Kept the runtime API responsive during blocking work (#1043)

AGENTS.md

@@ -130,6 +130,25 @@ Key modules:
 - `pandoc.rs` - Integration with Pandoc for document conversion
 - `log.rs` - Logging infrastructure using `flexi_logger`
 
+**Runtime API handlers never block.** All calls of the .NET app share one HTTP/2 connection, and the
+task driving it may wait on exactly the Tokio worker which a blocking handler occupies, so a single
+blocking handler can hold up the whole app. Therefore:
+
+- File and disk access, the OS keyring, waiting for a `std::sync::Mutex`, and CPU-bound work such as
+  loading a tokenizer or scanning text run in `tokio::task::spawn_blocking`. Follow
+  `run_qdrant_edge_request` in `qdrant_edge_database.rs`, `prepare_image` and `prepare_image_sync` in
+  `image.rs`, or `sanitize_batch` in `prompt_injection/api.rs`.
+- When an API offers a callback, as the file dialogs do, await the callback instead of calling the
+  blocking variant; see `await_dialog` in `file_actions.rs`.
+- Short, bounded work, such as a single metadata lookup or writing a log line, may stay on the worker.
+- When the blocking task fails, answer with an error, never with an empty value the app could take
+  for a valid answer.
+- Never keep a `std::sync::MutexGuard` alive across an `.await`, not even with an explicit `drop`
+  before it: the future of the handler is then no longer `Send`, and Axum refuses it. Scope the guard
+  in a block instead.
+- `runtime_api::test_support::assert_runtime_stays_free` tests a handler which waits for a lock. Run
+  such a test with `#[tokio::test(flavor = "multi_thread", worker_threads = 1)]`.
+
 ### .NET App (`app/MindWork AI Studio/`)
 **Entry point:** `app/MindWork AI Studio/Program.cs`
 

app/MindWork AI Studio/wwwroot/changelog/v26.10.1.md

@@ -26,6 +26,7 @@
 - Fixed AI Studio sometimes rebuilding the index of a data source from scratch when the configuration of your organization was applied while the data source was being updated.
 - Fixed justified texts being hyphenated by the rules of English even when AI Studio shows another language, such as German. Screen readers now also know which language AI Studio uses.
 - Fixed the log file of AI Studio showing the wrong time for some entries when AI Studio was busy. Each entry now shows when it actually happened.
+- Fixed AI Studio pausing for a moment while it opened the local index of a data source, such as a mailbox. Other work, for example applying the configuration of your organization, had to wait until then.
 - Updated the code contributions on the supporters page, which now thank everyone who has contributed code to AI Studio so far.
 - Upgraded several libraries to improve security.
 - Upgraded to Rust v1.99.0

runtime/src/clipboard.rs

@@ -77,6 +77,26 @@ fn release_clipboard<B>(clipboard: &mut Option<B>) -> bool {
 
 /// Sets the clipboard text to the provided encrypted text.
 pub async fn set_clipboard(_token: APIToken, encrypted_text: String) -> Json<SetClipboardResponse> {
+    //
+    // The clipboard backend of the operating system may take a moment to answer, a failed write is
+    // retried with a fresh backend, and the lock waits for a write still running. None of this
+    // belongs on a runtime worker, because all calls of the app share one HTTP/2 connection.
+    //
+    match tokio::task::spawn_blocking(move || set_clipboard_sync(encrypted_text)).await {
+        Ok(response) => response,
+        Err(error) => {
+            let issue = format!("The clipboard worker failed: {error}");
+            error!(Source = "Clipboard"; "{issue}");
+            Json(SetClipboardResponse {
+                success: false,
+                issue,
+            })
+        },
+    }
+}
+
+/// Performs the blocking part of [`set_clipboard`].
+fn set_clipboard_sync(encrypted_text: String) -> Json<SetClipboardResponse> {
     let encrypted_text = EncryptedText::new(encrypted_text);
 
     // Decrypt this text first:

runtime/src/environment.rs

@@ -1,4 +1,5 @@
 use crate::api_token::APIToken;
+use axum::http::StatusCode;
 use axum::Json;
 use log::{debug, error, info, warn};
 use serde::Serialize;
@@ -641,35 +642,35 @@ struct EnterpriseSourceData {
     encryption_secret: String,
 }
 
-pub async fn read_enterprise_env_config_id(_token: APIToken) -> String {
-    debug!("Trying to read the effective enterprise configuration ID.");
-    resolve_effective_enterprise_config_source()
-        .configs
-        .into_iter()
-        .next()
-        .map(|config| config.id)
-        .unwrap_or_default()
-}
-
-pub async fn read_enterprise_env_config_server_url(_token: APIToken) -> String {
-    debug!("Trying to read the effective enterprise configuration server URL.");
-    resolve_effective_enterprise_config_source()
-        .configs
-        .into_iter()
-        .next()
-        .map(|config| config.server_url)
-        .unwrap_or_default()
-}
-
-pub async fn read_enterprise_env_config_encryption_secret(_token: APIToken) -> String {
+pub async fn read_enterprise_env_config_encryption_secret(_token: APIToken) -> Result<String, (StatusCode, String)> {
     debug!("Trying to read the effective enterprise configuration encryption secret.");
-    resolve_effective_enterprise_secret_source().encryption_secret
+    read_enterprise_sources(|| resolve_effective_enterprise_secret_source().encryption_secret).await
 }
 
 /// Returns all enterprise configurations from the effective source.
-pub async fn read_enterprise_configs(_token: APIToken) -> Json<Vec<EnterpriseConfig>> {
+pub async fn read_enterprise_configs(_token: APIToken) -> Result<Json<Vec<EnterpriseConfig>>, (StatusCode, String)> {
     info!("Trying to read the effective enterprise configurations.");
-    Json(resolve_effective_enterprise_config_source().configs)
+    read_enterprise_sources(|| Json(resolve_effective_enterprise_config_source().configs)).await
+}
+
+/// Reads the enterprise sources on the blocking pool.
+///
+/// Every call reads the policy files and, on Windows, the registry again. That is quick most of
+/// the time, but a slow disk must not hold up every other call of the app, because they all share
+/// one HTTP/2 connection.
+///
+/// A failure is answered as an error, never as an empty value: the app would read an empty value
+/// as "no configuration of your organization" and act on it.
+async fn read_enterprise_sources<T, F>(read: F) -> Result<T, (StatusCode, String)>
+where
+    T: Send + 'static,
+    F: FnOnce() -> T + Send + 'static,
+{
+    tokio::task::spawn_blocking(read).await.map_err(|error| {
+        let issue = format!("The enterprise configuration worker failed: {error}");
+        error!("{issue}");
+        (StatusCode::INTERNAL_SERVER_ERROR, issue)
+    })
 }
 
 pub fn resolve_external_http_custom_root_certificate_policy() -> ExternalHttpCustomRootCertificatePolicy {

runtime/src/file_actions.rs

@@ -123,26 +123,25 @@ pub async fn select_directory(
     Query(query): Query<SelectDirectoryQuery>,
     previous_directory: Option<Json<PreviousDirectory>>,
 ) -> Json<DirectorySelectionResponse> {
-    let main_window_lock = MAIN_WINDOW.lock().unwrap();
-    let main_window = match main_window_lock.as_ref() {
-        Some(window) => window,
-        None => {
-            error!(Source = "Tauri"; "Cannot open directory dialog: main window not available.");
-            return Json(DirectorySelectionResponse {
-                user_cancelled: true,
-                selected_directory: String::from(""),
-            });
-        }
+    let dialog = MAIN_WINDOW
+        .lock()
+        .unwrap()
+        .as_ref()
+        .map(|w| w.dialog().file().set_parent(w).set_title(&query.title));
+
+    let Some(mut dialog) = dialog else {
+        error!(Source = "Tauri"; "Cannot open directory dialog: main window not available.");
+        return Json(DirectorySelectionResponse {
+            user_cancelled: true,
+            selected_directory: String::from(""),
+        });
     };
 
-    let mut dialog = main_window.dialog().file().set_parent(main_window).set_title(&query.title);
     if let Some(previous) = previous_directory {
         dialog = dialog.set_directory(previous.path.clone());
     }
 
-    drop(main_window_lock);
-
-    let folder_path = dialog.blocking_pick_folder();
+    let folder_path = await_dialog(|answer| dialog.pick_folder(answer)).await;
     match folder_path {
         Some(path) => {
             match path.into_path() {
@@ -203,7 +202,7 @@ pub async fn select_file(
     }
 
     // Show the file dialog and get the selected file path:
-    let file_path = file_dialog.blocking_pick_file();
+    let file_path = await_dialog(|answer| file_dialog.pick_file(answer)).await;
     match file_path {
         Some(path) => match path.into_path() {
             Ok(pb) => {
@@ -262,7 +261,7 @@ pub async fn select_files(
     }
 
     // Show the file dialog and get the selected file path:
-    let file_paths = file_dialog.blocking_pick_files();
+    let file_paths = await_dialog(|answer| file_dialog.pick_files(answer)).await;
     match file_paths {
         Some(paths) => {
             let converted: Vec<String> = paths.into_iter().filter_map(|p| p.into_path().ok()).map(|pb| pb.to_string_lossy().to_string()).collect();
@@ -314,7 +313,7 @@ pub async fn save_file(_token: APIToken, payload: Json<SaveFileOptions>) -> Json
     }
 
     // Displays the file dialogue box and select the file:
-    let file_path = file_dialog.blocking_save_file();
+    let file_path = await_dialog(|answer| file_dialog.save_file(answer)).await;
     match file_path {
         Some(path) => match path.into_path() {
             Ok(pb) => {
@@ -882,6 +881,26 @@ fn apply_filter<R: tauri::Runtime>(file_dialog: FileDialogBuilder<R>, filter: &O
     }
 }
 
+/// Shows a file dialog and waits for the user's answer without holding up the runtime.
+///
+/// A dialog stays open until the user answers it, which can take minutes. The blocking variants of
+/// the dialog API would hold a runtime worker that long, and with it every other call of the app,
+/// because they all share one HTTP/2 connection. The plugin shows the dialog on the main thread and
+/// hands the answer to a callback, so this function awaits that callback instead.
+async fn await_dialog<T: Send + 'static>(show: impl FnOnce(Box<dyn FnOnce(Option<T>) + Send>)) -> Option<T> {
+    let (sender, receiver) = tokio::sync::oneshot::channel();
+    show(Box::new(move |answer| {
+        // The receiver is only gone when the request was dropped, and then nobody waits for the answer:
+        let _ = sender.send(answer);
+    }));
+
+    receiver.await.unwrap_or_else(|_| {
+        // The plugin drops the callback without calling it when it cannot reach the main thread:
+        error!(Source = "Tauri"; "The file dialog ended without an answer.");
+        None
+    })
+}
+
 fn split_save_file_path(file_path: &str) -> (Option<PathBuf>, Option<String>) {
     let path = Path::new(file_path);
     let directory = path
@@ -1047,6 +1066,21 @@ mod tests {
     use super::*;
     use std::fs;
 
+    #[tokio::test]
+    async fn a_dialog_answer_arrives_through_its_callback() {
+        let answer = await_dialog(|answer| {
+            std::thread::spawn(move || answer(Some("Quarterly briefing.html")));
+        }).await;
+
+        assert_eq!(answer, Some("Quarterly briefing.html"));
+    }
+
+    #[tokio::test]
+    async fn a_dialog_without_an_answer_counts_as_cancelled() {
+        let answer = await_dialog::<String>(drop).await;
+        assert_eq!(answer, None);
+    }
+
     #[test]
     fn save_file_options_accept_the_previous_file_contract() {
         let options: SaveFileOptions = serde_json::from_str(

runtime/src/file_data.rs

@@ -524,7 +524,7 @@ pub async fn extract_data(
                 // whole file, and a failure here means we cannot answer the request at all.
                 //
                 let tokenizer = if query.include_token_count {
-                    match crate::tokenizer::get_tokenizer(&query.tokenizer_path) {
+                    match crate::tokenizer::get_tokenizer_off_worker(query.tokenizer_path.clone()).await {
                         Ok(tokenizer) => Some(tokenizer),
                         Err(e) => {
                             let error = ExtractionError::new(ExtractionErrorCode::InvalidRequest, format!("The tokenizer could not be loaded: {e}"));

runtime/src/prompt_injection/api.rs

@@ -45,13 +45,30 @@ pub struct SanitizeBatchResponse {
     pub results: Vec<SanitizeResponse>,
 }
 
-pub async fn sanitize(_token: APIToken, Json(request): Json<SanitizeRequest>) -> Json<SanitizeResponse> {
-    let (sanitized_text, report) = sanitize_text(&request.text);
+pub async fn sanitize(
+    _token: APIToken,
+    Json(request): Json<SanitizeRequest>,
+) -> Result<Json<SanitizeResponse>, (StatusCode, String)> {
+    //
+    // A single web page or retrieval context can still be large, and scanning it is CPU-bound. Like
+    // a batch, it goes to the blocking pool, so it does not stall every other call the app makes
+    // meanwhile.
+    //
+    tokio::task::spawn_blocking(move || {
+        let (sanitized_text, report) = sanitize_text(&request.text);
 
-    Json(SanitizeResponse {
-        sanitized_text,
-        findings: report.findings,
-        redacted_count: report.redacted_count,
+        Json(SanitizeResponse {
+            sanitized_text,
+            findings: report.findings,
+            redacted_count: report.redacted_count,
+        })
+    })
+    .await
+    .map_err(|error| {
+        (
+            StatusCode::INTERNAL_SERVER_ERROR,
+            format!("The prompt injection filter failed: {error}"),
+        )
     })
 }
 

runtime/src/qdrant_edge_database.rs

@@ -473,15 +473,25 @@ fn qdrant_edge_base_path() -> QdrantEdgeResult<PathBuf> {
 }
 
 pub async fn qdrant_edge_info(_token: APIToken) -> Json<QdrantEdgeServiceInfo> {
-    let status = QDRANT_EDGE_STATUS.lock().unwrap();
-    let current_status = status.status;
-    let unavailable_reason = status.unavailable_reason.clone();
-    drop(status);
+    let (current_status, mut unavailable_reason) = {
+        let status = QDRANT_EDGE_STATUS.lock().unwrap();
+        (status.status, status.unavailable_reason.clone())
+    };
 
-    let database_guard = QDRANT_EDGE_DATABASE.lock().unwrap();
-    let database_info = database_guard
-        .as_ref()
-        .and_then(|database| database.info().ok());
+    //
+    // The info needs the lock of the database, which another request may hold for seconds while
+    // it loads a store, and it reads the stores directory. Both belong on the blocking pool, see
+    // run_qdrant_edge_request.
+    //
+    let database_info = match tokio::task::spawn_blocking(read_qdrant_edge_info).await {
+        Ok(database_info) => database_info,
+        Err(error) => {
+            let reason = format!("The Qdrant Edge worker failed: {error}");
+            error!(Source = "Qdrant Edge"; "Reading the Qdrant Edge info failed: {reason}");
+            unavailable_reason = Some(reason);
+            None
+        },
+    };
 
     let is_available = current_status == QdrantEdgeStatus::Available && database_info.is_some();
     Json(QdrantEdgeServiceInfo {
@@ -495,40 +505,48 @@ pub async fn qdrant_edge_info(_token: APIToken) -> Json<QdrantEdgeServiceInfo> {
     })
 }
 
+fn read_qdrant_edge_info() -> Option<QdrantEdgeInfo> {
+    QDRANT_EDGE_DATABASE
+        .lock()
+        .unwrap()
+        .as_ref()
+        .and_then(|database| database.info().ok())
+}
+
 pub async fn ensure_qdrant_edge_store(_token: APIToken, Json(request): Json<EnsureQdrantEdgeStoreRequest>) -> Json<QdrantEdgeResponse<QdrantEdgeEnsureStoreResult>> {
-    execute_qdrant_edge_request(|database| {
+    run_qdrant_edge_request(move |database| {
         database.ensure_store_exists(&request.store_name, &request.data_source_name, request.vector_size)
-    })
+    }).await
 }
 
 pub async fn insert_qdrant_edge_embedding(_token: APIToken, Json(request): Json<InsertQdrantEdgeEmbeddingRequest>) -> Json<QdrantEdgeResponse<()>> {
-    execute_qdrant_edge_request(|database| {
+    run_qdrant_edge_request(move |database| {
         database.insert_embedding(&request.store_name, request.points)
-    })
+    }).await
 }
 
 pub async fn search_qdrant_edge_embeddings(_token: APIToken, Json(request): Json<SearchQdrantEdgeEmbeddingRequest>) -> Json<QdrantEdgeResponse<Vec<QdrantEdgeSearchResult>>> {
-    execute_qdrant_edge_request(|database| {
+    run_qdrant_edge_request(move |database| {
         database.search_embedding(&request.store_name, request.vector, request.max_matches, request.point_ids)
-    })
+    }).await
 }
 
 pub async fn delete_qdrant_edge_embedding_by_file(_token: APIToken, Json(request): Json<DeleteQdrantEdgeEmbeddingByFileRequest>) -> Json<QdrantEdgeResponse<()>> {
-    execute_qdrant_edge_request(|database| {
+    run_qdrant_edge_request(move |database| {
         database.delete_embedding_by_file(&request.store_name, &request.file_path)
-    })
+    }).await
 }
 
 pub async fn optimize_qdrant_edge_store(_token: APIToken, Json(request): Json<OptimizeQdrantEdgeStoreRequest>) -> Json<QdrantEdgeResponse<()>> {
-    execute_qdrant_edge_request(|database| {
+    run_qdrant_edge_request(move |database| {
         database.optimize_store(&request.store_name)
-    })
+    }).await
 }
 
 pub async fn delete_qdrant_edge_store(_token: APIToken, Json(request): Json<DeleteQdrantEdgeStoreRequest>) -> Json<QdrantEdgeResponse<()>> {
-    execute_qdrant_edge_request(|database| {
+    run_qdrant_edge_request(move |database| {
         database.delete_store(&request.store_name)
-    })
+    }).await
 }
 
 pub fn start_qdrant_edge_database<R: tauri::Runtime>(app_handle: tauri::AppHandle<R>) {
@@ -569,6 +587,32 @@ pub fn stop_qdrant_edge_database() {
     set_qdrant_edge_unavailable("Qdrant Edge was stopped.".to_string());
 }
 
+/// Runs one request against the database on the blocking pool.
+///
+/// A request waits for the lock of the database and may load a store from disk, which together
+/// can take seconds. On a runtime worker, that would hold up every other call of the app as well:
+/// they all share one HTTP/2 connection, and the task driving it may be waiting in the LIFO slot of
+/// exactly this worker, where no other worker can take it over.
+async fn run_qdrant_edge_request<T, F>(operation: F) -> Json<QdrantEdgeResponse<T>>
+where
+    T: Serialize + Send + 'static,
+    F: FnOnce(&mut QdrantEdgeDatabase) -> QdrantEdgeResult<T> + Send + 'static,
+{
+    match tokio::task::spawn_blocking(move || execute_qdrant_edge_request(operation)).await {
+        Ok(response) => response,
+        Err(error) => {
+            let issue = format!("The Qdrant Edge worker failed: {error}");
+            error!(Source = "Qdrant Edge"; "Qdrant Edge request failed: {issue}");
+            Json(QdrantEdgeResponse {
+                success: false,
+                issue,
+                issue_code: "",
+                data: None,
+            })
+        },
+    }
+}
+
 fn execute_qdrant_edge_request<T, F>(operation: F) -> Json<QdrantEdgeResponse<T>>
 where
     T: Serialize,
@@ -1033,6 +1077,7 @@ fn validate_store_name(store_name: &str) -> QdrantEdgeResult<()> {
 #[cfg(test)]
 mod tests {
     use super::*;
+    use crate::runtime_api::test_support::assert_runtime_stays_free;
     use qdrant_edge::WalOptions;
 
     #[test]
@@ -1254,6 +1299,27 @@ mod tests {
         fs::remove_dir_all(test_directory).unwrap();
     }
 
+    #[tokio::test(flavor = "multi_thread", worker_threads = 1)]
+    async fn a_request_waiting_for_the_database_leaves_the_runtime_free() {
+        let request = optimize_qdrant_edge_store(
+            APIToken::from_hex_text("test"),
+            Json(OptimizeQdrantEdgeStoreRequest { store_name: SEARCH_STORE.to_string() }),
+        );
+
+        let response = assert_runtime_stays_free(&QDRANT_EDGE_DATABASE, request).await;
+
+        // No database runs in the tests, so the request ends as soon as it gets the lock.
+        assert!(!response.success);
+    }
+
+    #[tokio::test(flavor = "multi_thread", worker_threads = 1)]
+    async fn the_info_waiting_for_the_database_leaves_the_runtime_free() {
+        let request = qdrant_edge_info(APIToken::from_hex_text("test"));
+
+        let info = assert_runtime_stays_free(&QDRANT_EDGE_DATABASE, request).await;
+        assert!(!info.is_available);
+    }
+
     const SEARCH_STORE: &str = "rag_6cc665a82b1e4d42bc748015b7b391ec";
     const POINT_X: &str = "0b5f1e8a-3c2d-4e6f-9a1b-7c8d9e0f1a2b";
     const POINT_Y: &str = "1c6a2f9b-4d3e-4f70-8b2c-8d9e0f1a2b3c";

runtime/src/runtime_api.rs

@@ -64,8 +64,6 @@ pub fn start_runtime_api() {
         .route("/system/runtime/info", get(crate::environment::get_runtime_info))
         .route("/system/language", get(crate::environment::read_user_language))
         .route("/system/username", get(crate::environment::read_user_name))
-        .route("/system/enterprise/config/id", get(crate::environment::read_enterprise_env_config_id))
-        .route("/system/enterprise/config/server", get(crate::environment::read_enterprise_env_config_server_url))
         .route("/system/enterprise/config/encryption_secret", get(crate::environment::read_enterprise_env_config_encryption_secret))
         .route("/system/enterprise/configs", get(crate::environment::read_enterprise_configs))
         .route("/retrieval/fs/extract", get(crate::file_data::extract_data))
@@ -113,3 +111,51 @@ fn install_rustls_crypto_provider() {
         let _ = rustls::crypto::aws_lc_rs::default_provider().install_default();
     });
 }
+
+#[cfg(test)]
+pub(crate) mod test_support {
+    use std::sync::Mutex;
+    use std::time::{Duration, Instant};
+
+    /// How long the lock stays held while the request waits for it.
+    const LOCK_DURATION: Duration = Duration::from_secs(1);
+
+    /// How long the probe may wait for the runtime. Far below the lock duration, so a request which
+    /// blocks the worker while it waits fails the test every time.
+    const PROBE_LIMIT: Duration = Duration::from_millis(200);
+
+    /// Starts a request while another thread holds the given lock, as if a long operation were
+    /// running, and checks that a probe still gets through the runtime meanwhile. Returns the
+    /// response once the lock is free again.
+    ///
+    /// Call it from a test with a single worker, `#[tokio::test(flavor = "multi_thread",
+    /// worker_threads = 1)]`. A request which blocks that worker while it waits then holds up the
+    /// probe, just as in the app it holds up the one HTTP/2 connection which all calls share.
+    pub(crate) async fn assert_runtime_stays_free<T, F>(lock: &'static Mutex<T>, request: F) -> F::Output
+    where
+        T: Send + 'static,
+        F: Future + Send + 'static,
+        F::Output: Send + 'static,
+    {
+        let (locked_sender, locked_receiver) = std::sync::mpsc::channel();
+        let lock_holder = std::thread::spawn(move || {
+            let guard = lock.lock().unwrap();
+            locked_sender.send(()).unwrap();
+            std::thread::sleep(LOCK_DURATION);
+            drop(guard);
+        });
+
+        locked_receiver.recv().unwrap();
+        let request = tokio::spawn(request);
+
+        let probe_started = Instant::now();
+        tokio::spawn(async {}).await.unwrap();
+        let probe_duration = probe_started.elapsed();
+
+        lock_holder.join().unwrap();
+        let response = request.await.unwrap();
+
+        assert!(probe_duration < PROBE_LIMIT, "The probe waited {probe_duration:?} for the runtime.");
+        response
+    }
+}

runtime/src/secret.rs

@@ -78,8 +78,37 @@ pub fn init_secret_store() {
     }
 }
 
+/// Runs one request against the credential store on the blocking pool.
+///
+/// The credential store of the operating system may take a moment to answer, and on macOS and Linux
+/// it may wait until the user unlocks it, which can take minutes. On a runtime worker, that would
+/// hold up every other call of the app, because they all share one HTTP/2 connection.
+async fn run_secret_store_request<T, F>(request: F) -> Result<T, String>
+where
+    T: Send + 'static,
+    F: FnOnce() -> T + Send + 'static,
+{
+    tokio::task::spawn_blocking(request).await.map_err(|error| {
+        let issue = format!("The secret store worker failed: {error}");
+        error!(Source = "Secret Store"; "{issue}");
+        issue
+    })
+}
+
 /// Stores a secret in the secret store using the operating system's keyring.
 pub async fn store_secret(_token: APIToken, request: Json<StoreSecret>) -> Json<StoreSecretResponse> {
+    match run_secret_store_request(move || store_secret_sync(request)).await {
+        Ok(response) => response,
+        Err(issue) => Json(StoreSecretResponse {
+            success: false,
+            issue,
+            issue_code: SecretStoreIssueCode::Unknown,
+        }),
+    }
+}
+
+/// Performs the blocking part of [`store_secret`].
+fn store_secret_sync(request: Json<StoreSecret>) -> Json<StoreSecretResponse> {
     let user_name = request.user_name.as_str();
     let decrypted_text = match ENCRYPTION.decrypt(&request.secret) {
         Ok(text) => text,
@@ -145,6 +174,19 @@ pub struct StoreSecretResponse {
 
 /// Retrieves a secret from the secret store using the operating system's keyring.
 pub async fn get_secret(_token: APIToken, request: Json<RequestSecret>) -> Json<RequestedSecret> {
+    match run_secret_store_request(move || get_secret_sync(request)).await {
+        Ok(response) => response,
+        Err(issue) => Json(RequestedSecret {
+            success: false,
+            secret: EncryptedText::new(String::from("")),
+            issue,
+            issue_code: SecretStoreIssueCode::Unknown,
+        }),
+    }
+}
+
+/// Performs the blocking part of [`get_secret`].
+fn get_secret_sync(request: Json<RequestSecret>) -> Json<RequestedSecret> {
     let user_name = request.user_name.as_str();
     let service = format!("mindwork-ai-studio::{}", request.destination);
     let entry = match Entry::new(service.as_str(), user_name) {
@@ -227,6 +269,19 @@ pub struct RequestedSecret {
 
 /// Deletes a secret from the secret store using the operating system's keyring.
 pub async fn delete_secret(_token: APIToken, request: Json<RequestSecret>) -> Json<DeleteSecretResponse> {
+    match run_secret_store_request(move || delete_secret_sync(request)).await {
+        Ok(response) => response,
+        Err(issue) => Json(DeleteSecretResponse {
+            success: false,
+            was_entry_found: false,
+            issue,
+            issue_code: SecretStoreIssueCode::Unknown,
+        }),
+    }
+}
+
+/// Performs the blocking part of [`delete_secret`].
+fn delete_secret_sync(request: Json<RequestSecret>) -> Json<DeleteSecretResponse> {
     let user_name = request.user_name.as_str();
     let service = format!("mindwork-ai-studio::{}", request.destination);
     let entry = match Entry::new(service.as_str(), user_name) {

runtime/src/tokenizer.rs

@@ -104,33 +104,58 @@ pub fn set_default_tokenizer_path(app_handle: tauri::AppHandle) {
 }
 
 pub async fn token_count(_token: APIToken, req: Json<SetTokenText>) -> Json<TokenizerResponse> {
-    match get_token_count(&req.tokenizer_path, &req.text) {
+    match run_tokenizer_work(move || get_token_count(&req.tokenizer_path, &req.text)).await {
         Ok(count) => Json(TokenizerResponse::available(count)),
         Err(e) => Json(TokenizerResponse::unavailable(e)),
     }
 }
 
 pub async fn validate_tokenizer(_token: APIToken, payload: Json<TokenizerPath>) -> Json<TokenizerResponse> {
-    match handle_tokenizer_validate(&PathBuf::from(payload.file_path.clone())) {
+    match run_tokenizer_work(move || handle_tokenizer_validate(&PathBuf::from(payload.file_path.clone()))).await {
         Ok(count) => Json(TokenizerResponse::available(count)),
         Err(e) => Json(TokenizerResponse::unavailable(e)),
     }
 }
 
 pub async fn store_tokenizer(_token: APIToken, payload: Json<TokenizerStorage>) -> Json<TokenizerResponse> {
-    match handle_tokenizer_store(&payload) {
+    match run_tokenizer_work(move || handle_tokenizer_store(&payload).map_err(|e| e.to_string())).await {
         Ok(dest_path) => Json(TokenizerResponse::stored(dest_path)),
-        Err(e) => Json(TokenizerResponse::unavailable(e.to_string())),
+        Err(e) => Json(TokenizerResponse::unavailable(e)),
     }
 }
 
 pub async fn delete_tokenizer(_token: APIToken, payload: Json<TokenizerDelete>) -> Json<TokenizerResponse> {
-    match handle_tokenizer_delete(&payload) {
+    match run_tokenizer_work(move || handle_tokenizer_delete(&payload).map_err(|e| e.to_string())).await {
         Ok(_) => Json(TokenizerResponse::stored(String::new())),
-        Err(e) => Json(TokenizerResponse::unavailable(e.to_string())),
+        Err(e) => Json(TokenizerResponse::unavailable(e)),
     }
 }
 
+/// Gets the tokenizer for the given path, loading it on the blocking pool when it is not cached yet.
+pub async fn get_tokenizer_off_worker(path: String) -> Result<Arc<Tokenizer>, String> {
+    run_tokenizer_work(move || get_tokenizer(&path)).await
+}
+
+/// Runs tokenizer work on the blocking pool.
+///
+/// Loading a tokenizer reads and parses its file, which takes most of a second for a large one, and
+/// it waits for the storage lock while another request copies or deletes a tokenizer. Encoding a
+/// long text is CPU-bound as well. On a runtime worker, any of this would hold up every other call
+/// of the app, because they all share one HTTP/2 connection.
+async fn run_tokenizer_work<T, F>(work: F) -> Result<T, String>
+where
+    T: Send + 'static,
+    F: FnOnce() -> Result<T, String> + Send + 'static,
+{
+    tokio::task::spawn_blocking(work)
+        .await
+        .unwrap_or_else(|error| {
+            let issue = format!("The tokenizer worker failed: {error}");
+            error!(Source = "Tokenizer"; "{issue}");
+            Err(issue)
+        })
+}
+
 fn handle_tokenizer_validate(path: &PathBuf) -> Result<usize, String> {
     validate_tokenizer_file(path)
 }
@@ -240,7 +265,7 @@ fn tokenizer_cache() -> &'static RwLock<HashMap<PathBuf, Arc<Tokenizer>>> {
     TOKENIZERS.get_or_init(|| RwLock::new(HashMap::new()))
 }
 
-pub fn get_tokenizer(path: &str) -> Result<Arc<Tokenizer>, String> {
+fn get_tokenizer(path: &str) -> Result<Arc<Tokenizer>, String> {
     let resolved_path = resolve_tokenizer_path(path)?;
     let tokenizer_path = fs::canonicalize(&resolved_path)
         .map_err(|e| format!("Could not resolve tokenizer file '{}': {e}", resolved_path.display()))?;
@@ -303,3 +328,31 @@ fn load_tokenizer_from_file(path: &PathBuf) -> Result<Tokenizer, String> {
     Tokenizer::from_file(path)
         .map_err(|e| format!("Failed to load tokenizer from '{}': {e}", path.display()))
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::runtime_api::test_support::assert_runtime_stays_free;
+
+    #[tokio::test(flavor = "multi_thread", worker_threads = 1)]
+    async fn a_token_count_waiting_for_the_storage_leaves_the_runtime_free() {
+        //
+        // Any file will do: the request waits for the storage lock before it reads the tokenizer.
+        // Since this file is no tokenizer, the request fails once it gets the lock.
+        //
+        let tokenizer_path = std::env::temp_dir().join(format!("ai-studio-tokenizer-{}.json", std::process::id()));
+        fs::write(&tokenizer_path, "This is not a tokenizer.").unwrap();
+
+        let request = token_count(
+            APIToken::from_hex_text("test"),
+            Json(SetTokenText {
+                text: "Hello, world!".to_string(),
+                tokenizer_path: tokenizer_path.to_string_lossy().to_string(),
+            }),
+        );
+
+        let response = assert_runtime_stays_free(&TOKENIZER_STORAGE_LOCK, request).await;
+        fs::remove_file(&tokenizer_path).unwrap();
+        assert!(!response.success);
+    }
+}