diff --git a/lib/apps/fabro-server/tests/it/api/mod.rs b/lib/apps/fabro-server/tests/it/api/mod.rs
index 7fcdf254a..8833cfa21 100644
--- a/lib/apps/fabro-server/tests/it/api/mod.rs
+++ b/lib/apps/fabro-server/tests/it/api/mod.rs
@@ -8,6 +8,7 @@ mod events;
 mod install;
 mod install_openai_compatible;
 mod mcp_servers;
+mod petri_store;
 mod routing;
 mod run_files;
 mod runs;
diff --git a/lib/apps/fabro-server/tests/it/api/petri_store.rs b/lib/apps/fabro-server/tests/it/api/petri_store.rs
new file mode 100644
index 000000000..9c8d6d157
--- /dev/null
+++ b/lib/apps/fabro-server/tests/it/api/petri_store.rs
@@ -0,0 +1,307 @@
+//! The Petri run store over HTTP: Petri's store conformance suite against
+//! `HttpRunStore` talking to a loopback server, the operator release through
+//! the server's store, a lost reply to an append, and two workers contending
+//! for one run's lease.
+
+use std::collections::HashMap;
+use std::future;
+use std::sync::atomic::{AtomicUsize, Ordering};
+use std::sync::{Arc, Mutex};
+use std::time::{Duration, Instant};
+
+use axum::Router;
+use axum::extract::{Request, State};
+use axum::http::{Method, StatusCode};
+use axum::middleware::{self, Next};
+use axum::response::Response;
+use fabro_client::{Client, Credential, apply_bearer_token_auth};
+use fabro_petri::HttpRunStore;
+use fabro_petri::petri::{Access, LogId, OwnerId, RunKey, RunLogs, RunStore, StoreError};
+use fabro_petri::test_support::run_store::{self, conformance, stale_owner_conformance};
+use fabro_server::server::{AppState, RouterOptions, build_router_with_options};
+use fabro_server::test_support::{test_app_state, test_auth_mode};
+use fabro_types::RunId;
+use tokio::net::TcpListener;
+use tokio::runtime::Handle;
+use tokio::task;
+
+/// A loopback server over `state`, its router wrapped by `wrap`.
+async fn serve(state: Arc<AppState>, wrap: impl FnOnce(Router) -> Router) -> String {
+    let listener = TcpListener::bind("127.0.0.1:0")
+        .await
+        .expect("a loopback listener binds");
+    let addr = listener.local_addr().expect("the listener has an address");
+    let router = wrap(build_router_with_options(
+        state,
+        &test_auth_mode(),
+        RouterOptions::default(),
+    ));
+    tokio::spawn(async move {
+        let _ = axum::serve(listener, router).await;
+    });
+    format!("http://{addr}")
+}
+
+/// A worker's client: the run's worker token as its bearer, and a request
+/// timeout after which a reply counts as lost.
+async fn worker_client(base_url: &str, token: &str, timeout: Duration) -> Client {
+    let http = apply_bearer_token_auth(fabro_http::HttpClientBuilder::new().no_proxy(), token)
+        .expect("the bearer header builds")
+        .build()
+        .expect("the test HTTP client builds");
+    Client::builder()
+        .transport(base_url, http)
+        .credential(Credential::Worker(token.to_string()))
+        .request_timeout(timeout)
+        .connect()
+        .await
+        .expect("the worker client connects")
+}
+
+/// The Petri key of a Fabro run: its id.
+fn petri_key(run_id: RunId) -> RunKey {
+    RunKey::new(run_id.to_string())
+}
+
+/// Wait until the server's store shows no lease holder on `key`: a dropped
+/// handle's release travels to the server on its own, and only the store
+/// that dropped it awaits that. Another worker starts after the first is
+/// gone, which is what this waits for.
+async fn wait_until_released(store: &fabro_petri::SqliteRunStore, key: &RunKey) {
+    for _ in 0..500 {
+        if store.owner(key).await.expect("reads the lease").is_none() {
+            return;
+        }
+        tokio::time::sleep(Duration::from_millis(10)).await;
+    }
+    panic!("the lease on {key} was not released");
+}
+
+/// One worker's view of one run: a store over a client whose token names
+/// the run.
+async fn worker_store(state: &AppState, base_url: &str, run_id: RunId) -> HttpRunStore {
+    let token = state.test_issue_worker_token(&run_id);
+    HttpRunStore::new(worker_client(base_url, &token, Duration::from_secs(5)).await)
+}
+
+/// The suite opens runs under keys of its own (`lifecycle`, `drop`), while a
+/// key over the API is a Fabro run id that the worker's token names. This
+/// adapter gives each suite key a Fabro run of its own, with a token minted
+/// for that run alone: the least a worker holds.
+struct WorkerRuns {
+    state:    Arc<AppState>,
+    base_url: String,
+    runs:     Mutex<HashMap<RunKey, Arc<WorkerRun>>>,
+}
+
+struct WorkerRun {
+    run_id: RunId,
+    store:  HttpRunStore,
+}
+
+impl WorkerRuns {
+    fn new(state: Arc<AppState>, base_url: String) -> Self {
+        Self {
+            state,
+            base_url,
+            runs: Mutex::default(),
+        }
+    }
+
+    async fn run(&self, key: &RunKey) -> Arc<WorkerRun> {
+        if let Some(run) = self.runs.lock().expect("runs lock").get(key) {
+            return Arc::clone(run);
+        }
+        let run_id = RunId::new();
+        let store = worker_store(&self.state, &self.base_url, run_id).await;
+        let run = Arc::new(WorkerRun { run_id, store });
+        self.runs
+            .lock()
+            .expect("runs lock")
+            .insert(key.clone(), Arc::clone(&run));
+        run
+    }
+
+    /// The Fabro run a suite key was given, once opened.
+    fn run_id(&self, key: &RunKey) -> RunId {
+        self.runs
+            .lock()
+            .expect("runs lock")
+            .get(key)
+            .expect("the suite opens a key before it releases it")
+            .run_id
+    }
+}
+
+#[async_trait::async_trait]
+impl RunStore for WorkerRuns {
+    async fn open(&self, key: &RunKey, access: Access) -> Result<Arc<dyn RunLogs>, StoreError> {
+        let run = self.run(key).await;
+        run.store.open(&petri_key(run.run_id), access).await
+    }
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn the_http_store_passes_the_conformance_suite() {
+    let state = test_app_state();
+    let base_url = serve(Arc::clone(&state), |router| router).await;
+    let runs: Arc<dyn RunStore> = Arc::new(WorkerRuns::new(state, base_url));
+    conformance(|| Arc::clone(&runs)).await;
+}
+
+/// An operator release through the server's own store ends the worker's
+/// lease from outside: the worker's handle turns stale and the next writer
+/// takes the run. The release is async, so the suite's synchronous closure
+/// blocks on it in place.
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn an_operator_release_through_the_server_makes_the_worker_stale() {
+    let state = test_app_state();
+    let base_url = serve(Arc::clone(&state), |router| router).await;
+    let runs = WorkerRuns::new(Arc::clone(&state), base_url);
+    let release = |key: &RunKey| {
+        let key = petri_key(runs.run_id(key));
+        let store = state.test_petri_run_store();
+        task::block_in_place(|| Handle::current().block_on(store.release_lease(&key)))
+            .expect("the operator releases the lease");
+    };
+    stale_owner_conformance(&runs, release).await;
+}
+
+/// Swallow the reply of the first append the server commits: the handler
+/// runs, its transaction commits, and the response never leaves. That is
+/// a lost reply as the worker sees it.
+async fn swallow_one_append_reply(
+    State(swallowed): State<Arc<AtomicUsize>>,
+    request: Request,
+    next: Next,
+) -> Response {
+    let is_append = request.method() == Method::POST && request.uri().path().ends_with("/records");
+    let response = next.run(request).await;
+    if is_append
+        && response.status() == StatusCode::NO_CONTENT
+        && swallowed
+            .compare_exchange(0, 1, Ordering::SeqCst, Ordering::SeqCst)
+            .is_ok()
+    {
+        future::pending::<()>().await;
+    }
+    response
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn a_lost_reply_to_an_append_is_safe_to_retry() {
+    let state = test_app_state();
+    let swallowed = Arc::new(AtomicUsize::new(0));
+    let base_url = serve(Arc::clone(&state), {
+        let swallowed = Arc::clone(&swallowed);
+        move |router| {
+            router.layer(middleware::from_fn_with_state(
+                swallowed,
+                swallow_one_append_reply,
+            ))
+        }
+    })
+    .await;
+    let run_id = RunId::new();
+    let token = state.test_issue_worker_token(&run_id);
+    let timeout = Duration::from_millis(500);
+    let store = HttpRunStore::new(worker_client(&base_url, &token, timeout).await);
+    let key = petri_key(run_id);
+    let logs = store
+        .open(&key, Access::Create {
+            owner: OwnerId::new("worker"),
+        })
+        .await
+        .expect("the worker creates the run");
+
+    let batch = [
+        run_store::record(0, "execution.started"),
+        run_store::record(1, "step.started"),
+    ];
+    let started = Instant::now();
+    logs.append(&LogId::Coordinator, &batch)
+        .await
+        .expect("the append succeeds once the resend is answered");
+    assert_eq!(
+        swallowed.load(Ordering::SeqCst),
+        1,
+        "the server committed one append whose reply was swallowed"
+    );
+    assert!(
+        started.elapsed() >= timeout,
+        "the first attempt waited out the request timeout"
+    );
+    assert_eq!(
+        logs.read(&LogId::Coordinator).await.expect("reads"),
+        batch.to_vec(),
+        "the log holds each record once"
+    );
+}
+
+/// Two workers with owners of their own never hold one run's lease at the
+/// same time, in either order, and the server's store reports who holds it.
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn two_workers_cannot_both_hold_a_run_lease() {
+    let state = test_app_state();
+    let base_url = serve(Arc::clone(&state), |router| router).await;
+    let run_id = RunId::new();
+    let key = petri_key(run_id);
+    let first_worker = worker_store(&state, &base_url, run_id).await;
+    let second_worker = worker_store(&state, &base_url, run_id).await;
+    let first = OwnerId::new("first");
+    let second = OwnerId::new("second");
+    let server_store = state.test_petri_run_store();
+
+    let held = first_worker
+        .open(&key, Access::Create {
+            owner: first.clone(),
+        })
+        .await
+        .expect("the first worker creates");
+    assert_eq!(
+        server_store.owner(&key).await.expect("reads"),
+        Some(first.clone())
+    );
+    let refused = second_worker
+        .open(&key, Access::Write {
+            owner: second.clone(),
+        })
+        .await
+        .err()
+        .expect("the second worker is refused while the first holds the lease");
+    assert!(
+        matches!(&refused, StoreError::Leased { owner, .. } if *owner == first),
+        "{refused}"
+    );
+    assert!(
+        refused.to_string().contains(&run_id.to_string()),
+        "the message names the run: {refused}"
+    );
+
+    drop(held);
+    wait_until_released(server_store, &key).await;
+    let taken = second_worker
+        .open(&key, Access::Write {
+            owner: second.clone(),
+        })
+        .await
+        .expect("the second worker takes the run once the first releases");
+    assert_eq!(
+        server_store.owner(&key).await.expect("reads"),
+        Some(second.clone())
+    );
+    let refused = first_worker
+        .open(&key, Access::Write {
+            owner: first.clone(),
+        })
+        .await
+        .err()
+        .expect("the first worker is refused in turn");
+    assert!(
+        matches!(&refused, StoreError::Leased { owner, .. } if *owner == second),
+        "{refused}"
+    );
+
+    drop(taken);
+    wait_until_released(server_store, &key).await;
+}