mtproto-client-behavior-matrix

# MTProto Client Behavior Matrix Use this skill when behavior differs by client platform (iOS/Android/Desktop) or when tuning relay/handshake timeouts. ## Evidence Policy - Do not publish behavior claims without evidence. - Accept only: - reproducible local captures/logs, or - direct client source links pinned to a tag and commit. - Mark each claim as `source-backed` or `field-capture`. ## iOS (Telegram iOS) Version snapshot: - Repo/tag: `TelegramMessenger/Telegram-iOS` `build-26855` - Commit: `b16d9acdffa9b3f88db68e26b77a3713e87a92e3` Source-backed behavior: - TCP connect timeout: `12s` - https://github.com/TelegramMessenger/Telegram-iOS/blob/b16d9acdffa9b3f88db68e26b77a3713e87a92e3/submodules/MtProtoKit/Sources/MTTcpConnection.m#L980 - Response watchdog base: `MTMinTcpResponseTimeout = 12.0` - https://github.com/TelegramMessenger/Telegram-iOS/blob/b16d9acdffa9b3f88db68e26b77a3713e87a92e3/submodules/MtProtoKit/Sources/MTTcpConnection.m#L576 - Response timeout includes payload-dependent term and resets on partial reads - https://github.com/TelegramMessenger/Telegram-iOS/blob/b16d9acdffa9b3f88db68e26b77a3713e87a92e3/submodules/MtProtoKit/Sources/MTTcpConnection.m#L1339 - https://github.com/TelegramMessenger/Telegram-iOS/blob/b16d9acdffa9b3f88db68e26b77a3713e87a92e3/submodules/MtProtoKit/Sources/MTTcpConnection.m#L1398 - Transport-level watchdog: `20s` - https://github.com/TelegramMessenger/Telegram-iOS/blob/b16d9acdffa9b3f88db68e26b77a3713e87a92e3/submodules/MtProtoKit/Sources/MTTcpTransport.m#L312 - Reconnect backoff steps (`1s`, then `4s`, then `8s`) - https://github.com/TelegramMessenger/Telegram-iOS/blob/b16d9acdffa9b3f88db68e26b77a3713e87a92e3/submodules/MtProtoKit/Sources/MTTcpConnectionBehaviour.m#L66 Field-capture behavior (proxy-side observations): - Pre-warms multiple idle sockets. - Can split 64-byte obfuscation handshake across TLS records. - May delay first payload after `ServerHello`. Proxy implications: - Assemble handshake bytes until full 64-byte payload is complete. - Keep generous handshake stage timeout; tighten only after relay is active. - Do not treat short idle prewarmed sockets as protocol failure. ## Android (Telegram Android) Version snapshot: - Repo/ref: `DrKLO/Telegram` `master` (snapshot: `12.6.4 (6666)`) - Commit: `009e97356f966bb81eceba113d210230bf383122` Source-backed behavior: - Enables `TCP_NODELAY`, switches socket to `O_NONBLOCK`, uses `connect(..., EINPROGRESS)` with edge-triggered epoll. - https://github.com/DrKLO/Telegram/blob/009e97356f966bb81eceba113d210230bf383122/TMessagesProj/jni/tgnet/ConnectionSocket.cpp#L618 - Connect path chooses address family/static flags and sets per-type logical timeouts (`Proxy=5s`, `Generic=8/12s`, `Upload=25/40s`, `Push=20/30s`). - https://github.com/DrKLO/Telegram/blob/009e97356f966bb81eceba113d210230bf383122/TMessagesProj/jni/tgnet/Connection.cpp#L276 - https://github.com/DrKLO/Telegram/blob/009e97356f966bb81eceba113d210230bf383122/TMessagesProj/jni/tgnet/Connection.cpp#L368 - Timeout model is logical/internal (`setTimeout` / `checkTimeout`). - https://github.com/DrKLO/Telegram/blob/009e97356f966bb81eceba113d210230bf383122/TMessagesProj/jni/tgnet/ConnectionSocket.cpp#L1105 - https://github.com/DrKLO/Telegram/blob/009e97356f966bb81eceba113d210230bf383122/TMessagesProj/jni/tgnet/ConnectionSocket.cpp#L1115 - Explicit connection-type split (`Generic`, `Download`, `Upload`, `Push`, `Temp`, `Proxy`) and multiple parallel slots. - https://github.com/DrKLO/Telegram/blob/009e97356f966bb81eceba113d210230bf383122/TMessagesProj/jni/tgnet/Defines.h#L68 - https://github.com/DrKLO/Telegram/blob/009e97356f966bb81eceba113d210230bf383122/TMessagesProj/jni/tgnet/Defines.h#L26 Proxy implications: - Expect several concurrent connection attempts from one device/session. - Treat connect churn as normal; optimize for quick accept + cheap close path. ## Desktop (Windows/Linux/macOS via Telegram Desktop) Version snapshot: - Repo/tag: `telegramdesktop/tdesktop` `v6.7.2` - Commit: `085c4ba65d1f8aa13abf0fd7fc8489f094552542` Source-backed behavior: - Builds multiple test connections and chooses by priority. - https://github.com/telegramdesktop/tdesktop/blob/085c4ba65d1f8aa13abf0fd7fc8489f094552542/Telegram/SourceFiles/mtproto/session_private.cpp#L1010 - Wait-for-connected starts at `1000ms` and grows on failures. - https://github.com/telegramdesktop/tdesktop/blob/085c4ba65d1f8aa13abf0fd7fc8489f094552542/Telegram/SourceFiles/mtproto/session_private.cpp#L34 - https://github.com/telegramdesktop/tdesktop/blob/085c4ba65d1f8aa13abf0fd7fc8489f094552542/Telegram/SourceFiles/mtproto/session_private.cpp#L1236 - TCP/HTTP transport full-connect timeout: `8s`. - https://github.com/telegramdesktop/tdesktop/blob/085c4ba65d1f8aa13abf0fd7fc8489f094552542/Telegram/SourceFiles/mtproto/connection_tcp.cpp#L21 - https://github.com/telegramdesktop/tdesktop/blob/085c4ba65d1f8aa13abf0fd7fc8489f094552542/Telegram/SourceFiles/mtproto/connection_http.cpp#L18 - Resolver path uses per-IP timeout `4000ms` and scales by resolved IP count. - https://github.com/telegramdesktop/tdesktop/blob/085c4ba65d1f8aa13abf0fd7fc8489f094552542/Telegram/SourceFiles/mtproto/connection_resolving.cpp#L16 - After first success, may wait `2000ms` for a better candidate. - https://github.com/telegramdesktop/tdesktop/blob/085c4ba65d1f8aa13abf0fd7fc8489f094552542/Telegram/SourceFiles/mtproto/session_private.cpp#L33 Linux/macOS note: - In reviewed MTProto sources, timeout/retry logic is shared with Windows path (no observed OS-specific branching there). - Validate package/channel specifics separately (Snap/DEB/Flatpak/macOS build IDs) before asserting behavioral deltas. Proxy implications: - Expect candidate racing and early cancellation patterns. - Keep reconnect path cheap and non-blocking. - Do not overfit timeout policy to a single desktop platform. ## Practical Use Checklist - If one platform fails and others pass, first compare known timeout model for that client. - Verify whether failures happen pre-handshake, in 64-byte obfuscation phase, or after relay starts. - Inspect whether client behavior is expected parallel racing vs actual proxy regression. - Update this skill only with pinned sources or reproducible captures.