The partition-healing / DHT-rendezvous algorithm that used to live in
`src/gossip.rs` is now its own crate, published as
`iroh-gossip-rendezvous` 0.1.0. Bridge depends on it for swarm
discovery; the 760-line gossip module shrinks to ~75 lines of glue.
`dht-partition-healing.md` is reduced to a pointer at the new
crate's `PROTOCOL.md` — the normative spec moved with the code.
### Dependency changes
- **New:** `iroh-gossip-rendezvous = "0.1"` (path + version for local
dev; publish strips the path).
- **Bumped:** `iroh` and `iroh-gossip` 0.97 → 0.98 (major).
- **Removed:** direct deps on `mainline`, `sha2`, `zeroize` — all
now transitive via `iroh-gossip-rendezvous`.
- `donglora-client` spec loosened from `"1.0.0"` to `"1"` so patch
and minor releases of the client propagate automatically.
BREAKING CHANGE: any direct users of `bridge::gossip` internals
relied on the in-bridge partition-healing types that have moved.
Import from `iroh_gossip_rendezvous::…` instead.
The 1.0 bridge running on the async Dongle API from donglora-client 1.0.0 and the Protocol v2 wire format.
* Radio loop is now a tokio task, not a dedicated std::thread. Everything is async end to end.
* [radio] config schema: tx_power renamed to tx_power_dbm (plain i8, no "max" keyword), cad field removed (CAD is a per-TX flag in v1.0), default sync_word changed 0x3444 to 0x1424 to match the SX126x register encoding of MeshCore private sync byte.
* TX retry with backoff (new [tx] config section). Every TX runs through Dongle::tx_with_retry, retrying TX_DONE(CHANNEL_BUSY) and ERR(EBUSY) per PROTOCOL.md §6.10 / §C.5.5. Defaults: 3 attempts, 20-100 ms randomized backoff, doubling to 500 ms cap, 5 s per-attempt timeout. Configurable via max_attempts, backoff_min_ms, backoff_max_ms, backoff_cap_ms, backoff_multiplier, per_attempt_timeout_secs.
* TUI packet log shows retry state: in-flight retries render with a cycle marker in yellow and an N/M attempt counter, terminal TX failures render in red, post-retry success shows final attempt count.
* Stats gain tx_retries and tx_failures counters surfaced in StatsSnapshot.
* Measured airtime from TX_DONE replaces the theoretical estimate, feeding the rate limiter with the real on-air duration.
* The explicit 2 s liveness ping loop is gone. The Dongle runs a 500 ms internal keepalive, matching spec §3.4 inactivity window with 2x margin.
Once the bridge connects via mux, all subsequent reconnects use mux-only
mode (connect_mux_auto). Previously, each reconnect re-ran the full
mux→USB fallback chain, which could steal the serial port from the mux
during a brief disconnect.
Radio reconnection:
- Add connect_mux_auto() to donglora-client — mux-only connect that never
silently falls back to USB serial (matches Python client's sticky mux)
- Fix codec.rs read_frame() to handle WouldBlock alongside TimedOut
- Add liveness ping (2s) to detect dead mux connections
- Treat transmit errors as fatal (trigger reconnect)
- Faster backoff: 250ms initial, 5s max, resets after stable connection
TUI improvements:
- Redesign packet log: RF+Net columns show full packet path (RX→TX),
hash column with dedup ellipsis, fixed-width columns throughout
- Move activity charts to right column (fills remaining width)
- 1-second sparkline buckets instead of 10-second
- Radio panel shows "(disconnected)" with red border when mux is down
- Header: status pills left, clock with TZ right-justified, Ctrl-C quits
- Local time via jiff crate with proper timezone abbreviation
DHT strategy:
- Alone: read every 3s, merge-publish every 30s (preserves existing peers)
- Connected: publish authoritative [self + neighbors], purging stale peers
- Publish immediately on first neighbor join
Signal handling:
- SIGINT and SIGTERM both trigger graceful shutdown via cancellation token
- All paths go through swarm.shutdown() to update DHT before exit
Replace manually configured peer connections with a passphrase-based gossip
swarm. Peers sharing the same passphrase automatically discover each other
via mainline DHT and join an iroh-gossip topic.
Key changes:
- Add gossip.rs: iroh-gossip swarm with DHT bootstrap, heartbeat, and
graceful shutdown that publishes neighbor list
- Add setup.rs: interactive config wizard replacing generate-config
- Add rate_limit.rs: token-bucket TX rate limiter derived from LoRa air time
- Remove network.rs: old direct P2P connection management
- Config: remove [[peers]] section, add passphrase to [bridge], update
radio defaults to MeshCore US/Canada recommended values
- Packet: rename Envelope to GossipFrame, simplify wire format (drop
version byte and seq number)
- Router: unify dedup into single content-hash cache, add rate limiting
- Identity: use ephemeral keypair per launch instead of persistent secret key
- Radio: improve config negotiation (GetConfig-first strategy), add device
info tracking
- TUI: complete redesign with network/swarm panel, sparkline activity
charts, color-coded RSSI/SNR, and shutdown overlay
iroh QUIC datagrams for peer-to-peer networking with NAT traversal,
donglora-client for radio access (direct USB or mux), ratatui TUI
with four-panel dashboard (peers, radio config, stats, packet log).
- Packet envelope with origin tracking, sequence numbers, RSSI/SNR
- Dual dedup: iroh-level (origin+seq) and radio echo (blake3 hash)
- SNR filtering per PROTOCOL.md (forward GOOD+MARGINAL, drop rest)
- Auto-generate config + identity on first run
- Graceful mux config fallback (adopts locked config if SetConfig rejected)
- WouldBlock workaround for donglora-client socket timeout bug
- Reconnect with exponential backoff on dongle disconnect
- CLI: show-id, generate-config subcommands, --log-only headless mode
swaits