fips/docs/design/fips-nostr-discovery.md

FIPS Nostr-Mediated Discovery and NAT Traversal

Nostr-mediated discovery lets FIPS nodes find each other, and if necessary, punch through UDP NAT, using public Nostr relays as the signaling channel. A node publishes its reachable transport endpoints to a small set of relays under its own Nostr identity (which is also its FIPS identity), and peers resolve those endpoints at dial time by npub. For peers behind UDP NAT, the same relay channel carries an encrypted offer/answer exchange, and STUN supplies the reflexive address used for a coordinated hole-punch.

Nostr discovery is unconditionally compiled into the fips binary on every supported platform and ships in every stock packaging artifact (.deb, AUR, systemd tarball, OpenWrt .ipk, macOS .pkg, Windows .zip). It is runtime-opt-in: the YAML configuration defaults to disabled (node.discovery.nostr.enabled: false), so the discovery runtime stays dormant — and opens no relay connections — until an operator flips the flag and supplies a relay list. When disabled, nodes behave exactly as before: only the static peers[] addresses are used.

Role

The feature adds three capabilities on top of FIPS's static peer model:

• Advertising. A node publishes the transport endpoints it wants peers to use (direct UDP, direct TCP, a Tor onion, or the special udp:nat rendezvous token) as a signed Nostr event. The advert is anchored to the node's FIPS identity key — a peer that knows the npub knows the advert is authentic.
• Lookup. When dialing a configured peer marked via_nostr, or any peer in policy: open mode, the node fetches that peer's advert from the configured relays and appends the advertised endpoints to its dial list. Static addresses are always tried first.
• UDP NAT hole-punch. When both sides of a connection have UDP NAT endpoints, the advert carries enough information to run a STUN-based offer/answer exchange over encrypted (NIP-59) Nostr events. Each side observes its reflexive address via STUN, exchanges candidate pairs through the relay, and both sides send UDP probes at a shared punch time. On the first successful probe, the punch socket is handed to FMP and becomes a normal UDP transport.

When to use it

• You run a public node and want peers who know your npub to reach you without you distributing an address list out-of-band.
• You want to reach a peer behind UDP NAT without deploying a relay or running Tor on both sides. The peer advertises udp:nat and you dial by npub.
• You want zero-touch peer discovery within a known application namespace (policy: open), subject to an admission budget.
• You want to advertise a Tor onion so peers don't need to know the .onion address out-of-band.

Skip the feature when every peer is already reachable through a stable static address (a LAN mesh, a pre-configured test bed, or a deployment where operators distribute peers[] blocks directly). The feature adds relay dependencies, STUN round-trips for NAT cases, and a small ambient background of relay traffic; none of that is useful when you already know where peers are.

Scenarios and configuration

For end-to-end operator recipes — each of the five activation scenarios (advertise a directly-reachable UDP node, advertise a Tor onion node, look up a configured peer by npub without advertising, NAT hole-punch between two configured peers, and open discovery within an app namespace) — see ../how-to/enable-nostr-discovery.md. The full configuration knob tables, per-transport keys, and startup validation rules live in ../reference/configuration.md under node.discovery.nostr.*. The Kind 37195 advert event format is in ../reference/nostr-events.md. The rest of this document covers the design of the discovery runtime itself.

Under the covers

The rest of this document describes how the feature works inside the node. For the generic protocol shape (event tags, NIP usage, on-the- wire offer/answer schema, failure-suppression machinery), see port-advertisement-and-nat-traversal.md.

Overview

The discovery runtime is a background task group started during node initialization when nostr.enabled is true. It maintains a single nostr-sdk client connected to the union of advert_relays and dm_relays, and runs four loops: advert publication, advert subscription (for open discovery and cache warming), DM subscription (for incoming offers and answers), and a periodic advert-cache prune. Discovery has no CLI surface; all operations are driven by the configuration and by connection attempts made by the rest of the node.

                    +-----------------------+
                    |   Discovery runtime   |
                    +-----------------------+
                       |       |       |
        advert publish |       | DM sub (offers, answers)
                       |       |
                       v       v
              +-------------------------+
              |   Nostr relay pool      |  (advert_relays ∪ dm_relays)
              +-------------------------+
                       ^       ^
    advert fetch/cache |       | encrypted signaling
                       |       |
   +----------------+  |       |  +--------------------+
   | connect_peer   |--+       +->|  offer / answer    |
   |  (node side)   |             |  handler           |
   +----------------+             +--------------------+
           |                                |
           v                                v
      +---------+                    +--------------+
      |  STUN   |<-- same socket --->|  UDP punch   |
      +---------+                    +--------------+
                                            |
                                            v
                                   adopt_established_traversal()
                                            |
                                            v
                                      FMP IK handshake
                                      on adopted socket

Phase 1 — Advertisement

Adverts are published as Nostr kind 37195 parameterized replaceable events (FIPS-specific, in the application-defined replaceable range 30000–39999; the digits visually spell FIPS — 7=F, 1=I, 9=P, 5=S). The d tag is hardcoded to the wire-format identifier fips-overlay-v1 (or fips-overlay-v1-next on the next branch), so each node has a single, in-place-updatable advert under its identity. The configurable app value populates a separate protocol tag, which scopes adverts within a relay set without splitting them across multiple d-tag streams. The event is signed with the node's FIPS identity key; there is no separate Nostr key. A NIP-40 expiration tag is set to now + advert_ttl_secs, and a version tag carries the protocol version. The advert content is a JSON document shaped as OverlayAdvert (see ../reference/nostr-events.md for the schema).

Publication happens on startup, again whenever the set of advertised endpoints changes (for example, when a Tor onion hostname first becomes available), and on a refresh timer every advert_refresh_secs. If the advertise flag is turned off, the previous advert event is deleted using a NIP-9 kind 5 delete event. Advert publication is fan-out: the same event is sent to every relay in advert_relays with no explicit failover — relay redundancy is implicit.

For a UDP or TCP transport with public: true, the address advertised follows a fixed precedence: an operator-supplied external_addr wins; otherwise a non-wildcard bound local_addr is used directly; otherwise — only for UDP — the runtime asks stun_servers for the reflexive address of the bound socket and advertises that. TCP has no STUN equivalent, so wildcard-bound TCP without external_addr produces a loud WARN and the endpoint is omitted from the advert.

Phase 2 — Lookup

When the node decides to dial a peer that is eligible for Nostr resolution (a via_nostr peer, or any peer under policy: open), it issues a Nostr REQ filtered by author = peer_pubkey, kind = 37195, #d = fips-overlay-v1. The fetch is time-bounded (~2 s) and runs against all configured advert_relays in parallel. The first valid advert wins; adverts whose protocol tag does not match the local app value are rejected at validation.

Results are kept in an in-memory cache keyed by author npub. Cache entries carry the advert's expiration time; a periodic prune drops expired entries, and an LRU-by-expiry eviction enforces advert_cache_max_entries. A parallel long-lived subscription on the advert relays populates the cache passively, so open-discovery candidates do not require per-dial fetches.

On cache hit, advert endpoints are appended to the peer's static address list with lower priority; the static list is tried first.

Phase 3 — Offer/Answer signaling

For any endpoint shaped as udp:nat, dialing triggers an offer/answer exchange before the first packet is sent. Signaling events are Nostr kind 21059 (ephemeral, not stored by conforming relays), gift-wrapped per NIP-59 and encrypted with NIP-44, so only the intended recipient can decrypt the payload.

The initiator performs STUN first (see Phase 4), then builds a TraversalOffer containing:

• A unique sessionId and a random nonce (used to correlate the answer).
• Its reflexive address (if STUN succeeded).
• Its list of local (private) addresses for same-LAN paths.
• The STUN server it used, for informational reporting only.
• An expiresAt equal to now + signal_ttl_secs.

The offer is sealed to the recipient's npub and published to the peer's preferred signaling relays — the node first tries to resolve the peer's NIP-17 DM relay list (kind 10050), and falls back to dm_relays if the inbox-relays fetch fails. Each side also publishes its own inbox relay list on startup so dialers can discover it.

On the receiving side, an inbound semaphore bounds concurrent offer processing at max_concurrent_incoming_offers. When the semaphore is full, the offer is dropped with a warn log; this is the primary guard against offer-spam from a misbehaving or compromised relay. A sessionId replay cache (bounded by seen_sessions_max_entries, with entries valid for replay_window_secs) rejects duplicates.

The responder runs its own STUN query and replies with a TraversalAnswer carrying its reflexive and local addresses plus a PunchHint { startAtMs, intervalMs, durationMs } that tells both sides when to begin probing and how aggressively. If the responder has no usable addresses at all, it replies with accepted: false and a reason string.

Phase 4 — UDP hole-punch

Each side runs STUN (parsing XOR-MAPPED-ADDRESS from the response, all other attributes ignored) on the same UDP socket it will later use for punching and for the adopted FMP transport. This is critical: NAT state is per-socket, so the punch has to reuse the socket that taught the NAT about this binding.

Given its own reflexive + local addresses and the peer's, each side builds a candidate-pair plan that tries, in priority order:

1. Reflexive ↔ reflexive. The classic STUN path. Tried first because it is the only candidate that's reliable across arbitrary network topologies — host candidates from one peer that happen to be reachable from the other (via a corporate VPN, a Tailscale subnet route, or overlapping private address space) will succeed at the socket layer in the punch but fail in the FMP handshake when the return path doesn't match.
2. LAN ↔ LAN. If both sides share a /24 prefix, same-subnet private addresses are likely reachable directly. Only fires when both peers shared local host candidates (which requires share_local_candidates to be enabled — off by default).
3. Mixed. Reflexive on one side, local on the other — catches hairpin and one-side-public scenarios.

At startAtMs both sides begin sending 24-byte probe packets on the candidate pair(s) at intervalMs cadence for up to durationMs. A probe carries a 4-byte magic (NPTC), a 4-byte sequence, and the first 16 bytes of SHA256(sessionId); both sides can compute the same session hash independently from the public sessionId, so no shared secret is needed on the punch path itself. On receiving a valid probe, a side replies with an NPTA ack. The first valid probe or ack seen from the far side records the working remote address and completes the attempt.

On timeout (attempt_timeout_secs as overall bound, punch_duration_ms as probe window), both sides issue NIP-9 deletes for their offer and answer events and report failure up to the discovery runtime's BootstrapEvent::Failed channel.

Phase 5 — Adoption

On success, the discovery runtime emits BootstrapEvent::Established carrying the session id, the punch socket, and the learned remote address. adopt_established_traversal() in the node lifecycle takes the socket, registers it with the UDP transport layer as a new transport instance, and calls initiate_connection() with the peer's FIPS identity as the expected remote. FMP's Noise IK handshake runs on the same socket — there is no "promote link" step between punch and handshake; the punch socket is the FMP socket.

From that moment on, the connection is a normal FMP link and is subject to the usual liveness (MMP heartbeats), rekey, and removal behavior. A link-dead event does not re-enter the discovery runtime automatically; reconnection relies on auto_reconnect and the same dial path that triggered the original punch.

Auto-connect semantics

Discovery does not itself initiate connections. It only supplies addresses. Dial attempts originate from the existing peer-connection machinery:

• Configured peers (peers[] with connect_policy: auto_connect) are dialed on startup and on retry. When via_nostr is set, advert endpoints are appended to the dial list with lower priority than static entries.
• Open discovery peers are assembled from the advert cache, fenced by the peer ACL, and enqueued into a bounded retry queue sized by open_discovery_max_pending. There is no event-driven "connect on every advert" — a peer re-enters the queue only when its prior attempt has drained.
• Manual dials (fipsctl connect) can target any configured peer and use the same dial path, including Nostr resolution if configured.

Rate limits and safeguards

Mechanism	Default	What it prevents	Behavior at limit
Offer semaphore (max_concurrent_incoming_offers)	16	CPU and memory exhaustion from offer spam on DM relays.	Warn log, offer dropped.
Advert cache (advert_cache_max_entries)	2048	Memory growth from ambient advert traffic under policy: open.	LRU-by-expiry eviction.
Seen-sessions (seen_sessions_max_entries)	2048	Replay of stale sessionId values.	Oldest entry evicted.
Signal TTL (signal_ttl_secs)	120 s	Indefinite in-flight offers on relays.	Expired offers rejected at validation.
Open discovery queue (open_discovery_max_pending)	64	Unbounded retry queue under ambient advert load.	New candidates skipped until the queue drains.
Punch window (punch_duration_ms)	10 s	Endless probe traffic after one side has given up.	Attempt declared failed; sockets discarded.
Failure-streak threshold (failure_streak_threshold)	5	Repeated traversal attempts against a peer that keeps failing.	Peer enters extended cooldown.
Extended cooldown (extended_cooldown_secs)	1800 s	Tight retry loops after a failure streak.	Per-peer suppression for the cooldown window.
WARN log throttle (warn_log_interval_secs)	300 s	Log floods from a peer that fails on every attempt.	One WARN per peer per interval; the rest demote to debug.
Failure-state cap (failure_state_max_entries)	4096	Memory growth from per-peer failure tracking.	LRU eviction.

The load-shedding mechanisms (max_concurrent_incoming_offers and the failure-streak / extended-cooldown pair) are deliberately conservative so that a misbehaving relay cannot flood the node with offers and a chronically unreachable peer cannot keep the traversal pipeline saturated. The remaining rows are capacity bounds.

Adverts also undergo a stale-advert sweep: cached entries whose expiresAt has passed are evicted on the periodic prune tick. Inbound signaling tolerates ±60 s of clock skew between sender and receiver, and the runtime maintains an NTP-style skew estimate per remote so that consistently-skewed relays don't trip the freshness check.

Relay model

All configured relays (advert + DM) are opened on a single nostr-sdk::Client at startup. Publication is fan-out: the same event is sent to every relay in the target list, with no explicit retry or relay selection. Redundancy is implicit — a downed relay simply means its copy of the advert or signal is unavailable, while other relays still serve the same data.

For signaling specifically, the node prefers the recipient's NIP-17 DM relays when available (the recipient publishes its DM relay list as a kind 10050 event to its own DM relays on startup) and falls back to the local dm_relays list otherwise. This keeps the common case off the sender's DM relays when those are different from the recipient's, at the cost of one extra NIP-17 fetch per offer.

There is no per-relay rate limiting or health check. The relay model assumes that an operator chooses relays they trust to be best-effort available and that outright misbehavior is handled at the offer semaphore and replay-cache layers downstream.

Security and threat model

• Relay operators can observe metadata. They see which npubs publish adverts, to whom offers are sent, and the timing of that traffic. The contents of offer and answer events are NIP-59/NIP-44 sealed — only the intended recipient decrypts them. Adverts are public by design.
• STUN servers see the node's public IP and port. Only the STUN servers listed in the node's own stun_servers are ever contacted for reflexive discovery. Peer-advertised STUN values are informational; a malicious peer cannot steer this node to a chosen STUN target. See the doc comment on node.discovery.nostr.stun_servers.
• The FIPS identity key signs adverts. Compromise of fips.key is compromise of the node's Nostr identity — an attacker can publish adverts on behalf of the node. The recovery path is the same as for any identity compromise: rotate the key and re-advertise. There is no separate Nostr keypair to rotate independently.
• Tor advertising leaks timing via clearnet relays. When a Tor-only node advertises its onion address, the advert itself is published on clearnet WebSocket relays. Operators who want full unlinkability between the advertising identity and the node's IP must route relay traffic through Tor as well — for example by running fips inside a network namespace with a Tor SOCKS proxy as its only egress, or by pointing advert_relays and dm_relays at onion relay endpoints.
• Open discovery accepts anyone publishing on the same app. Admission control is the peer ACL, not the discovery layer. Verify the ACL before enabling policy: open, and consider using a non-default app value to scope visibility.
• Nothing about discovery bypasses FMP. A successful punch yields a UDP socket with a claimed remote identity. That identity is not trusted until FMP's Noise IK handshake completes. A peer whose advert says "I am npub X at 1.2.3.4:5678" but whose FMP handshake presents a different static key is rejected at the mesh layer.

See also

• ../how-to/enable-nostr-discovery.md — operator activation recipes grouped under three capabilities (resolve, advertise, open) across five scenarios.
• ../tutorials/resolve-peers-via-nostr.md, ../tutorials/advertise-your-node.md, and ../tutorials/open-discovery.md — hand-held walkthroughs of the three capabilities, in pedagogical order.
• ../reference/configuration.md — full configuration reference, including all surrounding keys elided from the scenarios above.
• ../reference/nostr-events.md — Kind 37195 (overlay advert), Kind 21059 (gift-wrapped traversal signaling), Kind 10050 (NIP-17 inbox relay list).
• ../reference/security.md — consolidated security reference, including how the FIPS identity key signs both adverts and Noise handshakes.
• fips-transport-layer.md — UDP, TCP, and Tor transport mechanics; the punch socket is adopted as a normal UDP transport after handoff.
• fips-mesh-layer.md — FMP Noise IK handshake that runs on the adopted socket.
• port-advertisement-and-nat-traversal.md — generic protocol reference (event tags, NIP usage, on-the-wire offer/answer schema, failure-suppression machinery), with the FIPS-specific values called out as worked examples.