dsh-builtin-browser 0.4.0 → 0.4.2

package/lib/browser-electron/entry.js

@@ -17,8 +17,10 @@ import { join } from 'node:path';
 import { defaultHostMainPath, RemoteElectronViewHost } from './remote-host.js';
 import { DesktopBridgeViewHost } from './desktop-bridge-host.js';
 import { detectBrowser, searchSummary, SystemBrowserViewHost } from './system-browser.js';
+import { claimEphemeralProfile, ephemeralProfileName, sweepAbandonedEphemeralProfiles } from './ephemeral-profile.js';
 import { MissingSystemBrowserHost } from './missing-system-browser.js';
 import { SettingsStore } from './settings-store.js';
+import { sameOrigin } from './settings-route.js';
 export { ELECTRON_BROWSER_PROVIDER_ID, ElectronBrowserProvider, } from './provider.js';
 export { RemoteElectronViewHost, defaultHostMainPath } from './remote-host.js';
 /** Cordis plugin name used by loader diagnostics. */
@@ -83,7 +85,14 @@ export function apply(ctx, config) {
             ctx.logger?.info?.(`dsh-builtin-browser: ${note}`);
         }
         catch (error) {
-            // Never leave the surface without a provider.
+            // Never leave the surface without a provider. The half-built host is disposed FIRST: an
+            // adopt that got as far as constructing it may already hold a socket, a listener or a child
+            // process, and dropping it on the floor is how a failed upgrade becomes a leak that only
+            // shows up as a stray Electron in the task manager.
+            try {
+                host.dispose?.();
+            }
+            catch { /* the fallback below matters more than this */ }
             unregister = ctx.browser.registerBrowserProvider(new ElectronBrowserProvider(new RemoteElectronViewHost(defaultHostMainPath()), providerConfig));
             ctx.logger?.warn?.(`dsh-builtin-browser: could not adopt ${note} (${String(error)})`);
         }
@@ -118,15 +127,29 @@ export function apply(ctx, config) {
             }
         }
         else {
-            const home = process.env.DSH_HOME ?? homedir();
+            // Same profile root as settings-store, history-store and desktop-bridge-host: the
+            // `?.dsh` component used to be missing HERE only, so a desktop launched from a
+            // shortcut (no DSH_HOME in the environment) put its browser profile in
+            // `C:\Users\<user>\dsh-builtin-browser-host\` while its settings and history lived in
+            // `~/.dsh/dsh-builtin-browser-host/` — two half-profiles that never met, and a
+            // settings panel whose switches appeared not to apply.
+            const home = process.env.DSH_HOME ?? join(homedir(), '.dsh');
             // Login state lives in the browser profile, so the cookies setting decides where
             // that profile goes: a stable directory keeps the user signed in across restarts
             // (the point of using their own browser), while turning persistence off gets a
             // throwaway directory that is removed when the browser is released.
             const persist = settings.get().cookies.persist;
+            const profileRoot = join(home, BROWSER_PROFILE_DIR);
             const profileDir = persist
-                ? join(home, 'dsh-builtin-browser-host', `${detected.kind}-profile`)
-                : join(home, 'dsh-builtin-browser-host', `${detected.kind}-profile-ephemeral-${randomUUID()}`);
+                ? join(profileRoot, `${detected.kind}-profile`)
+                : join(profileRoot, ephemeralProfileName(detected.kind, randomUUID()));
+            if (!persist) {
+                // A process killed outright never runs its own cleanup (see ephemeral-profile.ts),
+                // so the sweep happens HERE — at the only moment that can still act on a profile
+                // whose owner is gone — and before this run claims a directory of its own.
+                sweepAbandonedEphemeralProfiles(profileRoot, detected.kind);
+                claimEphemeralProfile(profileDir);
+            }
             adopt(new SystemBrowserViewHost(detected, profileDir, [], persist ? undefined : profileDir), `will drive the installed ${detected.kind} (${detected.path}${persist ? '' : ', ephemeral profile'})`);
         }
     }
@@ -134,16 +157,41 @@ export function apply(ctx, config) {
     // replace it, which is exactly what this guard prevents.
     if (explicitChoice)
         return;
-    void DesktopBridgeViewHost.discover().then(sidebar => {
-        // No bridge (plain `dsh web`, an older desktop build, or a shell that already
-        // exited): keep self-hosting, which is what every surface has always done.
-        if (sidebar === undefined)
+    // The shell may not be up yet. `dsh web` starts before — or entirely without — the desktop app
+    // that hosts the bridge, and a one-shot discovery meant a bridge that appeared a second later
+    // was never adopted for the rest of the process's life. Retry a handful of times over roughly
+    // half a minute, then stop: this is a convenience, not a promise, and the session the user is
+    // already driving in the self-hosted window is not worth replacing minutes into their work.
+    // `adopt` is itself one-shot (its `upgraded` guard), so the only possible effect of a retry
+    // that lands late is adopting a bridge that showed up just after startup.
+    void (async () => {
+        for (const delay of [0, 2_000, 5_000, 15_000]) {
+            if (delay > 0)
+                await new Promise(resolve => setTimeout(resolve, delay));
+            let sidebar;
+            try {
+                sidebar = await DesktopBridgeViewHost.discover();
+            }
+            catch {
+                sidebar = undefined; // discovery is documented not to throw; never let it stop the retries
+            }
+            if (sidebar === undefined)
+                continue;
+            adopt(sidebar, 'driving the desktop sidebar browser');
             return;
-        adopt(sidebar, 'driving the desktop sidebar browser');
-    }).catch(() => { });
+        }
+    })();
 }
 /** Route the settings panel reads and writes. */
 const SETTINGS_ROUTE = '/dsh-builtin-browser/settings';
+/**
+ * The directory under the DSH home that holds this plugin's browser profile.
+ *
+ * Named once, because every consumer of it must agree: `settingsPath()` and the history
+ * store resolve their own copy from `$DSH_HOME`, and a profile written under a different
+ * root than the settings document is a pair that silently never meets.
+ */
+const BROWSER_PROFILE_DIR = 'dsh-builtin-browser-host';
 /**
  * Expose the settings document over the host's web server, which is how the
  * settings panel reads and writes it. A host without that service (headless /
@@ -169,31 +217,9 @@ function installSettingsRoute(ctx, settings) {
         // No web server on this surface: the settings file remains the interface.
     }
 }
-/**
- * Same-origin guard. Settings are machine-local configuration, so a page loaded
- * elsewhere must not be able to read or rewrite them.
- * @param request - the incoming request.
- * @returns whether the request may touch the settings document.
- */
-function sameOrigin(request) {
-    const site = String(request.headers['sec-fetch-site'] ?? '').toLowerCase();
-    if (site === 'cross-site')
-        return false;
-    const origin = String(request.headers.origin ?? '').trim();
-    if (origin === 'null')
-        return false;
-    if (origin === '')
-        return true;
-    try {
-        return new URL(origin).host.toLowerCase() === String(request.headers.host ?? '').trim().toLowerCase();
-    }
-    catch {
-        return false;
-    }
-}
 /**
  * One settings round-trip: `GET` reads the document, `PUT`/`POST` merges a
- * partial patch into it.
+ * partial patch into it. Admission lives in `settings-route.ts` so it can be tested.
  * @param request - the incoming request.
  * @param response - the response to write.
  * @param settings - the settings document.
@@ -206,11 +232,11 @@ async function handleSettingsRequest(request, response, settings) {
         });
         response.end(JSON.stringify(body));
     };
-    if (!sameOrigin(request)) {
+    const method = request.method ?? 'GET';
+    if (!sameOrigin(request, method !== 'GET')) {
         json(403, { ok: false, error: 'cross-origin request refused' });
         return;
     }
-    const method = request.method ?? 'GET';
     if (method === 'GET') {
         json(200, { ok: true, settings: settings.get(), path: settings.path() });
         return;

package/lib/browser-electron/ephemeral-profile.d.ts

@@ -0,0 +1,37 @@
+/** Test seams; production passes neither. */
+export interface SweepOptions {
+    /** The clock, for the age check. */
+    readonly now?: number;
+    /** Liveness test for an owner pid. */
+    readonly alive?: (pid: number) => boolean;
+}
+/**
+ * The directory name an ephemeral profile for `kind` gets, before it is claimed.
+ *
+ * The sweep matches on the same shape, so the two must agree: a name built anywhere else
+ * is a directory the sweep will never find.
+ * @param kind - the detected browser product (`chrome`, `edge`, ...).
+ * @param id - a per-run unique suffix.
+ * @returns the directory NAME, not a path.
+ */
+export declare function ephemeralProfileName(kind: string, id: string): string;
+/**
+ * Mark an ephemeral profile directory as this process's, so a later run can tell whether
+ * the process that created it is still there.
+ * @param dir - the profile directory (created if it does not exist yet).
+ * @param pid - the claiming process, defaulting to this one.
+ * @param now - the claim time, defaulting to the current clock.
+ * @returns true when the marker was written.
+ */
+export declare function claimEphemeralProfile(dir: string, pid?: number, now?: number): boolean;
+/**
+ * Delete the ephemeral profiles under `root` that belong to processes which are gone.
+ *
+ * Called before a new one is claimed, because that is the only moment that can still act
+ * on a profile whose owner was killed: nothing runs when the signal lands.
+ * @param root - the plugin's profile root (`<DSH_HOME>/dsh-builtin-browser-host`).
+ * @param kind - the browser product, which selects the directory-name prefix.
+ * @param options - test seams for the clock and the liveness test.
+ * @returns the names of the directories that were removed.
+ */
+export declare function sweepAbandonedEphemeralProfiles(root: string, kind: string, options?: SweepOptions): string[];

package/lib/browser-electron/ephemeral-profile.js

@@ -0,0 +1,146 @@
+/**
+ * The ephemeral browser profile: a throwaway directory under the DSH home, used when the
+ * user has turned profile persistence off, and deleted when the browser is released.
+ *
+ * Deleting on release is not enough by itself. A process killed outright — a DSH restart
+ * that ends in SIGKILL, "End task" in a task manager, a crashed host — never runs its
+ * cleanup, so the profile it was using survives on disk: the user asked for no profile to
+ * be kept and the machine keeps one anyway, holding cookies and history. Nothing can run
+ * at SIGKILL time, so the cleanup has to happen on the NEXT run instead, which is what
+ * this module provides.
+ *
+ * Ownership is a marker file inside the directory naming the process that claimed it. A
+ * directory whose owner is gone AND which has not been written to for a while is
+ * abandoned. The age check is what keeps the sweep away from a profile an orphaned
+ * browser (a child that outlived the parent which was killed) may still be using: a live
+ * Chromium writes to its profile continuously, so a directory that has not changed in an
+ * hour is not in use.
+ */
+import { mkdirSync, readFileSync, readdirSync, rmSync, statSync, writeFileSync } from 'node:fs';
+import { join } from 'node:path';
+/** The marker file, named once because both the claim and the sweep read it. */
+const OWNER_FILE = '.dsh-ephemeral-owner';
+/**
+ * How long a profile is left alone regardless of ownership. An orphaned browser keeps
+ * writing to its profile, so this is what distinguishes "abandoned" from "in use by a
+ * process this one did not start".
+ */
+const ABANDONED_AFTER_MS = 60 * 60 * 1000;
+/**
+ * Whether a process with this pid exists.
+ *
+ * A signal of 0 performs the permission and existence checks without delivering anything.
+ * EPERM is a live process owned by someone else — alive, not gone.
+ * @param pid - the pid to test.
+ * @returns true when the process exists.
+ */
+function processAlive(pid) {
+    try {
+        process.kill(pid, 0);
+        return true;
+    }
+    catch (error) {
+        return error.code === 'EPERM';
+    }
+}
+/**
+ * The directory name an ephemeral profile for `kind` gets, before it is claimed.
+ *
+ * The sweep matches on the same shape, so the two must agree: a name built anywhere else
+ * is a directory the sweep will never find.
+ * @param kind - the detected browser product (`chrome`, `edge`, ...).
+ * @param id - a per-run unique suffix.
+ * @returns the directory NAME, not a path.
+ */
+export function ephemeralProfileName(kind, id) {
+    return `${kind}-profile-ephemeral-${id}`;
+}
+/**
+ * Mark an ephemeral profile directory as this process's, so a later run can tell whether
+ * the process that created it is still there.
+ * @param dir - the profile directory (created if it does not exist yet).
+ * @param pid - the claiming process, defaulting to this one.
+ * @param now - the claim time, defaulting to the current clock.
+ * @returns true when the marker was written.
+ */
+export function claimEphemeralProfile(dir, pid = process.pid, now = Date.now()) {
+    try {
+        mkdirSync(dir, { recursive: true });
+        writeFileSync(join(dir, OWNER_FILE), JSON.stringify({ pid, at: now }));
+        return true;
+    }
+    catch {
+        // The browser creates the directory itself when this fails, and cleanup is best
+        // effort throughout: a marker that could not be written must not stop the launch.
+        return false;
+    }
+}
+/**
+ * The claimed owner of an ephemeral profile, when it has one.
+ * @param dir - the profile directory.
+ * @returns the owner pid, or undefined when there is no readable marker.
+ */
+function readOwner(dir) {
+    try {
+        const parsed = JSON.parse(readFileSync(join(dir, OWNER_FILE), 'utf8'));
+        return typeof parsed.pid === 'number' ? parsed.pid : undefined;
+    }
+    catch {
+        // No marker: a run from before this existed, or one killed before it wrote one. The
+        // age check has already applied, and a directory with no claimant that old is
+        // abandoned by any reading.
+        return undefined;
+    }
+}
+/**
+ * Delete the ephemeral profiles under `root` that belong to processes which are gone.
+ *
+ * Called before a new one is claimed, because that is the only moment that can still act
+ * on a profile whose owner was killed: nothing runs when the signal lands.
+ * @param root - the plugin's profile root (`<DSH_HOME>/dsh-builtin-browser-host`).
+ * @param kind - the browser product, which selects the directory-name prefix.
+ * @param options - test seams for the clock and the liveness test.
+ * @returns the names of the directories that were removed.
+ */
+export function sweepAbandonedEphemeralProfiles(root, kind, options = {}) {
+    const now = options.now ?? Date.now();
+    const alive = options.alive ?? processAlive;
+    const prefix = `${kind}-profile-ephemeral-`;
+    const removed = [];
+    let names;
+    try {
+        names = readdirSync(root);
+    }
+    catch {
+        // No profile root yet: nothing has ever been written here, so nothing to sweep.
+        return removed;
+    }
+    for (const name of names) {
+        if (!name.startsWith(prefix))
+            continue;
+        const dir = join(root, name);
+        let ageMs;
+        try {
+            ageMs = now - statSync(dir).mtimeMs;
+        }
+        catch {
+            continue;
+        }
+        // Recent writes mean a browser is still working in there. A killed DSH can leave its
+        // Chromium child running, and that child's profile is not garbage.
+        if (ageMs < ABANDONED_AFTER_MS)
+            continue;
+        const owner = readOwner(dir);
+        if (owner !== undefined && alive(owner))
+            continue;
+        try {
+            rmSync(dir, { recursive: true, force: true });
+            removed.push(name);
+        }
+        catch {
+            // In use, or not ours to remove. Leaving it is the safe outcome and the next run
+            // tries again — the alternative is deleting a profile out from under a live browser.
+        }
+    }
+    return removed;
+}

package/lib/browser-electron/history-store.d.ts

@@ -34,6 +34,22 @@ export declare function visitedHistoryPath(): string;
  */
 export declare class HistoryStore {
     private readonly file;
+    /**
+     * File size at which the next rewrite happens.
+     *
+     * Advanced after each rewrite so the cost is paid once per PRUNE_SLACK appends
+     * rather than on every append: asking "is it over the cap?" was true from the first
+     * over-cap write onwards, which is what made every later navigation rewrite the file.
+     */
+    private pruneThreshold;
+    /**
+     * Appends since the last rewrite.
+     *
+     * prune() runs after every navigation, and the answer is almost always "nothing to
+     * do" — but working that out from the file costs a full parse. Counting instead makes
+     * the common case a comparison.
+     */
+    private appendedSincePrune;
     private readonly maxEntries;
     private readonly maxAgeMs;
     /**
@@ -69,6 +85,15 @@ export declare class HistoryStore {
      * the count cap. Cheap no-op until the file grows past the cap, so callers can
      * run it after every append.
      */
+    /**
+     * How far past the cap the file may grow before it is rewritten.
+     *
+     * Scaled to the cap rather than fixed: a fixed 500 was meaningless when the cap is
+     * small (a caller asking for three entries would never reach 503), and the slack is
+     * there to batch rewrites, not to overrule the caller's bound.
+     * @returns the number of extra entries tolerated before a rewrite.
+     */
+    private slack;
     prune(): void;
     /** Parse the whole file, skipping blank and damaged lines. */
     private readAll;

package/lib/browser-electron/history-store.js

@@ -12,9 +12,10 @@
  * instead of poisoning the whole file.
  * @module dsh-browser/browser-electron/history-store
  */
-import { appendFileSync, existsSync, mkdirSync, readFileSync, writeFileSync } from 'node:fs';
+import { appendFileSync, existsSync, mkdirSync, readFileSync } from 'node:fs';
 import { homedir } from 'node:os';
 import { dirname, join } from 'node:path';
+import { writeFileAtomic } from './settings-store.js';
 /** Default entry cap — mirrors the documented retention rule. */
 const DEFAULT_MAX_ENTRIES = 5_000;
 /** Default age cap in days. */
@@ -37,6 +38,22 @@ export function visitedHistoryPath() {
  */
 export class HistoryStore {
     file;
+    /**
+     * File size at which the next rewrite happens.
+     *
+     * Advanced after each rewrite so the cost is paid once per PRUNE_SLACK appends
+     * rather than on every append: asking "is it over the cap?" was true from the first
+     * over-cap write onwards, which is what made every later navigation rewrite the file.
+     */
+    pruneThreshold;
+    /**
+     * Appends since the last rewrite.
+     *
+     * prune() runs after every navigation, and the answer is almost always "nothing to
+     * do" — but working that out from the file costs a full parse. Counting instead makes
+     * the common case a comparison.
+     */
+    appendedSincePrune = 0;
     maxEntries;
     maxAgeMs;
     /**
@@ -46,6 +63,8 @@ export class HistoryStore {
     constructor(file = visitedHistoryPath(), limits = {}) {
         this.file = file;
         this.maxEntries = limits.maxEntries ?? DEFAULT_MAX_ENTRIES;
+        // The first rewrite is due once the file has grown PRUNE_SLACK past the cap.
+        this.pruneThreshold = this.maxEntries + this.slack();
         this.maxAgeMs = (limits.maxAgeDays ?? DEFAULT_MAX_AGE_DAYS) * 24 * 60 * 60 * 1000;
     }
     /** Whether the history file exists yet (diagnostics and tests). */
@@ -66,6 +85,7 @@ export class HistoryStore {
         catch {
             // Best-effort: history is not worth failing a navigation over.
         }
+        this.appendedSincePrune++;
     }
     /**
      * Read visits, newest first. Unparseable lines are skipped; entries older
@@ -108,7 +128,22 @@ export class HistoryStore {
      * the count cap. Cheap no-op until the file grows past the cap, so callers can
      * run it after every append.
      */
+    /**
+     * How far past the cap the file may grow before it is rewritten.
+     *
+     * Scaled to the cap rather than fixed: a fixed 500 was meaningless when the cap is
+     * small (a caller asking for three entries would never reach 503), and the slack is
+     * there to batch rewrites, not to overrule the caller's bound.
+     * @returns the number of extra entries tolerated before a rewrite.
+     */
+    slack() {
+        return Math.min(PRUNE_SLACK, Math.max(1, Math.floor(this.maxEntries / 2)));
+    }
     prune() {
+        // Cheap first: past the cap but without enough new entries to justify a rewrite.
+        // Reading the file to reach this conclusion was the remaining per-navigation cost.
+        if (this.appendedSincePrune === 0)
+            return;
         const all = this.readAll();
         const cutoff = Date.now() - this.maxAgeMs;
         const fresh = all.filter(page => page.at >= cutoff);
@@ -117,14 +152,26 @@ export class HistoryStore {
         // is worth a full rewrite.
         if (kept.length === all.length)
             return;
-        if (all.length - kept.length < PRUNE_SLACK && fresh.length <= this.maxEntries)
+        // Only rewrite once the file has grown PRUNE_SLACK past what we last wrote. Asking
+        // "is the file over the cap?" instead was true from the first over-cap append
+        // onwards, so every single navigation after that rewrote the whole file.
+        if (all.length < this.pruneThreshold)
             return;
         try {
             const body = kept.map(page => `${JSON.stringify(page)}\n`).join('');
-            writeFileSync(this.file, body);
+            // Atomic, like the settings document. A plain writeFileSync truncates first, and this is
+            // the whole history in one call: a TerminateProcess mid-write — the browser host is
+            // killed exactly that way on Windows — left a partial file, and the reader drops a
+            // half-written TAIL rather than the middle, so everything after the tear was lost.
+            writeFileAtomic(this.file, body);
+            // The file now holds only what was kept, so the next rewrite is due once
+            // PRUNE_SLACK more entries have been appended.
+            this.pruneThreshold = kept.length + this.slack();
+            this.appendedSincePrune = 0;
         }
         catch {
-            // Best-effort: an unpruned file still reads correctly.
+            // Best-effort: an unpruned file still reads correctly, and the atomic write leaves the
+            // previous one in place when it fails.
         }
     }
     /** Parse the whole file, skipping blank and damaged lines. */