Layer 7 · The DIG Node peer network

Canonical references: dig-gossip (the peer transport, discovery, and gossip layer — TLS-WebSocket peers, peer_id = SHA-256(TLS SPKI DER), address manager, introducer + peer-exchange), dig-relay (the rendezvous / hole-punch coordinator / circuit relay serving the RelayMessage wire), dig-nat (the connect(peer) NAT-traversal ladder), dig-dht (the Kademlia DHT with provider records that locate which peers hold content), and dig-constants (DIG_RELAY_URL). This layer is how DIG Nodes find and reach each other; the dig RPC is what they speak once connected.

This is the normative anchor for DIG Node ↔ Node communication. Every peer-facing crate — dig-nat, dig-relay, dig-gossip, dig-dht, and dig-node — conforms to the contracts below. Where a statement is a wire contract it is fixed at the field/byte level; a conforming reimplementation must reproduce it exactly.

The thesis: authenticated peers, direct when possible, relay only as a last resort

• Every peer link is mutually authenticated. All node↔node traffic runs over mutual-TLS (mTLS). A peer's identity is the hash of its TLS public key; there is no unauthenticated peer channel.
• Direct paths are preferred. A node tries, in a fixed order, to open a direct connection (already reachable → UPnP → NAT-PMP → PCP → relay-coordinated hole-punch) before it ever relays. When any direct path succeeds, no relay is used.
• The relay is the last resort, never the trust anchor. relay.dig.net bridges bytes only when every direct strategy fails. It forwards opaque, end-to-end-authenticated payloads by peer id and can read none of them.

The relay has four distinct roles

relay.dig.net is not a single service — it fills four separate roles, three of which are low-bandwidth signalling. Keeping them distinct matters because only the fourth carries a peer's data stream:

#	Role	Bandwidth	What the relay does	Where
1	STUN server	tiny	Answers a Binding request so a node learns its public reflexive IP:port (RFC 5389).	§3
2	Introducer	small	Registers a node's presence and returns known-peer lists for rendezvous/discovery.	§4a
3	Hole-punch signalling	small	Brokers a hole punch between two NAT'd peers — relays their candidate-address exchange and coordinates the simultaneous-open timing — after which the peers connect directly. The relay carries only the coordination messages, never the data.	§5
4	Relayed (TURN-like) transport	full	Proxies all of a peer connection's data when no direct path exists. High-bandwidth, last resort only.	§6

Roles 1–3 are how the relay helps two peers connect directly; role 4 is the only one where the peer's stream flows through the relay. A node always prefers role 3 (broker a direct link) over role 4 (proxy the whole stream), because role 3 costs the relay almost nothing while role 4 consumes real bandwidth. Whatever the tier, once connected the link is mTLS with peer_id = SHA-256(SPKI) (§1) — the relay is never the trust anchor.

0 · The two RPC tiers — mTLS peer/control vs anonymous public read

A DIG Node's RPC is served on two distinct tiers with two distinct authentication models, because two different callers need it: other nodes (which present a client certificate) and browsers/agents (which cannot present a client certificate but must still be able to READ content). One dig-node process serves both; the tier a method lives on is a frozen contract every serve-layer and every RPC client conforms to.

	PEER / CONTROL tier	PUBLIC READ tier
Purpose	node↔node: discovery, DHT, PEX, availability-for-sync, PUSH/WRITE, config/control	browser/agent content retrieval only
Auth	mutual TLS — client cert REQUIRED; peer_id = SHA-256(TLS SPKI DER) (§1); write routes additionally per-request BLS-signed (§21.9)	none — anonymous, no client cert, no bearer for reads
Transport	dig-nat mTLS mux (peer RPC + DHT + PEX streams) on the P2P port DIG_PEER_PORT (default 9444); the §21 authenticated HTTPS write routes	plain HTTPS JSON-RPC, CORS-enabled, on the public read listener (network-wide: rpc.dig.net)
Browser-reachable	No — an anonymous caller cannot open it	Yes — this is the browser/agent read path
Mutation / identity	write, peer, config, control all live here	read-only — no mutation, no peer/config method reachable
Integrity trust	mTLS peer identity	client-side self-verification (merkle inclusion proof + chain-anchored root pin) — the server is untrusted (Verification & provenance)
Miss behavior	method-specific error	decoy-on-miss, never 404 (blind host)

The boundary invariant

Two rules, jointly, are the contract:

1. No peer / write / config / control method is reachable without mTLS. Every method that mutates state, exchanges peer information, moves the DHT/PEX, or reconfigures the node is served only on the mTLS peer/control tier. On the anonymous public-read listener these methods do not exist — an anonymous caller that names one receives -32601 (method not found), the same as any unimplemented method. A node MUST NOT honor a write/peer/control method on a connection that did not complete the mTLS handshake.
2. Content read requires no mTLS. The read methods — dig.getContent and the read side of dig.getProof / dig.getCapsule / dig.getManifest / dig.getMetadata, plus the §21 GET routes (content / proof / roots / descriptor) — are served anonymously so a browser works. They are read-only: they never mutate, never reveal peer/config state, and never accept a write.

Which tier every method sits on is enumerated in §7a.

Why this is secure

Splitting authentication by tier does not weaken the network, because the public tier is read-only and client-verified:

• The read server is already untrusted. Every read is verified client-side against the CHIP-0035 chain-anchored root (dig.getContent streaming contract) — a merkle inclusion proof under the on-chain root plus per-chunk AES-256-GCM-SIV authentication. A malicious or anonymous read server can only serve bytes that either verify (correct content) or fail the proof/tag (detected, discarded). Anonymity of the reader changes nothing: the content is public ciphertext, keyed by retrieval_key = SHA-256(URN), and the AES key is derived client-side and never sent. There is nothing to authenticate a reader for.
• A miss is indistinguishable. A content miss returns a deterministic decoy, byte-shaped like a hit, never a 404 — so anonymous access leaks no presence/absence oracle.
• Everything that could harm the network stays mTLS-gated. Advancing a root (PUSH/WRITE), announcing a provider record, injecting peers, or reconfiguring the node all require the client certificate (and, for writes, a per-request BLS signature). An attacker with no certificate can read public ciphertext and nothing more — exactly the capability a browser needs and no more.
• The two never blur. Because the anonymous listener simply does not implement the peer/write/control methods, there is no "forgot to check auth" path: an unauthenticated write is not a rejected request, it is a nonexistent method.

How one process serves both

A dig-node runs two listeners:

• The mTLS peer/control listener binds the P2P port (DIG_PEER_PORT, default 9444) with rustls CERT_REQUIRED; it carries the dig-nat mux (peer RPC, DHT, PEX). The §21 authenticated write/push routes (transport & push) are the control tier's HTTPS face (per-request BLS auth, middleware order auth-THEN-rate-limit).
• The public read listener serves the anonymous JSON-RPC read subset over plain HTTPS with Access-Control-Allow-Origin: *, no credentials, OPTIONS → 204. On the public network this listener is rpc.dig.net. A dig-node run purely for a local consumer keeps this listener on loopback (127.0.0.1, default read port 9778); a node that wants to serve the wider network exposes the read listener on its public interface (TLS-fronted, as rpc.dig.net is).

The recommended endpoint split (design-first call): the public read tier is the same JSON-RPC endpoint shape as the network profile — one POST endpoint speaking JSON-RPC 2.0 — but the anonymous serve layer answers only the read subset; the peer/write/control methods return -32601 there. This reuses the existing node-profile / dig.methods gating (an agent already gates on dig.methods rather than assuming one uniform surface) instead of inventing a parallel protocol, and it is exactly what a browser already speaks to rpc.dig.net. The peer/control tier is not a JSON-RPC-over-HTTPS surface at all — it is the dig-nat mTLS mux — so it cannot be reached by an anonymous HTTP client even by accident.

rpc.dig.net IS the public read tier

Today rpc.dig.net (the dig RPC network profile) is exactly this anonymous, CORS-enabled, decoy-on-miss, client-verified public read tier. This section names the tier the browser has always used and states the invariant that keeps the peer/write/control surface off it.

Browser specifics — fetch + verify without mTLS

A browser (or an agent with no client certificate) reads content like this, needing no mTLS at any step:

1. CORS. The public read listener answers a cross-origin POST with Access-Control-Allow-Origin: * and no credentials (Access-Control-Allow-Credentials is never set — the content is public ciphertext, there is nothing to send credentials for), and answers the OPTIONS preflight with 204. So a page on any origin can fetch() the read endpoint directly.
2. Read. POST JSON-RPC dig.getContent (windowed by offset/length) to the public read endpoint; reassemble the ciphertext by total_length (streaming contract). No client certificate, no bearer token.
3. Verify — the server is untrusted. Verify the first window's inclusion_proof against the caller-supplied chain-anchored root (the CHIP-0035 singleton's current on-chain metadata.root_hash, resolved independently of the serving node), then split by chunk_lens and AES-256-GCM-SIV-open each chunk with the client-derived key. A wrong/forged byte fails the proof or the authentication tag and is rejected. This is the same read-crypto every DIG read client runs; the anonymity of the transport does not relax it.

Because the read tier is CORS-* + anonymous, the exact same fetch works from a web page, a service worker, the extension, the SDK, or a headless agent — none of which can present a client certificate.

1 · Peer identity + mTLS

A peer is identified by the SHA-256 of the DER-encoded SubjectPublicKeyInfo of the certificate it presents in the TLS handshake:

peer_id = SHA-256( SubjectPublicKeyInfo DER )        // 32 bytes

• peer_id is a Bytes32 — 32 raw bytes, rendered as 64 lower-case hex on any text surface (dig-gossip types/peer.rs: pub type PeerId = Bytes32, peer_id_from_tls_spki_der).
• The hashed input is the full SubjectPublicKeyInfo ASN.1 sequence (algorithm identifier + subject public-key bit string) lifted from the peer's X.509 leaf certificate — not the bare public-key bit string. Both sides recover the other's peer_id from the certificate exchanged during the handshake, so identity is bound to key material: impersonation requires the private key.

mTLS is mandatory

All peer-to-peer connections use mutual TLS over a WebSocket (wss://). This is a hard requirement — plaintext and server-only TLS are never accepted for a peer link.

• Both endpoints present a certificate. The dialing peer presents its cert to the listener; the listener presents its cert to the dialer. Each derives the other's peer_id from the presented cert.
• Self-signed node certificates are expected. Peer identity is verified by the peer_id hash, not by a certificate authority — the CA chain is not the trust root here, the key hash is. A node generates its certificate on first run and reuses it thereafter.
• The listener requires a client certificate (CERT_REQUIRED). A peer that presents no certificate, or whose TLS handshake fails, is dropped — there is no fallback to a weaker transport.
• After the TLS handshake, peers exchange a Handshake carrying the network_id (the network genesis challenge as lower-case hex), the protocol version, the node's declared listen port, and its node type + capabilities. A network_id mismatch, or a protocol version below the minimum-compatible floor, ends the connection. The Handshake is the Chia-streamable message (big-endian) — this layer speaks the Chia peer protocol, not a bespoke framing.
• Unauthenticated peer traffic is rejected. A message received before a completed mTLS handshake + Handshake exchange is not processed.

The relay link is authenticated differently

A node's link to the relay is a standard server-authenticated wss:// connection (the relay presents a TLS certificate; the node does not present one to the relay). Peer↔peer identity is never delegated to the relay — end-to-end payloads carried over the relay remain authenticated by the peer protocol itself, so a relay cannot forge a peer.

Identity rotation

A node MAY rotate its network identity (regenerate its certificate, hence its peer_id) on an interval to reduce long-term linkability. Rotation is a network-layer concern only — it is independent of any consensus/validator identity and does not disturb address-book entries, which are keyed by IP:port, not by peer_id.

2 · Connection establishment — the NAT-traversal ladder

A node reaches a peer through a single abstract operation — connect(peer) — which attempts the strategies below in order and returns the first that yields a working, mTLS-authenticated link. This ordered ladder is the contract dig-nat implements; every strategy above "relayed" produces a direct peer link (no relay in the data path).

#	Strategy	What it is	Relay's role	Result
a	DIRECT	The peer is already reachable — publicly routable, or a port is forwarded to it. Dial its advertised address.	none	Direct link
b	UPnP / IGD	Ask the local gateway (Internet Gateway Device) to map an external port to the node via UPnP, then advertise the mapped address.	none	Direct link
c	NAT-PMP	Request a port mapping from the gateway via NAT Port Mapping Protocol.	none	Direct link
d	PCP	Request a mapping via the Port Control Protocol (RFC 6887), NAT-PMP's successor.	none	Direct link
e	RELAY-COORDINATED HOLE PUNCH	Neither peer is directly reachable: the relay signals only — it relays the candidate-address exchange and coordinates a simultaneous open so both sides punch through their NATs. The data stream then flows peer-to-peer (see §5).	signalling only (low bandwidth)	Direct link
f	RELAYED / TURN transport	Every direct strategy failed: the relay proxies all of the connection's data as an untrusted bridge (see §6).	carries the data (high bandwidth)	Relayed link

Rules:

• Attempt in order; stop at the first success. A node does not skip ahead to the relay while an earlier, cheaper strategy can still yield a direct path.
• Prefer hole-punch signalling (e) over full relaying (f). Both involve the relay, but they are not the same tier: in (e) the relay only brokers the introduction and the stream goes peer-to-peer, whereas in (f) the relay carries every byte. A node falls to (f) only when the hole punch of (e) fails — this saves relay bandwidth, since a brokered direct link costs the relay almost nothing. A successful (e) is a direct link (strategy result "Direct"), authenticated by the same mTLS peer_id as every other tier.
• Strategies (b)–(d) run once at startup / on address change, not per connection: they establish inbound reachability so future dials land as (a) DIRECT for peers dialing this node. A node that obtains a stable external mapping via UPnP/NAT-PMP/PCP advertises it as a candidate address so peers dial it directly.
• Candidate addresses. A node advertises the set of addresses at which it may be reachable — its configured/observed listen address, any UPnP/NAT-PMP/PCP-mapped external address, and its STUN-derived reflexive address (§3). Peers dial candidates most-direct-first and IPv6-first (§2a).
• Reflexive discovery precedes hole-punch. Before requesting a hole-punch a node learns its public reflexive IP:port via STUN (§3) and supplies it as the external_addr in the coordination exchange.

2a · Address family — IPv6-first, IPv4-fallback

All peer communication is IPv6-first, with IPv4 used only as a fallback. This applies wherever a node binds, advertises, or dials a peer address.

• Bind — dual-stack. A node's peer listener binds the IPv6 unspecified address [::] as a dual-stack socket, so the single socket accepts both native IPv6 connections and IPv4 connections (via IPv4-mapped-IPv6) on the same port. It does not bind an IPv4-only wildcard.
• Advertise — IPv6 first. A node advertises its real, directly-dialable candidate addresses ordered IPv6-first: a routable (global-unicast) IPv6 address precedes the IPv4 fallback. The wildcard bind address is never advertised (it is not dialable). A node with no routable address advertises no direct candidate and is reached via the relay-coordinated tiers (§2e/f).
• Dial — happy-eyeballs, IPv6 preferred. When dialing a peer with several candidate addresses, a node tries the peer's IPv6 candidate(s) first and falls back to IPv4 only when IPv6 fails or times out. The full candidate list is preserved through the dial path so every family is tried in IPv6-first order — not collapsed to one address.

IPv4 remains a fully supported fallback (many networks are still IPv4-only); it is the fallback, never the default, wherever IPv6 is possible.

3 · STUN — reflexive address discovery

relay.dig.net also serves as a STUN server (RFC 5389 Binding request/response) so a node behind NAT learns the public IP:port its traffic appears to originate from — its reflexive address.

• Endpoint. The STUN Binding service is co-located with the relay at relay.dig.net, served on the standard STUN port 3478 (RFC 5389). A node sends a STUN Binding request and reads its reflexive transport address from the XOR-MAPPED-ADDRESS attribute of the Binding success response. A node derives the STUN host from its configured relay endpoint (DIG_RELAY_URL), so pointing a node at a private relay also points its STUN at that host.
• Feeds candidate advertisement. The reflexive address discovered here is added to the node's candidate-address set (§2) and supplied as the external_addr in a hole-punch request (§5). It is how a NAT'd node tells a peer where to punch to.
• Advisory, not authenticated. A reflexive address is a hint used to attempt a direct path; a peer link is trusted only after the mTLS handshake over it succeeds (§1). STUN never grants trust — it only tells a node where it appears to live.

4 · Peer discovery — introducer + gossip

A node fills its address book from two complementary sources. Both yield candidate peers to dial; neither is a trust anchor (every dialed peer is authenticated by mTLS).

4a · Introducer (via the relay)

relay.dig.net acts as an introducer: nodes connected to the same relay can enumerate each other, and a node can register itself so others discover it. This uses the relay RelayMessage wire (§6):

• Ask for peers. Send get_peers (optionally scoped to a network_id); the relay replies with peers, a list of RelayPeerInfo (each: peer_id, network_id, protocol_version, connected_at, last_seen).
• Register presence. A node that holds a relay reservation (register, §6) is itself returned to other nodes' get_peers — registration is the introducer advertisement.
• Live notifications. While registered, a node receives peer_connected / peer_disconnected for same-network peers, so its view stays fresh without polling.

A node MAY additionally use a dedicated introducer over the peer protocol (RequestPeersIntroducer → RespondPeersIntroducer, a peer_list of TimestampedPeerInfo{host, port, timestamp}), and register with it via register_peer{ip, port, node_type} → register_ack{success}.

4b · Gossip peer-exchange (node ↔ node)

Nodes also ask each other for peers, so discovery does not depend on any single rendezvous. Over an established peer link a node exchanges:

• RequestPeers (no fields) → RespondPeers carrying a peer_list of TimestampedPeerInfo{host, port, timestamp} (Chia-streamable, big-endian). Received lists are bounded (per-response and lifetime caps) and merged into the address book.

The peer RPC methods in §7 expose this same peer-exchange over the node's JSON-RPC surface, so an agent or a non-gossip client can drive discovery through the documented node profile. The continuous, incremental, first-hand peer gossip that keeps address books fresh is PEX (§4d) — it generalizes this snapshot exchange with an anti-flood, delta-based model over both an mTLS mux stream and the relay (RLY-008).

4c · Content discovery — the DHT

§4a and §4b find peers; the DHT finds which peers hold a specific piece of content. It is a Kademlia distributed hash table whose provider records map a content key to the peer_ids that hold that content. A node consults the DHT to locate holders before it fetches: it looks up the content, gets back the holders' peer_ids and candidate addresses, then confirms and fetches from them with dig.getAvailability + dig.fetchRange. The DHT locates peers; the NAT ladder reaches them and the peer RPC moves the bytes.

Every node both serves the DHT (holds a slice of the routing table and of the global provider records, and answers lookups) and advertises its own held inventory as provider records, so content is findable without any central index.

The keyspace — one 256-bit XOR metric for nodes and content

Kademlia places nodes and content in a single 256-bit keyspace and measures closeness by XOR distance (Maymounkov & Mazières). DIG maps into it as follows — a frozen contract every implementation reproduces:

• A node's key IS its peer_id. peer_id = SHA-256(TLS SubjectPublicKeyInfo DER) (§1) is already a uniform 256-bit value, so the DHT node id and the peer id are one and the same.

• A content key is SHA-256(domain-tag ‖ canonical bytes) over a fixed, domain-separated byte encoding. The one-byte domain tag makes the three granularities distinct points even when they share a store_id, so a store-level record and a resource-level record never collide:

  Content	Tag	Canonical bytes hashed	Answers
  store	0x01	store_id (32 B)	does a peer serve this store?
  root / capsule	0x02	store_id ‖ root (64 B)	does a peer have this generation store_id:root?
  resource	0x03	store_id ‖ root ‖ retrieval_key (96 B)	does a peer have this resource in the capsule?

  All hashes are the raw 32-byte forms in the fixed field order shown; the leading tag byte is part of the frozen key derivation and is never renumbered. These granularities match the dig.getAvailability has_store / has_root / has_resource shapes, so a lookup and an availability check speak of the same content.

• Distance is XOR. d(a, b) = a XOR b, compared big-endian (smaller = closer). A key's routing-table bucket index is 255 − leading_zeros(distance) — the position of the most-significant set bit, i.e. the length of the shared prefix with this node's id. This gives 256 k-buckets, least-recently-seen ordered with the standard ping-and-replace eviction (long-lived nodes resist eviction). One iterative lookup engine (α-parallel, converging on the k closest peers) serves both find_node and find_providers.

The DHT RPC — a distinct framed wire

The DHT RPC is not a dig.* JSON-RPC 2.0 method. It rides an authenticated dig-nat mTLS stream (§1): each RPC opens a logical stream and writes a u32 big-endian length prefix + a type-tagged JSON body — byte-identical framing to the dig-nat / relay control messages (§6), so a node speaks one framing across the whole peer network. The framed body is bounded (a length prefix over the cap is rejected, never allocated). There are exactly four methods:

Method	Request	Response
find_node	{ "type":"find_node", "target":"<64hex>" }	{ "type":"nodes", "nodes":[Contact] } — the k peers the responder knows closest to target
find_providers	{ "type":"find_providers", "content_key":"<64hex>" }	{ "type":"providers", "providers":[ProviderRecord], "closer":[Contact] } — providers held locally plus the k closer peers
add_provider	{ "type":"add_provider", "record":ProviderRecord }	{ "type":"add_provider_ok" } — the record was accepted + stored
ping	{ "type":"ping", "nonce":<uint> }	{ "type":"pong", "nonce":<uint> } — liveness; the responder echoes the nonce

A responder that cannot answer returns the error envelope { "type":"error", "code":<uint>, "message":<str> } — advisory: a lookup treats it like an unreachable peer and walks on. find_providers always returns closer contacts (even when providers are already found), because more providers may live nearer the key — this is what lets an iterative lookup keep converging.

The two wire shapes:

Contact        = { "peer_id":"<64hex>",
                   "addresses":[ { "host":str, "port":uint,
                                   "kind":"direct"|"mapped"|"reflexive"|"relay" } ] }

ProviderRecord = { "content_key":"<64hex>",
                   "provider_peer_id":"<64hex>",
                   "addresses":[ { "host":str, "port":uint, "kind":… } ],
                   "expires_at":<unix-seconds> }

The addresses[] shape (and the kind tokens direct/mapped/reflexive/relay, most-direct-first) is byte-compatible with the L7 dig.getPeers addresses (§7), so a returned Contact or ProviderRecord drops straight into a dial target for the NAT ladder. content_key is the 64-hex content key derived above; provider_peer_id is the holder's peer_id.

Provider-record lifecycle — soft state, TTL'd + republished

A provider record is soft state, not a permanent entry, so an offline holder ages out automatically. These rules are normative:

• Announce on hold. When a node gains content it serves, it PUTs a ProviderRecord (via add_provider) at the k nodes closest to that content key — binding the content key to its own peer_id and candidate addresses.
• Absolute expiry. expires_at is set to now + TTL in absolute Unix seconds. A record at or after its expires_at is treated as absent.
• Republish before expiry. The holder re-announces (a fresh record with a new expires_at) on an interval strictly shorter than the TTL, so its records never expire while it is online.
• Withdraw on removal. A node that no longer holds content stops announcing it; the record then ages out on its TTL (no explicit delete is required).
• GC drops the expired. A responder discards expired records on read and does not return them.
• Inbound RPC populates the routing table bidirectionally. On every inbound DHT RPC, the responder folds the mTLS-verified caller (its Contact) into its own routing table — every request is evidence the caller is alive, so a node that queries you teaches you about itself. The caller identity MUST come from the authenticated transport, never from a field the caller sets.

How a node uses the DHT

• On content-want (a user asks for store_id, store_id:root, or a specific resource): derive the matching content key, run find_providers, then reach each returned provider over the NAT ladder and fetch via dig.getAvailability + dig.fetchRange. The DHT is step 1 of the multi-source download — it finds the candidate holders the download then fans out across.
• On inventory-change (the node gains or loses content it serves): add_provider for each new content key, and stop announcing what it no longer holds. Run republish on the configured interval.
• Bootstrap the routing table from existing discovery — the gossip peer pool (§4b) or the relay introducer (§4a) — then a self-lookup (find_node on the node's own id) fills the table. The DHT never hard-depends on a live relay.

4d · PEX — peer exchange

PEX (Peer Exchange) is the continuous, low-overhead peer-discovery gossip that keeps every node's address book fresh: a node advertises the peers it has first-hand knowledge of, and receives the same from its links. It is on the PEER/CONTROL tier — it runs only over authenticated links (node↔node mTLS, or the node's registered relay connection) — and every received entry is a hint the receiver independently verifies, never an authenticated fact: the only proof a peer exists is a completed mTLS handshake with it. PEX generalizes and supersedes the §4b RequestPeers/RespondPeers snapshot exchange with an incremental, first-hand, anti-flood model. The normative contract is the dig-pex specification (dig-pex/SPEC.md); this section pins its wire and its two transport bindings.

The four messages

All four are type-tagged JSON, sharing the uniform DIG peer-network convention. Unknown fields are ignored on receive.

Message	Shape	Role
pex_handshake	{ "type":"pex_handshake", "version":1, "network_id":str, "interval":uint, "flags":[str] }	First message per direction; declares wire version, network, the sender's minimum data-message spacing, and its capability flags.
pex_snapshot	{ "type":"pex_snapshot", "peers":[PeerEntry] }	The first data message per direction — a capped (≤ 200) first-hand known-peer set so a fresh link warms up in one message. Exactly one per direction.
pex_delta	{ "type":"pex_delta", "added":[PeerEntry], "dropped":["<64hex>"] }	The periodic change message relative to what this link was already told (≤ 50 added, ≤ 50 dropped). Empty deltas are never sent.
pex_error	{ "type":"pex_error", "code":uint, "message":str }	Advisory error envelope, either direction. Named pex_error (not error) on both bindings so it never collides with the relay's RLY error code space.

A PeerEntry's addresses[] is byte-compatible with the dig.getPeers / DHT Contact addresses (host / port / kind ∈ direct|mapped|reflexive|relay), so a PEX-learned entry drops straight into a dial target. The identity attributed to a PEX entry's source is always the transport identity (mTLS peer_id or the registered relay identity) — never a wire field. The pex_error code space: 1 PEX_BAD_MESSAGE, 2 PEX_UNSUPPORTED_VERSION, 3 PEX_RATE_VIOLATION, 4 PEX_OVERSIZED, 5 PEX_NETWORK_MISMATCH, 6 PEX_PROTOCOL_VIOLATION (distinct from the relay's error codes 1..4).

Two transport bindings

• Node ↔ Node — a dig-nat mux stream. PEX rides one self-identifying logical stream per advertising direction on the established dig-nat mTLS session, framed u32-BE length prefix + JSON body (byte-identical to the DHT wire), bounded by PEX_MAX_FRAME = 262144 (256 KiB) — a length prefix over the cap is rejected before allocation. The stream's identity is the connection's mTLS peer_id.
• Relay → Node — the RLY-008 binding. PEX rides the relay RelayMessage WebSocket as additive bare-JSON text frames (no length prefix — the WebSocket delimits messages; the same 256 KiB payload bound applies). The relay is the introducer: after register/register_ack (RLY-001), a PEX-capable node sends its pex_handshake; the relay then sends its own handshake, a pex_snapshot of its registered same-network peers, and periodic pex_deltas as registrations come and go. The relay's entries carry via: "introducer", the registrant's observed public address (kind: "reflexive") when known, and a relay-only flag when the relay knows no direct path — registration IS the relay's first-hand evidence. The relay is introducer-only: it MUST NOT fold node-sent PEX entries into its registry (a PEX hint must never impersonate a registration), and a pex_handshake from an unregistered connection is answered with the relay's error code 1 (NOT_REGISTERED). All relay PEX is scoped to the node's registered network_id.

Why PEX is safe on both tiers

PEX only runs over already-authenticated links, and its entries are hints, so a malicious source can at worst suggest peers — never inject a trusted or reachable one (the receiver only trusts a peer after its own mTLS handshake). A first-hand rule (a node advertises only peers it directly knows, via ∈ direct|relay|introducer) stops re-gossip amplification; per-source attribution + a strike-based mute (3 violations → mute the direction) bound a chatty or lying source; every inbound frame is size- and rate-capped before allocation. This is why PEX sits comfortably on the mTLS peer/control tier: the relay link is server-authenticated wss://, but a node treats even the relay's introductions as hints it verifies before trusting.

5 · Relay-coordinated hole-punching (signalling only)

This is the relay's third role (above) and a distinct message flow from the relayed/TURN data path of §6. When two nodes are both behind NAT, the relay signals only: it relays their candidate-address exchange and coordinates a simultaneous open — each side learns the other's reflexive address and dials it at the same moment, so both NATs see the outbound connection as solicited and let the peer's packets in. The relay carries only these small coordination messages; the data stream then flows peer-to-peer. The resulting link is direct (and mTLS, exactly like every other tier).

This is why the ladder prefers strategy (e) over (f): the hole-punch flow is a low-bandwidth introduction, not a data proxy. A node falls to full relayed transport (§6) only after the hole punch fails.

Ordered procedure — the hole-punch signalling wire, aligned to the relay RelayMessage hole-punch messages (RLY-007). Both peers already hold a relay reservation (§6):

1. A and B are each connected to the relay (reservations held).
2. A learns its reflexive IP:port via STUN (§3)  — its candidate to punch to.
3. A → relay:  hole_punch_request     { peer_id: A, target_peer_id: B, external_addr: A_reflexive }
4. relay → B:  hole_punch_coordinate  { peer_id: A, external_addr: A_reflexive }   // candidate carried to B
5. B learns its own reflexive IP:port (§3) and both sides SIMULTANEOUSLY dial:
      A dials B_reflexive,  B dials A_reflexive.                                   // the punch (peer-to-peer)
6. On the first direct link to complete: run the mTLS handshake (§1),
   migrate the data stream to the DIRECT link, and drop the relay path for this pair.
7. If both dials fail: keep any relay data path and retry after the hole-punch backoff.
   A → relay:  hole_punch_result      { peer_id, success }   // informational

Only the three hole_punch_* messages (steps 3, 4, 7) cross the relay — the candidate-address exchange and the coordinated-punch trigger. The peer's actual traffic (step 5 onward) never touches the relay. external_addr is serialized as the canonical "IP:port" string. A successful hole-punch is a direct connection — it satisfies strategy (e) of the ladder and keeps the data stream off the relay entirely.

6 · The relay wire (RelayMessage) + relayed / TURN transport

This section defines the relay's full RelayMessage wire, including its fourth role (above): relayed (TURN-like) transport, in which the relay proxies all of a peer connection's data. This is the high-bandwidth, last-resort tier (strategy (f) of the ladder) — entered only after direct, UPnP/NAT-PMP/PCP, and the hole-punch signalling of §5 have all failed. It is a distinct flow from the low-bandwidth signalling roles: here the peer's bytes flow through the relay (relay_message / broadcast), whereas STUN, introducer, and hole-punch carry only control messages.

The relay is a stateless rendezvous / circuit bridge speaking JSON messages over a secure WebSocket (wss://). The default endpoint is DIG_RELAY_URL = wss://relay.dig.net:9450 (override with the DIG_RELAY_URL environment variable; off disables the reservation). It exposes a plaintext health check at GET /health on port 9451.

Every message is a JSON object with a type discriminator. The message family is RLY-001..RLY-007:

ID	Message(s)	Direction	Shape
RLY-001	register	node → relay	{ "type":"register", "peer_id":str, "network_id":str, "protocol_version":uint }
	register_ack	relay → node	{ "type":"register_ack", "success":bool, "message":str, "connected_peers":uint }
	unregister	node → relay	{ "type":"unregister", "peer_id":str }
RLY-002	relay_message	node → relay → node	{ "type":"relay_message", "from":str, "to":str, "payload":[uint], "seq":uint }
RLY-003	broadcast	node → relay → nodes	{ "type":"broadcast", "from":str, "payload":[uint], "exclude":[str] }
RLY-005	get_peers	node → relay	{ "type":"get_peers", "network_id":str|null }
	peers	relay → node	{ "type":"peers", "peers":[RelayPeerInfo] }
	peer_connected	relay → node	{ "type":"peer_connected", "peer":RelayPeerInfo }
	peer_disconnected	relay → node	{ "type":"peer_disconnected", "peer_id":str }
RLY-006	ping / pong	either way	{ "type":"ping", "timestamp":uint } / { "type":"pong", "timestamp":uint }
RLY-007	hole_punch_request	node → relay	{ "type":"hole_punch_request", "peer_id":str, "target_peer_id":str, "external_addr":"IP:port" }
	hole_punch_coordinate	relay → node	{ "type":"hole_punch_coordinate", "peer_id":str, "external_addr":"IP:port" }
	hole_punch_result	node → relay	{ "type":"hole_punch_result", "peer_id":str, "success":bool }
RLY-008	pex_handshake / pex_snapshot / pex_delta / pex_error	either way	The PEX messages, carried as additive WebSocket text frames beside RLY-001..RLY-007 (§4d).
—	error	relay → node	{ "type":"error", "code":uint, "message":str }

where RelayPeerInfo = { "peer_id":str, "network_id":str, "protocol_version":uint, "connected_at":uint, "last_seen":uint } (connected_at/last_seen are unix seconds). peer_id fields are the 64-hex rendering of the peer_id. payload is a JSON array of byte values (0..255).

PEX binding (RLY-008). The relay is also the PEX introducer: PEX messages (§4d) ride this same WebSocket as additive top-level frames whose pex_* type tags do not collide with any RLY-001..RLY-007 tag — so the binding is purely additive and no existing RLY message changes shape. A frame is classified by peeking its "type" for the pex_ prefix before the RLY parse. PEX on the relay is gated on the node's pex_handshake after register (a legacy node sees the wire exactly as before), scoped to the node's registered network_id, and the relay's entries are registration-backed (via: "introducer") — node-sent PEX data is anti-flood-checked but never enters the introducer registry. PEX-level errors use pex_error (its own code space, §4d), keeping the relay's error envelope's RLY codes distinct. Full contract: §4d and the dig-pex specification (dig-pex/SPEC.md §10.2).

Reservation (RLY-001). A node opens the WebSocket and sends register; the relay records the reservation and replies register_ack. The reservation is held for the life of the connection — there is no fixed TTL — but a connection idle past the relay's idle timeout is reaped, so a node sends periodic ping (RLY-006) to keep it alive and reconnects on drop.

Relayed transport (RLY-002 / RLY-003). relay_message forwards payload to the single peer to; broadcast fans out to every same-network peer except from and any id in exclude. The relay re-stamps from to the sender's registered peer_id (a node cannot spoof another's id) and forwards the payload verbatim without inspecting it. Routing is scoped to the sender's network_id — the relay never bridges across networks. This is the last-resort data path when strategies (a)–(e) all fail.

Relay error codes (RLY error).

code	Name	Meaning
1	NOT_REGISTERED	A message arrived before the connection completed register.
2	BAD_MESSAGE	The frame was not valid relay JSON.
3	PEER_NOT_FOUND	A relay_message / hole-punch named a to / target_peer_id not registered on this network.
4	CAPACITY	The relay is at its connection cap; the registration was refused.

Health. GET http://<relay>:9451/health returns { "status":"ok", "connected_peers":uint, "uptime_secs":uint, "version":str } with HTTP 200 while serving — the reachability probe for the relay itself.

7a · Tier map — which method is on which tier

The implementers' contract for the dual-transport tiers. The dig-node serve layer and every RPC client build to this: a method's tier is frozen. Anonymous callers reach only the PUBLIC-READ column; everything in the PEER/CONTROL column returns -32601 on the anonymous listener and is served only over the mTLS peer/control transport.

Surface / method	Tier	Auth	Transport
dig.getContent (read)	PUBLIC READ	none (anonymous, CORS-*)	plain HTTPS JSON-RPC
dig.getProof / dig.getProofStatus (read)	PUBLIC READ	none	plain HTTPS JSON-RPC
dig.getCapsule (alias dig.getModule) (read)	PUBLIC READ	none	plain HTTPS JSON-RPC
dig.getManifest / dig.getMetadata / dig.listCapsules (read)	PUBLIC READ	none	plain HTTPS JSON-RPC
dig.health / dig.methods (discovery)	PUBLIC READ	none	plain HTTPS JSON-RPC
§21 GET content / proof / roots / descriptor	PUBLIC READ	none	plain HTTPS REST
dig.getPeers / dig.announce / dig.getNetworkInfo	PEER / CONTROL	mTLS	dig-nat mTLS
dig.getAvailability / dig.listInventory	PEER / CONTROL	mTLS	dig-nat mTLS
dig.fetchRange (peer sync / multi-source)	PEER / CONTROL	mTLS	dig-nat mTLS
DHT find_node / find_providers / add_provider / ping (§4c)	PEER / CONTROL	mTLS	dig-nat mTLS stream
PEX pex_handshake / pex_snapshot / pex_delta / pex_error (§4d)	PEER / CONTROL	mTLS (node↔node) or registered relay identity (RLY-008)	dig-nat mTLS stream / relay WebSocket
§21 PUSH / WRITE: module/upload · module PUT · module/complete · tombstone	PEER / CONTROL	mTLS + per-request BLS (§21.9)	authenticated HTTPS
node config / control (cache.*, control.*, dig.stage, dig.getAnchoredRoot)	CONTROL (loopback)	local authorization (loopback-only)	local FFI / loopback HTTP

Notes:

• dig.getAvailability / dig.fetchRange are PEER/CONTROL, not public read: they are the node↔node sync/multi-source surface, ridden over the mTLS mux. The browser/agent public read path is the dig RPC (dig.getContent et al.) — a browser does not fan byte-ranges across peers; that is a node-side operation (multi-source download).
• dig.getAnchoredRoot and cache.* are local control (loopback / in-process FFI, node profile), not exposed on the public read listener. A browser resolves the anchored root from its own chain source, not from the serving node.
• The read subset is identical in shape to the network profile — this is why one JSON-RPC endpoint serves both a browser and the local consumer; only the set of methods the anonymous listener answers differs.

7 · Peer RPC methods (node profile)

The node profile of the dig RPC exposes the peer network over JSON-RPC 2.0, so an agent can inspect and drive discovery without speaking the binary peer protocol. These are additive to the existing node methods; they appear in the node's dig.methods catalogue and in openrpc-node.json. All are node-profile only (absent from the network profile) and, like every dig RPC method, are POST JSON-RPC 2.0 with by-name params.

dig.getPeers

Return the peers this node currently knows, each with its peer_id and candidate addresses — the peer-exchange of §4b over RPC.

• params: { "network_id"?: str, "limit"?: uint } — optional network filter; limit caps the returned list.
• result:

{
  "peers": [
    {
      "peer_id": "<64hex>",
      "addresses": [ { "host": "203.0.113.7", "port": 9444, "kind": "direct" } ],
      "network_id": "DIG_MAINNET",
      "last_seen": 1719763200,
      "via": "direct"
    }
  ]
}

Each addresses[] entry is a candidate address: { host: str, port: uint, kind: "direct"|"reflexive"|"mapped"|"relay" } where kind records how the address was learned (advertised/observed direct, STUN reflexive, UPnP/NAT-PMP/PCP mapped, or relay-reachable). via is how this node currently reaches the peer ("direct" or "relay").

dig.announce

Advertise this node (its peer_id + candidate addresses) to a target peer, and offer to establish or upgrade a connection — the RPC face of the introducer/announce path in §4.

• params:

{
  "peer_id": "<64hex>",
  "addresses": [ { "host": "198.51.100.4", "port": 9444, "kind": "mapped" } ],
  "network_id": "DIG_MAINNET",
  "target"?: "<64hex>"
}

peer_id + addresses describe the announcing node; the optional target is a specific peer to announce to (omit to announce to the relay/introducer as a general registration).

• result: { "accepted": bool, "known_peers": uint } — whether the announcement was accepted and the resulting size of the recipient's peer view.

dig.getNetworkInfo

Report this node's own network posture — its identity, reachability, candidate addresses, and relay-reservation state. This is the self-describe surface for discovery and the STUN-derived reflexive address.

• params: none.
• result:

{
  "peer_id": "<64hex>",
  "network_id": "DIG_MAINNET",
  "listen_addr": "[2001:db8::7]:9444",
  "reflexive_addr": "203.0.113.1:9444",
  "candidate_addresses": [ "[2001:db8::7]:9444", "203.0.113.7:9444" ],
  "reachability": "direct",
  "relay": { "url": "wss://relay.dig.net:9450", "reserved": true, "connected_peers": 42 }
}

candidate_addresses is the node's real, directly-dialable candidate endpoints, ordered IPv6-first (§2a): a routable IPv6 address (when the host has one) precedes the IPv4 fallback. It never contains the wildcard bind address. listen_addr is the primary (first, IPv6-preferred) dialable candidate — not the socket's wildcard bind target. reflexive_addr is the STUN-discovered public address (§3) or null if not yet learned; reachability is "direct" (a direct inbound path exists — publicly reachable or a working UPnP/NAT-PMP/PCP mapping) or "relayed" (only reachable through the relay); relay.reserved reflects the RLY-001 reservation state.

Peer RPC error codes

Alongside the standard JSON-RPC codes and the shared -32004 (resource unavailable), the peer methods add these node-profile codes:

Code	Name	Meaning
-32006	PEER_UNREACHABLE	No connection to the named peer could be established — every traversal strategy (direct, UPnP/NAT-PMP/PCP mapping, relay-coordinated hole-punch, and relayed fallback) failed, or the peer is not registered on this network.
-32007	RANGE_NOT_SATISFIABLE	The requested byte range lies outside the resource (offset ≥ its length). Returned by dig.fetchRange.
-32008	CONTENT_REDIRECT	This node does not hold the requested content, but it located peers that do — a redirect, not a not-found (§9a). error.data.redirect names the holder(s).

See the full error catalog.

8 · Streaming-first content transport

Peer RPC is streaming-first for data, not buffer-the-whole-payload. A connect(peer) link (§2) is a multiplexed stream transport: it carries many concurrent, independent logical streams over the single mTLS connection, so a node can run several content transfers (and control calls) in parallel without head-of-line blocking between them.

• Control methods are message-style. The small methods — dig.getPeers, dig.announce, dig.getNetworkInfo — request and return a single JSON object. They fit one logical message; no streaming needed.
• Data methods are chunk-streamed. Any method that returns a large or content-bearing payload (dig.fetchRange, §9) delivers it as an ordered stream of chunk frames on its own logical stream. The caller reads frames incrementally and reassembles — it never has to hold the whole resource in memory to begin using it.
• Backpressure. The stream transport applies flow control: a slow reader slows the sender rather than forcing the sender to buffer unboundedly. A caller that stops reading pauses the transfer; a caller that reads faster receives faster.
• Framing. Each data frame carries { offset, length, bytes, complete } — the frame's start offset within the requested range, its byte length, the raw ciphertext bytes, and whether it is the final frame. Frames arrive in ascending offset order and tile the requested range exactly. A caller reassembles by offset and stops on complete.
• Cancellation. Closing the logical stream cancels just that transfer; the connection and any sibling streams are unaffected.

This mirrors the dig RPC streaming contract (window/offset/next_offset reassembly) but over the peer stream transport rather than repeated JSON-RPC POSTs — the same incremental, verify-as-you-go model.

9 · Byte-range content fetch + multi-source download

A node can request a specific byte range [offset, offset+length) of a content resource or an entire .dig capsule, and receive only those bytes, streamed. This is the primitive behind multi-source download: a client splits a resource into ranges and fetches different ranges from different peers simultaneously, verifies each independently, and reassembles — the same multisource + range + integrity + resume model implemented by the dig-download-utility reference.

Availability first — ask before you fetch

Before requesting any content, a client asks candidate peers whether they actually hold it, so it fans ranges only at peers that can serve them. Availability is a small control RPC (message-style, not streamed), and it is batchable — a downloader checks several peers × several items in one call each — at all three granularities:

• store — does the peer serve this store_id at all (and which roots does it hold)?
• root — does the peer have this specific generation (store_id, root)?
• capsule / resource — does the peer have this specific immutable capsule store_id:root, or a specific resource within it?

dig.getAvailability

Ask one peer about many items at once. The granularity of each item is inferred from which fields it carries.

• params:

{
  "items": [
    { "store_id": "<64hex>" },
    { "store_id": "<64hex>", "root": "<64hex>" },
    { "store_id": "<64hex>", "root": "<64hex>", "retrieval_key": "<64hex>" }
  ]
}

• store_id only → has_store: does the peer serve the store.

• store_id + root → has_root: does the peer have that generation (the capsule store_id:root).

• store_id + root + retrieval_key → has_resource: does the peer have that resource within the capsule.

• result: one answer per item, positionally aligned with items:

{
  "items": [
    { "available": true, "roots": ["<64hex>", "<64hex>"] },
    { "available": true, "chunk_count": 40, "total_length": 10485760, "complete": true },
    { "available": true, "total_length": 262144, "chunk_count": 1, "complete": true }
  ]
}

Per-item fields (present where cheap for the peer to answer):

• available (bool, always) — whether the peer holds the queried item.
• roots (store granularity) — the generation roots the peer currently holds for the store, newest-first.
• total_length + chunk_count (root/resource granularity) — the resource/capsule ciphertext length and its chunk count, so the caller can plan its ranges without a probe fetch.
• complete (bool) — whether the peer holds the full resource/capsule (true) or only part of it (false); a partial holder can still serve the ranges it has.

dig.listInventory

Enumerate what a peer serves — the discovery variant.

• params: { "store_id"?: "<64hex>", "limit"?: uint } — omit store_id to list the stores the peer serves; supply it to list the roots the peer holds for that store.
• result:

{ "stores": ["<64hex>", "..."] }

or, when store_id is given:

{ "store_id": "<64hex>", "roots": ["<64hex>", "..."] }

Enumeration is best-effort discovery — a peer MAY cap or omit it (privacy / size); dig.getAvailability is the authoritative per-item check.

dig.fetchRange

Stream a byte range of a resource or capsule from this peer.

• params:

{
  "store_id": "<64hex>",
  "retrieval_key": "<64hex>",
  "root": "<64hex>",
  "capsule": false,
  "offset": 0,
  "length": 4194304
}

• Resource identity. For a content resource: store_id + retrieval_key (+ optional root, defaulting to the chain-anchored tip). For a whole capsule / .dig: set capsule: true and identify it by store_id (+ optional root); retrieval_key is then omitted. The capsule identity is <store_id>[:<root>].

• Range. offset (bytes into the resource ciphertext, default 0) and length (bytes to return). length is clamped to the node's window (3 MiB); a request whose range is not chunk-aligned is widened to whole-chunk boundaries (see integrity below), so the response may return slightly more than asked.

• result: a stream (§8) of dig.fetchRange frames. Beyond the base frame fields, the first frame (offset == range start) carries the verification metadata for the range:

{
  "offset": 0,
  "length": 262144,
  "bytes": "<base64 ciphertext>",
  "complete": false,
  "total_length": 10485760,
  "chunk_lens": [262144, 262144, 131072],
  "chunk_index": 0,
  "inclusion_proof": "<base64 merkle proof>",
  "root": "<64hex>"
}

• total_length — the full resource ciphertext length (so a client can plan its ranges).
• chunk_lens — the per-chunk ciphertext lengths of the whole resource, in order (first frame only) — identical to the dig RPC chunk_lens. This is how a client maps a byte range to the chunk(s) that cover it.
• chunk_index — the index (into chunk_lens) of the first chunk in this frame.
• inclusion_proof — the merkle inclusion proof of the whole resource against the capsule's generation root (first frame only), relayed verbatim (Merkle inclusion proofs). For capsule: true the capsule self-verifies on install, so inclusion_proof is null (as with dig.getCapsule).

Per-range integrity — verify a range without the whole file

A range fetched from one peer is independently verifiable against the capsule's on-chain merkle root, so a single peer cannot forge a range and multi-source pieces always reassemble correctly. Integrity aligns exactly with the existing digstore content model: a resource is a sequence of AES-256-GCM-SIV chunks; the resource commits to the generation merkle root as a single leaf (resource_leaf = SHA-256(concatenated chunk ciphertexts)); chunk_lens fixes the chunk boundaries.

A requested range maps to whole chunk(s) — the node widens the range to chunk boundaries — so each returned chunk is a complete, verifiable unit. A client verifies a fetched range as follows:

1. Split by chunk_lens. Using the first frame's chunk_lens and chunk_index, cut the reassembled range bytes into the exact chunk(s) it covers.
2. Verify the resource against the root. The inclusion_proof proves resource_leaf (= SHA-256 of the whole resource ciphertext, reconstructed from all chunks in chunk_lens order) is included under the caller-supplied chain-anchored root (Verification & provenance) — the node is never the trust anchor. A client that already holds this proof (from an earlier range/peer) reuses it; the proof is the same regardless of which peer or range served the bytes.
3. Bind each chunk to the committed resource. Because chunk_lens fixes each chunk's length and the resource leaf is SHA-256 over the concatenation of all chunk ciphertexts in order, a chunk delivered for a given chunk_index is correct iff, when placed at its chunk_lens offset, the whole-resource hash still matches the proven resource_leaf. A chunk from a bad source yields a resource hash that does not match the proof — detected without downloading the whole file.
4. Decrypt. AES-256-GCM-SIV-open each verified chunk; a wrong key/salt or corrupted bytes fails the authentication tag (DIG_ERR_DECRYPT_TAG).

Detecting a bad source: any of a chunk-length mismatch against chunk_lens, a failed whole-resource inclusion proof, or a decryption-tag failure marks the serving peer's range as invalid — the client discards it and retries that range from a different peer (-32006 / a fresh source), penalizing the bad peer.

The multi-source download pattern (normative)

1. DISCOVER  — locate candidate holders in the DHT: find_providers(content_key)
               for the store/root/resource (§4c), backed by the introducer +
               dig.getPeers peer discovery (§4).
2. QUERY     — dig.getAvailability (batch) against the candidates: has_store /
   AVAILABILITY   has_root / has_capsule (§9 availability). Keep only peers that
               actually HOLD the resource; read total_length + chunk_count to plan.
3. PLAN      — partition the resource into chunk-aligned ranges (using chunk_lens
               from the first range frame or the availability chunk_count).
4. FAN OUT   — request DIFFERENT ranges from DIFFERENT holders CONCURRENTLY over the
               multiplexed stream transport (§8), respecting each peer's backpressure.
5. VERIFY    — verify each returned range independently against the chain-anchored
               root (§9 integrity) as it arrives; do not trust unverified bytes.
6. RETRY     — a failed, timed-out, missing, or mismatched range is re-requested from
               ANOTHER holder (ranges are independent); penalize the bad source.
7. REASSEMBLE— place verified ranges by offset into the full resource; decrypt per chunk.

Step 2 is the gate: a downloader never fans a range at a peer it has not confirmed holds the content, and a partial holder (complete: false) is used only for the ranges it actually has.

• Concurrency across sources is what makes this fast: N peers each serve a slice of the file in parallel.
• Resume. Because each range is independently addressable and independently verifiable, an interrupted download resumes per range — a client re-requests only the ranges it has not yet verified, from any peer that holds the resource. No range already verified is refetched.
• Any source, one root. Every range — whichever peer served it — is verified against the same on-chain generation root, so mixing sources never weakens integrity.

Range fetch error codes

Code	Name	Meaning
-32004	RESOURCE_UNAVAILABLE	This peer does not hold the resource/capsule at the requested root, and it could not locate any peer that does (a genuine not-found). When it CAN locate a holder it returns the redirect (-32008) instead.
-32007	RANGE_NOT_SATISFIABLE	The requested offset/length lies outside the resource (offset >= total_length), or the range is otherwise unsatisfiable.
-32008	CONTENT_REDIRECT	This node does not hold the content but located peers that do — see redirect-on-miss.

9a · Redirect-on-miss — never a silent not-found when a holder exists

A content read (dig.getContent, dig.fetchRange, or a peer range stream) may reach a node that does not hold the requested store / capsule / resource locally. Rather than dead-ending on a bare not-found, the node consults the DHT (find_providers, §4c) for that content key. If a holder exists, the node does one of two things — never a silent 404 while a provider exists:

• REDIRECT (default). The node answers with the -32008 CONTENT_REDIRECT error whose error.data.redirect names the holder(s), so the caller re-requests against a holder. Cheap and stateless — the node does not fetch the bytes. The redirect object:

  {
    "code": -32008,
    "message": "content not held by this node; re-request against a provider in data.redirect",
    "data": {
      "redirect": {
        "content": { "store_id": "<64hex>", "root": "<64hex>", "retrieval_key": "<64hex>" },
        "providers": [
          { "peer_id": "<64hex>", "addresses": [ { "host": "…", "port": 9444, "kind": "direct" } ] }
        ],
        "redirect_depth": 1,
        "max_redirects": 4
      }
    }
  }

  Each provider's addresses[] is byte-compatible with the dig.getPeers / DHT Contact address shape (host / port / kind), so a redirect target drops straight into a dial. The content object echoes the requested item at whatever granularity it was named (store_id [+ root [+ retrieval_key]]).

• FETCH-THROUGH (opt-in, DIG_NODE_ON_MISS=fetch). Instead of redirecting, the node pulls the resource from the holder(s) via the multi-source download (verified against the chain-anchored root), caches it, and serves it directly — transparent to the caller, one round-trip, and the node now holds it. If the fetch fails the node falls back to the redirect (still never a silent 404 while a holder exists).

The redirect is a READ-TIER response. It carries only what the caller needs to reach a holder — it exposes no peer/write/control surface — so it is served on both the mTLS peer/control tier and, where the read path is public (§0), the anonymous read tier.

Bounded hops (no loops). The redirect carries a redirect_depth (the number of redirects already followed) and the max_redirects cap. A caller echoes redirect_depth in its re-request params; a node that receives a request already at/over the cap answers the plain not-found instead of another redirect, so a set of nodes can never bounce a caller in a loop. A node never redirects a caller back to itself (its own peer_id is excluded from the providers).

10 · The relay-last-fallback invariant

A node uses the relay to carry peer↔peer DATA (role 4, TURN-like) ONLY when none of DIRECT, UPnP, NAT-PMP, PCP, or hole-punch succeeds. Concretely:

• If port-forwarding, UPnP, NAT-PMP, or PCP yields a working direct path — inbound or outbound — no relay carries the data. The relay may still fill its low-bandwidth signalling roles (STUN, introducer, hole-punch coordination), but the peer's bytes flow directly.
• Prefer hole-punch signalling over full relaying. The relay-coordinated hole punch (§5, role 3) and full relayed transport (§6, role 4) are different tiers: a node always attempts the hole punch first because it costs the relay only the coordination messages while the stream goes peer-to-peer. Full relayed transport — where the relay carries every byte — is entered only when the hole punch fails. This keeps relay bandwidth minimal: the relay brokers the introduction; the data does not flow through it unless there is no other way.
• A relay reservation held for reachability does not mean traffic is relayed: a reserved node whose peer can reach it directly (including via a brokered hole punch) is served over the direct link.
• The relayed/TURN transport (§6) is entered only after strategies (a)–(e) of the ladder are exhausted, and a node continues to attempt a hole-punch upgrade so a relayed pair can be promoted to a direct peer-to-peer link when conditions allow, dropping the relay from the data path.

The relay is an untrusted bridge: it forwards end-to-end-authenticated payloads by peer_id and can read none of them. Trust always rests on the mTLS peer identity (§1), never on the relay.

11 · Conformance

The peer network is implemented by several crates that must interoperate byte-for-byte. The frozen shapes a reimplementation MUST match:

Surface	Frozen shape	What it pins
peer_id	SHA-256(SubjectPublicKeyInfo DER) → Bytes32, 64-hex on text surfaces	the identity every peer derives for every other peer; a mismatch means no interop
mTLS handshake	wss:// + client cert required + Chia Handshake (hex network_id, protocol version, node type)	that a peer link is authenticated and network-scoped before any message is processed
Two RPC tiers	PEER/CONTROL = mTLS-authenticated (peer/DHT/PEX/availability-for-sync/write/config); PUBLIC-READ = anonymous, CORS-*, read-only, client-verified, decoy-on-miss (§0, tier map §7a). No peer/write/config method reachable without mTLS; content read requires no mTLS	that a browser can retrieve+verify content with no client cert while every mutation/identity/peer surface stays mTLS-gated — the boundary invariant is uniform across every serve layer + client
RelayMessage wire	the RLY-001..RLY-008 JSON shapes in §6, type-tagged, payload as a byte array, from re-stamped, network_id-scoped; RLY-008 = the additive PEX binding	that any relay + any node speak the same rendezvous/hole-punch/relayed/PEX wire
PEX	the four pex_* messages (§4d) with the dig-pex/SPEC.md shapes; two bindings — dig-nat mux stream (u32-BE+JSON, PEX_MAX_FRAME 262144) and the relay RLY-008 (additive WS frames, introducer-only, registration-backed via:"introducer"); entries are hints (first-hand rule, via provenance, strike-mute); addresses[] byte-compatible with dig.getPeers/DHT Contact	that peer gossip is anti-flood + first-hand across both transports, entries are verified-before-trusted, and PEX-learned contacts drop straight into a dial target
Relay error codes	1..4 (NOT_REGISTERED/BAD_MESSAGE/PEER_NOT_FOUND/CAPACITY)	deterministic relay-side failure signalling
Relay roles	the four roles are distinct: STUN + introducer + hole-punch signalling are low-bandwidth control; only relayed/TURN transport carries data	that a node prefers brokering a direct link over proxying the stream
Peer exchange	RequestPeers→RespondPeers of TimestampedPeerInfo{host, port, timestamp} (Chia-streamable, big-endian)	that nodes discover peers from each other identically
Peer RPC	dig.getPeers / dig.announce / dig.getNetworkInfo + -32006; dig.getAvailability / dig.listInventory; dig.fetchRange + -32007, generated into openrpc-node.json	the machine surface an agent drives; CI-diffable against a live node
Availability	dig.getAvailability batch per-item answers at store / root / capsule granularity (available + roots/total_length/chunk_count/complete)	that a downloader can confirm a peer HOLDS content (and plan ranges) before any fetch
Streaming + range	dig.fetchRange streams RangeFrame{offset,length,bytes,complete}; first frame carries total_length + chunk_lens + chunk_index + inclusion_proof; ranges are chunk-aligned	that data streams (not buffered), and a single-peer range verifies against the chain-anchored root — so multi-source pieces reassemble and can't be forged
Range integrity	a range maps to whole chunk(s); each verifies via chunk_lens + the whole-resource inclusion_proof against the on-chain root (same as merkle-proofs)	that any peer's range is independently verifiable + a bad source is detectable without the whole file
NAT ladder	the ordered strategies DIRECT → UPnP → NAT-PMP → PCP → hole-punch (relay signalling only) → RELAYED/TURN (relay carries data), relay-data-last	that every connect(peer) implementation prefers direct, prefers hole-punch signalling over full relaying, and proxies the stream only as a last resort
DHT content key	SHA-256(tag ‖ canonical bytes) with tags 0x01 store (store_id), 0x02 root/capsule (store_id ‖ root), 0x03 resource (store_id ‖ root ‖ retrieval_key); node id = peer_id; distance = XOR; bucket = 255 − leading_zeros (§4c)	that every node derives the identical content key for the same content, and places nodes + content in one 256-bit keyspace, so a provider record announced by one implementation is found by another
DHT RPC wire	the four type-tagged methods find_node / find_providers / add_provider / ping (+ the error envelope), u32-BE length-prefixed JSON over an authenticated dig-nat mTLS stream (same framing as the relay control messages), find_providers always returning closer (§4c)	that any node's DHT speaks the same locate-the-holders wire; dig-nat/dig-dht/dig-node conform
DHT shapes	Contact { peer_id:<64hex>, addresses:[{host,port,kind}] } and ProviderRecord { content_key, provider_peer_id, addresses, expires_at }, addresses[] byte-compatible with dig.getPeers; provider records are TTL'd (absolute expires_at), republished before expiry, GC'd when stale, and every inbound RPC folds the mTLS-verified caller into the routing table	that returned contacts/records drop straight into a dial target and that provider state is soft state that ages out

A reimplementation of any peer crate conforms iff it reproduces these — the same discipline that keeps the read path parity-locked.

Related

• The dig RPC — the PUBLIC READ tier: what a browser/agent reads over anonymous JSON-RPC; the node profile the peer RPC extends
• §21 transport & push — the §21 GET (public read) vs PUSH/WRITE (mTLS peer/control) split
• BLS signatures & DSTs — the node/attestation signatures carried over peer links; the per-request write auth on the control tier
• Conformance & parity — the cross-implementation parity discipline
• Verification & provenance — why a read server (public tier) is never a trust anchor
• The blind host model — decoy-on-miss on the public read tier
• Error codes — the full catalog incl. the peer -32006