Files
zapret-autotune/deploy/managed-target/README.md
T
2026-10-10 09:45:34 +03:00

87 lines
5.6 KiB
Markdown

# Isolated managed-test definition — deployment deferred
This definition is **not autonomous acceptance-ready**: `managed monitor` and
entrypoint recovery integration exist in code but are not live verified. Do not build/pull/start it as part of this
change. Existing containers, production and Xray configuration are out of scope.
For a later approved deployment, from this directory:
1. Copy `.env.example` to `.env` and restrict it to `0600`; it sets only the
host loopback SOCKS port. Generate a **new independent** secret locally:
`(umask 077; set -C; openssl rand -hex 32 > ss_password)`.
This refuses to overwrite an existing file. Never copy existing secrets or
put the password in `.env`, environment variables or command arguments.
`ss_password`, `.env` and state are gitignored and excluded from the build
context. Do not commit the secret file. Compose mounts it read-only at
`/run/secrets/ss_password`; keep the host file owned by the invoking user
with mode `0600` (no deploy-time chown or world-readable permissions).
Use only this newly generated regular file at `./ss_password`, not a symlink
or a reused secret. Do not mount its parent directory or other host secrets.
2. Provision a **new empty** `./target-state`, owned by host UID 0 (runtime
user), mode `0700`. Rootless/user-remapped Docker needs the mapped owner.
Never reuse mounts/state. Compose refuses a missing bind source. Retain
state for recreation; deletion loses identity, journal, backups and DB.
3. Orchestrator may validate without deploying:
`docker compose --env-file .env -f compose.yaml config --quiet`.
Rendered Compose config contains only the secret file path, not its value. Build/run require
separate approval and runtime validation gates.
Offline static guard (from repository root):
`python3 deploy/managed-target/check_static.py`. It pins the capability allowlist
and single-secret/single-state mount shape; it does not prove runtime readability.
Only SOCKS TCP/UDP is published on host
`127.0.0.1:${ZAPRET_AUTOTUNE_HOST_SOCKS_PORT:-11080}` for same-host Xray;
no Xray edits are supplied. SS port `8388` is internal, never published.
The server binds `0.0.0.0:8388` only inside the dedicated bridge namespace;
ss-local binds `0.0.0.0:1080`. Startup requires a hex-only secret of at least
32 characters and writes JSON configs to the dedicated target-owned tmpfs
`/run/zapret-autotune-ss` (directory `0700`, files `0600`). Both Shadowsocks
processes receive only `-c <config_path>`, never a password in argv or env.
Docker administrators/root can still read mounted secrets and runtime configs;
file-backed Compose secrets are not encrypted storage.
The dedicated bridge is not shared with existing services. No host networking,
Docker socket, existing mounts, host Lua/fakes or privileged mode; all
capabilities are dropped except `NET_ADMIN`, `NET_RAW`, `SETUID`, `SETGID`
and `DAC_OVERRIDE`.
`NET_ADMIN` covers nft/NFQUEUE; `NET_RAW` covers nfqws2 raw sends. Zapret2's
`nfqws2 --user=tpws` needs `SETUID`/`SETGID` for UID/GID switching, then retains
only `CAP_NET_RAW` and `CAP_NET_ADMIN`. `DAC_OVERRIDE` lets the root entrypoint
read the invoking user's mode-`0600` file-backed secret despite its different
host UID; it does not make the host file world-readable or writable through
the read-only secret mount. This capability is container-wide, not restricted
to one pathname: its safe scope depends on exposing only
`./ss_password` → `/run/secrets/ss_password` (read-only) and the fresh
`./target-state` → `/var/lib/zapret-autotune/target` bind, plus dedicated runtime
tmpfs. Never add host directories, other secrets or existing state to this
container. No default-all capability set is used. Rootless/user-remapped Docker
must validate secret readability separately; this does not bypass user-namespace
or host filesystem restrictions.
Go builder matches `Dockerfile`; runtime digest is pinned and image-owned
Lua/fakes retained. Startup: ss-server/ss-local → internal listener 1080
(30-second bound) → guarded managed init on fresh state (status on recreation)
→ zapret2 start using target config → one installed managed monitor child. Drift/pending state refuses startup;
no automatic adoption of unknown state. SIGTERM/SIGINT cancels and waits the monitor
before stopping target-local zapret and the entrypoint's Shadowsocks children.
Monitor fatal exits stop the entrypoint; no restart policy hides fail-closed exits.
Monitor probes Player API via owned loopback SOCKS; unresolved GGC is not playback health.
Inside target, `zapret-autotune health youtube` uses
`socks5://127.0.0.1:1080`, never the host port. It is one-shot Player API health,
not playback or scheduling.
The supplied remote read-only inspection verified shadowsocks-libev 3.3.5
and `ss-server -h` supports `-c <config_file>`; runtime JSON loading remains
unverified. Before deployment acceptance, validate Lua/fakes,
`ZAPRET_CONFIG` binding and full zapret start/stop/restart with only
the explicit `NET_ADMIN`, `NET_RAW`, `SETUID`, `SETGID`, `DAC_OVERRIDE` capability set (including
the `tpws` user drop). Prove target-local routing,
NFQUEUE activation/rollback, signal cleanup and persistent recreation on the
new target only. Compose syntax, build and runtime integration are unverified
by this producer. Do not widen capabilities or reuse existing config/network
as a workaround. Autonomous acceptance also needs tests/build and live verification
of monitor/recovery and crash reconciliation, including shutdown grace sufficient
for bounded restoration (which may exceed the defined 45s). Docker configuration
has not been changed by this code integration increment.