#sipkit

RFC 4475 torture corpus   49 messages extracted mechanically from the RFC text
                          45 verdicts asserted, 4 role-dependent (not asserted)
                          all 45 asserted verdicts match
                          every accepted message round-trips byte-for-byte *
Robustness                >4,000 generated inputs (random, truncated, mutated,
                          spliced) with zero panics
Coverage                  50 uncovered lines in the library (about 97% of 1,698
                          executable lines), measured with `moon coverage analyze`
Tests                     79, green on wasm · wasm-gc · js · native
Dependencies              none
Panics in the library     none — no `unwrap`, `panic` or `abort` outside tests
API documentation         112 of 112 public functions carry a doc comment

#Why this exists

#What it does

#What it deliberately does not do

• No transaction or dialog layer. No INVITE retransmission, no ACK matching, no Route set computation. RFC 4475 §3.2.1 is therefore checked as a diagnostic (branch identifier present), not as transaction behaviour.
• No SDP. Content-Type: application/sdp bodies are returned as opaque octets. coke089/moon-sdp is the natural companion.
• No authentication. Authorization / WWW-Authenticate are parsed as text; no Digest computation, no credential storage.
• No TLS/DTLS transport. moonbitstack/moontls owns that layer.
• No extension support. Require and Proxy-Require with any value are reported as unsupported, because the library implements none.
• Role-dependent verdicts are not asserted. Where RFC 4475 says "a proxy forwards but an endpoint replies 415", the library refuses to guess the caller's role.

#Conformance evidence

python3 tools/extract_rfc4475.py --rfc rfc4475.txt --out-dir .
python3 tools/gen_tests.py
moon test

• lwsdisp (§3.1.1.6) — From: caller<sip:caller@example.com> has no LWS before <. The grammar forbids it, but the RFC identifies this as a specification bug and requires acceptance. sipkit accepts it.
• esc02 (§3.1.1.5) — % is not an escape everywhere. C%6Fntact is not Contact, and branch=z9hG4bK209%fzsnel234 is well formed. sipkit compares header names literally and accepts a stray %.

#Robustness

#Quick start

moon add BenPaoBaMingWangXing/sipkit

let input : Bytes =
  b"INVITE sip:bob@example.com SIP/2.0\r\nTo: <sip:bob@example.com>\r\nFrom: <sip:alice@example.com>;tag=1\r\nCall-ID: a1\r\nCSeq: 1 INVITE\r\nVia: SIP/2.0/UDP host;branch=z9hG4bK1\r\nMax-Forwards: 70\r\nContent-Length: 0\r\n\r\n"

let a = analyze(input)
println(a.verdict())            // well-formed
let msg = a.message().unwrap()
println(ascii_of(msg.verb().unwrap().view()))   // INVITE
assert_eq(to_bytes(msg).length(), input.length())

match analyze(b"garbage").syntax() {
  Some(pair) => println("rejected: \{pair.0} at byte \{pair.1}")
  None => ()
}

#Diagnostics

for d in analyze(input).diagnostics() {
  println("\{d}")   // error:content-length-mismatch@214 declared 152 but body is 0
}

#Examples

moon run examples/parse_dump    # parse an INVITE, print every recovered field
moon run examples/diagnose      # the three verdicts, side by side
moon run examples/uri           # URI decomposition, including the awkward cases

start line
  method     INVITE
  uri.scheme sip
  uri.user   bob
  uri.host   example.com
  uri.port   5060
  uri.params 1
    ;transport=tcp
headers (9), in wire order
  line 2  Via: SIP/2.0/UDP alice-pc.example.com:5060;branch=z9hG4bK776asdhds
  ...
round-trip identical: true

--- semantic-rejected
    verdict: semantic-rejected
    [error] content-length-mismatch at byte 236
             declared 42 but body is 5
    [error] cseq-method-mismatch at byte 120
             CSeq method INVITE != start line OPTIONS

--- well-formed, with a warning
    verdict: well-formed
    [warning] missing-required-header at byte 0
             request has no Max-Forwards header

--- sip:al%41ice@example.com
    user     al%41ice
    user*    alAice   (decoded once)

--- sip:al%2541ice@example.com
    user*    al%41ice   (decoded once)      <- not `alAice`: decoding is single-pass

--- sip:user;par=u%40example.net@example.com
    user     user;par=u%40example.net
    user*    user;par=u@example.net         <- the ; belongs to the user part

#Design decisions

#Ecosystem position

#Data sources and licensing

• Source code: Apache-2.0 (see LICENSE).
• RFC 3261 / RFC 4475 octets reproduced in generated fixtures are used under the IETF Trust Legal Provisions (see NOTICE). No RFC text is bundled: rfc4475.txt is fetched by the reproduction steps and is deliberately absent from the repository.
• Generated files (rfc4475_test.mbt, rfc4475_vectors.json) contain only message octets and section numbers; provenance is recorded in NOTICE.

#Reproduce

python3 tools/extract_rfc4475.py --rfc rfc4475.txt --out-dir .   # needs rfc4475.txt
python3 tools/gen_tests.py
moon check
moon fmt --check
moon test --target wasm
moon test --target wasm-gc
moon test --target js
moon coverage analyze        # line-level coverage; ~50 uncovered lines in the library

#Known limitations

• → full §25.1 SIP-date since 2026-10-27: fixed 29-octet shape, weekday/month whitelists, day-in-month with leap years, HH≤23/MM≤59/SS≤60 (60 as leap second), and weekday/date agreement via Zeller's congruence — Mon, 01 Jan 2010 is rejected on the weekday although every token looks plausible.

1. : in the user part. RFC 4475 §3.1.1.2's message puts : inside the user part; RFC 3261 §25.1's user-unreserved does not include it (: is the user/password separator). The ABNF wins: the user ends at the first colon, the rest is the password.
2. Parameter namespace on a bare addr-spec. Contact: sip:a@b;unknownparam without angle brackets makes ;-parameters both URI parameters and contact parameters, and the grammar cannot separate them. RFC 3261 §20.10 admits this and recommends always using <>. The parameter stays reachable and user/host are unaffected.
3. mpart01 sample data. The RFC prints Content-Length: 553 for a hex body that is 14 octets longer. Framing by Content-Length (the correct behaviour) leaves those 14 octets as trailing.