git + credential 协议(stdin 读 `protocol/host/path`,stdout 回用户名/口令), + 口令只进 git 子进程内存。 +- `chv-guest/chv-authd.service`(新):systemd 单元(root 运行,单元外)。 +- `chv-guest/chv-sync.sh`:`boot` 在 authd socket 存在时改配 + `credential.helper` 走 authd,不再把真 token 写进 + `.git-credentials`;同时配 `credential.useHttpPath=true` + (git 默认不把 path 发给 helper,没有 path helper 无法推断 + 仓库与操作、每次都拒绝——新测试抓到的真问题);socket 不存在 + 时走旧路径 + stderr 警告(过渡兼容,见诚实边界第 5 条)。 +- `cloud/tests/test_chv_muse_p3.py`(新):替身过期、真 token 泄露面、 + SSRF 挡截、deny/ask 路径、seed 拆分、wake 预审、helper 端到端 + (真 git `credential fill` + 假 authd socket);全量 cloud 套件全过。 +- 镜像集成(只留接入说明,不改 `chv-guest/build.sh`,与 P1 同款规矩, + 合并时由 P2 一并带进镜像): + `--copy-in chv-guest/authd.py:/usr/local/bin/`、 + `--copy-in chv-guest/git-credential-chv-authd:/usr/local/bin/`、 + `--copy-in chv-guest/chv-authd.service:/etc/systemd/system/`、 + `--copy-in cloud/chv/authd.py:/usr/local/lib/chv-authd/cloud/chv/`、 + `--copy-in cloud/chv/sentinel.py:/usr/local/lib/chv-authd/cloud/chv/`、 + `--copy-in cloud/__init__.py:/usr/local/lib/chv-authd/cloud/`、 + `--copy-in cloud/chv/__init__.py:/usr/local/lib/chv-authd/cloud/chv/`、 + `--run-command 'chmod +x /usr/local/bin/chv-authd /usr/local/bin/git-credential-chv-authd'`、 + `--run-command 'systemctl enable chv-authd || true'`、 + `99-chv-sync.cfg` 的 runcmd 在 `chv-sync boot` 之前加 + `systemctl start chv-authd || true`。 + +## 3. 诚实边界(本轮不宣称的东西) + +1. **seed 文件仍是 0600 root 落盘**:真 token 在 `/etc/chv/authd-seed` + 里仍以文件形式存在(cloud-init 的唯一投递通道就是写文件),且 seed + ISO 本身挂在客机上。真正让它对 agent 不可见的是 P2 的 nspawn 单元 + 隔离(单元内 root 映射为非特权用户,读不到单元外的 0600 文件);在 + P2 落地之前、codex 仍以客内 root 跑的镜像上,root agent 理论上仍能 + 读到它。本轮堵住的是“进程环境/argv/日志/`.git-credentials` 落盘” + 这条泄露面,不是“客内 root 万能”——后者是 P2 的边界,不是 P3 能修的。 +2. **本期默认的“替身”不是 Gitea 侧可独立校验的新 token**:按次释放给 + git 子进程的值仍是真 token,只是只进 git 进程内存、不进 agent。 + 真铸 token(Gitea API 铸短期专用 token、到期吊销)已实现辅助函数 + 与假客户端测试,但默认关闭——它要求真 token 有建 token 的权限, + 现有测试 token 只有 `write:repository`,铸 token 会失败;开它须另配 + 有铸 token 权限的 token,并做真机验证。不得把本期默认写成“Muse 同款 + 替身令牌已落地”。 +3. **SUB2API_API_KEY 不在本轮**:网关 key 仍走 guest-env 进 codex 环境 + (H2 既有行为)。按 Muse 模型它也该走 privsep worker(由 worker 代 + codex 调网关、codex 不见 key),那是网关调用链改造,超出本轮范围, + 列为后续,不静默声称已覆盖。 +4. **eBPF 污点传播未做**:Sentinel 本轮是决策层(wake 预审 + authd 放 + 凭证前评估),不是网络层强制。客内任意进程的直连仍由 Tinyproxy + 白名单 + 宿主防火墙(第三棒)兜底;eBPF cgroup 拦截 + LSM 污点传播 + 是后续工作,不得写成“出口已全部经 Sentinel”。 +5. **过渡兼容留了旧路径**:`chv-sync boot` 在 authd socket 不存在时仍走 + 旧的 `.git-credentials` 落盘(只多一句 stderr 警告),为的是 lite9/ + lite10 旧镜像不被一次改断。旧镜像退役后应删掉该回落,改为无 socket + 即 fail-closed;删除前不得声称“旧路径已关闭”。 +6. **无真机验证**:本轮全部是代码 + 本地单测(含真 git helper 端到端, + 但 authd 守护进程本身未在真 VM 里跑过)。真机待验证清单:authd 在 + lite11 镜像里启动时序(须在 `chv-sync boot` 之前就绪)、SO_PEERCRED + 放行 chv 用户、helper 在 pull/push 下的实际表现、wake 预审对真实 + Gitea 主机解析的判定、铸 token 的真实 API 行为。测试机(WSL)须等 + 第三棒 V-1~V-9 让出后再做。