v0.2.1

Fix: reqresp completes on empty/short responses instead of hanging to timeout

The reqresp requester only ever reacted to response bytes and ignored the stream FIN. When a responder closed the stream with no chunk — the libp2p "I don't have it" reply (zeam blocks_by_root for a root not in its DB, an empty blocks_by_range, the genesis anchor root every node requests) — the request never completed and hung until the embedder's request timeout (8 s in zeam), retrying forever. In a 3-node devnet this was a relentless ~60/min blocks_by_root retry storm that stalled finalization.

Fix: the requester now completes via zquic v1.7.42's rawAppStreamFullyReceived (FIN seen and all bytes up to the final size contiguously reassembled) — empty/complete responses finish in milliseconds; large in-flight responses are never truncated by a FIN that races ahead of the cwnd-queued payload.

• Bumps zquic to v1.7.42.
• Tests: empty-response-ends-fast regression; the 300 KB reqresp test still passes (no truncation).

Full diff: v0.2.0...v0.2.1

v0.2.0

Highlights

This release resolves the zeam devnet fork end-to-end. Live-validated: a 3-node devnet (with a deliberately delayed sync node) converges on a single head with all nodes finalizing; zero KnownInvalidBlock, zero ACK-starvation teardowns.

Fixes (since v0.1.98)

• zquic v1.7.41 (#247) — splits oversized stream writes into per-packet pending entries. Previously any QUIC stream write larger than one 1-RTT packet that queued under congestion/flow control was silently dropped (no retransmit). libp2p reqresp blocks_by_range bodies (~250 KB beam blocks) were lost, wedging a delayed node's initial sync into a permanent fork; the KnownInvalidBlock symptom was the same bug surfacing as truncated-block verify failures.
• non-blocking QUIC dials (#245, v0.1.98) — handleDial no longer blocks the single drive thread (which deadlocked simultaneous mutual dials in the Initial handshake).

Tests

• #248 — large (~300 KB) reqresp regression test over QuicRuntime loopback (guards the oversized-frame fix).
• Mutual-dial regression test.

Full diff: v0.1.99...v0.2.0

v0.1.99

deps: bump zquic to v1.7.41 (#247).

Fixes silent data loss on any QUIC stream write larger than one 1-RTT packet that is queued under congestion/flow control — the oversized pending entry failed to serialize on drain and was discarded, never retransmitted. libp2p reqresp blocks_by_range bodies (~200 KB/block) were lost, wedging a zeam delayed-node's initial sync into a permanent fork. Small writes (gossip chunks, status, single small blocks) were unaffected.

Full diff: v0.1.98...v0.1.99

v0.1.98

Highlights

quic: non-blocking outbound dials — fixes an Initial-handshake deadlock that forked zeam devnets (#245)

QuicRuntime.handleDial previously spun a blocking drive loop for up to 20s per dial on the single drive thread, and never called pollAccept during it. Two peers dialing each other simultaneously (an all-to-all gossip mesh at boot) would both wedge in the QUIC Initial handshake — neither accepted the other's inbound — time out at 20s (stalled_phase=initial), and starve established connections of ACKs. Downstream a zeam 3-node devnet forked early and never finalized.

Dials are now advanced non-blocking by driveLoop each tick alongside the listener, pollAccept, established outbounds, gossip, and host ticks. Live-validated: a 3-node zeam devnet that previously sat at finalized=slot 0 now justifies steadily (slot 0→18 in lockstep with wall-clock). Adds a simultaneous mutual dial regression test.

Also included since v0.1.97

• #239 re-enable previously-skipped tests
• #242 connection-manager: protect/unprotect, expiry-based trim grace, jittered reconnect backoff
• #243 remove permanently-skipped handshake-sizes test
• #244 swarm: hook deadlines, event-queue backpressure policy, deferred command queue (#212)

Full diff: v0.1.97...v0.1.98

v0.1.97

Maintenance and developer-experience release on top of v0.1.96 — no protocol changes; the public API is unchanged. Full cross-impl interop matrix (zig × go-libp2p × rust-libp2p) green.

Tests

• Re-enabled the QUIC TLS remote peer-id loopback test — it now mints a libp2p TLS certificate (RFC 0001) in memory instead of the deprecated std.fs.cwd file I/O, and verifies the dialer recovers the listener's peer id over a real loopback handshake. (#239, #235)
• Moved the 60s sustained-gossipsub soak out of zig build test into an opt-in zig build soak-test (gated by test_options.enable_soak_tests), keeping the default test run fast. Fixed its assertion: gossipsub does not run the topic validator on the publisher's own messages, so only receivers are checked.

Release process

• Removed release-please; releases are now cut manually. A release is: tag the commit + gh release create vX.Y.Z --target main --notes …, bumping .version in build.zig.zon and the README install snippet in the same change. build.zig.zon .version is now accurate again (it had drifted to 0.1.87 under the old automation). (#238)

Docs

• README Repository layout section + per-folder READMEs for src/{core,primitives,protocols,transport,security}, reflecting the v0.1.96 layout reorg. (#237)

v0.1.96

Feature-heavy release: four new protocol areas, expanded discovery/NAT support, and a repository-layout reorg. Public API is additive — existing zig_libp2p.* exports are unchanged. Full cross-impl interop matrix (zig × go-libp2p × rust-libp2p) green.

New protocols

• Rendezvous /rendezvous/1.0.0 — namespace-scoped peer discovery: client (register / unregister / discover with cookie paging) and server with a registration store. Registrations are bound to the transport peer via signed peer records. (#234, #209)
• mDNS LAN discovery — _p2p._udp.local multicast (IPv4 + IPv6), dnsaddr TXT ingestion, peer_discovered events, Host auto-registerKnownPeer. Discovered dial targets are restricted to private/LAN ranges by default (allow_public_addrs opt-in). (#229, #230, #207)
• gossipsub v1.1 peer scoring — per-topic time-in-mesh / delivery / invalid-message scoring with decay and gossip/publish/graylist/PX thresholds + opportunistic graft. Opt-in via peer_scoring_enabled. (#232, #199)

Kademlia DHT

• Host lifecycle wiring: AutoNAT-driven mode promotion, periodic provider republish, routing-table eviction on disconnect. (#228, #203)
• Pluggable record validators with longest-prefix matching + an IPNS validator implementing the spec (IpnsEntry protobuf, DAG-CBOR data, Ed25519 signatureV2, monotonic sequence, EOL expiry), verified byte-for-byte against a go/boxo reference record. (#231, #233, #198)

NAT traversal & relay

• AutoNAT vote aggregation + active probing (sliding-window quorum, real dial-back verification, observed-IP amplification guard). (#226, #206)
• DCUtR auto-trigger on relayed connections with retry/backoff. (#225, #205)
• Circuit Relay v2 reservation auto-refresh + /p2p-circuit dial path. (#224, #204)
• Identify Push auto-triggered on advertisement changes, streamed over the QUIC runtime. (#222, #223, #202)

Repository & docs

• Repository layout rationalized into core/ · primitives/ · protocols/, QUIC transport split, vendored deps moved to vendor/, build/ helper split + opt-in soak-test. No public-API or import-path changes (@import("zig_libp2p") → src/root.zig). (#236)
• README modernized (libp2p-style) with a spec-coverage matrix; links repointed to the blockblaz org. (#221, #227)

Notes for embedders

• gossipsub peer scoring and mDNS public-addr acceptance are both opt-in; defaults are unchanged.
• The layout reorg keeps every existing zig_libp2p.* export; no consumer changes required.

v0.1.95

gossipsub: FANOUT + FANOUT_TTL (#200)

Adds gossipsub v1.1 FANOUT so a node can publish to topics it is not subscribed to (e.g. a validator publishing to an attestation subnet it doesn't follow).

• New per-topic fanout peer set keyed by topic; on a fanout publish, up to mesh_n connected peers are selected (direct-peer / score / backoff aware) and reused until they disconnect or the entry expires.
• fanout_ttl_ms config (libp2p FANOUT_TTL, default 60 s); heartbeat prunes stale fanout entries.
• subscribe() promotes existing fanout peers straight into the topic mesh.
• Subscribed-topic local publishes now forward directed to mesh peers (spec-aligned with inbound forwarding) instead of broadcasting to all connected peers.
• Owns its topic keys (consistent with the subs/mesh key-ownership model from v0.1.94's interop SIGSEGV fix).
• 4 new unit tests (target selection, reuse-before-TTL, TTL prune, subscribe promotion).

Closes #200. Full cross-impl interop matrix (zig × go-libp2p × rust-libp2p) green.

Note for embedders: publishing on a subscribed topic is now mesh-directed rather than broadcast — confirm your mesh forms before relying on propagation.

v0.1.94

Bumps zquic to v1.7.40: fixes the u5 integer-overflow panic in the 1-RTT PN-candidate sweep that crashed QUIC nodes on undecryptable datagrams >31 bytes. (#220)

v0.1.72

Bump zquic to v1.7.24: the client now decodes the QUIC packet Length field and Handshake CRYPTO frame offset/length with permissive (non-minimal-tolerant) varint decoding on the receive path. Fixes the zeam↔lantern (ngtcp2/AWS-LC) handshake stalling at stalled_phase=initial — strict varint decode was silently dropping the trailing coalesced Handshake packet. (Full zquic↔ngtcp2 libp2p interop has a separate open item: AWS-LC rejects zquic's client cert flight with TLS bad_certificate.)

v0.1.41

v0.1.41

First tagged release since v0.1.17 — rolls up everything merged to main between v0.1.17 and v0.1.41. Highlights from PR #192 (the bug fix that motivated this release):

• gossipsub: inbound IHAVE now generates IWANT for unseen ids (libp2p v1.1 §3.4); previously dropped silently, breaking lazy-gossip recovery against rust- and go-libp2p peers.
• gossipsub: GRAFT during active PRUNE back-off now docks the sender's behaviour score (rust-libp2p P7 weight).
• autonat: v2 amplification cost is now uniformly sampled across [min, max]; the previous std.math.clamp(min, 1, max) always returned the minimum.

See the full changelog below for the 0.1.18 → 0.1.41 history covering security fixes, the /ws WebSocket transport, libp2p-TLS UTCTime parsing fix, AutoNAT v2 + relay (circuit v2) + DCUtR + Kademlia DHT scaffolding, persistent per-peer /meshsub/1.1.0 streams, and matching zquic protocol fixes pulled in along the way.

---

0.1.41 (2026-06-11)

Fixed

• gossipsub: handle inbound IHAVE control messages by emitting IWANT for
  unseen ids (libp2p gossipsub v1.1 §3.4). Previously zig-libp2p decoded IHAVE
  and dropped it on the floor, so lazy-gossip recovery against rust- and
  go-libp2p peers was effectively dead: a peer outside our mesh could announce
  message ids but we never pulled them, breaking mesh healing after a partition
  and degrading dissemination redundancy. The new
  runtime.handleIHaveOffer filters against pull_fifo + recent_seen,
  caps the number of ids accepted per RPC (max_ihave_ids_per_rpc) and the
  number requested in the resulting IWANT (max_iwant_ids_per_rpc) — both
  defaulting to 5000 to match rust-libp2p — and exposes
  iwantTxCount / iwantIdsRequestedCount / ihaveIdsCappedCount.

• gossipsub: dock the sender's behaviour score (configurable via
  graft_during_backoff_score_delta, default -50) when an inbound GRAFT
  arrives during the peer's own active PRUNE back-off window. rust-libp2p
  models this as the P7 weight; we previously refused the GRAFT with a
  PRUNE+remaining-backoff but applied no penalty, so a misbehaving peer
  could flood-GRAFT freely. The unsubscribe-cooldown branch keeps its
  original behaviour (no penalty — that path doesn't represent the same
  flood pattern).

• autonat: v2 amplification cost is now uniformly sampled across the
  full [amplification_min_bytes, amplification_max_bytes] range from a
  SplitMix64 step over the peer nonce. The previous shape
  std.math.clamp(min, 1, max) always returned exactly the minimum
  (30 KiB by default) and silently ignored amplification_max_bytes,
  making the cost trivially predictable to a client and leaving the
  spec-mandated range knob dead.

0.1.40 (2026-06-11)

Fixed

• transport/quic_runtime: detect outbound QUIC connection close as soon as
  CONNECTION_CLOSE is received, not after the 3×PTO draining deadline. zquic
  flips conn.draining = true on CONNECTION_CLOSE receipt but keeps
  conn.phase at .connected until the drain timer reaps the slot — the
  client side never reaps at all today. The previous detector only checked
  phase == .closed, so a remote-initiated close (e.g. rust-libp2p ending the
  session due to gossipsub mesh degrade or transport error) was silent on the
  zeam side: outbound_by_peer retained the dead entry, gossip publishes
  drained into the void, and connection_manager never scheduled a redial.
  Now detectOutboundConnectionClose treats phase == .closed || draining
  as terminal and fires host.onConnectionClosed immediately, mirroring the
  listener-side close path.

0.1.39 (2026-06-11)

Fixed

• transport/quic_runtime: bind the persistent /meshsub/1.1.0 publish stream
  to the outbound (locally-dialed) QUIC connection only. When a rust-libp2p
  peer dials first, opening gossip publish on that inbound leg delivers a few
  frames then stalls once the dialer leg comes up; zeam→ethlambda gossip now
  migrates to the outbound connection and replays SUBSCRIBE there.

0.1.22 (2026-06-10)

Fixed

• transport/quic_runtime: respond to inbound /ipfs/id/push/1.0.0 streams from
  rust-libp2p identify (ethlambda opens push after the initial identify exchange).
  Previously these streams got na → ProtocolNegotiationFailed at startup.
  Also log accumulated byte count on other inbound handshake failures for easier
  diagnosis.

0.1.21 (2026-06-10)

Fixed

• transport/multistream_negotiate, transport/stream_multistream: thread the
  framing detected from the multistream offer line through every subsequent
  token read. The previous per-token first-byte auto-detection
  ('/' ⇒ legacy, otherwise delimited) collided with go-multistream delimited
  framing whenever the varint length byte equalled 0x2F = '/' — i.e. a token
  of total wire length 47 bytes. The lean consensus protocol id
  /leanconsensus/req/blocks_by_root/1/ssz_snappy has a 46-byte body, so its
  delimited wire form starts with '/' and got mis-classified as a legacy
  line. The responder mis-parsed the offer and replied na even though the
  protocol was supported, and the initiator mis-parsed the ack the same way.
  In practice rust-libp2p (ethlambda) observed The remote supports none of the requested protocols for every blocks_by_root request to a
  zig-libp2p peer, blocking chain sync. status and blocks_by_range were
  unaffected because their wire lengths (39 and 48) don't collide with '/'.

Fixed

• transport/quic_runtime: drain the per-stream multistream-select tail
  into the protocol dispatch accumulators (gossipsub, req/resp, relay) before
  reading new bytes from the raw recv buffer. rust-libp2p and go-libp2p
  routinely flush the protocol ack and the first application bytes in a
  single QUIC STREAM frame; those bytes landed in ms_tail after
  negotiation and were never forwarded, so inbound blocks_by_root /
  status requests timed out on the responder (#184).

0.1.19 (2026-06-10)

Fixed

• transport/quic_runtime: persist the multistream-select accumulator across
  drive ticks for inbound streams. Previously a partial responder negotiation
  (DialFailed) lost the bytes the helper had already pulled from the raw
  reader, so the second attempt mis-parsed and the peer saw na for
  legitimate protocols (e.g. rust-libp2p blocks_by_root requests).

0.1.18 (2026-06-10)

Fixed

• transport/quic_runtime: use go-multistream delimited framing on outbound
  req/resp and publish streams so rust-libp2p responders accept the handshake
  (#184). Replace the
  legacy two-newline inbound pre-buffer with incremental responder negotiation.
  Answer inbound /ipfs/id/1.0.0 and /ipfs/ping/1.0.0 streams from
  rust-libp2p peers.