diff --git a/Cargo.lock b/Cargo.lock index d95ae5b5..abcdcced 100644 --- a/Cargo.lock +++ b/Cargo.lock @@ -1673,6 +1673,7 @@ dependencies = [ "clap-markdown", "clap_complete", "forgejo-api", + "futures-util", "hive-agent-sock", "hive-core-agent-sock", "hive-host-sock", diff --git a/docs/knowledge.md b/docs/knowledge.md index 5f2b00d1..604ad8be 100644 --- a/docs/knowledge.md +++ b/docs/knowledge.md @@ -38,18 +38,23 @@ create a new document. hive-c0re maintains the local clone at `/var/lib/hyperhive/knowledge` via two paths: -1. **Forgejo push webhook** — `ensure_webhook` registers a push - hook on `internal/knowledge` at startup pointing at - `https:///webhook/knowledge` (routed through the - gateway, avoiding the Forgejo SSRF guard that blocks loopback - delivery). On any push to main (including merge commits) hive-c0re - runs `git pull` so agents see the new content on their next turn. - The endpoint is protected by an auto-generated HMAC secret that - hive-c0re verifies on every delivery. +1. **Swarm event** — the swarm controller holds the single push hook on + `internal/knowledge` (see `docs/swarm/README.md` § Swarm-wide forge + webhooks). On any push to main, including merge commits, it sends an + event to every hive over the swarm queue and each hive runs `git + pull`, so agents see the new content on their next turn. - On a swarm, a **second** hook on the same repo points at the - swarm controller — see `docs/swarm/README.md` § Swarm-wide forge - webhooks. Both are expected; neither should be deleted. + A hive that is offline when the event is sent does not get it on + reconnect — the periodic pull below is what closes that gap. So one + hive briefly showing older `/knowledge` content than another is + expected, and resolves by itself within the fallback interval. + + **Do not add a per-hive hook.** A webhook has exactly one target + URL, so a second registration against the same repo does not add a + recipient — it takes delivery away from whoever registered first. + Earlier versions had each hive register its own; hive-c0re now + removes its own leftover at startup, so no operator step is needed + to migrate. 2. **Periodic pull** — a background task in `hive-c0re::main` pulls on a fixed cadence as a fallback (webhook missed, c0re diff --git a/docs/swarm/README.md b/docs/swarm/README.md index 2338b0fd..853a44cc 100644 --- a/docs/swarm/README.md +++ b/docs/swarm/README.md @@ -356,18 +356,23 @@ itself — a `push` hook on `internal/knowledge` and a `pull_request` hook on the `agent-configs` org, both under `https:///webhook/forge/`. -**Two hooks exist per swarm-wide repo: each hive's own, plus the -controller's.** Both are expected — **do not delete either.** Removing a -hive's hook stops that hive acting on knowledge pushes and config PRs; -removing the controller's just gets recreated on its next start. - The controller **interprets** a delivery and sends hives a specific message — *the knowledge repo changed*, *deploy agent `foo` at rev `abc123`* — rather than forwarding forge payloads for each hive to re-derive. Approval happens once, at the swarm level: a hive receives a -decision, not an event to adjudicate. Today the controller logs each -verified delivery and sends nothing, because the swarm→hive channel does -not exist yet; the hive-side hooks are what act in the meantime. +decision, not an event to adjudicate. + +**`internal/knowledge` is on that path.** The controller's is the only +hook on it: hives no longer register their own, and each removes its +leftover at startup. A webhook has exactly one target URL, so per-hive +registration never added a recipient — it took delivery away from +whichever hive registered before it. + +**The `agent-configs` org is not yet.** Each hive still registers its own +`pull_request` hook there, so that repo has two — the hive's and the +controller's — and **both are expected; do not delete either.** Removing +a hive's stops it acting on config PRs; removing the controller's just +gets recreated on its next start. Nothing to configure. The hooks are registered only when this host also serves the swarm UI vhost — that is what publishes the endpoint, and a diff --git a/hive-c0re/Cargo.toml b/hive-c0re/Cargo.toml index 6ffc9549..68440cc2 100644 --- a/hive-c0re/Cargo.toml +++ b/hive-c0re/Cargo.toml @@ -8,6 +8,9 @@ readme = "README.md" workspace = true [dependencies] +# For `StreamExt::next` on the swarm-event subscription in `swarm_status`. +# Workspace-level, same version swarm-controller already uses — not a second copy. +futures-util.workspace = true anyhow.workspace = true # Named directly only for the client type the swarm status publisher passes # around; the connect itself lives in `swarm-queue-client` below. diff --git a/hive-c0re/src/dashboard/mod.rs b/hive-c0re/src/dashboard/mod.rs index 9fc491b2..1a820d13 100644 --- a/hive-c0re/src/dashboard/mod.rs +++ b/hive-c0re/src/dashboard/mod.rs @@ -184,7 +184,6 @@ pub async fn serve( .routes(routes!(schedules::post_schedule_resume)) .routes(routes!(schedules::post_schedule_fire_now)) .routes(routes!(schedules::post_rebuild_queue_cancel)) - .routes(routes!(webhook::post_webhook_knowledge)) .routes(routes!(webhook::post_webhook_config_pr)) .routes(routes!(approvals::post_approve)) .routes(routes!(approvals::post_deny)) @@ -430,7 +429,6 @@ mod router_build_probe { .routes(routes!(schedules::post_schedule_resume)) .routes(routes!(schedules::post_schedule_fire_now)) .routes(routes!(schedules::post_rebuild_queue_cancel)) - .routes(routes!(webhook::post_webhook_knowledge)) .routes(routes!(webhook::post_webhook_config_pr)) .routes(routes!(approvals::post_approve)) .routes(routes!(approvals::post_deny)) diff --git a/hive-c0re/src/dashboard/webhook.rs b/hive-c0re/src/dashboard/webhook.rs index 248c6c03..ce993841 100644 --- a/hive-c0re/src/dashboard/webhook.rs +++ b/hive-c0re/src/dashboard/webhook.rs @@ -1,12 +1,18 @@ //! Forgejo webhook endpoints. //! -//! - **`/webhook/knowledge`** — push events on `internal/knowledge` trigger a -//! `git pull` on the local clone so agents see up-to-date docs. //! - **`/webhook/config-pr`** — `pull_request` events on any `agent-configs/*` //! repo queue a [`hive_sh4re::approvals::ApprovalKind::MergeConfigPr`] approval row //! so the operator can review + approve the merge from the dashboard. //! -//! Both endpoints are reached via the gateway (HTTPS, public domain URL) so +//! There was a second endpoint here, `/webhook/knowledge`, which pulled the +//! local `internal/knowledge` clone on push. It is gone along with the +//! per-hive registration that fed it: a webhook has exactly one target URL, +//! so every hive registering one against the shared repository was +//! last-writer-wins. The swarm controller now holds the single registration +//! and addresses an event to each hive over the queue, which +//! [`crate::workers::knowledge`] documents. +//! +//! The endpoint is reached via the gateway (HTTPS, public domain URL) so //! Forgejo's SSRF guard does not block delivery. Each delivery is verified //! against the `X-Hub-Signature-256` HMAC header Forgejo attaches; the //! shared secret is auto-generated at startup and persisted to @@ -42,104 +48,6 @@ fn verify_hmac(state: &AppState, headers: &HeaderMap, body: &Bytes) -> Result<() crate::webhook_secret::verify_signature(secret, body, sig).map_err(|e| e.to_string()) } -// ── knowledge webhook ────────────────────────────────────────────────────────── - -/// Minimal Forgejo push-webhook payload — only the fields we care about. -#[derive(Deserialize)] -pub(super) struct PushWebhookPayload { - #[serde(rename = "ref")] - git_ref: Option, - repository: Option, -} - -#[derive(Deserialize)] -pub(super) struct PushWebhookRepo { - full_name: Option, -} - -/// POST `/webhook/knowledge` — Forgejo push webhook for -/// `internal/knowledge`. -/// -/// Runs `git pull` on the local clone so agents see up-to-date documents -/// on their next turn. -/// -/// Expected Forgejo webhook configuration: -/// - URL: `https:///webhook/knowledge` -/// - Content type: `application/json` -/// - Event: "Push" (fires on merge commits to main as well) -/// - Secret: auto-generated HMAC key (see [`crate::webhook_secret`]) -/// -/// The gateway routes `/webhook/` → hive-c0re; the HMAC secret protects -/// the endpoint from unauthenticated callers. -#[utoipa::path( - post, - path = "/webhook/knowledge", - request_body( - content = String, - content_type = "application/json", - description = "Forgejo push-webhook payload, taken as raw bytes \ - (not a typed extractor) so HMAC verification runs \ - over the exact wire bytes before any JSON parsing" - ), - responses( - (status = 200, description = "processed (pull triggered or ignored)", body = String), - (status = 400, description = "invalid JSON payload"), - (status = 401, description = "bad or missing HMAC signature"), - (status = 503, description = "HMAC secret unavailable at startup"), - ), - tag = "webhook" -)] -pub(super) async fn post_webhook_knowledge( - State(state): State, - headers: HeaderMap, - body: Bytes, -) -> Response { - if let Err(e) = verify_hmac(&state, &headers, &body) { - tracing::warn!("webhook/knowledge: HMAC verification failed: {e}"); - let status = if e.contains("unavailable") { - StatusCode::SERVICE_UNAVAILABLE - } else { - StatusCode::UNAUTHORIZED - }; - return (status, e).into_response(); - } - - let payload = match serde_json::from_slice::(&body) { - Ok(p) => p, - Err(e) => { - tracing::warn!("webhook/knowledge: JSON parse error: {e}"); - return (StatusCode::BAD_REQUEST, "invalid JSON").into_response(); - } - }; - - let expected_repo = format!("{}/{}", crate::knowledge::ORG, crate::knowledge::REPO); - let full_name = payload - .repository - .as_ref() - .and_then(|r| r.full_name.as_deref()) - .unwrap_or(""); - if full_name != expected_repo { - tracing::debug!( - full_name, - "webhook/knowledge: ignoring push from unexpected repo" - ); - return (StatusCode::OK, "ignored").into_response(); - } - let git_ref = payload.git_ref.as_deref().unwrap_or(""); - if git_ref != "refs/heads/main" { - tracing::debug!(git_ref, "webhook/knowledge: ignoring non-main push"); - return (StatusCode::OK, "ignored").into_response(); - } - tracing::info!("webhook/knowledge: pull triggered by push to {expected_repo}"); - let coord = state.coord.clone(); - tokio::spawn(async move { - if let Err(e) = crate::knowledge::pull(&coord).await { - tracing::warn!(error = ?e, "webhook/knowledge: pull failed"); - } - }); - (StatusCode::OK, "ok").into_response() -} - // ── config-PR webhook ────────────────────────────────────────────────────────── /// Minimal Forgejo `pull_request`-webhook payload. diff --git a/hive-c0re/src/forge/mod.rs b/hive-c0re/src/forge/mod.rs index 5d5bb182..4da942b1 100644 --- a/hive-c0re/src/forge/mod.rs +++ b/hive-c0re/src/forge/mod.rs @@ -483,8 +483,7 @@ pub async fn ensure_all() { /// existing hooks first and skips creation when one is already targeting the /// correct URL. /// -/// `hive_domain` is the public domain name of the hive (e.g. -/// `pr1ma.darkest.space`); the webhook URL is +/// `hive_domain` is the public domain name of the hive; the webhook URL is /// `https:///webhook/config-pr` (routed through the gateway, /// avoiding the Forgejo SSRF guard that blocks loopback delivery). /// @@ -495,8 +494,9 @@ pub async fn ensure_all() { /// An org-level hook covers every repo in `agent-configs` automatically, /// so no per-repo setup is needed as new agents are provisioned. /// -/// Called at startup alongside `knowledge::ensure_webhook`. No-op when the -/// core token is absent (forge not yet provisioned). +/// Called at startup beside `knowledge::remove_webhook`, its opposite: that +/// repo's one hook is the controller's now, this one has not moved yet. No-op +/// when the core token is absent (forge not yet provisioned). /// /// # Errors /// diff --git a/hive-c0re/src/job_queue/exec.rs b/hive-c0re/src/job_queue/exec.rs index 51578c0e..5d55b9cf 100644 --- a/hive-c0re/src/job_queue/exec.rs +++ b/hive-c0re/src/job_queue/exec.rs @@ -179,10 +179,18 @@ async fn run_matrix_sweep() -> Result<()> { } } -/// Boot-time Forgejo webhook registration as a DAG node — see +/// Boot-time Forgejo webhook management as a DAG node — see /// [`NodeKind::WebhookRegister`]. Mirrors the guard chain the /// `tokio::spawn` block it replaced used: no-op (not an error) when the /// HMAC secret, core token, or hive domain aren't available yet. +/// +/// The node now does one of each: it still registers the config-PR hook, +/// and it *removes* the knowledge one. A knowledge push is delivered to +/// the swarm controller, which addresses an event to each hive over the +/// queue — so a hive holding its own registration is holding a shared +/// resource only one party can own. The removal runs every boot rather +/// than behind a marker because it is already idempotent: it is a no-op +/// the moment the hook is gone. async fn run_webhook_register() -> Result<()> { let Ok(webhook_secret) = crate::webhook_secret::load_or_generate() else { tracing::debug!("webhook secret unavailable; skipping hook registration"); @@ -198,10 +206,8 @@ async fn run_webhook_register() -> Result<()> { tracing::debug!("HYPERHIVE_HIVE_DOMAIN unset; skipping webhook registration"); return Ok(()); }; - if let Err(e) = - crate::workers::knowledge::ensure_webhook(&token, &domain, &webhook_secret).await - { - tracing::warn!(error = ?e, "knowledge: ensure_webhook failed"); + if let Err(e) = crate::workers::knowledge::remove_webhook(&token, &domain).await { + tracing::warn!(error = ?e, "knowledge: remove_webhook failed"); } if let Err(e) = crate::forge::ensure_config_pr_webhook(&token, &domain, &webhook_secret).await { tracing::warn!(error = ?e, "forge: ensure_config_pr_webhook failed"); diff --git a/hive-c0re/src/main.rs b/hive-c0re/src/main.rs index 9efe1d51..5e5d6f06 100644 --- a/hive-c0re/src/main.rs +++ b/hive-c0re/src/main.rs @@ -483,7 +483,7 @@ async fn cmd_serve( // A no-op on a standalone hive (no queue env, logged once) — see // swarm_status, which owns the whole task including its own decision // not to start. - swarm_status::spawn(coord.shutdown_rx()); + swarm_status::spawn(std::sync::Arc::clone(&coord), coord.shutdown_rx()); // Per-agent events.sqlite + bash-tasks file cleanup now runs // agent-side in the harness (`hive_agent::vacuum`): the files are // agent-owned, so host-side deletes hit PermissionDenied / readonly-db diff --git a/hive-c0re/src/swarm_status.rs b/hive-c0re/src/swarm_status.rs index 822b8564..d4310b91 100644 --- a/hive-c0re/src/swarm_status.rs +++ b/hive-c0re/src/swarm_status.rs @@ -31,6 +31,9 @@ use std::time::Duration; use anyhow::{Context, Result}; +// `Subscriber` is a `Stream`, so reading the next event needs the extension +// trait — there is no inherent `next()` on it. +use futures_util::StreamExt as _; use crate::stats::sweep_health::{self, SweepHealth}; @@ -62,7 +65,10 @@ const FAILURES_BEFORE_BANNER: u32 = 3; /// makes it a hard error; it is bannered here rather than swallowed, /// because the failure it otherwise produces is a hive that looks fine /// and silently never reports. -pub fn spawn(mut shutdown: tokio::sync::watch::Receiver) { +pub fn spawn( + coord: std::sync::Arc, + mut shutdown: tokio::sync::watch::Receiver, +) { let cfg = match swarm_queue_client::QueueConfig::from_env(ENV_PREFIX) { Ok(Some(cfg)) => cfg, Ok(None) => { @@ -123,6 +129,14 @@ pub fn spawn(mut shutdown: tokio::sync::watch::Receiver) { } }; + // The hive's ONE queue connection, now serving both directions: + // status goes up, swarm events come down. A second `connect` would + // double the auth-callout traffic against authelia and give the two + // paths independent reconnect state, so one could be serving while + // the other was still down. `async_nats::Client` is a handle, so the + // clone is cheap. + tokio::spawn(drain_swarm_events(client.clone(), coord, shutdown.clone())); + let mut health = SweepHealth::new("swarm_status_publish", "warn", FAILURES_BEFORE_BANNER); loop { match publish(&client, &hive).await { @@ -160,6 +174,84 @@ pub fn spawn(mut shutdown: tokio::sync::watch::Receiver) { }); } +/// Listen on the swarm's knowledge-event subject and act on what arrives. +/// +/// The controller decides *what a forge delivery means* and addresses the +/// result here; this end does not know a forge exists. Today the one event is +/// **the knowledge repository changed**, and the response is the pull this +/// daemon already runs at boot. +/// +/// # There is no payload, and that is deliberate +/// +/// The event carries nothing. The webhook handler this replaces read two +/// fields from Forgejo and used neither — both were filters — then ran +/// `git pull`, which re-derives everything from the repository. So it is an +/// edge trigger, and reading a body here would invent a contract nobody owes. +/// +/// # What a missed message costs +/// +/// Core NATS, so delivery is at-most-once: a hive that is down when the +/// controller publishes never hears it, and its knowledge stays as of its last +/// pull until it next boots. **That is not a regression** — a webhook delivery +/// to a hive that is down is lost identically, and this daemon pulls at startup +/// regardless. `JetStream` would require this end to *publish* to +/// `$JS.API.CONSUMER.CREATE.`, which the callout policy does not grant, +/// so durability would cost a grant on both sides to remove a failure the boot +/// pull already covers. +/// +/// ⚠️ **A refused subscription is indistinguishable from a quiet one.** NATS +/// reports an authorization violation asynchronously on the connection, not as +/// an error from `subscribe`, so this task cannot tell "no events published" +/// from "not allowed to hear them". If a hive stops picking up knowledge +/// changes, the server log is the thing that knows why — nothing here will say. +async fn drain_swarm_events( + client: async_nats::Client, + coord: std::sync::Arc, + mut shutdown: tokio::sync::watch::Receiver, +) { + // One subject for the whole swarm, so this hive's own name never enters + // it: the controller publishes once and core NATS fans out to whoever is + // subscribed. + let subject = swarm_queue_client::KNOWLEDGE_SUBJECT; + let mut sub = match client.subscribe(subject).await { + Ok(sub) => sub, + Err(e) => { + // Warn rather than a boot banner: the hive is fully functional + // without this, it just falls back to learning about knowledge + // changes at its next boot. + tracing::warn!( + %subject, error = %e, + "swarm events: subscribe failed; this hive will not hear knowledge changes" + ); + return; + } + }; + tracing::info!(%subject, "swarm events: listening"); + + loop { + tokio::select! { + msg = sub.next() => { + if msg.is_none() { + // The subscription ended — the connection went away for + // good. Returning is right: `async-nats` reconnects + // underneath a live subscription, so a closed stream is + // not a blip this should spin on. + tracing::warn!(%subject, "swarm events: subscription closed"); + return; + } + tracing::info!(%subject, "swarm events: knowledge change announced, pulling"); + if let Err(e) = crate::workers::knowledge::pull(&coord).await { + tracing::warn!(error = ?e, "swarm events: knowledge pull failed"); + } + } + _ = shutdown.changed() => { + tracing::info!("swarm events: shutdown signal received"); + return; + } + } + } +} + /// Offer one snapshot: this hive's current readiness, under its own key. async fn publish(client: &async_nats::Client, hive: &str) -> Result<()> { // An unconnected client does not fail a JetStream request, it hangs diff --git a/hive-c0re/src/workers/knowledge.rs b/hive-c0re/src/workers/knowledge.rs index dbd249a9..c2e36079 100644 --- a/hive-c0re/src/workers/knowledge.rs +++ b/hive-c0re/src/workers/knowledge.rs @@ -6,15 +6,18 @@ //! contribute by forking the repo and opening PRs — they never write //! to the bind-mounted path inside the container. //! -//! hive-c0re maintains the local clone. A Forgejo webhook notifies it -//! on push to main so agents always see an up-to-date snapshot. The -//! webhook is auto-created by [`ensure_webhook`] at startup. A +//! hive-c0re maintains the local clone. It learns that the repository +//! moved from the **swarm controller**, which owns the one Forgejo +//! webhook and addresses an event to each hive over the queue; a //! periodic pull in `main.rs` provides a fallback cadence. - -use std::collections::BTreeMap; +//! +//! A hive used to register that webhook itself, pointing at its own +//! `/webhook/knowledge`. A webhook has exactly one target URL, so with +//! more than one hive that was last-writer-wins rather than idempotent — +//! every hive but the most recent silently stopped receiving deliveries. +//! [`remove_webhook`] is the migration off it. use anyhow::{Context, Result}; -use forgejo_api::structs::{CreateHookOption, CreateHookOptionConfig, CreateHookOptionType}; use crate::coordinator::Coordinator; use crate::forge::{core_auth_header, forge_git_url}; @@ -142,103 +145,71 @@ async fn seed_readme(core_token: &str) -> Result<()> { } } -/// Ensure a Forgejo push webhook for `internal/knowledge` exists and -/// points at hive-c0re's `/webhook/knowledge` endpoint. Idempotent — -/// lists existing hooks first and skips creation when one is already -/// targeting the correct URL. +/// Delete this hive's own `internal/knowledge` push webhook if it is +/// still registered, so the swarm controller is the only party holding +/// one. /// -/// `hive_domain` is the public domain name of the hive; the webhook URL is -/// `https:///webhook/knowledge` (routed through the gateway, -/// avoiding the Forgejo SSRF guard that blocks loopback delivery). +/// # Why this is a migration and not just a deletion /// -/// `webhook_secret` is the HMAC secret Forgejo will attach as -/// `X-Hub-Signature-256` on each delivery; hive-c0re verifies this header -/// in the dashboard webhook handler (`post_webhook_knowledge`). +/// Not registering any more fixes nothing on a hive that has already +/// run: the hook it created persists on the forge, so the contention +/// this removes would survive on exactly the deployments that have it +/// while fresh installs looked fixed. The hive that created a hook is +/// the one that removes it. /// -/// Called at startup alongside [`ensure_local_clone`]. No-op when the -/// core token is absent (forge not yet provisioned). -pub async fn ensure_webhook( - core_token: &str, - hive_domain: &str, - webhook_secret: &str, -) -> Result<()> { +/// # It removes only its OWN hook, never a neighbour's +/// +/// The match is the full URL, not the `/webhook/knowledge` suffix. A +/// hook with that suffix and a different base belongs to *another hive* — +/// one that may not have been upgraded yet — and deleting it would break +/// its knowledge sync until it was. Reaping a neighbour's registration is +/// the very behaviour this issue is about; doing it in the name of fixing +/// it would just invert the direction. +/// +/// (The predecessor did reap by suffix, to clear loopback hooks left by +/// an older single-hive layout. That was safe when a hive was alone on +/// its forge and is not safe now.) +/// +/// A listing failure is an error rather than a silent skip: there is no +/// create attempt left to fall through to, so swallowing it would leave +/// the hook in place with nothing said. The caller logs and continues — +/// boot does not depend on this. +pub async fn remove_webhook(core_token: &str, hive_domain: &str) -> Result<()> { // The typed client carries no per-request timeout, so each call is // wrapped in one: this runs as a detached startup task, and a forge - // that accepts connections but never answers would otherwise hang - // it forever (and the hourly pull fallback masks the missing hook). + // that accepts connections but never answers would otherwise hang it + // forever. const HTTP_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(10); - let target_url = format!("https://{hive_domain}/webhook/knowledge"); + let own_url = format!("https://{hive_domain}/webhook/knowledge"); let client = crate::forge::api(core_token)?; - // List existing hooks — skip creation if ours is already there. - // Best-effort like the raw-HTTP predecessor: a listing failure - // falls through to the create attempt. - let listed = tokio::time::timeout(HTTP_TIMEOUT, client.repo_list_hooks(ORG, REPO).all()) - .await - .map_err(anyhow::Error::from) - .and_then(|r| r.map_err(anyhow::Error::from)); - match listed { - Ok(hooks) => { - let already_exists = hooks.iter().any(|h| { - h.config - .as_ref() - .and_then(|c| c.get("url")) - .map(String::as_str) - == Some(target_url.as_str()) - }); - if already_exists { - tracing::debug!(%target_url, "knowledge: push webhook already configured"); - return Ok(()); - } - // Delete stale hooks that point at our path but a different base - // (e.g. old loopback hooks from before the SSRF-bypass migration). - for h in &hooks { - let hook_url = h - .config - .as_ref() - .and_then(|c| c.get("url")) - .map_or("", String::as_str); - if hook_url.ends_with("/webhook/knowledge") - && hook_url != target_url - && let Some(id) = h.id - { - tracing::info!(hook_url, "knowledge: deleting stale webhook (wrong base)"); - let _ = tokio::time::timeout( - HTTP_TIMEOUT, - client.repo_delete_hook(ORG, REPO, id).send(), - ) - .await; - } - } - } - Err(e) => { - tracing::debug!(error = %e, "knowledge: listing hooks failed; attempting create"); - } - } - - // Create the webhook. - let mut additional = BTreeMap::new(); - additional.insert("secret".to_owned(), webhook_secret.to_owned()); - - let hook = CreateHookOption { - active: Some(true), - authorization_header: None, - branch_filter: None, - config: CreateHookOptionConfig { - content_type: "json".to_owned(), - url: url::Url::parse(&target_url).context("parse webhook target url")?, - additional, - }, - events: Some(vec!["push".to_owned()]), - r#type: CreateHookOptionType::Forgejo, - }; - tokio::time::timeout(HTTP_TIMEOUT, client.repo_create_hook(ORG, REPO, hook)) + let hooks = tokio::time::timeout(HTTP_TIMEOUT, client.repo_list_hooks(ORG, REPO).all()) .await .map_err(anyhow::Error::from) .and_then(|r| r.map_err(anyhow::Error::from)) - .with_context(|| format!("create webhook for {ORG}/{REPO}"))?; - tracing::info!(%target_url, "knowledge: push webhook created"); + .with_context(|| format!("list webhooks for {ORG}/{REPO}"))?; + + for h in &hooks { + let hook_url = h + .config + .as_ref() + .and_then(|c| c.get("url")) + .map_or("", String::as_str); + if hook_url == own_url + && let Some(id) = h.id + { + tokio::time::timeout(HTTP_TIMEOUT, client.repo_delete_hook(ORG, REPO, id).send()) + .await + .map_err(anyhow::Error::from) + .and_then(|r| r.map_err(anyhow::Error::from)) + .with_context(|| format!("delete webhook {id} for {ORG}/{REPO}"))?; + tracing::info!( + %own_url, + "knowledge: removed this hive's push webhook — the swarm controller owns it now" + ); + } + } Ok(()) } diff --git a/swarm-controller/src/status.rs b/swarm-controller/src/status.rs index eaff6038..5090cd7a 100644 --- a/swarm-controller/src/status.rs +++ b/swarm-controller/src/status.rs @@ -133,6 +133,23 @@ impl StatusReader { } } + /// A handle on the queue connection this reader holds. + /// + /// The controller has exactly **one** connection to the swarm queue and + /// more than one thing to do with it: status is read out of a KV bucket, + /// swarm events are published on a subject. Handing out a clone is cheap — + /// `async_nats::Client` is a handle, not a socket — and is strictly better + /// than opening a second connection, which would double the auth-callout + /// traffic and give the two paths independent reconnect state, so one could + /// be serving while the other was still down. + /// + /// That this lives on the *status* reader is an accident of who constructed + /// the connection first, not a claim that events are a kind of status. + #[must_use] + pub fn queue_client(&self) -> async_nats::Client { + self.client.clone() + } + /// Reads [`STALE_AFTER_ENV`], falling back to /// [`DEFAULT_STALE_AFTER`]. A zero or unparseable value takes the /// default rather than failing startup — same rule as `load_hives`: diff --git a/swarm-controller/src/webhook.rs b/swarm-controller/src/webhook.rs index 5be60268..ce4f9098 100644 --- a/swarm-controller/src/webhook.rs +++ b/swarm-controller/src/webhook.rs @@ -170,13 +170,14 @@ pub(super) enum DeliveryKind { /// The route prefix a registered `target_url` must point at. /// /// ⚠️ Deliberately **not** `/webhook/knowledge` or `/webhook/config-pr`, the -/// paths the per-hive receivers use. Both hive-side registrars delete any -/// hook whose URL ends with *their* path but has a different base — see -/// `hive-c0re`'s `forge::ensure_config_pr_webhook` and -/// `workers::knowledge::ensure_webhook`. A swarm-level hook under those -/// paths would therefore be deleted by every hive on every boot, and the -/// symptom is a hook that silently stops existing. `webhook_urls_survive_the_hive_side_reapers` -/// pins that. +/// paths the per-hive receivers use. A hive-side registrar deletes any hook +/// whose URL ends with *its* path but has a different base — see +/// `hive-c0re`'s `forge::ensure_config_pr_webhook`. A swarm-level hook under +/// such a path would therefore be deleted by every hive on every boot, and +/// the symptom is a hook that silently stops existing. +/// `webhook_urls_survive_the_hive_side_reapers` pins that, and its own doc +/// records why `/webhook/knowledge` stays in the check even though the +/// knowledge registrar no longer reaps. const ROUTE_PREFIX: &str = "/webhook/forge/"; impl DeliveryKind { @@ -365,9 +366,75 @@ pub(super) async fn post_webhook_forge( bytes = body.len(), "webhook: verified delivery" ); + + // Only the knowledge delivery is acted on. Deploy coordination is a + // separate concern with its own issue — a hive does not want to hear that a + // config PR was opened, it wants to be told when to rebuild from main, and + // that is a decision the controller makes after a merge rather than a relay + // of this delivery. Written as a condition rather than a match arm holding + // an empty body, which would claim this is where that path is handled. + if kind == DeliveryKind::Knowledge { + announce_knowledge_change(&state).await; + } + (StatusCode::OK, "ok").into_response() } +/// Tell every hive in the swarm that the knowledge repository changed. +/// +/// The event carries **no payload**, because there is nothing to carry: the +/// hive-side handler this replaces read two fields from Forgejo's webhook and +/// used neither — both were filters — and then ran `git pull`, which re-derives +/// everything from the repository itself. So what crosses the queue is an edge +/// trigger, and adding fields to it would invent a contract nobody reads. +/// +/// One publish to one shared subject, not one per hive: every subscriber gets +/// the same empty event, so the roster is never consulted and this daemon does +/// not need to know who the hives are in order to tell them. +/// +/// # Failure +/// +/// Returns nothing and fails soft. A missed announcement costs a hive stale +/// knowledge until its next boot pull — the same cost as a webhook delivery to +/// a hive that happened to be down, which is what this replaces. +/// +/// ⚠️ A **permission** failure cannot be observed here. `publish` hands the +/// message to the connection's buffer, and a NATS authorization violation is +/// reported asynchronously on the connection rather than as an error on this +/// call — it reaches a client as a timeout, or as nothing at all. The `flush` +/// below therefore proves the bytes left this process, and nothing more; if +/// hives stop hearing events, the server log is the place that knows why. +async fn announce_knowledge_change(state: &AppState) { + let Some(status) = state.status.as_ref() else { + // Verified, accepted, and dropped. Worth a warning rather than + // silence: the forge will report a 200 and nobody would otherwise + // learn that the delivery reached a controller with nowhere to put it. + tracing::warn!( + "webhook: knowledge delivery verified but no swarm queue is \ + configured; no hive will be told" + ); + return; + }; + let client = status.queue_client(); + let subject = swarm_queue_client::KNOWLEDGE_SUBJECT; + + // One publish, not one per hive: every subscriber gets the same empty + // event, so the roster is not consulted at all. The controller does not + // need to know who the hives are in order to say the repository moved. + if let Err(e) = client.publish(subject, Vec::new().into()).await { + tracing::warn!(%subject, error = %e, "webhook: publishing the knowledge event failed"); + return; + } + // Logged after the flush rather than after the publish: `publish` only + // hands the message to the client's write buffer, so a line printed there + // would claim delivery this end cannot yet know about. + if let Err(e) = client.flush().await { + tracing::warn!(%subject, error = %e, "webhook: flushing the knowledge event failed"); + return; + } + tracing::info!(%subject, "webhook: knowledge event published"); +} + #[cfg(test)] mod tests { use super::{DeliveryKind, Refusal, load_or_generate_at, secret_path_from, verify}; @@ -567,17 +634,22 @@ mod tests { } } - /// A cross-daemon invariant with nothing else to enforce it: both - /// per-hive registrars in `hive-c0re` **delete** hooks whose URL ends - /// with their own path but carries a different base — see - /// `forge::ensure_config_pr_webhook` and `knowledge::ensure_webhook`. + /// A cross-daemon invariant with nothing else to enforce it: a per-hive + /// registrar in `hive-c0re` **deletes** hooks whose URL ends with its + /// own path but carries a different base. A controller URL matching such + /// a suffix would be deleted by every hive on every boot — the swarm hook + /// would simply cease to exist, with the cause in a different daemon's + /// startup sweep. Serving these under `/webhook/forge/` is what avoids + /// it, and this is the only place that says so in a form that fails. /// - /// While the swarm-level hooks live alongside the per-hive ones, a - /// controller URL matching either suffix would be deleted by every hive - /// on every boot: the swarm hook would simply cease to exist, with the - /// cause in a different daemon's startup sweep. Serving these under - /// `/webhook/forge/` is what avoids it, and this is the only place that - /// says so in a form that fails. + /// `forge::ensure_config_pr_webhook` still reaps that way. The knowledge + /// registrar no longer does — it was replaced by a removal that matches + /// the full URL, so it cannot touch another hive's hook. **The + /// `/webhook/knowledge` arm is kept anyway**, because the hazard is not + /// this repository's current code: it is whatever is *deployed*, and a + /// hive still running the previous version reaps by suffix until it is + /// upgraded. Drop that arm once no such hive can exist, not when the + /// source stops mentioning it. #[test] fn webhook_urls_survive_the_hive_side_reapers() { for suffix in ["/webhook/knowledge", "/webhook/config-pr"] { diff --git a/swarm-nats-auth/src/policy.rs b/swarm-nats-auth/src/policy.rs index 4f3a76b2..1eb3e083 100644 --- a/swarm-nats-auth/src/policy.rs +++ b/swarm-nats-auth/src/policy.rs @@ -234,6 +234,19 @@ impl Policy { // stays for a named/durable consumer. format!("$JS.API.CONSUMER.CREATE.{stream}"), format!("$JS.API.CONSUMER.CREATE.{stream}.>"), + // The knowledge event. One writer, many readers: the controller is + // the only publisher and every hive subscribes, so this is one + // literal subject rather than a per-hive family — named from the + // crate the publisher and the subscriber also name, so a rename + // cannot leave the grant pointing at a subject nobody uses. + // + // This is the reader's only non-JetStream subject, and without it + // the controller cannot emit the event at all. Worth stating because + // the symptom is unhelpful: a refused publish reaches the client as + // a **timeout**, so the visible failure is a hive that never hears + // about a change, with nothing in the controller's log to say a + // permission was the reason. + swarm_queue_client::KNOWLEDGE_SUBJECT.to_owned(), ]); subjects } @@ -295,6 +308,40 @@ mod tests { assert!(!p.publish.iter().any(|s| s.contains("beta"))); } + #[test] + fn a_reader_may_publish_the_knowledge_event() { + let p = policy().permissions("swarm-controller").expect("a reader"); + assert!( + p.publish + .contains(&swarm_queue_client::KNOWLEDGE_SUBJECT.to_owned()), + "the controller is the only publisher of this event; without the \ + grant its publish is refused, and a refusal arrives as a timeout" + ); + } + + #[test] + fn a_hive_may_not_publish_the_knowledge_event_to_anyone_including_itself() { + // The controller *interprets* what a delivery means; a hive receives + // that verdict. A hive able to publish here could tell every other hive + // in the swarm — or itself — that the knowledge repo changed when it did + // not, which is an unauthenticated write into someone else's control + // path wearing an event's shape. + // + // One writer and many readers makes this arm matter MORE, not less: with + // a single shared subject a forged event reaches the whole swarm, where + // a per-hive subject would have reached one. + let p = policy() + .permissions("hive-alpha") + .expect("a hive is admitted"); + assert!( + !p.publish + .iter() + .any(|s| s == swarm_queue_client::KNOWLEDGE_SUBJECT), + "a hive must not publish the knowledge event: {:?}", + p.publish + ); + } + #[test] fn a_hive_grant_never_includes_the_jetstream_wildcard() { // `$JS.API.>` also covers `$JS.API.STREAM.DELETE.KV_hive-status`, with diff --git a/swarm-queue-client/src/lib.rs b/swarm-queue-client/src/lib.rs index 1c03a317..c48f981b 100644 --- a/swarm-queue-client/src/lib.rs +++ b/swarm-queue-client/src/lib.rs @@ -156,6 +156,15 @@ pub fn chain(error: &dyn std::error::Error) -> String { /// which is the disagreement this module exists to prevent. pub mod status; +/// The subject the swarm controller publishes on when the hive-wide knowledge +/// repository has changed. One writer, many readers — every hive subscribes. +/// +/// Here rather than in a module of its own for the same reason as the bucket +/// name above: three crates must agree on the string, and the one that agrees +/// hardest — the auth-callout responder, which decides whether a publish is +/// permitted at all — speaks neither `jetstream` nor `kv`. +pub const KNOWLEDGE_SUBJECT: &str = "$SWARM.knowledge"; + /// Only the fields this needs; authelia returns several. #[derive(serde::Deserialize)] struct TokenResponse {