Owner 2026-11-17: the plain list in the picker was unusable, make a proper table. The picker now reuses NodesView + NodesTableModel: flag, protocol badge, ping, speed, subscription groups (switchable), header sort, the pinned server drawn as the active row. «Закрепить выбранный», double click or Enter pins; «Авто» drops the pin; Esc goes back. The model is filled only when the picker opens. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
18 KiB
Service slots on Windows (design draft, 2026-11-16)
Status (2026-11-16): phases 1-5 in code, not yet released or tried on win10. Modules: xray_fluent/slots/
(catalog.py, config_writer.py, runtime_plan.py, assigner.py, rescue.py),
application/slot_runtime_service.py, ui/services_page.py; tests tests/test_service_slots_*.py (51).
Owner decisions (2026-11-16)
-
Windows gets slots, generated behind an explicit switch (AskUserQuestion, option "Generate behind a switch").
-
"Each service has its own logic, because it is slightly different, although not always": one shared slot base (routing, selector, failover, marks) plus a per-service policy only where the service really differs (YouTube: slow-video rescue, dead-cluster tracker, proven servers; AI: names resolved through the slot's server; Telegram: its own server).
-
"On Windows it must be made for beginners, so it is convenient to use": one plain option per service with working defaults; the slot picks its server itself; a pin is optional. No JSON knowledge needed.
-
"I quickly picked the option, and the program put everything into the config itself": ticking a service WRITES the slot selector and the slot rules into the active native sing-box JSON. They are visible and editable on the «Маршрутизация» pages; unticking removes exactly what the app wrote.
-
Helpers (2026-11-16): «one per server» — a slot may use any server; each distinct Hysteria2 server in use runs its own
hysteria.exe(up to 4), VLESS shares the one Xray sidecar. -
Scope of the first release (2026-11-16): «all three at once» — YouTube, AI and Telegram slots together.
-
Place (2026-11-16): a card «Серверы по сервисам» on the main page under the server card: a tick per service and «Авто» or a pinned server, as on the phone. REPLACED 2026-11-17 (owner, with a screenshot: "they should be in a separate tab and in a proper full-screen menu"; with thirty or more servers the card's ComboBoxes overflowed the window, scrolled slowly and lagged): the «Сервисы» page, see "What is built".
Point 4 is what keeps slots inside the routing contract: the active raw JSON stays the source of truth, the change is a user action from the GUI (like the existing "restore stock" actions), and nothing is injected at launch.
What Windows has today (survey 2026-11-16, file:line from that day)
- One routing exit: the template's
proxyoutbound. The runtime replaces it (engines/singbox/runtime_planner.py~433-500): aselectortaggedproxyover the local provider__app_nodeswhen the selected server runs on sing-box itself, otherwise the single native outbound. - Only servers whose core is sing-box go into that provider. VLESS, VMess and Trojan go to ONE Xray sidecar (hybrid mode:
proxybecomes a selector over two loopback SOCKS relays into Xray,runtime_planner.py~715-890, the Xray pool and its balancer inoutbound_pool_service.py). Hysteria2 goes to ONEhysteria.exesidecar behind a SOCKS relay. AWG to Amnezia. The controller owns one manager of each (controller.py~415-418). - Hot switch: Clash API
select_singbox_outbound(api_port, "proxy", tag)(selector_api.py), Xray balancer override;transition_engine.pypicks proxy_update / tun_hot_swap / proxy_hot_swap / reconnect. - YouTube:
diagnostics/dead_video_cluster.py(pure) feedssmart_switch_service.py, which switches the MAIN server. There is no per-service server. - Templates:
data/templates/sing-box/default.jsonsends the YouTube names (googlevideo.com, youtube.com, ...) toproxy; no Telegram or AI rules;finalisproxy.
Model
The slot model is the same as Android's (pure, tested, no Qt): SlotSpec(id, title, matchers, pool, pin, policy).
Built-in slots: youtube, ai, telegram; everything unmatched stays on proxy (the main server). Matchers on
Windows are process_name (YouTube desktop clients are browsers, so YouTube matches by domain; Telegram Desktop and
forks by process name and the Telegram IP ranges; AI by domain: chatgpt.com, openai.com, claude.ai, anthropic.com,
gemini.google.com, ...). A browser tab on YouTube is caught by domain, which is the main Windows case.
What the app writes into the active sing-box JSON when a slot is ticked
{"type": "selector", "tag": "zapret-slot-youtube", "outbounds": ["proxy"], "default": "proxy",
"interrupt_exist_connections": true}
and one route rule, placed right after the always-proxy names and before the RU-direct rules:
{"domain_suffix": ["googlevideo.com", "youtube.com", "youtu.be", "ytimg.com", "ggpht.com",
"youtube-nocookie.com", "youtubei.googleapis.com"], "action": "route", "outbound": "zapret-slot-youtube"}
- At rest the selector has only
proxy, so the file is valid on its own; the runtime adds the server pool (providers: ["__app_nodes"]) at launch, the same app-owned contract as theproxyreplacement. - Apps, names and IP ranges are separate rules (sing-box ANDs different fields inside one rule), in the order AI, YouTube, Telegram, as on Android.
- Tags starting with
zapret-slot-are app-written; untick removes exactly those blocks (and nothing the user wrote). default: proxykeeps the slot working when the app is not running its slot logic (the slot simply follows the main server): the JSON is valid and safe on its own.- The slot's DNS rule (names resolved through the slot's server, Android 0.6.2 lesson: a googlevideo address chosen by one node can be unreachable from another) is written the same way, only when the profile's DNS goes through the node.
- Template sync: these are user-made edits of
outbounds/route/dns, so the update merge keeps them verbatim.
The hard part: several native-core helpers at once
A slot on a different server than the main one needs that server's native core to carry traffic at the same time:
| Slot server's core | What it needs |
|---|---|
| sing-box (Shadowsocks, SOCKS, HTTP, ...) | nothing new: it is already a member of __app_nodes |
| Xray (VLESS, VMess, Trojan; Reality / XHTTP) | the same ONE Xray sidecar, one more loopback relay inbound per slot routed to that server's outbound |
| Hysteria2 | one more hysteria.exe per distinct Hysteria2 server in use (up to 4 with 3 slots) |
| AWG (Amnezia) | one more Amnezia instance per distinct server (heaviest; WireGuard adapters) |
Each helper ends in a loopback SOCKS relay that becomes a provider member __app_nodes/node_<hash>, so the slot
selector switches it with the same Clash API call as the main server. This keeps the architecture basis: every protocol
on its own native core, sing-box only routes.
Phase 3 plan: how a slot reaches its server at runtime (2026-11-16, after reading the launch path)
Facts: the provider file __app_nodes (constants.SINGBOX_PROVIDER_FILE) is written once at launch
(controller.py ~1377) and holds every sing-box-core server of the pool; it is built only when the MAIN server is a
sing-box-core server (runtime_planner._plan_runtime_outbound ~433-490). Helpers: one HysteriaManager
(controller.py ~417), one Xray sidecar, one Amnezia.
Plan:
- One stable relay per slot. When a slot is ticked, the runtime adds a loopback SOCKS outbound
zapret-slot-relay-<id>(own port and one-session credentials, like the Hysteria sidecar'sproxyrelay) and makes it a member of the slot selector. Behind it runs the slot server's own native core:hysteria.exefor Hysteria2,xray.exefor VLESS, Amnezia for AWG (owner 2026-11-16: «one per server»). Changing the slot's server to another server of the same kind restarts only that helper on the same port: sing-box is not restarted and the main traffic is not touched. (Change against the first draft: VLESS gets its own Xray per slot instead of a second inbound in the main Xray sidecar, so a slot move never restarts the main server's helper.) - sing-box-core servers need no helper: the provider is built whenever slots are on (not only when the main
server is a sing-box-core one) and the slot selector gets
providers: ["__app_nodes"]; a move is one Clash APIselecton the slot selector. - Members of a slot selector at runtime:
proxy(follow the main server), the slot relay, the provider. The selected member is pinned through the Clash API before TUN/system proxy is exposed, as forproxytoday. - Planner output:
SlotRuntimePlan(slot_id, member_tag, helper: kind + node + relay port/credentials | None), a pure function with tests; the controller owns the helpers' lifecycle (start after sing-box, stop with the session, health like the main Hysteria sidecar). - Assignment (which server per slot): pure, as on Android (
SlotPlanner): pin first, then servers that worked for the service (speed history), avoid the main server and other slots while the pool allows, skip servers that failed recently. The slot's failover and the YouTube rescue (dead cluster, slow video) move the SLOT, not the main server, when the YouTube slot is on.
Phases
- Pure slot module + tests (spec, assignment from the pool, pin, failure marks) shared in idea with Android.
- JSON writer: tick/untick writes/removes the
zapret-slot-*blocks; tests for order, idempotence, user-edited JSON kept,sing-box checkon the result. - Helpers: per-slot Xray relay inbound; then extra Hysteria sidecars; AWG last or never.
- Runtime: per-slot selection through the Clash API, per-slot failover, YouTube rescue and dead-cluster tracker on the YouTube slot instead of the main server.
- UI: one card «Серверы по сервисам» with an option per service, «Авто» or a pinned server.
- Release, wiki guide (Zapret/premium.md, premium/apps.md), screenshots on win10.
What is built (2026-11-16)
- Page «Сервисы» in the navigation under «Серверы» (
ServicesPage, since 0.8.85; it replaced the main-page cardServiceSlotsCard): a row per service with its switch, «Сервер: Авто» or «Закреплён: …», «сейчас: …» while connected and «Выбрать сервер». The button opens a full-page picker for that one service. Since 0.8.86 (owner 2026-11-17: the plain list was "nonsense", make a proper table) the picker is the same virtual table as «Серверы» (NodesView+NodesTableModel): flag and name, protocol badge, ping, speed; groups by subscription (switchable: country, protocol, none); sorted by ping, a header click sorts by name, protocol, ping or speed. The pinned server is drawn as the active row. «Закрепить выбранный» (or a double click / Enter) pins the selected server, «Авто» drops the pin, Esc goes back. The table model is filled only when the picker opens; runtime refreshes (slot_runtime.changed) only update the three rows (the old card refilled three ComboBoxes with every server on each of them, which made it lag). A switch callsAppController.set_service_slots, which writes the slot blocks withapply_slotsand saves throughapply_singbox_config_text(the routing editor's own path: validation, save, reconnect when connected); the editor page reloads the saved JSON. Pins live inAppSettings.service_slot_pins. - Launch:
_plan_runtime_singbox/_plan_proxy_runtime_singboxcall_attach_slot_plans:assign_slots(pin first, then the best free server: alive, faster, lower ping; never the main server or another slot's while the pool allows; AWG skipped) andattach_service_slots(relay per slot, provider members when the main server is a sing-box-core one, a native out otherwise, the helpers'process_path -> direct). - Helpers:
SlotRuntimeServicestarts after sing-box is up and stops onsingbox.stopped. Each helper runs from its own executable copydata/runtime/slots/<core>-slot-<id>.exewith its own config (the Hysteria and Xray managers gotexe_path/config_file): the main sidecars' orphan cleanup kills by executable path and their stale-config cleanup globs by name, so with the shared defaults a slot helper would kill or delete the main one's. A helper that does not start leaves its slot onproxy, is logged, and that server is skipped for the slot until the app restarts or the user pins it again. - Masking: a slot's Xray helper gets «Маскировка» of its server like the main path (mux, XHTTP connection
rotation, fragmentation:
helper_config(helper, masking)masks the pool outbound); packet tricks are a sing-box route option of the main path's server socket, so the carried plan is dropped and the log says[slots] <id>: маскировка: пакеты-обманки маскировки у сервера слота не применяются. Hysteria2 and sing-box-core slot servers are not masked (masking has nothing for them). - Hot moves: every slot has a stable loopback relay, so moving a slot to a Hysteria2 or VLESS server only starts a
helper behind the same port, and moving to a provider server is one Clash API
select; sing-box is not restarted and other apps are not touched. A pin that needs a restart (a sing-box-core server outside the provider) reconnects. - YouTube rescue: with the YouTube slot on, a dead YouTube cache cluster (
dead_video_cluster.py) moves the YouTube slot (slots/rescue.py: best free hot-reachable server, the dead one remembered, 90 s cooldown, back to the main server when nothing is free) instead of switching the main server; honours «авто-переключение» and «переключение при низкой скорости», not «сервер выбран вручную» (that is the main server's pick). Log[slots] youtube: кластер … не отвечает на A, слот YouTube -> B; status «… YouTube переведён на B, остальные программы не тронуты. Перезапустите видео …». - A bad pinned server gives way (owner decision 2026-11-17): "the program should choose the best server on its own,
even if the user explicitly selected another one, if the server selected by the user is bad". A pin is honoured only
while its server did not fail for the slot (helper failure, dead YouTube cluster) and did not fail its last check
(
is_alive is False); otherwise the slot takes the best free server with the reason «закреплённый сервер не отвечал». The pin stays saved and wins again once a check says the server is good; pinning a server whose check failed moves nothing and does not reconnect (log[slots] <id>: закреплённый сервер не отвечает, выбран лучший). - One machine is one server: a provider machine often comes as several entries (VLESS gRPC, VLESS WS, Hysteria2,
each with its own host name). The assigner counts entries whose host keys meet as one server (
host_keys: the host name plus the addresses of this session's DNS snapshot,bootstrap_dns.cached_addresses, stable for the session so a background refresh never changes a built plan). «Авто» never puts a slot on the main server's machine or on another slot's machine through a sibling entry; with no other machine free the slot follows the main server. The YouTube rescue never moves to a sibling of the machine the video died on (same address, same cache cluster). A pin is honoured as it is. No new DNS lookup is made for this: the addresses come from the trusted bootstrap cache that the server checks fill (each ping resolves its server by DoH and remembers the answer), and the app reads that cache once per run, so servers first checked in this run are grouped from the next start of the app. - Live e2e on win10 (2026-11-16, source
ff37b8c): headless AppController in proxy mode against five loopback servers of the real cores (threehysteria.exe serverwith a self-signed certificate andpinSHA256, twoxray.exeVLESS over TCP). Main server hy1. Ticking the three services gave AI -> hy2, YouTube -> hy3 (bothhysteria-slot-*.exe) and Telegram -> vl1 (xray-slot-telegram.exe), each selector on its own relay. Through the local SOCKS: YouTube 204, ChatGPT 200, Telegram 200, Wikipedia 200. Each Hysteria server saw exactly one client, so every service really used its own server. Pinning YouTube to vl2 while connected moved it Hysteria -> Xray behind the same relay, with the same sing-box PID and no reconnect, and traffic kept flowing. Disconnect left no slot helper running. Not covered: real censorship and the provider-member path (no sing-box-only server in the set). - Live e2e on real vpnbot nodes (2026-11-16, source
0f53c7d): a synthetic vpnbot test person, 12 entries (7 VLESS incl. one Reality with ML-KEM VLESS encryption, 5 Hysteria2 with salamander obfs) on 6 machines. Its link serves Mihomo YAML only, so the entries were converted to links and imported (slots_e2e.py --links). Main server HY2 Finaltelekom; the three slots gotxray-slot-*.exehelpers on their relays; YouTube 204, ChatGPT 200, Telegram 200, main 200 in 1-2 s; a hot pin move kept the sing-box PID. The run showed the gap fixed above: the main server, all three slots and the move target were five entries of ONE machine (same address), so nothing was really separated. After the fix (source46c6e13), two app runs as a restart: run 1 pinged the pool (11 of 12 names in the trusted cache, all 11 already frozen empty in that run's snapshot, as designed), run 2 put main Finaltelekom, YouTube Rusina, AI Nithesina, Telegram us-vdsryzen-8 and the hot-move target Hiphopgen on five different machines. The same runs found and verified the fix of a python.exe access violation (a finished worker step's deadline QTimer collected on a worker thread,async_steps.TransitionStep.release). - Not yet: slow-video rescue (Windows has no slow-video detector, only dead clusters), AWG slot servers (
slot_capablekeeps them out of the picker list, and an old AWG pin falls back with the reason «серверы AmneziaWG для сервисов пока не поддерживаются»), a speed history per service, per-slot traffic on the page, Clash/Mihomo YAML subscriptions (neither app imports them).