Repository navigation
v0.3.7
-
Added managed Linux desktop capture with
auto_route, optional TCPauto_redirect, independent DNS interception, and best-effort systemd-resolved integration. Capture supports single-stack or temporarily offline underlays and retains network ownership across reloads; proxy transports and DIRECT traffic still require native connectivity. This is host TCP/UDP capture, not gateway forwarding. Requires network-administration privileges andnft; network-topology changes require a restart. See managed TUN requirements and limitations. -
Added routing
networkpredicates for TCP/UDP andprotocolpredicates for STUN/DTLS. Classification runs only when a listener's rules useprotocol, inspects the first inner UDP datagram, and pins the selected route for that flow/destination; fixed-target streams can then wait up to 300 ms. These spoofable hints do not provide complete WebRTC detection, shared NAT mappings, or call-wide exit affinity. See UDP protocol routing. -
shoes --versionandshoes versionnow report the build target and actual Rust compiler on separate lines; useshoes -Vfor one-line output. Non-official builds include a 12-digit commit hash, with.dirtyfor tracked changes or+unknownwhen no revision is available. Official release versions remain plain. Docker releases use separatesha-<commit>-releaseimages instead of reusing development binaries. -
XHTTP H1 packet-up clients without application byte caps now collect one bounded successor POST while waiting for the previous acknowledgement instead of accumulating only an input pipe of data. HTTP/1 pipelining and pacing are unchanged. Explicit byte caps retain the previous ACK gating to avoid staging data into memory needed by a download before its ACK.
-
Raised the XHTTP H2 client connection receive-window default to 64 MiB (servers retain 8 MiB), improving headroom for concurrent downloads without changing the 4 MiB stream default. Added opt-in
pool.h2_receive_admissionto limit downloads to the connection/stream window ratio: 16 per carrier with default windows, up to 48 across the default pool. Byte-budget clamps can reduce this to one per carrier. Admission remains disabled by default because lingering packet-up download tails retain slots and can exhaust them during ordinary connection churn. Without it, larger windows mitigate but do not eliminate paused-reader stalls; application-global memory exhaustion can still backpressure sessions. -
XHTTP H1 split-mode clients now establish their two carrier connections concurrently, retaining atomic connection admission and cancellation on setup failure.
-
Updated ML-KEM to 0.3.4 for secret-key debug redaction, constant-time equality/reduction hardening, and improved zeroization. Updated AWS-LC-RS to 1.19.0 (bundling AWS-LC 5.12.0), plus the BLAKE2 and Linux netlink dependencies, while retaining protocol formats, maintained forks, and existing Rust/mobile platform requirements.
-
Added
conditionsfor general-purpose routing expressions, including nestedmasksaddress leaves. Existing top-levelmasks/masksyntax remains supported without warnings. Config serialization keeps plain address expressions undermasksand emits other expressions underconditions; controller rule and connection diagnostics now explicitly wrap address leaves inmasks. See rule configuration. -
Added opt-in
global_settings.ntpfor application-local time correction using unauthenticated NTP. It adjusts protocol timestamps and rustls certificate validation with an offset, never writes the OS clock, and retains samples across unchanged reloads; see NTP settings and limitations. -
Added
generate-tendns-txtfor offline encrypted A/AAAA/CNAME publications using a key supplied only throughTENDNS_KEY. Publishing requires replacing old values and splitting values longer than 255 bytes into strings within one TXT record; see CLI operations. Tendns resolution in Shoes is not yet active. -
Added
global_settings.default_dnsfor listeners, TUN entries, and WireGuard inbounds. Omitted or nulldnsinherits it; explicitdnsobjects replace it completely. Inline fake-IP defaults share one mapping pool. Independent bootstrap and service lookups retain their own DNS configuration. -
Added resolver-scoped static
hostsmappings in named DNS groups, inline DNS settings, and default DNS. They override upstream and Fake DNS results without changing TLS SNI, certificate identity, or XHTTP Host. Mapped A/AAAA queries with no address in that family return an empty answer rather than falling back upstream; remotely resolved proxy destinations remain remote. Static real addresses do not get Fake DNS hostname restoration. See static hosts and precedence. -
Added a native Hysteria2 client with shared HTTP/3 authentication, TCP streams, bounded UDP fragmentation and demultiplexing, and connection-owned reconnect lifetimes. Native QUIC clients must be outermost at chain hop 0; existing stream-based proxies can follow them. Brutal and 0-RTT are not supported.
-
Added a TUIC v5 client with TCP, native and QUIC-stream UDP relay modes, session-exporter authentication, heartbeats, and bounded association teardown.
-
Added optional Salamander obfuscation for Hysteria2 clients, including coalesced UDP receive handling and transport MTU accounting.
-
Added connection-owned Hysteria2 port hopping with fixed or randomized intervals, protected sockets, and receive overlap across hops.
-
TUIC authentication now buffers bounded pre-auth commands without consuming the stream credit needed for AUTH. Native client connections release closed generations promptly and allow a bounded drain when the last retired flow closes.
-
Added
global_settingsfor process-wide options, with resource limits and the optionalrule_cache_dirin one flat block.global_limitsremains a compatibility alias and logs a warning that it will be deprecated in the next version; rename it toglobal_settings. Only one block is allowed across all config files, including when using the alias. -
Breaking: removed ordered
type: fallbackclient groups andhealth_check. These configurations are rejected; supported group types areround-robin,selector, andurl-test. Address-level connection fallback and protocol authentication fallbacks are unchanged. -
Breaking: action-free file/source groups now use
rule_source_group, referenced withmasks: {rule_source_group: name}. Migrate existing file-backedrule_groupdeclarations andmasks: {rule_group: name}references.rule_groupnow contains policy objects with actions (omitted actions still meanallow). -
Added remote rule sources with explicit download routing and process-owned refresh. All runtime sources use retained in-memory rules; an optional
rule_cache_diradds best-effort persistence with atomic file replacement and generated filenames whenpathis omitted. Startup creates missing cache directories and their parents; existing directories may be owned by another user and accessed through group permissions or ACLs, but must not be world-writable. Reloads reuse compiled sources. Startup fetches missing or invalid cached content through supported configured bootstrap routes without fallback. Cache-save failures warn without rejecting valid rules; unsuccessful saves may require another download after restart.checkand--dry-runremain offline. Sources can be composed undermasksor written flat alongside a policy'saction. See remote rule sources. -
Added opt-in WebSocket early-data sending through path suffixes or headers, including sing-box and Xray interoperability. Default servers automatically accept up to 2048 decoded bytes from
Sec-WebSocket-Protocol; static subprotocol selectors suppress automatic reception. Other nonempty subprotocol tokens now become decoded bytes or strict errors, not ordinary negotiation. Early-data acceptance is not negotiated: configure matching carriers, since mismatched custom headers or paths can silently lose the prefix despite a successful upgrade. Use sing-box 1.14.3 or newer for client-side early data to avoid losing initial response bytes. WebSocket handshakes now enforce a 32 KiB header-block limit, requireUpgrade/Connectionupgrade response tokens, and reject malformed response status codes and duplicate accept/subprotocol response headers, including on plain upgrades. See WebSocket compatibility and configuration. -
Added default-on AnyTLS client session reuse for TCP and UDP-over-TCP, with
idle_session_check_interval_secs,idle_session_timeout_secs,min_idle_session, anddisable_reusecontrols. Warm requests skip carrier and wrapper handshakes. After a clean peer FIN, late upload bytes are discarded so buffered responses can drain without affecting the next lease. See AnyTLS client configuration. -
Added the outermost
sudoku_maskTCP client/server wrapper with Xray FinalMask interoperability, stacked outer masks, TLS/REALITY, WebSocket, native VLESS encryption/Vision, and XHTTP carriers. Masks inside other protocols are rejected. UDP forwarded inside TCP remains supported; native UDP/QUIC masking is not supported. Sudoku is obfuscation, not transport security; see configuration and costs. -
Any effective manual selector change now clears all URL-test samples and incumbents and schedules fresh checks, including for unrelated idle lazy groups; active rounds finish first and obsolete in-flight results remain fenced.
-
Changed automatic health checks to start each member's timeout after probe admission, so rounds can exceed their interval under load. Manual API delay requests retain their request-wide deadlines.
-
Client groups now accept
client_chainandclient_chainsas aliases with the same syntax as routes:[a, b]is sequential;[{chain: [a]}, {chain: [b]}]declares alternatives. Groups preserve whole-path choices of different lengths across routing, DNS and handshake/carrier chains. Legacyclient_proxy/client_proxiesretain alternative-member behavior with deprecation warnings. Round-robin weights immediate alternatives instead of flattened leaves. -
Fixed NaiveProxy session reuse across different selected prefixes. Each handler caches up to 16 per-prefix sessions, pruning closed sessions before least-recently-used eviction without interrupting existing tunnels.
-
Breaking: DNS
client_chain: [a, b]now means sequential hops, matching routes, rather than alternatives. DNS also accepts theclient_chainsalias. Use explicit chain wrappers or a named group for alternatives; no legacy DNS interpretation is retained. -
Fixed lost TLS writer wakeups and Vision framing with small TLS buffers and deferred KeyUpdate replies. Fixed certificate verification for TLS/QUIC hops without SNI, including IP-address hops, to use the configured hop identity instead of a placeholder name. Kept verification identity separate from wire SNI for protocols nested inside XHTTP.
-
Hardened protocol frame lengths, padding, WebSocket upgrade validation, and terminal proxy failure responses. Corrected Vision/XUDP handling and per-destination UDP routing, and preserved bounded TUIC traffic pending authentication.
-
Bounded HTTP/3 buffered frames and incremental unknown-frame skipping through the maintained h3 fork, including DNS-over-HTTP/3. The fork defaults decoded headers to 64 KiB; XHTTP continues to set its own configured header limit.
-
Fixed TCP/DoT DNS queries stalling behind bursts of unsolicited or stale responses through a Hickory receive-wakeup backport. Buffered replies keep progressing without extra retries or longer deadlines.
-
Fixed DNS resolver refresh after transport failures, stale macOS config reload notifications, and duplicated group ancestry in Clash connection chains.
-
Refreshed the smoltcp TCP recovery fork used by TUN and WireGuard, retaining retransmission ACK sequencing and incorporating upstream timer and FIN recovery fixes.
-
Published the configuration documentation site with English and Simplified Chinese references and tracked examples.
-
Raised the default
global_settings.quic_socket_bufferrequest to 8 MiB and extended the setting from QUIC listeners to outbound QUIC proxies and DNS-over-HTTP/3. Buffer tuning remains best-effort; unsatisfied or unverifiable requests now warn at most once per receive/send direction per process, including SOCKS5 UDP's existing 2 MiB request. -
Added Linux ARMv7 hard-float GNU and static musl release builds, with QEMU smoke checks for the packaged binaries.
-
Added a static
mips-unknown-linux-muslCLI release build for big-endian MIPS32r2, o32, soft-float CPUs with 4 KiB kernel pages. Shoes and the pinned upstream GotaTun revision use portable 64-bit atomics where native support is unavailable. This Tier 3 build uses Rust 1.99.1rust-src/build-stdand the pinned OpenWrt GCC 13.3.0/musl 1.2.5 toolchain. Validation uses QEMU user mode, not physical routers or an OpenWrt kernel; MIPS TUN and independent-peer WireGuard operation have not been validated. Little-endian MIPS, MIPS64, pre-R2/R6 CPUs, MIPS shared libraries and MIPS container images are not included. The executable is approximately 32 MiB; check available storage. -
Hardened binary releases with pinned build tooling and cross images, native macOS smoke checks, tag/version validation, SHA-256 checksums, and GitHub build-provenance attestations. macOS binaries are not yet Developer ID signed or notarized.
-
Reduced release binary dependencies without removing supported protocols. Linux firewall schema validation is test-only, Quinn's unused platform verifier is disabled, and release tracing is capped at info level.
This release is distributed through GitHub Releases and GHCR, not crates.io. Building from source requires Rust 1.99.1 or newer, with the additional Tier 3 build requirements above for MIPS; cargo install shoes still installs the previously published crates.io version. Nine Linux/macOS tarballs include LICENSE, THIRD_PARTY_NOTICES.md and licenses/, which distributors must retain along with the corresponding source information. Container images support linux/amd64 and linux/arm64; native mobile interfaces remain source-built, with no AAR or XCFramework archives.
Full changelog: v0.3.6...v0.3.7