Hello everyone.
Part one was about what Chute reads from Surge, Shadowrocket, Clash/mihomo and sing-box, and what it refuses to pretend to read. Part two was about the big gaps that closed: proxy chains, random groups, SUB-RULE, plain DNS over TCP.
This one is about imports that did not fail.
They returned a profile. The profile loaded. The traffic went somewhere other than where the line said.
An import has two jobs, and the second is harder to see: read the line, and keep its meaning. A line read with the wrong meaning still produces a policy, still connects, and still looks fine in a list. Every failure in this post has that shape.
A Name Is Also Punctuation
A policy line is a name, an equals sign, and fields separated by commas. So a node named US=1 is not a node named US=1; it is a line the parser refuses. A node named Fast=direct, x is worse: it comes back as a policy named Fast, of type DIRECT. The node is gone, and a policy that sends traffic around the proxy has taken its place.
Provider nodes are where these names come from. A subscription is written by someone else, with someone else’s idea of what a name can contain.
Every import path now shares one naming rule: a comma becomes a space, an equals sign becomes a hyphen, and a name already taken gets a counter. All names are registered before any line is written, so two nodes cannot quietly become one, and a chain that named a renamed sibling follows it.
The visible consequence is a rename. HK, 01 becomes HK 01 — two spaces, because the comma was a separator everywhere else. A selection persisted under the old name is lost once, and a filter’s regular expression meets the rewritten name. That is the honest cost of a name that cannot change the shape of a line.
Spaces That Were Part of the Value
AmneziaWG describes its camouflage as a chain of tags, and the tags are separated by spaces: <b 0x…> <r 16>. A policy line drops spaces outside quotes, so the same chain written without quotes comes back as <b0x…><r16> — one token that no tag parser can split.
Three writers had the same bug from three directions: the wg-quick import wrote the chain into its section, the mihomo conversion quoted it only when it contained a comma, and the section exporter wrote what the editor had set. All three quote it now — and only the chain keys, because every other section key is read without quotes.
The same section had a smaller version of the same problem. A WebSocket, XHTTP or v2ray-plugin header given as a list became an empty value, while the server name taken from the same Host used the first entry. Headers now take the first entry too. And mihomo refuses a whole profile over a group use or proxies that is not a list; we used to ignore it without a word. Now it is reported.
ShadowTLS Under Shadowsocks
ShadowTLS is usually a front for Shadowsocks: the client speaks Shadowsocks, wrapped in a TLS handshake that belongs to someone else’s site. Two import paths got the wrapping wrong, and neither said so.
mihomo’s plugin: shadow-tls on a Shadowsocks node was reported as unsupported, and the node was written bare. The stream reached the ShadowTLS port unwrapped, which is not a session the server can read. sing-box’s version was subtler: a Shadowsocks outbound with a detour to a ShadowTLS outbound became an underlying-proxy chain. A chain between policies carries SOCKS5, not the Shadowsocks stream the server is waiting for. And sing-box’s own example — server and port on the ShadowTLS outbound only — lost the Shadowsocks outbound entirely, as missing a server.
Now mihomo’s plugin options map to the Shadowsocks line’s own ShadowTLS keys — password, cover name, version — with the default v2 converted to v3 and said out loud, skip-cert-verify kept, and fingerprint and ALPN reported if they cannot cross. sing-box’s Shadowsocks line dials the ShadowTLS outbound’s server and port, and takes its password, TLS name, fingerprint and insecure flag; the ShadowTLS outbound’s own detour becomes the line’s upstream.
The three paths — mihomo, an ss:// link with SIP003 plugin options, and sing-box — were checked against a real ShadowTLS v3 server fronting both aes-128-gcm and 2022-blake3-aes-128-gcm. All three connect.
The ECH That Went Out in the Clear
Encrypted Client Hello hides the name of the site you are visiting inside a second, encrypted handshake. The client needs the server’s ECH configuration before it can do that. A share link can carry one in an ech= value; a profile can point at a DNS name where it can be fetched.
A share link’s value was written into the policy exactly as it came. The handshake reads standard base64; a value in URL-safe base64, or one without padding, did not parse. The socket then asked DNS for the configuration, and when DNS had none it sent GREASE only — the real server name in the clear, with nothing said to the user. The import looked perfect.
The importer now decodes the value and writes it back as standard base64 with padding; a value that was already standard is left alone.
Two more ECH cases came out of the same review. A + is a normal character in standard base64, but it is also the separator in the name+https://dns lookup form, so a pinned configuration containing one was dropped and looked up over DNS instead. A value is now a configuration only if it decodes to a length-prefixed ECHConfigList. And mihomo’s query-server-name — the name the configuration is looked up under — was written as the public name sent in the clear, replacing the outer server name the server expects. It is now reported rather than written; sing-box’s equivalent had been ignored, and is now reported too.
Share Links That Meant the Opposite
A share link is a small language with a large number of dialects. Most of the work in this round was making the dialects mean what they meant where they came from.
- v2rayN’s vmess JSON writes
"tls":"none"for a node without TLS — x-ui and 3x-ui write it for every node without TLS — and we read it as a value, so TLS was turned on.insecureandallowInsecurewere ignored, and the XHTTP and gRPC modes v2rayN keeps intypewere lost. - SIP002 userinfo is percent-encoded. A SIP022 2022 key whose base64 ends in
=arrives as%3Dand was refused; a password writtenpa%5Fsswas imported with the%5Fstill in it. Userinfo is now percent-decoded before anything else, and base64 is tried only when there is no colon. - A TLS node without
sninow takes the firsthost— for gRPC, theauthority— the way v2rayN does. CDN-fronted nodes had been sending the IP, or no name at all. - Hysteria 2’s
pinSHA256and Xray’spcsare the certificate pin. Withinsecure=1and the pin never imported, there was nothing left to check, and any certificate was accepted. That one is not a compatibility bug; it is a security one, and it is fixed. - REALITY without
pbkis refused instead of being written as plain TLS to the decoy. - New forms:
ssr://, vmess’s URI form, Shadowrocket’s base64 vmess and vless. Read or reported:ech,spx, gRPCmode=multi, XHTTPextra, TUICallow_insecure, Xray’stype=http(HTTP/2) andheaderType=http,pqv, port-hopping ranges. A subscription now names the link kinds it skipped.
One structural fix sits under all of them. name = scheme, <link> — the form the Scheme and QR Scan editors write — had its own parser for six schemes and rejected everything else as an unsupported scheme. A vless:// link imported cleanly as a share link and failed the moment it was entered as a policy. Every share link now goes through the same reading, so the two ways in cannot disagree about what a link means.
WireGuard’s allowed-ips
wg-quick files now import as a policy and its section, with the AllowedIPs deciding the rules and a popup that says where those rules send traffic. mihomo’s amnezia-wg-option makes an AmneziaWG node.
A WireGuard peer’s allowed-ips had a quoting bug worth its own paragraph, because it was silent in the worst way. The line writer quotes a value when the text before the equals sign is an option key it knows, and quotes the whole field otherwise. allowed-ips was not in the table, so the field went out as allowed-ips="a, b". Read back, that is an option named 'allowed-ips: the peer’s list was never set, every destination went into the tunnel, and the unknown field was carried along and written again on every save. The key is in the table now.
What This Round Means
An import always writes a new profile; the old one stays where it is. If an import warned you off a few weeks ago, it is worth importing again — the line may now mean what it said.
The full reference for every mapping and every refusal is in the Chute Manual.
The common thread in all of this is not a parser bug. It is that a line can be read, accepted, and still mean something else. Compatibility is not the list of keys you understand; it is the meaning you preserve when the same key arrives in someone else’s dialect.
Thanks.
Chute Devs