Support Client ID Metadata Documents (CIMD) as a registration mode

.github/workflows/deploy-native.yml

@@ -245,6 +245,11 @@ jobs:
           # Unset -> endpoint 404s. An optional secret that neither caller
           # passes arrives here as the empty string.
           OPENAI_APPS_CHALLENGE_TOKEN: ${{ secrets.OPENAI_APPS_CHALLENGE_TOKEN }}
+          # Client ID Metadata Documents: advertised only where this GitHub
+          # Environment defines the OAUTH_CIMD_ENABLED variable as `1` (this job
+          # runs in that environment, so its variables resolve here). Unset
+          # arrives as the empty string, which the server reads as off.
+          OAUTH_CIMD_ENABLED: ${{ vars.OAUTH_CIMD_ENABLED }}
         run: deploy/native/deploy.sh
 
       # Confirm the host is actually running the commit we just shipped, rather

Cargo.toml

@@ -44,6 +44,7 @@ dist = false
 # codec, and candid lines can never drift between imcp2-core, imcp2, and
 # imcp2-local.
 [workspace.dependencies]
+httpdate = "1.0.3"
 imcp2-core = { path = "crates/imcp2-core", version = "0.4.0" }
 rmcp = "1.7"
 axum = "0.8"

README.md

@@ -692,6 +692,43 @@ its AS issuer is `<PUBLIC_URL>/mcp` and everything OAuth lives under it:
   honoured (intersected with `authorization_code`). A **hosted** `redirect_uri` is
   rejected unless its host is on the allow-list (see the Companion-control note
   below); loopback redirects are always accepted.
+- **Client ID Metadata Documents** — the MCP authorization spec's preferred
+  registration, advertised as `client_id_metadata_document_supported: true`. A
+  client may skip `/register` and use the https URL of its metadata document as
+  its `client_id`; `/mcp/oauth/authorize` fetches that document under the same
+  SSRF guard as app discovery (https only, public addresses only, pinned, no
+  redirects, 8 KiB cap, 5 s including DNS), requires its `client_id` to equal the URL, and checks the requested
+  redirect against the document's `redirect_uris` exactly as it would a DCR
+  registration's — hosted-redirect allow-list included, and checked before any
+  fetch, so a document can neither admit a redirect a DCR client couldn't
+  register nor make the server fetch a URL for a redirect it would refuse. A
+  hosted redirect must also be same-origin with the document URL (loopback
+  excepted), so a self-asserted document cannot point the code at another
+  party. Only clients that can authenticate as PUBLIC clients are accepted: a
+  `token_endpoint_auth_method` of `none` (or none given), or `none` among the
+  document's `token_endpoint_auth_methods_supported` — ChatGPT's case, since it
+  prefers `private_key_jwt` but lists `none`, which is what it uses here.
+  Documents must be served as `application/json`, and are cached (bounded; the
+  origin's remaining freshness — `max-age` less `Age` — honoured up to 24 h,
+  10 min when it sends none, not at all on `no-store`), so a directory client
+  connecting thousands of times mints no registrations. Concurrent requests for
+  one document share a single fetch, and at most eight fetches are in flight at
+  once per process (however many instances the binary mounts), two per host, so
+  one slow host cannot hold up the rest — and no more than sixty a minute per
+  process, thirty per vendor domain, so an origin that answers at once cannot be
+  made to answer without end. The fetch connects directly, never
+  through a proxy from the environment, so the address pin always binds. Claude and ChatGPT both select CIMD over DCR when it is
+  advertised — which it is only where `OAUTH_CIMD_ENABLED=1` is set (the deploy
+  template takes it from the GitHub Environment's variable of that name, so a
+  deploy never enables it by itself; to roll back, unset it and redeploy — the
+  value is read once at start-up, so the variable alone changes nothing — and
+  clients re-read the metadata within minutes and fall back to DCR). Only a
+  document on a vetted vendor origin is fetched at all — a host on or under an
+  allow-listed domain, default port (the trust policy of PR #143); any other URL
+  `client_id` is refused before any request and pointed at the allow-listing
+  contact. A document-intrinsic failure (no document there, not JSON, about
+  another URL) is remembered for a minute so a repeat is cheap; a transient one
+  is not.
 
 - `GET  /mcp/oauth/authorize` — validates the client + redirect, requires PKCE, sets
   the binding cookie, then redirects to II's handshake (with `registration_key`)

crates/imcp2-core/Cargo.toml

@@ -18,7 +18,7 @@ rmcp = { workspace = true, features = ["server", "macros"] }
 ic-agent = { workspace = true }
 candid = { workspace = true }
 candid_parser = { workspace = true }
-tokio = { workspace = true, features = ["macros", "rt-multi-thread", "sync", "net"] }
+tokio = { workspace = true, features = ["macros", "rt-multi-thread", "sync", "net", "time"] }
 serde = { workspace = true }
 serde_json = { workspace = true }
 tracing = { workspace = true }
@@ -31,6 +31,7 @@ reqwest = { workspace = true }
 getrandom = { workspace = true }
 urlencoding = { workspace = true }
 url = { workspace = true }
+httpdate = { workspace = true }
 # Only for the end-to-end canister-tool tests (src/e2e_canister_tools.rs),
 # behind the `e2e` feature so the default build compiles neither. `pocket-ic`
 # runs a real replica locally (needs the POCKET_IC_BIN server binary at
@@ -48,3 +49,8 @@ wat = { version = "1", optional = true }
 # `Identities::seed_app_identity` — are gated on `cfg(test)` as well, so they
 # exist only in this crate's own test binary, never in a library build.
 e2e = ["dep:pocket-ic", "dep:wat", "rmcp/client"]
+
+[dev-dependencies]
+# Synthetic responses for pinning `public_fetch`'s acceptance rules without a
+# network (`reqwest::Response: From<http::Response<_>>`).
+http = { workspace = true }

crates/imcp2-core/src/discover.rs

@@ -1350,6 +1350,9 @@ fn ipv4_is_global(ip: &Ipv4Addr) -> bool {
         || (o[0] == 100 && (o[1] & 0xc0) == 64) // 100.64.0.0/10 CGNAT (shared)
         || (o[0] == 192 && o[1] == 0 && o[2] == 0) // 192.0.0.0/24 IETF protocol
         || (o[0] == 198 && (o[1] & 0xfe) == 18) // 198.18.0.0/15 benchmarking
+        // 192.88.99.0/24, the deprecated 6to4 relay anycast block (RFC 7526), is not
+        // globally reachable — bar 192.88.99.2, the 6a44 relay anycast (RFC 6751).
+        || (o[0] == 192 && o[1] == 88 && o[2] == 99 && o[3] != 2)
         || o[0] >= 240) // 240.0.0.0/4 reserved
 }
 
@@ -1362,50 +1365,106 @@ fn ipv6_is_global(ip: &Ipv6Addr) -> bool {
         return ipv4_is_global(&v4);
     }
     let seg = ip.segments();
+    // 2001::/23 is IETF protocol assignments (RFC 2928): NOT globally reachable
+    // by default — Teredo (2001::/32), benchmarking (2001:2::/48), ORCHID and
+    // ORCHIDv2 (2001:10::/28, 2001:20::/28), and everything unassigned — with
+    // the IANA registry's globally reachable exceptions admitted by name, so a
+    // new assignment is refused until audited rather than accepted until noticed.
+    if seg[0] == 0x2001 && (seg[1] & 0xfe00) == 0 {
+        return ietf_protocol_assignment_is_global(&seg);
+    }
     !(ip.is_unspecified()                    // ::
         || ip.is_loopback()                  // ::1
         || ip.is_multicast()                 // ff00::/8
         || (seg[0] & 0xfe00) == 0xfc00       // fc00::/7 unique-local
         || (seg[0] & 0xffc0) == 0xfe80       // fe80::/10 link-local unicast
+        || (seg[0] & 0xffc0) == 0xfec0       // fec0::/10 site-local (deprecated, RFC 3879)
+        || (seg[0] == 0x0100 && seg[1] == 0 && seg[2] == 0 && seg[3] == 0) // 100::/64 discard-only (RFC 6666)
+        || (seg[0] == 0x3fff && (seg[1] & 0xf000) == 0)          // 3fff::/20 documentation (RFC 9637)
+        || seg[0] == 0x5f00                                      // 5f00::/16 SRv6 SIDs, not globally reachable (RFC 9602)
         || (seg[0] == 0x2001 && seg[1] == 0x0db8) // 2001:db8::/32 documentation
         // Transition mechanisms embed an IPv4 address deeper in the v6 space than
-        // `to_ipv4` decodes, so a NAT64/6to4/Teredo host would otherwise translate
-        // one of these to loopback/link-local/RFC1918/metadata (ICPBB-377). imcp2
-        // never needs to reach them, so refuse the prefixes outright.
+        // `to_ipv4` decodes, so a NAT64/6to4 host would otherwise translate one
+        // of these to loopback/link-local/RFC1918/metadata (ICPBB-377); Teredo is
+        // refused with the rest of 2001::/23 above. imcp2 never needs to reach
+        // them, so refuse the prefixes outright.
         || (seg[0] == 0x0064 && seg[1] == 0xff9b)  // 64:ff9b::/32 NAT64 (RFC 6052 WKP + RFC 8215 local-use)
-        || seg[0] == 0x2002                        // 2002::/16 6to4
-        || (seg[0] == 0x2001 && seg[1] == 0x0000)) // 2001::/32 Teredo
+        || seg[0] == 0x2002) // 2002::/16 6to4
+}
+
+/// The globally reachable exceptions inside `2001::/23` (IETF protocol
+/// assignments), per the IANA IPv6 Special-Purpose Address Registry. Everything
+/// else in the block — assigned to a non-routable use or not assigned at all —
+/// is refused.
+fn ietf_protocol_assignment_is_global(seg: &[u16; 8]) -> bool {
+    let anycast = seg[1] == 0x0001 && seg[2..7] == [0; 5] && (1..=3).contains(&seg[7]);
+    anycast // 2001:1::1 PCP (RFC 7723), 2001:1::2 TURN (RFC 8155), 2001:1::3 DNS-SD SRP (RFC 9665)
+        || seg[1] == 0x0003                       // 2001:3::/32 AMT (RFC 7450)
+        || (seg[1] == 0x0004 && seg[2] == 0x0112) // 2001:4:112::/48 AS112-v6 (RFC 7535)
+        || (seg[1] & 0xfff0) == 0x0030 // 2001:30::/28 Drone Remote ID (RFC 9374)
+}
+
+/// Why [`resolve_public_url`] returned no addresses: the URL itself is refused —
+/// it does not parse, is not https, names no host, or resolves to a non-public
+/// address (the SSRF guard) — or its host could not be resolved RIGHT NOW, which
+/// says nothing about the URL. A caller that remembers refusals (the OAuth
+/// server's client-metadata cache) must not remember a resolver outage as one.
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub(crate) enum ResolveError {
+    Refused(String),
+    Unresolved(String),
+}
+
+impl std::fmt::Display for ResolveError {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            Self::Refused(why) | Self::Unresolved(why) => f.write_str(why),
+        }
+    }
+}
+
+/// The discovery crawl reports every failure as a message, and tells these apart
+/// no further; `?` keeps working there.
+impl From<ResolveError> for String {
+    fn from(err: ResolveError) -> Self {
+        match err {
+            ResolveError::Refused(why) | ResolveError::Unresolved(why) => why,
+        }
+    }
 }
 
 /// Validate a user-supplied discovery URL against SSRF and return the parsed URL
 /// plus the socket addresses to PIN the client to. https only; every resolved
 /// address must be global. Async DNS (no blocking of the executor).
-async fn resolve_public_url(raw: &str) -> Result<(url::Url, Vec<SocketAddr>), String> {
-    let url = url::Url::parse(raw).map_err(|e| format!("invalid discovery URL {raw}: {e}"))?;
+pub(crate) async fn resolve_public_url(
+    raw: &str,
+) -> Result<(url::Url, Vec<SocketAddr>), ResolveError> {
+    let url = url::Url::parse(raw)
+        .map_err(|e| ResolveError::Refused(format!("invalid discovery URL {raw}: {e}")))?;
     if url.scheme() != "https" {
-        return Err(format!(
+        return Err(ResolveError::Refused(format!(
             "refusing to fetch {raw}: only https:// discovery targets are allowed (SSRF guard)"
-        ));
+        )));
     }
     let port = url.port_or_known_default().unwrap_or(443);
     let addrs: Vec<SocketAddr> = match url.host() {
         Some(url::Host::Ipv4(v4)) => vec![SocketAddr::new(IpAddr::V4(v4), port)],
         Some(url::Host::Ipv6(v6)) => vec![SocketAddr::new(IpAddr::V6(v6), port)],
         Some(url::Host::Domain(host)) => tokio::net::lookup_host((host, port))
             .await
-            .map_err(|e| format!("could not resolve {host}: {e}"))?
+            .map_err(|e| ResolveError::Unresolved(format!("could not resolve {host}: {e}")))?
             .collect(),
-        None => return Err(format!("refusing to fetch {raw}: no host")),
+        None => return Err(ResolveError::Refused(format!("refusing to fetch {raw}: no host"))),
     };
     if addrs.is_empty() {
-        return Err(format!("refusing to fetch {raw}: host did not resolve"));
+        return Err(ResolveError::Unresolved(format!("could not resolve {raw}: no addresses")));
     }
     if let Some(bad) = addrs.iter().find(|a| !ip_is_global(&a.ip())) {
-        return Err(format!(
+        return Err(ResolveError::Refused(format!(
             "refusing to fetch {raw}: it resolves to a non-public address ({}) — discovery is \
              restricted to public hosts (SSRF guard)",
             bad.ip()
-        ));
+        )));
     }
     Ok((url, addrs))
 }
@@ -1536,10 +1595,26 @@ enum Overflow {
 /// The shared read. `Err((partial, error))` carries what had arrived before the
 /// transfer failed, so the fail-soft caller can keep it and the strict one can
 /// report the failure.
-async fn read_capped_inner(
-    mut resp: reqwest::Response,
+pub(crate) async fn read_capped_inner(
+    resp: reqwest::Response,
     max: usize,
 ) -> Result<String, (String, String)> {
+    // Lossy by design for the crawl: a stray byte must not cost a whole bundle.
+    // A caller that treats the body as a statement (the CIMD fetch) reads the
+    // bytes and decodes strictly instead.
+    match read_capped_bytes(resp, max).await {
+        Ok(buf) => Ok(String::from_utf8_lossy(&buf).into_owned()),
+        Err((buf, e)) => Err((String::from_utf8_lossy(&buf).into_owned(), e)),
+    }
+}
+
+/// The shared read in bytes: up to `max` of the body, stopping — and so dropping
+/// the response and its connection — at the cap. `Err((partial, error))` carries
+/// what had arrived before the transfer failed.
+pub(crate) async fn read_capped_bytes(
+    mut resp: reqwest::Response,
+    max: usize,
+) -> Result<Vec<u8>, (Vec<u8>, String)> {
     let mut buf: Vec<u8> = Vec::new();
     loop {
         if buf.len() >= max {
@@ -1554,10 +1629,10 @@ async fn read_capped_inner(
                 }
             }
             Ok(None) => break,
-            Err(e) => return Err((String::from_utf8_lossy(&buf).into_owned(), e.to_string())),
+            Err(e) => return Err((buf, e.to_string())),
         }
     }
-    Ok(String::from_utf8_lossy(&buf).into_owned())
+    Ok(buf)
 }
 
 /// GET `url` and return up to `max` bytes of its body, distinguishing "this app
@@ -2794,6 +2869,7 @@ mod tests {
         // Publicly-routable addresses.
         assert!(g("8.8.8.8"));
         assert!(g("1.1.1.1"));
+        assert!(g("192.88.99.2"), "the 6a44 relay anycast is the reachable exception in its /24");
         assert!(g("2606:4700:4700::1111"));
         // Loopback / private / link-local / CGNAT / reserved / doc / bench, plus
         // IPv4 embedded in IPv6 as MAPPED (::ffff:…) and COMPATIBLE (::…) forms.
@@ -2808,12 +2884,23 @@ mod tests {
             "255.255.255.255",
             "192.0.2.1",
             "198.18.0.1",
+            "192.88.99.1", // 6to4 relay anycast, deprecated
             "240.0.0.1",
             "::1",
             "::",
             "fc00::1",
             "fd12::1",
             "fe80::1",
+            "fec0::1",     // site-local: deprecated, still routable on legacy networks
+            "100::1",      // discard-only
+            "2001:2::1",   // benchmarking
+            "2001:10::1",  // ORCHID (deprecated)
+            "2001:20::1",  // ORCHIDv2 (not routable)
+            "2001:1::4",   // unassigned inside 2001::/23 (IETF protocol assignments)
+            "2001:5::1",   // likewise
+            "2001:1ff::1", // the block's last /32, likewise
+            "3fff::1",     // documentation (RFC 9637)
+            "5f00::1",     // SRv6 SIDs (RFC 9602)
             "2001:db8::1",
             "::ffff:127.0.0.1",
             "::ffff:10.0.0.1", // IPv4-mapped private/loopback
@@ -2833,6 +2920,13 @@ mod tests {
         }
         // A real public v6 that merely starts with 0x2001 (not db8/Teredo) stays global.
         assert!(g("2001:4860:4860::8888"));
+        assert!(g("2001:200::1"), "just past 2001::/23");
+        // The globally reachable exceptions inside 2001::/23 stay global.
+        for good in
+            ["2001:1::1", "2001:1::2", "2001:1::3", "2001:3::1", "2001:4:112::1", "2001:30::1"]
+        {
+            assert!(g(good), "{good} is globally reachable per the IANA registry");
+        }
     }
 
     // The exact vectors from the finding, plus https-to-internal, are refused

crates/imcp2-core/src/lib.rs

@@ -32,6 +32,7 @@
 
 pub mod identities;
 pub mod iiconnect;
+pub mod public_fetch;
 pub mod skills;
 pub mod tools;