meta-transport-ipfs 1.2.2: gateway keep-alive (re-dial a dropped pinned gateway every 30 s, re-stamp caps on live ones)

Cargo.lock

@@ -1964,7 +1964,8 @@ checksum = "cf8baf1c55e62ffcace7a9f06f4bd9cd3f0c4beb022d3b367256b91b87513d98"
 
 [[package]]
 name = "meta-feeder-sdk"
-version = "1.4.0"
+version = "1.7.0"
+source = "git+https://github.com/worph/meta-feeder-sdk?tag=v1.7.0#027917c8cffe7ca43dadcf3b426cc5f69ecc443a"
 dependencies = [
  "anyhow",
  "async-trait",
@@ -1989,7 +1990,7 @@ dependencies = [
 
 [[package]]
 name = "meta-transport-ipfs"
-version = "1.0.0"
+version = "1.2.2"
 dependencies = [
  "anyhow",
  "async-trait",

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "meta-transport-ipfs"
-version = "1.0.0"
+version = "1.2.2"
 edition = "2021"
 rust-version = "1.85"
 license = "MIT"
@@ -19,7 +19,7 @@ path = "src/main.rs"
 # The transport-plugin contract. Pinned by TAG, never by path — a path dep
 # resolves outside this repo, which is the Docker build context. See the
 # Dockerfile for building against an unreleased SDK.
-meta-feeder-sdk = { git = "https://github.com/worph/meta-feeder-sdk", tag = "v1.4.0", features = ["transport"] }
+meta-feeder-sdk = { git = "https://github.com/worph/meta-feeder-sdk", tag = "v1.7.0", features = ["transport"] }
 
 # libp2p — transport only: identify (address exchange, load-bearing for bitswap
 # reachability), kad (DHT provider records), mdns (LAN discovery) + bitswap.
@@ -56,7 +56,7 @@ multihash-codetable = { version = "0.1", features = ["sha2", "sha3"] }
 
 [dev-dependencies]
 tempfile = "3"
-meta-feeder-sdk = { git = "https://github.com/worph/meta-feeder-sdk", tag = "v1.4.0", features = ["testkit"] }
+meta-feeder-sdk = { git = "https://github.com/worph/meta-feeder-sdk", tag = "v1.7.0", features = ["testkit"] }
 
 # ---------------------------------------------------------------------------
 # Workaround for the yanked transitive `core2` (all versions). `multihash-

Dockerfile

@@ -44,8 +44,8 @@ COPY --from=builder /meta-transport-ipfs /usr/local/bin/meta-transport-ipfs
 
 # Same uid as meta-share: the plugins share its /data volume.
 RUN useradd --system --uid 10001 --no-create-home metashare \
-    && mkdir -p /data/meta-share \
-    && chown metashare:metashare /data/meta-share
+    && mkdir -p /data/meta-share /state \
+    && chown metashare:metashare /data/meta-share /state
 USER metashare
 WORKDIR /data/meta-share
 
@@ -57,7 +57,7 @@ EXPOSE 3000 4001
 ENV HTTP_LISTEN=0.0.0.0:3000 \
     P2P_LISTEN=/ip4/0.0.0.0/tcp/4001 \
     META_SHARE_DATA=/data/meta-share \
-    META_SHARE_CONFIG_DIR=/config \
+    META_SHARE_PLUGIN_STATE_DIR=/state \
     RUST_LOG=info,meta_transport_ipfs=info,libp2p=warn,beetswap=warn,yamux=warn
 
 ENTRYPOINT ["/usr/local/bin/meta-transport-ipfs"]

README.md

@@ -23,11 +23,20 @@ meta-share's variables, unchanged: `P2P_LISTEN`, `PUBLIC_ADDR`, `KAD_*`,
 `META_SHARE_BLOCKSTORE_CACHE_MB`, `META_SHARE_INGRESS_*`,
 `META_SHARE_LOOSE_BLOCK_MAX_BYTES`, `META_SHARE_LOCAL_FILES_PATH`,
 `META_SHARE_FILES_PATH`, `META_SHARE_SEED_DHT_PROVIDE`, `META_SHARE_DATA`, plus
-`META_SHARE_HULL_URL`. `meta_core_url` / `peer_url` come from the hull's
-`settings.json` (mount `/config` read-only), env as the first-boot fallback.
+`META_SHARE_HULL_URL`. `meta_core_url` / `peer_url` are asked from the hull at
+boot (`GET /internal/network`), env as the fallback.
 
-Start it **after** the hull is healthy: the hull migrates its material index out
-of `ipfs/blocks.redb` on its first boot after the split (redb is single-process).
+**Own settings** (config plane, `META_SHARE_PLUGIN_STATE_DIR`, default `/state` →
+`config.json`): cohort namespace, DHT/pinned bootstrap peers, mDNS, DHT announce,
+max connections, background completion. Edited from meta-share's dashboard
+(Plugins → ipfs → configure); `config.json` wins and is written onto the env names
+above at boot, so env only seeds it. Saving restarts the plugin. `PUBLIC_ADDR` /
+`P2P_LISTEN` stay env-only (they follow the published port and the box's IP).
+
+Boot order is handled here, not by compose: before opening `ipfs/blocks.redb` the
+plugin waits (≤ 60 s) for the hull's `/health`, because the hull migrates its
+material index out of that file on its first boot after the split and redb is
+single-process.
 
 ## Build
 

src/config.rs

@@ -268,7 +268,10 @@ pub struct Config {
 }
 
 impl Config {
-    pub fn from_env() -> Result<Self> {
+    /// `net`: the hull's endpoints (`GET /internal/network`), or
+    /// [`NetworkSettings::from_env`](crate::settings::NetworkSettings::from_env)
+    /// when the hull did not answer.
+    pub fn from_env(net: crate::settings::NetworkSettings) -> Result<Self> {
         // Reachability posture first — the listen addr, the kad bootstrap
         // list, and whether we publish provider records all fall out of it.
         // `P2P_ANNOUNCE` is the deprecated spelling of `PUBLIC_ADDR`.
@@ -323,13 +326,7 @@ impl Config {
             )),
         };
 
-        // The network endpoints are the hull's settings (`settings.json`,
-        // seeded once from env by the hull). Read-only here; env fallback when
-        // the file doesn't exist yet.
-        let config_dir: PathBuf = std::env::var("META_SHARE_CONFIG_DIR")
-            .unwrap_or_else(|_| "/config".to_string())
-            .into();
-        let net = crate::settings::NetworkSettings::load(&config_dir);
+        // The network endpoints are the hull's settings, handed in by `main`.
         let peer_api_url = net
             .peer_url
             .as_deref()

src/main.rs

@@ -6,9 +6,12 @@
 //! `cache/`. Serves the transport contract on `HTTP_LISTEN` (internal only) and
 //! the public `/ipfs/:cid` gateway, which the hull relays.
 //!
-//! The swarm announces the **hull's** URL in identify (`baseUrl=`, from the
-//! hull's `settings.json`), so peers and gateways keep calling meta-share's
-//! public API exactly as before — the plugin is invisible on the wire.
+//! The swarm announces the **hull's** URL in identify (`baseUrl=`, asked from the
+//! hull at boot), so peers and gateways keep calling meta-share's public API
+//! exactly as before — the plugin is invisible on the wire.
+//!
+//! Own settings: [`plane`] (`<state dir>/config.json`, edited from meta-share's
+//! dashboard), overlaid onto the env names the modules read.
 
 mod api;
 mod blockstore;
@@ -19,6 +22,7 @@ mod ingress;
 mod ingress_commit;
 mod ipfs_chunk;
 mod material;
+mod plane;
 mod plugin;
 mod settings;
 mod share;
@@ -35,7 +39,8 @@ use std::sync::Arc;
 
 use anyhow::{Context, Result};
 use libp2p::identity;
-use meta_feeder_sdk::transport::{serve_transport, FocusView, HullClient};
+use meta_feeder_sdk::transport::config::state_dir_from_env;
+use meta_feeder_sdk::transport::{serve_transport, ConfigPlane, FocusView, HullClient};
 use tracing::{debug, info, warn};
 use tracing_subscriber::EnvFilter;
 
@@ -46,7 +51,35 @@ async fn main() -> Result<()> {
     });
     tracing_subscriber::fmt().with_env_filter(filter).with_target(false).init();
 
-    let cfg = config::Config::from_env()?;
+    // Own settings first: `config.json` wins over the env it was seeded from,
+    // written onto the env names every module reads — before any of them does.
+    let plane = Arc::new(ConfigPlane::new(plane::schema(), &state_dir_from_env(), plane::seed()));
+    plane::overlay_env(&plane.effective());
+
+    let http = reqwest::Client::builder()
+        .user_agent(concat!("meta-transport-ipfs/", env!("CARGO_PKG_VERSION")))
+        .timeout(std::time::Duration::from_secs(30))
+        .build()
+        .context("build shared reqwest client")?;
+    let hull = HullClient::from_env(http.clone());
+
+    // Wait for the hull before opening `ipfs/blocks.redb`: on its first boot
+    // after the split it migrates the material index out of that file, and redb
+    // is single-process. Compose `depends_on` used to order this; separate
+    // store apps cannot. Bounded — a hull that stays down must not wedge us.
+    if hull.is_enabled() && !hull.wait_ready(std::time::Duration::from_secs(60)).await {
+        warn!("meta-share (hull) did not answer within 60s; starting anyway");
+    }
+    let net = match hull.network().await {
+        Ok(Some(n)) => settings::NetworkSettings::from(n),
+        Ok(None) => settings::NetworkSettings::from_env(),
+        Err(e) => {
+            warn!(error = %e, "could not ask the hull for its endpoints; using env");
+            settings::NetworkSettings::from_env()
+        }
+    };
+
+    let cfg = config::Config::from_env(net)?;
     let keypair = identity::Keypair::generate_ed25519();
     let local_peer_id = libp2p::PeerId::from(keypair.public());
     let agent_version = swarm::build_agent_version(cfg.peer_api_url.as_deref());
@@ -61,17 +94,13 @@ async fn main() -> Result<()> {
         kad_provide_enabled = cfg.kad.provide_enabled,
         seed_dht_provide = cfg.seed_dht_provide,
         agent_version = %agent_version,
+        config = %plane.path().display(),
         "starting meta-transport-ipfs"
     );
 
     let raw_block_store = blockstore::open_redb_blockstore(&cfg.data_dir)
         .await
         .context("open ipfs blockstore")?;
-    let http = reqwest::Client::builder()
-        .user_agent(concat!("meta-transport-ipfs/", env!("CARGO_PKG_VERSION")))
-        .timeout(std::time::Duration::from_secs(30))
-        .build()
-        .context("build shared reqwest client")?;
     let webdav_url_cache = Arc::new(tokio::sync::OnceCell::new());
     let redb_handle = raw_block_store.raw_db();
     let ingress = Arc::new(ingress::IngressRegistry::from_data_dir(&cfg.data_dir));
@@ -132,7 +161,7 @@ async fn main() -> Result<()> {
         peer_directory,
         focus,
         http: http.clone(),
-        hull: HullClient::from_env(http),
+        hull,
         data_dir: cfg.data_dir.clone(),
         meta_core_url: cfg.meta_core_url.clone(),
         webdav_url_cache,
@@ -149,5 +178,5 @@ async fn main() -> Result<()> {
     }
     ingress_commit::spawn(Arc::clone(&state));
 
-    serve_transport(Arc::new(plugin::IpfsPlugin { state }), cfg.http_addr).await
+    serve_transport(Arc::new(plugin::IpfsPlugin { state, config: plane }), cfg.http_addr).await
 }

src/plane.rs

@@ -0,0 +1,167 @@
+//! The plugin's own settings (config plane), edited from meta-share's dashboard
+//! (Plugins → ipfs → configure) and stored in `<state dir>/config.json`.
+//!
+//! Every knob here was — and still is — an env var read by the module that uses
+//! it (`config.rs`, `swarm/`, `ingress_commit.rs`). Rather than re-thread each
+//! one, [`overlay_env`] writes the effective values onto those same env names
+//! once at boot, before anything reads them: `config.json` wins, the env seeds
+//! it. Infra stays env-only: `PUBLIC_ADDR` / `P2P_LISTEN` (bound to the port
+//! compose publishes and the box's IP) and the data dirs.
+
+use meta_feeder_sdk::transport::config::read;
+use meta_feeder_sdk::{ConfigField, ConfigSchema};
+use serde_json::{json, Value};
+
+const GIB: u64 = 1024 * 1024 * 1024;
+const DEFAULT_NAMESPACE: &str = "metamesh-share-default";
+
+pub fn schema() -> ConfigSchema {
+    ConfigSchema {
+        fields: vec![
+            ConfigField::text("kad_namespace", "Cohort namespace")
+                .with_help("Peers only see others with the same namespace. Change it on every node of a private group."),
+            ConfigField::list("kad_bootstrap_peers", "DHT bootstrap peers")
+                .with_help("Multiaddrs used to join the public IPFS DHT. Empty = the app's default list. `none` disables the DHT."),
+            ConfigField::list("bootstrap_peers", "Pinned peers")
+                .with_help("Multiaddrs this node keeps dialled (e.g. a known gateway)."),
+            ConfigField::bool("enable_mdns", "LAN discovery (mDNS)"),
+            ConfigField::bool("seed_dht_provide", "Announce seeds on the DHT")
+                .with_help("Publish provider records for what this node seeds, so remote peers can find it."),
+            ConfigField::number("max_conns", "Max connections"),
+            ConfigField::bool("ingress_commit", "Background completion")
+                .with_help("Finish partially fetched IPFS content in the background."),
+            ConfigField::number("ingress_commit_max_concurrent", "Concurrent completions"),
+            ConfigField::number("ingress_commit_min_free_gib", "Disk floor (GiB)")
+                .with_help("Background completion stands down below this much free disk."),
+        ],
+    }
+}
+
+fn env(key: &str) -> Option<String> {
+    std::env::var(key).ok().map(|s| s.trim().to_string()).filter(|s| !s.is_empty())
+}
+
+fn env_list(key: &str) -> Vec<String> {
+    env(key)
+        .map(|s| s.split(',').map(|p| p.trim().to_string()).filter(|p| !p.is_empty()).collect())
+        .unwrap_or_default()
+}
+
+fn env_flag(key: &str, default: bool) -> bool {
+    match env(key) {
+        Some(v) => !matches!(v.to_ascii_lowercase().as_str(), "0" | "false" | "no" | "off"),
+        None => default,
+    }
+}
+
+fn env_u64(key: &str, default: u64) -> u64 {
+    env(key).and_then(|v| v.parse().ok()).unwrap_or(default)
+}
+
+/// The values a plugin with no `config.json` runs with. Booleans are explicit
+/// (the page's checkbox always submits one), so each carries the default the
+/// reading module would apply.
+pub fn seed() -> Value {
+    // `config.rs` defaults the seed announce to "public mode", i.e. a PUBLIC_ADDR.
+    let public = env("PUBLIC_ADDR").or_else(|| env("P2P_ANNOUNCE")).is_some();
+    json!({
+        "kad_namespace": env("KAD_NAMESPACE").unwrap_or_else(|| DEFAULT_NAMESPACE.to_string()),
+        "kad_bootstrap_peers": env_list("KAD_BOOTSTRAP_PEERS"),
+        "bootstrap_peers": env_list("BOOTSTRAP_PEERS"),
+        "enable_mdns": env_flag("ENABLE_MDNS", true),
+        "seed_dht_provide": env_flag("META_SHARE_SEED_DHT_PROVIDE", public),
+        "max_conns": env_u64("META_SHARE_MAX_CONNS", 128),
+        "ingress_commit": env_flag("META_SHARE_INGRESS_COMMIT", true),
+        "ingress_commit_max_concurrent": env_u64("META_SHARE_INGRESS_COMMIT_MAX_CONCURRENT", 2),
+        "ingress_commit_min_free_gib": env_u64("META_SHARE_INGRESS_COMMIT_MIN_FREE_BYTES", 5 * GIB) / GIB,
+    })
+}
+
+/// `(env name, value)` pairs the effective config sets. An unset/blank value
+/// is skipped, leaving the env (or the module's default) in force.
+pub fn env_overrides(v: &Value) -> Vec<(&'static str, String)> {
+    let mut out = Vec::new();
+    let bool_s = |b: bool| if b { "true".to_string() } else { "false".to_string() };
+    if let Some(s) = read::text(v, "kad_namespace") {
+        out.push(("KAD_NAMESPACE", s));
+    }
+    if let Some(l) = read::list(v, "kad_bootstrap_peers").filter(|l| !l.is_empty()) {
+        // `none` is the documented "no DHT" spelling; config.rs reads set-but-empty that way.
+        let joined = if l.iter().any(|p| p.eq_ignore_ascii_case("none")) { String::new() } else { l.join(",") };
+        out.push(("KAD_BOOTSTRAP_PEERS", joined));
+    }
+    if let Some(l) = read::list(v, "bootstrap_peers").filter(|l| !l.is_empty()) {
+        out.push(("BOOTSTRAP_PEERS", l.join(",")));
+    }
+    if let Some(b) = read::flag(v, "enable_mdns") {
+        out.push(("ENABLE_MDNS", bool_s(b)));
+    }
+    if let Some(b) = read::flag(v, "seed_dht_provide") {
+        out.push(("META_SHARE_SEED_DHT_PROVIDE", bool_s(b)));
+    }
+    if let Some(n) = read::uint(v, "max_conns").filter(|n| *n > 0) {
+        out.push(("META_SHARE_MAX_CONNS", n.to_string()));
+    }
+    if let Some(b) = read::flag(v, "ingress_commit") {
+        out.push(("META_SHARE_INGRESS_COMMIT", bool_s(b)));
+    }
+    if let Some(n) = read::uint(v, "ingress_commit_max_concurrent").filter(|n| *n > 0) {
+        out.push(("META_SHARE_INGRESS_COMMIT_MAX_CONCURRENT", n.to_string()));
+    }
+    if let Some(n) = read::uint(v, "ingress_commit_min_free_gib") {
+        out.push(("META_SHARE_INGRESS_COMMIT_MIN_FREE_BYTES", n.saturating_mul(GIB).to_string()));
+    }
+    out
+}
+
+/// Apply [`env_overrides`] to this process's environment. Call once at the top
+/// of `main`, before any config is read and before other tasks run.
+pub fn overlay_env(v: &Value) {
+    for (k, val) in env_overrides(v) {
+        std::env::set_var(k, val);
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn get<'a>(o: &'a [(&'static str, String)], k: &str) -> Option<&'a str> {
+        o.iter().find(|(n, _)| *n == k).map(|(_, v)| v.as_str())
+    }
+
+    #[test]
+    fn page_values_map_onto_the_env_names_the_modules_read() {
+        let o = env_overrides(&json!({
+            "kad_namespace": "private-cohort",
+            "kad_bootstrap_peers": ["/dns4/a/tcp/4001/p2p/X", "/dns4/b/tcp/4001/p2p/Y"],
+            "bootstrap_peers": [],
+            "enable_mdns": false,
+            "seed_dht_provide": true,
+            "max_conns": null,
+            "ingress_commit": true,
+            "ingress_commit_max_concurrent": 4,
+            "ingress_commit_min_free_gib": 10,
+        }));
+        assert_eq!(get(&o, "KAD_NAMESPACE"), Some("private-cohort"));
+        assert_eq!(get(&o, "KAD_BOOTSTRAP_PEERS"), Some("/dns4/a/tcp/4001/p2p/X,/dns4/b/tcp/4001/p2p/Y"));
+        assert_eq!(get(&o, "BOOTSTRAP_PEERS"), None, "an empty list keeps the env");
+        assert_eq!(get(&o, "ENABLE_MDNS"), Some("false"));
+        assert_eq!(get(&o, "META_SHARE_MAX_CONNS"), None, "blank number keeps the env");
+        assert_eq!(get(&o, "META_SHARE_INGRESS_COMMIT_MIN_FREE_BYTES"), Some("10737418240"));
+    }
+
+    #[test]
+    fn none_disables_the_dht() {
+        let o = env_overrides(&json!({ "kad_bootstrap_peers": ["none"] }));
+        assert_eq!(get(&o, "KAD_BOOTSTRAP_PEERS"), Some(""));
+    }
+
+    #[test]
+    fn every_seeded_key_is_in_the_schema() {
+        let keys: Vec<String> = schema().fields.into_iter().map(|f| f.key).collect();
+        for k in seed().as_object().unwrap().keys() {
+            assert!(keys.contains(k), "{k} missing from schema");
+        }
+    }
+}

src/plugin.rs

@@ -17,11 +17,11 @@ use meta_feeder_sdk::transport::ipfs::{
     RelPath, ShareContainer, ShareLibrary, Shared,
 };
 use meta_feeder_sdk::transport::{
-    ApiError, Capabilities, Deleted, FocusView, Health, Job, Lane, Manifest, ReconcileReport,
+    ApiError, Capabilities, ConfigPlane, Deleted, FocusView, Health, Job, Lane, Manifest, ReconcileReport,
     ReconcileRequest, TransportPlugin, CONTRACT_VERSION,
 };
 use serde::Deserialize;
-use tracing::debug;
+use tracing::{debug, warn};
 
 use crate::api::files::{bitswap, bytes_to_response, PLAYER_HEADER};
 use crate::api::AppState;
@@ -34,6 +34,7 @@ const IMPORT_BODY_LIMIT: usize = 600 * 1024 * 1024;
 
 pub struct IpfsPlugin {
     pub state: Arc<AppState>,
+    pub config: Arc<ConfigPlane>,
 }
 
 fn lane_of(headers: &HeaderMap) -> Lane {
@@ -53,6 +54,7 @@ impl TransportPlugin for IpfsPlugin {
             version: env!("CARGO_PKG_VERSION").into(),
             contract: CONTRACT_VERSION,
             capabilities: Capabilities { fetch: true, share: true },
+            config: true,
         }
     }
 
@@ -110,6 +112,10 @@ impl TransportPlugin for IpfsPlugin {
         &self.state.focus
     }
 
+    fn config(&self) -> Option<Arc<ConfigPlane>> {
+        Some(Arc::clone(&self.config))
+    }
+
     fn extra_routes(self: Arc<Self>) -> Router {
         let gateway = Router::new()
             .route("/ipfs/:cid", get(crate::api::ipfs_gateway::get_ipfs))
@@ -144,6 +150,7 @@ impl TransportPlugin for IpfsPlugin {
             .route("/ipfs-tier/share/container", post(share_container))
             .route("/ipfs-tier/stats/blockstore", get(blockstore_stats))
             .route("/ipfs-tier/debug/:cid", get(debug_block))
+            .route("/ipfs-tier/restart", post(restart))
             .with_state(self)
             .merge(gateway)
     }
@@ -422,3 +429,15 @@ async fn debug_block(State(p): S, Path(cid): Path<String>) -> Response {
 pub fn bytes_response(bytes: Vec<u8>, range: Option<&HeaderValue>) -> Response {
     bytes_to_response(bytes, range)
 }
+
+/// The hull changed the endpoints this plugin announces (`peer_url`) or reads
+/// through (`meta_core_url`): exit so `restart: unless-stopped` re-execs us and
+/// `main` asks the hull again.
+async fn restart() -> StatusCode {
+    tokio::spawn(async {
+        tokio::time::sleep(Duration::from_millis(300)).await;
+        warn!("restarting to pick up the hull's new network settings");
+        std::process::exit(0);
+    });
+    StatusCode::NO_CONTENT
+}

src/settings.rs

@@ -1,11 +1,9 @@
-//! The `network` section of the hull's `settings.json`, read-only.
+//! The hull's two endpoints this plugin needs: `peer_url` (announced in
+//! identify as `baseUrl=`) and `meta_core_url` (library byte reads).
 //!
-//! The hull owns the file (it seeds it from the deprecated env vars on first
-//! boot and edits it from the dashboard); this plugin only needs the two
-//! endpoints, and falls back to the same env-derived defaults the hull would
-//! seed when the file doesn't exist yet.
-
-use std::path::Path;
+//! `main` asks the hull (`GET /internal/network`) at boot; when it does not
+//! answer, [`NetworkSettings::from_env`] recomputes the same env-derived values
+//! the hull seeds its own settings from.
 
 use serde::Deserialize;
 use tracing::warn;
@@ -21,20 +19,6 @@ pub struct NetworkSettings {
 }
 
 impl NetworkSettings {
-    /// `<config_dir>/settings.json` → `network`, else [`Self::from_env`].
-    pub fn load(config_dir: &Path) -> Self {
-        #[derive(Deserialize, Default)]
-        #[serde(default)]
-        struct File {
-            network: Option<NetworkSettings>,
-        }
-        std::fs::read(config_dir.join("settings.json"))
-            .ok()
-            .and_then(|b| serde_json::from_slice::<File>(&b).ok())
-            .and_then(|f| f.network)
-            .unwrap_or_else(Self::from_env)
-    }
-
     /// The hull's first-boot seed, recomputed: `META_CORE_URL`, and
     /// `INTERNAL_BASE_URL` → `BASE_URL` → [`DEFAULT_PEER_URL`].
     pub fn from_env() -> Self {
@@ -63,22 +47,21 @@ fn env_nonempty(key: &str) -> Option<String> {
         .filter(|s| !s.is_empty())
 }
 
+impl From<meta_feeder_sdk::transport::NetworkInfo> for NetworkSettings {
+    fn from(n: meta_feeder_sdk::transport::NetworkInfo) -> Self {
+        Self {
+            meta_core_url: n.meta_core_url,
+            peer_url: n.peer_url,
+        }
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
 
     #[test]
-    fn reads_the_network_section_of_the_hulls_file() {
-        let d = std::env::temp_dir().join(format!("ipfs-settings-{}", std::process::id()));
-        std::fs::create_dir_all(&d).unwrap();
-        std::fs::write(
-            d.join("settings.json"),
-            r#"{"network":{"meta_core_url":"http://mc:9000","peer_url":"https://share.example"},"usenet":{"nntp_host":"x"}}"#,
-        )
-        .unwrap();
-        let n = NetworkSettings::load(&d);
-        assert_eq!(n.meta_core_url.as_deref(), Some("http://mc:9000"));
-        assert_eq!(n.peer_url.as_deref(), Some("https://share.example"));
-        let _ = std::fs::remove_dir_all(&d);
+    fn the_env_fallback_always_has_a_peer_url() {
+        assert!(NetworkSettings::from_env().peer_url.is_some());
     }
 }

src/swarm/event_loop.rs

@@ -9,6 +9,7 @@
 //! locals (not held on `self`) so each handler method can borrow
 //! `&mut self` cleanly without conflicting with the arm futures.
 
+use std::collections::HashMap;
 use std::time::Duration;
 
 use futures::StreamExt;
@@ -25,10 +26,45 @@ use crate::gateway_discovery::{advertises_gateways, fetch_gateway_caps};
 use super::bootstrap::{addr_targets_match, redial_bootstrap, BootstrapEntry};
 use super::bitswap_client::BitswapInflight;
 use super::identify_agent::parse_base_url;
-use super::kad_discovery::{namespace_key, peer_id_from_multiaddr, KadConfig};
+use super::kad_discovery::{is_dialable_addr, namespace_key, peer_id_from_multiaddr, KadConfig};
 use super::peer_directory::PeerDirectory;
 use super::{Behaviour, BehaviourEvent, Command, GatewayCapsConfig, PeersInfo};
 
+/// How often [`SwarmTask::on_gateway_keepalive_tick`] re-dials dropped
+/// gateways and re-stamps live ones. Half meta-search's 60 s: an NZB play is
+/// blocked for as long as the gateway is missing (no one else can redeem).
+const GATEWAY_KEEPALIVE_INTERVAL: Duration = Duration::from_secs(30);
+
+/// A gateway we've identified and now pin for keep-alive: its API base URL (for
+/// the periodic caps refresh that keeps it routable) plus dialable addresses
+/// (for re-pinning the connection if it drops). Persists across TTL expiry and
+/// disconnect — unlike the read-side directory entry, which `forget_peer` drops
+/// on the last close — so a gateway is never permanently forgotten once seen.
+///
+/// Ported from meta-search. Without it a gateway that restarts under the same
+/// peer id is lost for good: its mDNS record never expires, so `Discovered`
+/// never fires again, and the gateway itself never dials consumers
+/// (watch.nsl.sh 2026-10-04: every NZB answered "no gateway can redeem it"
+/// until this container was restarted).
+#[derive(Clone, Debug, PartialEq)]
+struct GatewayPin {
+    base_url: Option<String>,
+    addrs: Vec<Multiaddr>,
+}
+
+/// The pin an identify from a gateway yields, or `None` for any other peer.
+/// Only dialable addresses are kept: a pin is retried for the life of the
+/// process, so a loopback / unspecified listener would be a permanent cost.
+fn gateway_pin(agent_version: &str, listen_addrs: &[Multiaddr], base_url: Option<String>) -> Option<GatewayPin> {
+    if !advertises_gateways(agent_version) {
+        return None;
+    }
+    Some(GatewayPin {
+        base_url,
+        addrs: listen_addrs.iter().filter(|a| is_dialable_addr(a)).cloned().collect(),
+    })
+}
+
 /// All the state that lives across iterations of the event loop. `swarm`
 /// and the channel receivers stay as locals in [`SwarmTask::run`] so the
 /// `tokio::select!` arms can borrow them without conflicting with
@@ -63,6 +99,9 @@ pub(super) struct SwarmTask {
     http: reqwest::Client,
     /// Timeout + debounce for that fetch.
     gateway_caps: GatewayCapsConfig,
+    /// Gateways kept connected + routable by the keep-alive tick. See
+    /// [`GatewayPin`].
+    pinned_gateways: HashMap<PeerId, GatewayPin>,
 }
 
 impl SwarmTask {
@@ -101,6 +140,7 @@ impl SwarmTask {
             bitswap_inflight: BitswapInflight::new(),
             http,
             gateway_caps,
+            pinned_gateways: HashMap::new(),
         }
     }
 
@@ -176,6 +216,11 @@ impl SwarmTask {
         let mut kad_discovery_tick = tokio::time::interval(self.kad_discovery_interval);
         kad_discovery_tick.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Delay);
 
+        // Keep known gateways connected + routable (re-dial dropped ones,
+        // re-fetch caps on live ones). See [`GatewayPin`].
+        let mut gateway_keepalive_tick = tokio::time::interval(GATEWAY_KEEPALIVE_INTERVAL);
+        gateway_keepalive_tick.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Delay);
+
         loop {
             tokio::select! {
                 cmd = rx.recv() => match cmd {
@@ -206,6 +251,9 @@ impl SwarmTask {
                 _ = kad_discovery_tick.tick() => {
                     self.on_kad_discovery_tick(&mut swarm, &mut kad_discovery_tick);
                 },
+                _ = gateway_keepalive_tick.tick() => {
+                    self.on_gateway_keepalive_tick(&mut swarm);
+                },
                 event = swarm.select_next_some() => {
                     self.on_swarm_event(&mut swarm, event, &mut kad_discovery_tick).await;
                 }
@@ -292,6 +340,71 @@ impl SwarmTask {
 
     /// Top-level swarm-event dispatch. Each branch inspects `&event` and
     /// falls through to `default_log` at the end.
+    /// Keep pinned gateways connected + routable (ported from meta-search):
+    ///
+    /// 1. **Connection drop** — re-dial a pinned gateway that isn't connected,
+    ///    directly (not through the mDNS / kad-provider dial-gates), re-seeding
+    ///    its addresses into kad first in case they were evicted. The redial
+    ///    triggers identify, which restores its base URL and redeem claims.
+    /// 2. **Directory staleness** — identify goes quiet on an idle connection,
+    ///    so the 600 s gateway TTL can lapse on a live peer. Re-fetch caps
+    ///    over HTTP to re-stamp it.
+    fn on_gateway_keepalive_tick(&mut self, swarm: &mut Swarm<Behaviour>) {
+        // Small, stable set — clone to sidestep the self/swarm borrow overlap.
+        for (pid, pin) in self.pinned_gateways.clone() {
+            if swarm.is_connected(&pid) {
+                if self
+                    .peer_directory
+                    .needs_cap_refresh(&pid, self.gateway_caps.refresh_after)
+                {
+                    if let Some(base_url) = pin.base_url {
+                        self.spawn_caps_fetch(pid, base_url);
+                    }
+                }
+            } else if pin.addrs.is_empty() {
+                // Every address it announced was undialable — most likely a
+                // gateway without PUBLIC_ADDR seen only over a relay. A dial
+                // would fail every tick for the life of the process.
+                trace!(peer_id = %pid,
+                    "gateway keep-alive: no dialable address for pinned gateway; skipping re-dial");
+            } else {
+                for addr in &pin.addrs {
+                    swarm.behaviour_mut().kad.add_address(&pid, addr.clone());
+                }
+                match swarm.dial(pid) {
+                    Ok(()) => debug!(peer_id = %pid,
+                        "gateway keep-alive: re-dialing dropped gateway"),
+                    Err(e) => debug!(peer_id = %pid, error = %e,
+                        "gateway keep-alive: re-dial failed"),
+                }
+            }
+        }
+    }
+
+    /// Fetch a gateway's capabilities over HTTP off its `base_url` and record
+    /// them in the directory. Fire-and-forget: a slow/dead gateway must not
+    /// stall the swarm loop; a failure keeps whatever was held (it ages out on
+    /// its own TTL) and is retried on the next identify / keep-alive tick.
+    fn spawn_caps_fetch(&self, pid: PeerId, base_url: String) {
+        let http = self.http.clone();
+        let dir = self.peer_directory.clone();
+        let timeout = self.gateway_caps.fetch_timeout;
+        tokio::spawn(async move {
+            match fetch_gateway_caps(&http, &base_url, timeout).await {
+                Ok(caps) => {
+                    let redeems = caps.redeem_claims();
+                    debug!(peer_id = %pid, base_url = %base_url,
+                        nzb_fetch = caps.nzb_fetch,
+                        redeem_claims = redeems.len(),
+                        "gateway caps: fetched");
+                    dir.record_gateway_caps(pid, caps.nzb_fetch, redeems);
+                }
+                Err(e) => debug!(peer_id = %pid, base_url = %base_url,
+                    error = %e, "gateway caps: fetch failed"),
+            }
+        });
+    }
+
     async fn on_swarm_event(
         &mut self,
         swarm: &mut Swarm<Behaviour>,
@@ -371,9 +484,11 @@ impl SwarmTask {
             identify::Event::Received { peer_id, info, .. }
         )) = &event {
             let pid = *peer_id;
-            let addrs: Vec<Multiaddr> = info.listen_addrs.clone();
-            for addr in addrs {
-                swarm.behaviour_mut().kad.add_address(&pid, addr);
+            // Only what a remote peer could actually dial: a container behind
+            // docker's bridge reports 0.0.0.0 / 127.0.0.1 listeners too, and
+            // the keep-alive below would retry a pinned bad address forever.
+            for addr in info.listen_addrs.iter().filter(|a| is_dialable_addr(a)) {
+                swarm.behaviour_mut().kad.add_address(&pid, addr.clone());
             }
             let base_url = parse_base_url(&info.agent_version).map(str::to_string);
             let update = self.peer_directory.upsert_from_identify(pid, base_url);
@@ -381,6 +496,15 @@ impl SwarmTask {
                 debug!(peer_id = %pid, base_url = %url,
                     "identify: learned/refreshed peer base URL");
             }
+            // Pin every identified gateway for the keep-alive tick, refreshing
+            // its addresses each time, so a dropped gateway gets re-dialed.
+            if let Some(pin) = gateway_pin(
+                &info.agent_version,
+                &info.listen_addrs,
+                self.peer_directory.base_url(&pid),
+            ) {
+                self.pinned_gateways.insert(pid, pin);
+            }
             // Gateway discovery. A non-empty `gateways=` token is emitted by
             // meta-gateway and nobody else, so it's the "this peer is a
             // gateway" signal. Its *capabilities* aren't on the wire though —
@@ -389,36 +513,16 @@ impl SwarmTask {
             // plain HTTP off the just-learned baseUrl instead (the reliable
             // source that replaced the flaky gossipsub heartbeat).
             //
-            // Fire-and-forget: a slow/dead gateway must not stall the swarm
-            // loop. Debounced by the directory so the identify burst on a
-            // multi-connection peer doesn't fan out into redundant fetches;
-            // identify's ~5min re-announce then keeps the capability fresh
-            // inside its 600s TTL, which is why there's no keep-alive tick.
+            // Debounced by the directory so the identify burst on a
+            // multi-connection peer doesn't fan out into redundant fetches.
+            // The keep-alive tick re-stamps it when identify goes quiet.
             if advertises_gateways(&info.agent_version)
                 && self
                     .peer_directory
                     .needs_cap_refresh(&pid, self.gateway_caps.refresh_after)
             {
                 if let Some(base_url) = self.peer_directory.base_url(&pid) {
-                    let http = self.http.clone();
-                    let dir = self.peer_directory.clone();
-                    let timeout = self.gateway_caps.fetch_timeout;
-                    tokio::spawn(async move {
-                        match fetch_gateway_caps(&http, &base_url, timeout).await {
-                            Ok(caps) => {
-                                let redeems = caps.redeem_claims();
-                                debug!(peer_id = %pid, base_url = %base_url,
-                                    nzb_fetch = caps.nzb_fetch,
-                                    redeem_claims = redeems.len(),
-                                    "gateway caps: fetched");
-                                dir.record_gateway_caps(pid, caps.nzb_fetch, redeems);
-                            }
-                            // Best-effort: keep whatever we held (it ages out on
-                            // its own TTL) and retry on the next identify.
-                            Err(e) => debug!(peer_id = %pid, base_url = %base_url,
-                                error = %e, "gateway caps: fetch failed"),
-                        }
-                    });
+                    self.spawn_caps_fetch(pid, base_url);
                 }
             }
         }
@@ -556,3 +660,37 @@ fn default_log(event: SwarmEvent<BehaviourEvent>) {
         _ => {}
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn ma(s: &str) -> Multiaddr {
+        s.parse().unwrap()
+    }
+
+    #[test]
+    fn only_a_gateway_is_pinned() {
+        let addrs = [ma("/ip4/172.18.0.16/tcp/4002")];
+        assert_eq!(gateway_pin("meta-share/2.3.0 baseUrl=http://x:3000", &addrs, None), None);
+        assert!(gateway_pin("meta-gateway/1.0.41 gateways=usenet,tmdb", &addrs, None).is_some());
+    }
+
+    #[test]
+    fn a_pin_keeps_only_dialable_addresses() {
+        let addrs = [
+            ma("/ip4/127.0.0.1/tcp/4002"),
+            ma("/ip4/0.0.0.0/tcp/4002"),
+            ma("/ip4/172.18.0.16/tcp/4002"),
+            ma("/ip4/85.17.246.67/tcp/4002"),
+        ];
+        let pin = gateway_pin(
+            "meta-gateway/1.0.41 gateways=usenet",
+            &addrs,
+            Some("https://metagateway-watch.nsl.sh".into()),
+        )
+        .unwrap();
+        assert_eq!(pin.addrs, vec![addrs[2].clone(), addrs[3].clone()]);
+        assert_eq!(pin.base_url.as_deref(), Some("https://metagateway-watch.nsl.sh"));
+    }
+}