Files
zapret-autotune/README.md
T
2026-10-10 09:16:58 +03:00

215 lines
12 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# zapret2-autotune
Go CLI for scanning and checking YouTube TLS bypass strategies with `blockcheckw`.
## Manual managed-test activation (not deployed)
`managed init|status|apply` requires exact role `managed-test`, target ID
`zapret-autotune-managed-v1`, a provisioned runtime marker and zapret executables,
and a private dedicated mount at `/var/lib/zapret-autotune/target`.
`ZAPRET_CONFIG` must equal `/var/lib/zapret-autotune/target/config`: the standalone
full config is allowed only in that fresh target volume, never over an existing
service config. Existing-service include fragments are not implemented.
The managed proxy must equal `socks5://127.0.0.1:$SOCKS_PORT` and that listener
must be owned by target-local `ss-local`; arbitrary loopback/host ports fail.
Init creates a private disabled baseline and identity/journal without overwriting
existing config. Status shows generation/hash/groups, not secrets or live health.
Apply accepts `--frontend-candidate ID --googlevideo-candidate ID` and requires
a qualified pair, fixed target init executable and explicit runtime-config
binding. It uses full target-local firewall/daemon restart and Player API post-check, with
atomic rollback on failure. See [MANAGED_TARGET.md](MANAGED_TARGET.md) for guards
and provisioning prerequisites. An isolated pinned deployment definition is in
[deploy/managed-target](deploy/managed-target/README.md); build/run are deferred.
No autonomous loop, remote activation or Xray changes are included. Image
compatibility and route/rollback gates are not yet remotely proven; adapter
tests/build are deferred to the orchestrator.
## One-shot Player API health
```sh
zapret-autotune health youtube
```
This makes one bounded Android Player API request through the explicitly
configured **loopback SOCKS5** endpoint. Healthy means only
`playabilityStatus.status == "OK"`; no streaming URL is required and no media is
fetched. **API health is not playback health.** Non-OK/missing status, malformed
responses and request failures return exit 1; invalid config returns exit 2.
Curl configuration and proxy bypasses are disabled; ambient proxy settings are
not used and there is no direct fallback. The request times out after 30 seconds
(35-second process deadline), with a 1 MiB response limit and sanitized errors.
| Environment variable | Default |
| --- | --- |
| `ZAPRET_AUTOTUNE_SOCKS_PROXY` | `socks5://127.0.0.1:11080` |
| `ZAPRET_AUTOTUNE_VIDEO_ID` | `dQw4w9WgXcQ` |
| `ZAPRET_AUTOTUNE_HEALTH_INTERVAL` | `5m` |
| `ZAPRET_AUTOTUNE_FAILURE_THRESHOLD` | `3` |
| `ZAPRET_AUTOTUNE_RECOVERY_COOLDOWN` | `30m` |
Configuration is parsed once per health invocation. The proxy must use `socks5`
with a loopback IP or `localhost`, port 1–65535, and no credentials/path/query.
Video IDs must be 11 characters; durations and the failure threshold must be
positive. Explicit empty values are invalid. Interval, consecutive-failure
threshold and cooldown are validated future recovery policy settings only:
**An injected autonomous scheduling core exists, but no monitor command, actual
recovery callback or live managed target is wired yet**.
The future managed target is a **brand-new prod-like zapret2-ss container in the
test environment only**, not the existing container. Its stable configurable
SOCKS endpoint must be loopback-accessible to this probe and same-host Xray.
The one-shot command does not persist health. This command does not create or change any container, apply
configuration, restart services or roll back anything. Xray configuration is
out of scope and will be configured separately by the user. Run this command in
the host network context where that loopback endpoint is reachable; Docker
`localhost` otherwise refers to the CLI container itself.
## Exact cache-host proposal
```sh
zapret-autotune tune googlevideo --video-id dQw4w9WgXcQ --proxy socks5://localhost:1080
```
Both flags are required. The validated Android 21.02.35 player request discovers
exactly one current cache hostname through the explicit proxy. `blockcheckw`
then scans only that hostname for TLS1.2/TLS1.3, checks the shortlist with
`--take 10 --passes 3`, ranks fully passing strategies by measured data-transfer
speed/latency, and checks the exact combined arguments 3/3 on both protocols.
The proxy is for discovery; blockcheckw uses its own network path, not that proxy.
No Selectel speedtest is used. This is not a YouTube playback/SABR test.
Only after successful transactional persistence does this command print one
inert `NFQWS2_OPT` profile filtered by the **exact discovered hostname**, never
the `googlevideo.com` wildcard. Cache hosts rotate; one host is not wildcard
evidence. GGC proposals use `--hostlist-domains=^<discovered-host>`: the `^`
prefix disables Zapret2's default subdomain matching, so untested subdomains
are not included. Only validated lowercase `*.googlevideo.com` hostnames are
accepted; commas and a leading `^` in the discovered hostname are rejected.
Schema-v3 candidates use group `googlevideo`, one exact endpoint, and
both endpoint×TLS measurements with raw JSON and timestamps. An additive v3
cache-run journal records failed attempts without candidates and successful
attempts atomically with their candidate/evidence; legacy frontend runs keep
their existing target constraint. Database failures emit no options.
`history` and `strategies youtube` include the cache group and exact endpoint;
`status` still describes the legacy frontend journal only.
Signed URLs are not printed. Scanner diagnostics are retained in restricted
artifact files rather than relayed to the terminal. No configuration is applied,
no profile is marked active, and no service is restarted. The staged pipeline is
scan/propose → guarded manual managed-test activation → future monitoring;
the deployment definition exists but is not deployed; the monitor scheduling core
is internal and **not wired to a command or actual recovery**. Manual
proposal-only recovery is available below.
No-argument `tune youtube` remains frontend-only and unchanged.
## Alternatives-first recovery (proposal only)
```sh
zapret-autotune recover googlevideo --video-id dQw4w9WgXcQ --proxy socks5://localhost:1080
```
Recovery discovers the current exact cache host once, then rechecks qualified
schema-v3 `googlevideo` candidates newest-first using their exact saved arguments.
Qualification covers every saved endpoint × TLS protocol; revalidation checks
the fresh host for combined TLS1.2+TLS1.3 3/3. The first passing scenario is saved
as a new immutable candidate/run scoped to that host and emitted as an exact-host
proposal, **not active**. No scan runs after a passing saved scenario.
Only no eligible candidates or complete, valid non-passing evidence permit the
normal tune scan on the already-discovered host. Database, discovery, cancellation,
runner and malformed/incomplete-report errors fail closed without a fallback scan.
This command does not apply configuration, restart services or start monitoring.
## Cache-host discovery
```sh
zapret-autotune discover youtube --video-id dQw4w9WgXcQ --proxy socks5://localhost:1080
```
Both flags are required. Use a proxy address reachable from the CLI; in Docker,
`localhost` is the container itself, not the Docker host. Proxy credentials,
cookies, and API credentials are not supported. This command requires `curl`
with `--json` support and POSTs to the YouTube player API through SOCKS5 using
Android client 21.02.35, SDK 30, Android 11. Requests time out after 30 seconds;
responses are limited to 1 MiB and must have playability status `OK` and a valid
HTTPS `serverAbrStreamingUrl` under `.googlevideo.com`.
Output is JSON containing only `videoId` and `hostnames`; signed streaming URLs
are never printed. Cache hosts may rotate. This is **cache-host discovery only**,
not verification of real video transfer: a discovered SABR URL can return
`sabr.malformed_config` on GET. Discovery does not fetch media, persist results,
apply production configuration, or restart services. Invalid input and failed
requests return nonzero with sanitized errors.
Build the Docker image with:
```sh
docker build -t zapret-autotune .
```
The `tune youtube` command scans and checks candidate strategies; it prints a
proposed `NFQWS2_OPT` and does not apply configuration or restart services.
Tune runs are recorded in pure-Go SQLite (no CGO). A verified combined candidate
and its TLS1.2/TLS1.3 3/3 pass results are committed together before options are
printed; failed scans/checks are recorded as failed runs without a candidate.
Database initialization or persistence errors return nonzero and emit no options.
`status` shows the latest recorded run, or explicitly says there is no recorded
run. Candidates are **PROPOSED, not active**; this journal does not track managed activation and
production health is not checked. A later failed run is shown instead of an older
proposal. An output-write failure can leave a verified proposal recorded, but
still returns nonzero.
Set `ZAPRET_AUTOTUNE_DB` to override the default database path
`/var/lib/zapret-autotune/autotune.db`. The parent directory is created if needed;
it must be writable. In Docker, mount that directory to retain history:
```sh
docker run --rm -v autotune-data:/var/lib/zapret-autotune zapret-autotune status
```
Schema version 5 is initialized transactionally on `tune` and `status`; subsequent
opens are idempotent and newer, unsupported schema versions are rejected.
Schema v4 adds a policy-state foundation keyed by stable service ID (not a cache
hostname or SOCKS address): last probe time, consecutive failures, recovery lock
and cooldown expiry. Internal operations accept injected time: success resets the
streak, failures increment it, and an atomic recovery acquisition requires the
configured threshold, an elapsed cooldown and no running recovery. Completion,
successful or failed, requires the exact per-acquisition ownership token returned
by acquisition, clears the lock and starts cooldown from completion time. Stale
or duplicate completion is rejected without changing the lock, token or cooldown,
even after a newer acquisition starts. State survives reopening the database.
The internal `runManagedMonitor` supervisor probes immediately and periodically,
persists observations, and invokes injected recovery only at threshold outside
cooldown. Recovery errors are reported without stopping monitoring; DB, lock and
finalization errors are fatal. Cooldown starts when recovery returns, including
failure or cancellation. Callbacks must honor cancellation and join their children.
Existing target guards run before a lifetime nonblocking `flock` in the private
target state directory. With that exclusive singleton held, one-shot startup
reconciliation uses schema-v5 PID/starttime identity: only a proven dead or
PID-reused owner permits clearing its exact lease and starting crash cooldown.
A live owner remains busy; legacy or unverifiable owners fail closed. Probe
history is preserved. The lock file is never unlinked.
The scheduler core has fake-time tests for scheduling, reset, cooldown,
cancellation, parallel starts, fatal DB errors, reopen and safe startup
reconciliation. **There is no actual recovery adapter,
monitor CLI/entrypoint wiring or live target yet**; container/Xray are untouched.
Manual managed activation remains separate. Verification of this increment is
deferred to the orchestrator.
Schema v3 adds immutable strategy candidate definitions (group, declared exact
endpoints, args and normalized-args fingerprint) and endpoint/protocol evidence.
Each newly successful tune saves its exact verified combined frontend profile for
`youtube.com`, both TLS measurements, raw measurement JSON and checked timestamps
in the same transaction as the existing run. Qualification requires every declared
endpoint to pass exactly 3/3 for both TLS1.2 and TLS1.3, with full success/coverage.
Existing v2 history is unchanged and is not backfilled into independent candidates.
Candidates have no DB active/installed state; managed activation uses its own
durable config journal. Autonomous scheduling is internal only; actual recovery
and live monitoring remain unwired.
The explicit `tune googlevideo` path adds exact-cache proposals only.
Proposed runs store a SHA-256 fingerprint of validated arguments with separator
whitespace normalized and argument order preserved; repeated proposals retain
separate history rows. Failed runs have no fingerprint. Version 1 history is
upgraded atomically, rejecting invalid proposed arguments without changing it.