Merge pull request #8 from Twest2/StepForge pr/05-capture-fixes

Fix region capture, power ownership, click-source reporting, strict timing (plan PR 5)

README.md

@@ -1,17 +1,26 @@
 # StepForge
 
-StepForge is a **fully offline**, open-source desktop app for Windows, with
+StepForge is a **local-first**, open-source desktop app for Windows, with
 Linux (WIP) builds. It captures step-by-step workflows as screenshots, lets
 you annotate and describe each step in a focused three-pane editor, and
 exports the result to Markdown, DOCX, PPTX, PDF, HTML (WIP), GIF (WIP),
 confluence (WIP), Wiki.js (WIP), and image bundles (WIP). The current
 reconmendations for exporting is Markdown and PDF.
 
-It is an independent offline desktop guide-capture tool inspired by publicly
-documented workflow patterns of commercial documentation tools like Folge. It contains no
-third-party branding, assets, or code from those tools, and it never talks to
-the network: no telemetry, no update checks, no license checks, no cloud, no
-remote AI.
+It is an independent desktop guide-capture tool inspired by publicly
+documented workflow patterns of commercial documentation tools like Folge. It
+contains no third-party branding, assets, or code from those tools.
+
+**Network and privacy contract.** StepForge has no telemetry, no update
+checks, no license checks, and no cloud. Guides never leave your machine on
+their own. The only outbound network feature is the **optional** AI
+integration: when *you* enable it and configure an [Ollama](https://ollama.com)
+endpoint, StepForge sends step screenshots and text to that endpoint to
+generate titles and descriptions. By default that endpoint must be **local
+(loopback)**; sending data to a remote host requires the explicit "Allow
+remote AI host" opt-in. See [docs/PRIVACY.md](docs/PRIVACY.md) for exactly
+what is collected and sent. Note that OCR (Tesseract) and its English language
+data are bundled production dependencies — Electron is not the only one.
 
 ## Overview
 

app/capture.js

@@ -198,11 +198,18 @@ class CaptureService {
     notify,
     screenApi = screen,
     textIntel = null,
+    powerPolicy = null,
   }) {
     this.store = store;
     this.settings = settings;
     this.getWindow = getWindow;
     this.notify = notify;
+    // Single owner of OS power/throttling state for the capture lifecycle.
+    // setRecording(true) is called exactly while a session is actively
+    // recording (session present and not paused); setRecording(false) whenever
+    // it pauses or ends. No-op by default so tests and non-Electron hosts work.
+    this.powerPolicy = powerPolicy || { setRecording() {} };
+    this._recordingPower = false;
     // Injectable for tests; the click/coordinate paths must never reach for
     // the global `screen` directly so coordinate handling stays testable.
     this.screen = screenApi;
@@ -213,6 +220,10 @@ class CaptureService {
     this.session = null; // { guideId, paused, count, intervalSec }
     this.intervalTimer = null;
     this.clickWatcher = null;
+    // Explicit trigger source rather than a bare boolean, so the UI can tell
+    // the truth about how clicks are being captured (or that they are not):
+    //   windows-hook | x11 | evdev-x11 | evdev-wayland | unavailable
+    this.clickSource = 'unavailable';
     this.frameLoopTimer = null;
     this.frameLoopRunning = false;
     this.frameWaiters = [];
@@ -240,6 +251,22 @@ class CaptureService {
     this.warmingUp = false;
   }
 
+  /**
+   * Reconcile OS power state with the actual recording state. Called after
+   * every session transition so there is exactly one owner: the blocker is
+   * held iff a session exists and is not paused. Idempotent.
+   */
+  syncPower() {
+    const recording = Boolean(this.session && !this.session.paused);
+    if (recording === this._recordingPower) return;
+    this._recordingPower = recording;
+    try {
+      this.powerPolicy.setRecording(recording);
+    } catch {
+      // power management is best-effort; never break capture over it
+    }
+  }
+
   state() {
     return this.session
       ? {
@@ -248,12 +275,21 @@ class CaptureService {
         guideId: this.session.guideId,
         count: this.session.count,
         intervalSec: this.session.intervalSec || 0,
-        clickCapture: Boolean(this.clickWatcher),
+        // clickCapture reflects any live global click source (xinput/evdev/
+        // Windows hook), not just the spawned-process watcher — evdev has no
+        // child process, so the old Boolean(this.clickWatcher) reported false
+        // while clicks were in fact being captured.
+        clickCapture: this.clickSource !== 'unavailable',
+        clickSource: this.clickSource,
         clickCaptureAvailable: this.clickCaptureAvailable(),
         clickFrameSource: this.streamBackend ? 'stream' : (this.frameLoopRunning ? 'loop' : 'idle'),
         strictClickFrames: this.strictClickFrames(),
       }
-      : { active: false, clickCaptureAvailable: this.clickCaptureAvailable() };
+      : {
+        active: false,
+        clickSource: this.clickSource,
+        clickCaptureAvailable: this.clickCaptureAvailable(),
+      };
   }
 
   /**
@@ -331,6 +367,9 @@ class CaptureService {
     this.session = { guideId, paused: true, count: 0, intervalSec: interval };
     if (this.settings.get('capture.captureOutsideClicks') !== false) this.startClickWatcher();
     this.applyInterval();
+    // A new session starts paused, so it must NOT hold the power blocker yet
+    // (it did before, leaking the blocker while idle). syncPower reconciles.
+    this.syncPower();
     this.notify('capture:state', this.state());
 
     // (Skipped for the dev screenshot hook, which needs a visible page.)
@@ -476,6 +515,9 @@ class CaptureService {
       this.stopFrameLoop();
       this.stopClickFrameBackend();
     }
+    // Recording only while unpaused: this is the one place tray/second-instance
+    // pauses previously bypassed, leaking the power blocker. syncPower owns it.
+    this.syncPower();
     if (this.rebuildTrayMenu) this.rebuildTrayMenu();
     this.notify('capture:state', this.state());
   }
@@ -555,6 +597,9 @@ class CaptureService {
     this.stopClickFrameBackend();
     this.destroySessionTray();
     this.session = null;
+    // A finished session never records: release the power blocker here so it
+    // can't outlive the recording (finish from any path — bar, tray, quit).
+    this.syncPower();
     if (this.hiddenForSession) {
       this.hiddenForSession = false;
       this.showWindow();
@@ -562,6 +607,23 @@ class CaptureService {
     this.notify('capture:state', this.state());
   }
 
+  /**
+   * Best-effort drain of clicks still encoding in the queue, for application
+   * shutdown. Resolves when the queue settles or the deadline passes so quit
+   * is never blocked indefinitely. Clicks captured mid-session are pinned to
+   * their guide id and stored even after the session ends (see onOsClick).
+   */
+  async drainPendingClicks(timeoutMs = 2000) {
+    try {
+      await Promise.race([
+        this.clickQueue.catch(() => {}),
+        new Promise((resolve) => setTimeout(resolve, Math.max(0, timeoutMs))),
+      ]);
+    } catch {
+      // best effort
+    }
+  }
+
   /**
    * True when the user is interacting with StepForge itself. Deliberately
    * based on cursor position over the visible window, not isFocused():
@@ -624,11 +686,22 @@ class CaptureService {
         if (result.ok) this.noteStepAdded(result.step, trigger, guideId);
         return result;
       }
-      // No usable frame: fall through to a one-off fresh shot — but only
-      // while still recording. After a stop, a fresh shot would show
-      // whatever replaced the user's workflow on screen.
-      clog('click@', clickAt, 'no frame qualified — falling back to a fresh (post-click) shot');
       if (!sessionLive) return { ok: false, reason: 'session ended before the fallback shot' };
+      // No usable pre-click frame. In strict mode a fresh shot now would be a
+      // POST-click frame — exactly what strict mode promises never to store.
+      // Skip with a visible diagnostic instead of silently labeling a
+      // post-click fallback as strict. Non-strict mode keeps the legacy
+      // fresh-shot fallback below.
+      if (this.strictClickFrames()) {
+        clog('click@', clickAt, 'strict mode — no pre-click frame; skipping rather than storing a post-click shot');
+        this.notify('capture:diagnostic', {
+          kind: 'strict-click-skipped',
+          guideId,
+          reason: 'No pre-click frame was ready for this click. In strict timing mode StepForge skips the shot rather than capturing the screen after the click.',
+        });
+        return { ok: false, reason: 'strict mode: no pre-click frame available for this click' };
+      }
+      clog('click@', clickAt, 'no frame qualified — falling back to a fresh (post-click) shot');
     }
 
     // For non-click triggers (interval, hotkey, manual) pull the latest frame
@@ -1009,6 +1082,7 @@ class CaptureService {
           ? ['stdbuf', '-oL', 'xinput', 'test-xi2', '--root']
           : ['xinput', 'test-xi2', '--root'];
         this.clickWatcher = spawn(argv[0], argv.slice(1), { stdio: ['ignore', 'pipe', 'ignore'] });
+        this.clickSource = 'x11';
         this.clickWatcher.stdout.on('data', (chunk) => {
           this.ingestClickWatcherChunk(chunk.toString(), 'linux');
         });
@@ -1025,6 +1099,11 @@ class CaptureService {
         // by other processes and a polling loop can miss short clicks under
         // load; WH_MOUSE_LL gives us one event for each button-down, with the
         // hook-time cursor position and timestamp.
+        //
+        // Raw typed-text capture is a keylogging surface, so the hook only
+        // emits printable CHAR events when the user explicitly opted in; by
+        // default the characters never even cross the process boundary.
+        const captureTypedText = this.settings.get('capture.captureTypedText') ? 'true' : 'false';
         const ps = `
 $ErrorActionPreference = 'Stop'
 Add-Type -TypeDefinition @'
@@ -1054,6 +1133,7 @@ public static class SFHook {
   private const uint PROCESS_POWER_THROTTLING_EXECUTION_SPEED = 0x1;
   private const uint HIGH_PRIORITY_CLASS = 0x00000080;
 
+  private static readonly bool CaptureTypedText = ${captureTypedText};
   private static IntPtr hook = IntPtr.Zero;
   private static IntPtr keyHook = IntPtr.Zero;
   private static LowLevelMouseProc proc = MouseHookCallback;
@@ -1306,8 +1386,10 @@ public static class SFHook {
             queue.Enqueue("KEY Escape " + unixMs); signal.Set();
           } else if (vk == 0x0D) {
             queue.Enqueue("KEY Enter " + unixMs); signal.Set();
-          } else {
-            // Map to Unicode character using current keyboard layout + shift state.
+          } else if (CaptureTypedText) {
+            // Map to Unicode character using current keyboard layout + shift
+            // state. Only reached when the user opted into typed-text capture;
+            // otherwise raw characters are never read or emitted.
             byte[] ks = new byte[256];
             GetKeyboardState(ks);
             var sb = new System.Text.StringBuilder(4);
@@ -1366,6 +1448,7 @@ public static class SFHook {
           stdio: ['ignore', 'pipe', 'pipe'],
           windowsHide: true,
         });
+        this.clickSource = 'windows-hook';
         this.clickWatcher.stdout.on('data', (chunk) => {
           this.ingestClickWatcherChunk(chunk.toString(), 'win32');
         });
@@ -1417,6 +1500,7 @@ public static class SFHook {
       this.clickWatcher = null;
     }
     this.stopEvdevWatcher();
+    this.clickSource = 'unavailable';
     this.clickWatcherBuf = '';
     this.linuxEvent = null;
     this.discardPendingRawClick();
@@ -1442,26 +1526,48 @@ public static class SFHook {
           buf = rest;
           for (const button of presses) this.onOsClick(Date.now(), null, button);
         });
-        // A device can disappear (unplugged); just drop that stream.
-        stream.on('error', () => {});
+        // A device can disappear (unplugged): drop that stream, and if it was
+        // the last live click source, fall back like any other watcher loss
+        // instead of silently going dark.
+        stream.on('error', () => this.handleEvdevStreamLoss(stream, 'device read error'));
+        stream.on('close', () => this.handleEvdevStreamLoss(stream, 'device closed'));
         this.evdevStreams.push(stream);
       } catch {
         // Node became unreadable between enumeration and open — skip it.
       }
     }
     if (!this.evdevStreams.length) {
-      console.error('[stepforge] no readable mouse input devices — add your user to the "input" group for per-click capture: sudo usermod -aG input "$USER" (then log out and back in)');
+      this.clickSource = 'unavailable';
+      console.error('[stepforge] no readable mouse input devices for per-click capture (see docs/linux for the least-privilege device-access setup)');
     } else {
+      // evdev carries no cursor position on Wayland (no marker); on X11 without
+      // xinput it is the click source but likewise has no root coordinates.
+      this.clickSource = this.onWayland() ? 'evdev-wayland' : 'evdev-x11';
       console.log(`[stepforge] per-click capture via evdev on ${this.evdevStreams.length} device(s)${this.onWayland() ? ' (Wayland: no click marker)' : ''}`);
     }
   }
 
+  /** One evdev device stream ended; drop it and fall back if none remain. */
+  handleEvdevStreamLoss(stream, reason) {
+    if (!this.evdevStreams) return; // already stopped deliberately
+    const idx = this.evdevStreams.indexOf(stream);
+    if (idx === -1) return;
+    this.evdevStreams.splice(idx, 1);
+    try { stream.destroy(); } catch { /* already gone */ }
+    if (this.evdevStreams.length === 0) {
+      this.evdevStreams = null;
+      this.clickSource = 'unavailable';
+      this.handleClickWatcherLoss(`evdev ${reason}`);
+    }
+  }
+
   stopEvdevWatcher() {
     if (!this.evdevStreams) return;
-    for (const stream of this.evdevStreams) {
+    const streams = this.evdevStreams;
+    this.evdevStreams = null; // clear first so close handlers no-op
+    for (const stream of streams) {
       try { stream.destroy(); } catch { /* already closed */ }
     }
-    this.evdevStreams = null;
   }
 
   /**
@@ -1790,6 +1896,11 @@ public static class SFHook {
     this._keyLastAt = now;
 
     if (type === 'CHAR') {
+      // Raw typed-text capture is off by default: it can record passwords or
+      // other secrets. Only buffer printable characters when the user has
+      // explicitly opted in via capture.captureTypedText. Shortcut/navigation
+      // keys below are unaffected.
+      if (!this.settings.get('capture.captureTypedText')) return;
       const ch = typeof data === 'number' ? String.fromCharCode(data) : String(data);
       this._keyBuffer = (this._keyBuffer + ch).slice(-200);
     } else if (type === 'KEY') {
@@ -1994,13 +2105,43 @@ public static class SFHook {
     const cropped = image.crop(rect);
     const size = cropped.getSize();
     if (!size.width || !size.height) return { ok: false, reason: 'empty selection' };
-    const step = await this.storeFrameAsStep(guideId, 'region', {
+    // storeFrameAsStep already returns { ok, step }; return it directly so
+    // callers see result.step.stepId — not result.step.step.stepId, which is
+    // what wrapping it in another { ok, step } produced (region selection and
+    // region auto-documentation both read result.step.stepId).
+    return this.storeFrameAsStep(guideId, 'region', {
       image: cropped,
       size,
       display,
       cursor: null,
     }, null, null);
-    return { ok: true, step };
+  }
+
+  /**
+   * Convert an overlay selection (display px) into an image-space crop rect,
+   * clamped to the image bounds. Returns null for an empty/invalid rect.
+   */
+  overlayRectToImageRect(rect, display, imgSize) {
+    if (!rect || !imgSize) return null;
+    const { width: iw, height: ih } = imgSize;
+    if (!iw || !ih) return null;
+    const sx = iw / display.bounds.width;
+    const sy = ih / display.bounds.height;
+    const toNum = (v) => (Number.isFinite(v) ? v : 0);
+    let x = Math.round(toNum(rect.x) * sx);
+    let y = Math.round(toNum(rect.y) * sy);
+    let w = Math.round(toNum(rect.w) * sx);
+    let h = Math.round(toNum(rect.h) * sy);
+    // Normalize negative-size drags (drawn up/left).
+    if (w < 0) { x += w; w = -w; }
+    if (h < 0) { y += h; h = -h; }
+    // Clamp to the image so image.crop can never read out of bounds.
+    x = Math.min(Math.max(0, x), iw);
+    y = Math.min(Math.max(0, y), ih);
+    w = Math.min(w, iw - x);
+    h = Math.min(h, ih - y);
+    if (w <= 0 || h <= 0) return null;
+    return { x, y, width: w, height: h };
   }
 
   /** Fullscreen overlay window that resolves with a crop rect (image px). */
@@ -2025,31 +2166,28 @@ public static class SFHook {
       });
       // The overlay may only display region.html; deny navigation/popups.
       require('./security').installWindowSecurity(overlay, 'region');
+      const { ipcMain } = require('electron');
       let settled = false;
+      // Idempotent cleanup: remove the IPC listener and close the overlay
+      // exactly once, whether the user picked, cancelled, closed, or the page
+      // failed to load. Previously the listener was only removed on a pick, so
+      // cancelling/closing leaked it (and the captured overlay/image refs).
       const finish = (rect) => {
         if (settled) return;
         settled = true;
+        ipcMain.removeListener('region:picked', onPick);
         if (!overlay.isDestroyed()) overlay.close();
         resolve(rect);
       };
-      const { ipcMain } = require('electron');
       const onPick = (event, rect) => {
         if (event.sender !== overlay.webContents) return;
-        ipcMain.removeListener('region:picked', onPick);
         if (!rect) return finish(null);
-        const imgSize = image.getSize();
-        const sx = imgSize.width / display.bounds.width;
-        const sy = imgSize.height / display.bounds.height;
-        finish({
-          x: Math.round(rect.x * sx),
-          y: Math.round(rect.y * sy),
-          width: Math.round(rect.w * sx),
-          height: Math.round(rect.h * sy),
-        });
+        finish(this.overlayRectToImageRect(rect, display, image.getSize()));
       };
       ipcMain.on('region:picked', onPick);
       overlay.on('closed', () => finish(null));
-      overlay.loadFile(path.join(__dirname, 'renderer', 'region.html'));
+      overlay.webContents.on('did-fail-load', () => finish(null));
+      overlay.loadFile(path.join(__dirname, 'renderer', 'region.html')).catch(() => finish(null));
     });
   }
 }

app/main.js

@@ -3,6 +3,7 @@
 const path = require('node:path');
 const fs = require('node:fs');
 const os = require('node:os');
+const { pathToFileURL } = require('node:url');
 const {
   app, BrowserWindow, ipcMain, dialog, shell, nativeTheme, globalShortcut,
   clipboard, nativeImage, screen, powerSaveBlocker, session, desktopCapturer,
@@ -204,7 +205,12 @@ function createWindow() {
           // Second scenario, reproducing the "I clicked many times but only
           // got two screenshots" report: a fast burst of clicks immediately
           // followed by finishing the session, so most clicks are still
-          // queued (frames still encoding) when the stop lands.
+          // queued (frames still encoding) when the stop lands. This scenario
+          // tests the queue DRAIN, not strict timing — 30ms-apart clicks
+          // outpace the frame sampler, so run it in balanced mode where every
+          // queued click stores. (Strict-mode skip-vs-store is covered by the
+          // marker scenario above and by unit tests.)
+          settings.set('capture.strictClickFrames', false);
           const burstGuide = store.createGuide({ title: 'burst selftest' });
           capture.startSession(burstGuide.guideId, { intervalSec: 0 });
           capture.stopClickWatcher();
@@ -228,6 +234,7 @@ function createWindow() {
           const burstSteps = store.getGuide(burstGuide.guideId).stepsOrder.length;
           console.log('CLICK-SELFTEST burst:', burstSteps, 'of', burstCount,
             burstSteps === burstCount ? 'OK — no clicks dropped on finish' : 'FAIL — clicks lost');
+          settings.set('capture.strictClickFrames', true); // restore for later scenarios
 
           // Helper: wait until armRecording has finished warming (window
           // hidden, buffer primed) so an injected click counts as a real
@@ -514,7 +521,12 @@ function setupIpc() {
   });
   h('step:imagePath', ({ guideId, stepId, which }) => {
     const p = store.stepImagePath(guideId, stepId, which || 'working');
-    return p && fs.existsSync(p) ? `file://${p}?v=${fs.statSync(p).mtimeMs}` : null;
+    if (!p || !fs.existsSync(p)) return null;
+    // pathToFileURL correctly encodes spaces, #, %, drive letters, etc.; the
+    // mtime is a cache-buster so the renderer reloads after an edit.
+    const url = pathToFileURL(p);
+    url.searchParams.set('v', String(fs.statSync(p).mtimeMs));
+    return url.href;
   }, {
     validate: (a) => c.id(a.guideId) && c.id(a.stepId)
       && (a.which === undefined || a.which === null || c.oneOf(a.which, ['original', 'working'])),
@@ -599,6 +611,12 @@ function setupIpc() {
     validate: (a) => c.string(a.text, 200000)
       && c.optionalString(a.guideTitle, 1000) && c.optionalString(a.stepTitle, 1000),
   });
+  // Cancel outstanding AI requests, e.g. when a guide/editor closes, so a
+  // slow response can't resolve against data the user has moved on from.
+  h('ai:cancel', ({ guideId = null } = {}) => {
+    textIntel.cancelInflight(guideId || null);
+    return true;
+  }, { validate: (a) => c.optionalId(a.guideId) });
   h('placeholders:globals:get', () => settings.getGlobalPlaceholders());
   h('placeholders:globals:set', (values) => settings.setGlobalPlaceholders(values));
 
@@ -645,43 +663,20 @@ function setupIpc() {
     }
     return result;
   }, { validate: (a) => c.id(a.guideId) });
-  let capturePowerBlocker = -1;
-  const startCapturePower = () => {
-    if (!powerSaveBlocker.isStarted(capturePowerBlocker)) {
-      capturePowerBlocker = powerSaveBlocker.start('prevent-app-suspension');
-    }
-  };
-  const stopCapturePower = () => {
-    if (powerSaveBlocker.isStarted(capturePowerBlocker)) {
-      powerSaveBlocker.stop(capturePowerBlocker);
-    }
-  };
-
-  // Opt every live Electron process (browser, GPU, the screen-capture utility,
-  // any renderers) out of EcoQoS for the duration of a recording. The hidden
-  // capture-worker renderer is created later, during warmup, so it opts itself
-  // out separately (see stream-backend.js); this covers the rest.
-  const keepCaptureProcessesResponsive = () => {
-    try {
-      keepProcessesResponsive(app.getAppMetrics().map((m) => m.pid));
-    } catch { /* metrics unavailable — best effort */ }
-  };
 
+  // Power/throttling state is owned entirely by the capture service's
+  // recording transitions (see createCapturePowerPolicy) so there is exactly
+  // one owner: it is held iff a session is actively recording, and paused,
+  // finished, tray, and second-instance transitions all release it correctly.
   h('capture:session', async ({ action, guideId, intervalSec }) => {
     if (action === 'start') {
       capture.startSession(guideId, { intervalSec: intervalSec ?? null });
-      startCapturePower();
-      keepCaptureProcessesResponsive();
     } else if (action === 'pause') {
       capture.togglePause(true);
-      stopCapturePower();
     } else if (action === 'resume') {
       capture.togglePause(false);
-      startCapturePower();
-      keepCaptureProcessesResponsive();
     } else if (action === 'finish') {
       capture.finishSession();
-      stopCapturePower();
     } else if (action === 'interval') {
       capture.setInterval(intervalSec);
     }
@@ -848,7 +843,7 @@ function setupIpc() {
     });
     const result = runExport(format, ast, previewDir, options || {});
     producedFiles.add(result.file);
-    return { ok: true, file: result.file, fileUrl: `file://${result.file}` };
+    return { ok: true, file: result.file, fileUrl: pathToFileURL(result.file).href };
   }, {
     validate: (a) => c.id(a.guideId) && validFormat(a.format)
       && (a.options === undefined || security.isPlainArgs(a.options)),
@@ -961,12 +956,36 @@ if (!gotLock) {
       }
     };
 
+    // Single owner of OS power/throttling state for recording. The capture
+    // service calls setRecording(true/false) on every recording transition;
+    // this holds a power-save blocker and opts live Electron processes out of
+    // EcoQoS while recording, and releases the blocker when recording stops.
+    const capturePowerPolicy = (() => {
+      let blocker = -1;
+      const keepResponsive = () => {
+        try { keepProcessesResponsive(app.getAppMetrics().map((m) => m.pid)); } catch { /* best effort */ }
+      };
+      return {
+        setRecording(recording) {
+          if (recording) {
+            if (!powerSaveBlocker.isStarted(blocker)) {
+              blocker = powerSaveBlocker.start('prevent-app-suspension');
+            }
+            keepResponsive();
+          } else if (powerSaveBlocker.isStarted(blocker)) {
+            powerSaveBlocker.stop(blocker);
+          }
+        },
+      };
+    })();
+
     capture = new CaptureService({
       store,
       settings,
       getWindow: () => mainWindow,
       notify: captureNotify,
       textIntel,
+      powerPolicy: capturePowerPolicy,
     });
 
     // Deny-by-default permission policy. The only grant in the entire app is
@@ -1019,6 +1038,19 @@ if (!gotLock) {
     });
   });
 
+  // Drain clicks still encoding in the capture queue before the app exits, so
+  // a fast burst immediately before quit is not lost. Defer the quit exactly
+  // once with a bounded deadline, then let it proceed.
+  let quitDrained = false;
+  app.on('before-quit', (event) => {
+    if (quitDrained || !capture) return;
+    quitDrained = true;
+    event.preventDefault();
+    // Stop new clicks from being queued, then wait for the queue to settle.
+    capture.stopClickWatcher();
+    capture.drainPendingClicks(2000).finally(() => app.quit());
+  });
+
   app.on('will-quit', () => {
     globalShortcut.unregisterAll();
     if (capture) {

app/preload.js

@@ -56,6 +56,7 @@ const api = {
     test: invoke('ai:test'),
     fillStep: invoke('ai:fillStep'),
     rewriteText: invoke('ai:rewriteText'),
+    cancel: invoke('ai:cancel'),
   },
   capture: {
     shoot: invoke('capture:shoot'),

app/renderer/app.js

@@ -172,7 +172,7 @@ class StepForgeApp {
       el('div.welcome', {},
         el('div.welcome-title', {},
           el('h1', {}, 'StepForge'),
-          el('p.muted', {}, 'Capture, annotate, and export step-by-step guides, fully offline.'),
+          el('p.muted', {}, 'Capture, annotate, and export step-by-step guides. Local-first, no telemetry.'),
         ),
         el('div.welcome-actions', {},
           el('button.welcome-btn.primary', {

app/renderer/editor.js

@@ -151,6 +151,11 @@ class GuideEditor {
 
   setActive(active) {
     this.active = Boolean(active);
+    // Leaving the editor cancels any in-flight AI request for this guide so a
+    // slow response can't resolve against a guide the user has closed.
+    if (!this.active && this.guideId) {
+      api.ai.cancel({ guideId: this.guideId }).catch(() => {});
+    }
   }
 
   setSettings(settings) {

app/text-intel.js

@@ -8,6 +8,7 @@ const {
   DEFAULT_CAPTURE_TITLES,
   buildCaptureTitle,
   normalizeOllamaHost,
+  validateOllamaHost,
   normalizeAiPatch,
   buildAiPrompt,
   applyAiPatchToStep,
@@ -74,9 +75,101 @@ class TextIntelService {
     this.workerQueue = Promise.resolve();
     this.ocrDataDir = path.join(dataDir, 'ocr', 'eng');
     this.modelCapabilityCache = new Map();
+    // In-flight AI request controllers, grouped by guide, so closing a guide
+    // (or quitting) can cancel outstanding requests instead of leaving them
+    // to resolve against stale data.
+    this.inflight = new Set();
+    // Bounded concurrency for AI network calls.
+    this.maxConcurrent = 2;
+    this.activeCount = 0;
+    this.waitQueue = [];
+  }
+
+  aiNetworkOptions() {
+    const ai = this.settings.get('ai') || {};
+    const timeoutMs = Number.isFinite(ai.timeoutMs) && ai.timeoutMs > 0 ? ai.timeoutMs : 60000;
+    const maxImageBytes = Number.isFinite(ai.maxImageBytes) && ai.maxImageBytes > 0
+      ? ai.maxImageBytes
+      : 12 * 1024 * 1024;
+    return {
+      allowRemote: Boolean(ai.allowRemoteHost),
+      attachScreenshots: ai.attachScreenshots !== false,
+      timeoutMs,
+      maxImageBytes,
+    };
+  }
+
+  // Resolve the endpoint against the local-first policy or throw a clear error.
+  resolveHost(host) {
+    const { allowRemote } = this.aiNetworkOptions();
+    const result = validateOllamaHost(host, { allowRemote });
+    if (!result.ok) {
+      const err = new Error(result.reason);
+      err.code = 'STEPFORGE_AI_HOST_BLOCKED';
+      throw err;
+    }
+    return result.host;
+  }
+
+  // fetch with a hard deadline and cooperative cancellation. Every AI network
+  // call goes through here so a dead endpoint can never hang the UI.
+  async fetchJson(url, { method = 'GET', body = null, guideId = null } = {}) {
+    const { timeoutMs } = this.aiNetworkOptions();
+    const controller = new AbortController();
+    if (guideId) controller._guideId = guideId;
+    this.inflight.add(controller);
+    const timer = setTimeout(() => controller.abort(new Error('timeout')), timeoutMs);
+    try {
+      const res = await this.fetch(url, {
+        method,
+        headers: body ? { 'Content-Type': 'application/json' } : undefined,
+        body: body ? JSON.stringify(body) : undefined,
+        signal: controller.signal,
+      });
+      return res;
+    } catch (err) {
+      if (controller.signal.aborted) {
+        // The abort reason distinguishes an explicit cancel from a timeout.
+        const reasonMsg = controller.signal.reason && controller.signal.reason.message;
+        const cancelled = reasonMsg === 'cancelled';
+        const wrapped = new Error(cancelled ? 'AI request cancelled.' : 'AI request timed out.');
+        wrapped.code = 'STEPFORGE_AI_ABORTED';
+        throw wrapped;
+      }
+      throw err;
+    } finally {
+      clearTimeout(timer);
+      this.inflight.delete(controller);
+    }
+  }
+
+  // Cancel in-flight AI requests, optionally scoped to one guide.
+  cancelInflight(guideId = null) {
+    for (const controller of [...this.inflight]) {
+      if (!guideId || controller._guideId === guideId) {
+        try { controller.abort(new Error('cancelled')); } catch { /* already settled */ }
+        this.inflight.delete(controller);
+      }
+    }
+  }
+
+  // Bounded-concurrency gate for AI network work.
+  async withConcurrency(fn) {
+    if (this.activeCount >= this.maxConcurrent) {
+      await new Promise((resolve) => this.waitQueue.push(resolve));
+    }
+    this.activeCount += 1;
+    try {
+      return await fn();
+    } finally {
+      this.activeCount -= 1;
+      const next = this.waitQueue.shift();
+      if (next) next();
+    }
   }
 
   async shutdown() {
+    this.cancelInflight();
     if (this.worker) {
       try {
         await this.worker.terminate();
@@ -374,8 +467,19 @@ public static class Win32 {
     if (!config.ollama.host) {
       return { ok: false, reason: 'Set an Ollama host first.' };
     }
-    const tagsUrl = new URL('/api/tags', `${config.ollama.host.replace(/\/+$/, '')}/`);
-    const res = await this.fetch(tagsUrl, { method: 'GET' });
+    let host;
+    try {
+      host = this.resolveHost(config.ollama.host);
+    } catch (err) {
+      return { ok: false, reason: err.message };
+    }
+    const tagsUrl = new URL('/api/tags', `${host.replace(/\/+$/, '')}/`);
+    let res;
+    try {
+      res = await this.fetchJson(tagsUrl, { method: 'GET' });
+    } catch (err) {
+      return { ok: false, reason: err.message };
+    }
     if (!res.ok) {
       return { ok: false, reason: `Ollama check failed (${res.status})` };
     }
@@ -383,7 +487,7 @@ public static class Win32 {
     const models = Array.isArray(data?.models) ? data.models.map((model) => model.name).filter(Boolean) : [];
     const installed = config.ollama.model ? models.includes(config.ollama.model) : false;
     const vision = installed ? await this.modelSupportsVision({
-      host: config.ollama.host,
+      host,
       model: config.ollama.model,
     }) : false;
     return {
@@ -391,7 +495,7 @@ public static class Win32 {
       installed,
       vision,
       models,
-      host: config.ollama.host,
+      host,
       model: config.ollama.model,
     };
   }
@@ -407,10 +511,9 @@ public static class Win32 {
     const url = new URL('/api/show', `${normalizedHost.replace(/\/+$/, '')}/`);
     let capabilities = [];
     try {
-      const response = await this.fetch(url, {
+      const response = await this.fetchJson(url, {
         method: 'POST',
-        headers: { 'Content-Type': 'application/json' },
-        body: JSON.stringify({ model: normalizedModel }),
+        body: { model: normalizedModel },
       });
       if (response.ok) {
         const payload = await response.json();
@@ -439,12 +542,12 @@ public static class Win32 {
     return fs.readFileSync(imagePath).toString('base64');
   }
 
-  async callOllamaText({ host, model, prompt, systemPrompt }) {
+  async callOllamaText({ host, model, prompt, systemPrompt, guideId = null }) {
     const url = new URL('/api/chat', `${host.replace(/\/+$/, '')}/`);
-    const response = await this.fetch(url, {
+    const response = await this.withConcurrency(() => this.fetchJson(url, {
       method: 'POST',
-      headers: { 'Content-Type': 'application/json' },
-      body: JSON.stringify({
+      guideId,
+      body: {
         model,
         stream: false,
         messages: [
@@ -452,8 +555,8 @@ public static class Win32 {
           { role: 'user', content: prompt },
         ],
         options: { temperature: 0.4 },
-      }),
-    });
+      },
+    }));
     if (!response.ok) throw new Error(`Ollama request failed (${response.status})`);
     const payload = await response.json();
     const content = payload?.message?.content;
@@ -461,16 +564,16 @@ public static class Win32 {
     return content.trim();
   }
 
-  async callOllama({ host, model, prompt, systemPrompt, images = [] }) {
+  async callOllama({ host, model, prompt, systemPrompt, images = [], guideId = null }) {
     const url = new URL('/api/chat', `${host.replace(/\/+$/, '')}/`);
     const userMessage = { role: 'user', content: prompt };
     if (Array.isArray(images) && images.length) {
       userMessage.images = images;
     }
-    const response = await this.fetch(url, {
+    const response = await this.withConcurrency(() => this.fetchJson(url, {
       method: 'POST',
-      headers: { 'Content-Type': 'application/json' },
-      body: JSON.stringify({
+      guideId,
+      body: {
         model,
         stream: false,
         format: 'json',
@@ -481,8 +584,8 @@ public static class Win32 {
         options: {
           temperature: 0.2,
         },
-      }),
-    });
+      },
+    }));
     if (!response.ok) {
       throw new Error(`Ollama request failed (${response.status})`);
     }
@@ -508,6 +611,13 @@ public static class Win32 {
       if (!config.ollama.host || !config.ollama.model) {
         return { ok: false, reason: 'Configure Ollama host and model in Settings.' };
       }
+      let host;
+      try {
+        host = this.resolveHost(config.ollama.host);
+      } catch (err) {
+        return { ok: false, reason: err.message };
+      }
+      const netOptions = this.aiNetworkOptions();
 
       const guide = this.store.getGuide(guideId);
       const step = this.store.getStep(guideId, stepId);
@@ -522,10 +632,20 @@ public static class Win32 {
         return { ok: false, reason: 'Block not found.' };
       }
 
-      const screenshotBase64 = step.image ? this.readStepImageBase64(guideId, stepId) : '';
+      // Only attach a screenshot when the user allows it, the model can use
+      // it, and it is within the size budget (a full 4K PNG base64-expands to
+      // tens of MB in the request body).
+      let screenshotBase64 = '';
+      if (netOptions.attachScreenshots && step.image) {
+        const candidate = this.readStepImageBase64(guideId, stepId);
+        const bytes = candidate ? Math.floor((candidate.length * 3) / 4) : 0;
+        if (candidate && bytes <= netOptions.maxImageBytes) {
+          screenshotBase64 = candidate;
+        }
+      }
       const screenshotAttached = Boolean(screenshotBase64)
         ? await this.modelSupportsVision({
-          host: config.ollama.host,
+          host,
           model: config.ollama.model,
         })
         : false;
@@ -588,11 +708,12 @@ public static class Win32 {
       });
 
       const raw = await this.callOllama({
-        host: config.ollama.host,
+        host,
         model: config.ollama.model,
         prompt,
         systemPrompt,
         images: screenshotAttached ? [screenshotBase64] : [],
+        guideId,
       });
       const patch = normalizeAiPatch(raw);
       const updated = applyAiPatchToStep(step, patch, { target, blockId });
@@ -610,6 +731,12 @@ public static class Win32 {
       if (!config.ollama.host || !config.ollama.model) {
         return { ok: false, reason: 'Configure Ollama host and model in Settings.' };
       }
+      let host;
+      try {
+        host = this.resolveHost(config.ollama.host);
+      } catch (err) {
+        return { ok: false, reason: err.message };
+      }
       const trimmed = normalizeWhitespace(text);
       if (!trimmed) return { ok: false, reason: 'No text to rewrite.' };
 
@@ -628,7 +755,7 @@ public static class Win32 {
       ].filter((l) => l !== null).join('\n');
 
       const result = await this.callOllamaText({
-        host: config.ollama.host,
+        host,
         model: config.ollama.model,
         prompt,
         systemPrompt: 'You are a documentation editor. Return only the improved text, nothing else.',

core/settings.js

@@ -45,6 +45,13 @@ const DEFAULT_SETTINGS = {
     // user is likely to click so the buffer holds frames of the now-visible
     // screen rather than the just-dismissed app window.
     postHideSettleMs: 150,
+    // Raw typed-text capture. When true, printable characters typed between
+    // captures are buffered and stored in step capture metadata (and can be
+    // sent to a configured AI host). This can record passwords or other
+    // secrets, so it is OFF by default and must be explicitly opted into.
+    // Shortcut detection (Ctrl+T, Enter, …) is unaffected — only raw
+    // character logging is gated here.
+    captureTypedText: false,
   },
   editor: {
     focusedViewDefaultForNewSteps: false,
@@ -52,6 +59,22 @@ const DEFAULT_SETTINGS = {
   },
   ai: {
     enabled: false,
+    // Auto-document captured steps in the background when a session capture
+    // lands. Requires enabled + a reachable model.
+    autoDoc: false,
+    // Local-first: only a loopback Ollama endpoint is contacted unless this
+    // is explicitly turned on. Turning it on sends screenshots and text to
+    // the configured remote host.
+    allowRemoteHost: false,
+    // Attach the step screenshot to AI requests (only for vision-capable
+    // models). Turning this off keeps requests text-only.
+    attachScreenshots: true,
+    // Per-request network deadline (ms). A dead endpoint fails fast instead
+    // of leaving UI actions pending forever.
+    timeoutMs: 60000,
+    // Skip attaching a screenshot larger than this many bytes (pre-base64)
+    // to avoid multi-hundred-MB request bodies.
+    maxImageBytes: 12 * 1024 * 1024,
     ollama: {
       host: 'http://127.0.0.1:11434',
       model: 'llama3.2:1b',

core/text-intel.js

@@ -538,6 +538,60 @@ function normalizeOllamaHost(host) {
   return `http://${raw.replace(/\/+$/, '')}`;
 }
 
+// A hostname/IP that refers to this machine only. StepForge is local-first:
+// by default the Ollama endpoint must be loopback so screenshots and text
+// never leave the device, unless the user explicitly opts into a remote host.
+function isLoopbackHost(host) {
+  const normalized = normalizeOllamaHost(host);
+  if (!normalized) return false;
+  let url;
+  try {
+    url = new URL(normalized);
+  } catch {
+    return false;
+  }
+  const name = url.hostname.toLowerCase().replace(/^\[|\]$/g, '');
+  if (name === 'localhost' || name === '::1' || name === '0.0.0.0' || name === '::') return true;
+  // IPv4 loopback block 127.0.0.0/8.
+  const m = /^(\d{1,3})\.(\d{1,3})\.(\d{1,3})\.(\d{1,3})$/.exec(name);
+  if (m && Number(m[1]) === 127 && m.slice(1).every((o) => Number(o) >= 0 && Number(o) <= 255)) {
+    return true;
+  }
+  // IPv4-mapped IPv6 loopback, e.g. ::ffff:127.0.0.1.
+  if (/^::ffff:127\.\d{1,3}\.\d{1,3}\.\d{1,3}$/.test(name)) return true;
+  return false;
+}
+
+/**
+ * Validate a configured Ollama endpoint against the local-first policy.
+ * Returns { ok, host, reason }. Remote hosts are rejected unless the caller
+ * passes allowRemote: true (the explicit ai.allowRemoteHost opt-in).
+ */
+function validateOllamaHost(host, { allowRemote = false } = {}) {
+  const normalized = normalizeOllamaHost(host);
+  if (!normalized) return { ok: false, host: '', reason: 'No Ollama host configured.' };
+  let url;
+  try {
+    url = new URL(normalized);
+  } catch {
+    return { ok: false, host: normalized, reason: 'Ollama host is not a valid URL.' };
+  }
+  if (url.protocol !== 'http:' && url.protocol !== 'https:') {
+    return { ok: false, host: normalized, reason: 'Ollama host must use http or https.' };
+  }
+  if (!allowRemote && !isLoopbackHost(normalized)) {
+    return {
+      ok: false,
+      host: normalized,
+      reason:
+        'Remote Ollama hosts are disabled. StepForge only contacts a local (loopback) ' +
+        'Ollama by default. Enable "Allow remote AI host" in AI settings to send ' +
+        'screenshots and text to this host.',
+    };
+  }
+  return { ok: true, host: normalized, reason: '' };
+}
+
 function normalizeAiLevel(level) {
   const key = normalizeWhitespace(level).toLowerCase();
   return AI_LEVEL_ALIASES.get(key) || (TEXTBLOCK_LEVELS.includes(key) ? key : 'info');
@@ -845,6 +899,8 @@ module.exports = {
   buildCaptureTitle,
   plainTextToHtml,
   normalizeOllamaHost,
+  isLoopbackHost,
+  validateOllamaHost,
   normalizeAiPatch,
   buildAiPrompt,
   applyAiPatchToStep,

docs/PRIVACY.md

@@ -0,0 +1,65 @@
+# StepForge privacy and network contract
+
+StepForge is **local-first**. Guides, screenshots, and settings live on your
+machine and are never uploaded on their own. This document describes exactly
+what data StepForge collects locally and the one situation in which data
+leaves your device.
+
+## What never happens
+
+- No telemetry or analytics.
+- No update checks, license checks, or "phone home".
+- No cloud storage or sync.
+- No dependency downloads at runtime (dependencies are installed only by you,
+  via `npm ci`).
+
+## Data StepForge collects locally
+
+When you capture a step, StepForge may record, **stored only on disk in your
+data directory**, capture context to help title and describe the step:
+
+- The screenshot image.
+- OCR text read from the region around your click (via the bundled Tesseract
+  engine — this runs locally, it is not a network call).
+- The foreground window title and application name.
+- The accessibility label/role/value of the clicked UI element (Windows).
+- Keyboard shortcuts you pressed (for example `Ctrl+T`).
+
+### Raw typed text is OFF by default
+
+StepForge can additionally record the **raw printable characters** you type
+between captures. Because this can capture passwords or other secrets, it is
+**disabled by default**. It is only recorded when you explicitly enable
+`capture.captureTypedText`, and even then the characters are used only to
+title the current step and are not retained beyond it. With the setting off,
+raw characters are never read or stored (on Windows they never even leave the
+keyboard-hook process).
+
+## The one outbound feature: optional AI
+
+StepForge has an **optional** AI integration that generates step titles and
+descriptions with a local large-language-model runtime
+([Ollama](https://ollama.com)). It is **off by default**. When you turn it on
+and configure an endpoint:
+
+- StepForge sends the step **screenshot** (only to vision-capable models, only
+  when "Attach screenshots" is on, and only if within the size limit) and the
+  step **text/capture context** to the configured Ollama endpoint.
+- By default the endpoint must be a **local (loopback) address** — for example
+  `http://127.0.0.1:11434`. StepForge refuses to send data to a non-loopback
+  host unless you explicitly enable **"Allow remote AI host"**. Enabling that
+  option means your screenshots and text are sent to the remote host you
+  configured; StepForge cannot control what that host does with them.
+- Every AI request has a timeout, can be cancelled (closing the guide cancels
+  in-flight requests), and runs under a bounded concurrency limit.
+
+## Bundled dependencies
+
+Beyond the Electron desktop shell, StepForge bundles the Tesseract OCR engine
+and its English language data as production dependencies. All OCR runs locally.
+
+## Where your data lives
+
+- Windows: `%APPDATA%\stepforge`
+- Linux: `~/.local/share/stepforge` (or `$XDG_DATA_HOME/stepforge`)
+- Override with the `STEPFORGE_DATA_DIR` environment variable.

package.json

@@ -2,7 +2,7 @@
   "name": "stepforge",
   "version": "0.3.2",
   "buildVersion": "0.3.2.1",
-  "description": "Fully offline desktop tool for capturing, annotating, and exporting step-by-step guides.",
+  "description": "Local-first desktop tool for capturing, annotating, and exporting step-by-step guides, with an optional user-configured local AI integration.",
   "main": "app/main.js",
   "author": "StepForge [email protected]",
   "license": "MPL-2.0",

tests/unit/ai-network.test.js

@@ -0,0 +1,195 @@
+'use strict';
+
+const test = require('node:test');
+const assert = require('node:assert/strict');
+
+const { makeTmpDir, rmrf } = require('./helpers');
+const { isLoopbackHost, validateOllamaHost } = require('../../core/text-intel');
+const { TextIntelService } = require('../../app/text-intel');
+const CaptureService = require('../../app/capture');
+
+function makeSettings(ai = {}) {
+  const data = {
+    ai: {
+      enabled: true,
+      allowRemoteHost: false,
+      attachScreenshots: true,
+      timeoutMs: 60000,
+      maxImageBytes: 12 * 1024 * 1024,
+      ollama: { host: 'http://127.0.0.1:11434', model: 'llama3.2:1b' },
+      ...ai,
+      ollama: { host: 'http://127.0.0.1:11434', model: 'llama3.2:1b', ...(ai.ollama || {}) },
+    },
+  };
+  return {
+    get(key) {
+      return key.split('.').reduce((acc, part) => (acc == null ? undefined : acc[part]), data);
+    },
+  };
+}
+
+// ---- loopback policy (pure core) -------------------------------------------
+
+test('isLoopbackHost recognizes local addresses and rejects remote ones', () => {
+  for (const host of [
+    'http://127.0.0.1:11434',
+    '127.0.0.1',
+    'localhost:11434',
+    'http://[::1]:11434',
+    'http://127.5.6.7',
+    '0.0.0.0',
+  ]) {
+    assert.equal(isLoopbackHost(host), true, `loopback: ${host}`);
+  }
+  for (const host of [
+    'http://10.0.0.5:11434',
+    'http://192.168.1.20',
+    'https://ollama.example.com',
+    'http://8.8.8.8',
+    'http://ollama.internal',
+  ]) {
+    assert.equal(isLoopbackHost(host), false, `remote: ${host}`);
+  }
+});
+
+test('validateOllamaHost blocks remote hosts unless explicitly allowed', () => {
+  assert.equal(validateOllamaHost('http://127.0.0.1:11434').ok, true);
+  const blocked = validateOllamaHost('http://10.0.0.5:11434');
+  assert.equal(blocked.ok, false);
+  assert.match(blocked.reason, /remote/i);
+  assert.equal(validateOllamaHost('http://10.0.0.5:11434', { allowRemote: true }).ok, true);
+  assert.equal(validateOllamaHost('').ok, false);
+  assert.equal(validateOllamaHost('ftp://127.0.0.1').ok, false);
+});
+
+// ---- host policy enforced by the service -----------------------------------
+
+test('AI connection test refuses a remote host without the opt-in', async (t) => {
+  const root = makeTmpDir('ai-remote-block');
+  t.after(() => rmrf(root));
+  let called = false;
+  const service = new TextIntelService({
+    store: { settingsDir: root },
+    settings: makeSettings({ ollama: { host: 'http://10.0.0.5:11434', model: 'llama3.2:1b' } }),
+    dataDir: root,
+    fetchImpl: async () => { called = true; return { ok: true, json: async () => ({}) }; },
+  });
+  const result = await service.testAiConnection();
+  assert.equal(result.ok, false);
+  assert.match(result.reason, /remote/i);
+  assert.equal(called, false, 'must not contact a blocked host at all');
+});
+
+test('AI connection test allows a remote host with the opt-in', async (t) => {
+  const root = makeTmpDir('ai-remote-allow');
+  t.after(() => rmrf(root));
+  const service = new TextIntelService({
+    store: { settingsDir: root },
+    settings: makeSettings({
+      allowRemoteHost: true,
+      ollama: { host: 'http://10.0.0.5:11434', model: 'llama3.2:1b' },
+    }),
+    dataDir: root,
+    fetchImpl: async (url) => {
+      const pathname = new URL(url).pathname;
+      if (pathname === '/api/tags') return { ok: true, json: async () => ({ models: [{ name: 'llama3.2:1b' }] }) };
+      if (pathname === '/api/show') return { ok: true, json: async () => ({ capabilities: ['vision'] }) };
+      throw new Error(`unexpected ${pathname}`);
+    },
+  });
+  const result = await service.testAiConnection();
+  assert.equal(result.ok, true);
+  assert.equal(result.host, 'http://10.0.0.5:11434');
+});
+
+// ---- request timeout / cancellation ----------------------------------------
+
+test('a hung endpoint times out instead of hanging forever', async (t) => {
+  const root = makeTmpDir('ai-timeout');
+  t.after(() => rmrf(root));
+  const service = new TextIntelService({
+    store: { settingsDir: root },
+    settings: makeSettings({ timeoutMs: 40 }),
+    dataDir: root,
+    fetchImpl: (url, init) => new Promise((resolve, reject) => {
+      // Never resolves on its own; only the abort signal ends it.
+      init.signal.addEventListener('abort', () => {
+        const err = new Error('aborted');
+        err.name = 'AbortError';
+        reject(err);
+      });
+    }),
+  });
+  const result = await service.testAiConnection();
+  assert.equal(result.ok, false);
+  assert.match(result.reason, /timed out|timeout/i);
+});
+
+test('cancelInflight aborts an outstanding request', async (t) => {
+  const root = makeTmpDir('ai-cancel');
+  t.after(() => rmrf(root));
+  const service = new TextIntelService({
+    store: { settingsDir: root },
+    settings: makeSettings({ timeoutMs: 60000 }),
+    dataDir: root,
+    fetchImpl: (url, init) => new Promise((resolve, reject) => {
+      init.signal.addEventListener('abort', () => {
+        const err = new Error('aborted');
+        err.name = 'AbortError';
+        reject(err);
+      });
+    }),
+  });
+  const pending = service.testAiConnection();
+  // Give the request a tick to register in the inflight set.
+  await new Promise((r) => setTimeout(r, 10));
+  assert.equal(service.inflight.size, 1);
+  service.cancelInflight();
+  const result = await pending;
+  assert.equal(result.ok, false);
+  assert.match(result.reason, /cancel/i);
+});
+
+// ---- typed-text capture is off by default ----------------------------------
+
+function makeCaptureService(settingsOverrides = {}) {
+  const settingsData = { 'capture.mode': 'fullscreen', ...settingsOverrides };
+  return new CaptureService({
+    store: {},
+    settings: { get: (k) => (k in settingsData ? settingsData[k] : null) },
+    getWindow: () => null,
+    notify: () => {},
+    screenApi: { getCursorScreenPoint: () => ({ x: 0, y: 0 }), getAllDisplays: () => [] },
+  });
+}
+
+test('raw typed text is ignored unless explicitly enabled', () => {
+  const off = makeCaptureService();
+  off.onKeyboardEvent('CHAR', 'h'.charCodeAt(0), Date.now());
+  off.onKeyboardEvent('CHAR', 'i'.charCodeAt(0), Date.now());
+  assert.equal(off.snapshotKeyContext().recentTyped, '', 'typed text must not be buffered by default');
+
+  const on = makeCaptureService({ 'capture.captureTypedText': true });
+  on.onKeyboardEvent('CHAR', 'h'.charCodeAt(0), Date.now());
+  on.onKeyboardEvent('CHAR', 'i'.charCodeAt(0), Date.now());
+  assert.equal(on.snapshotKeyContext().recentTyped, 'hi');
+});
+
+test('shortcut detection still works with typed text disabled', () => {
+  const off = makeCaptureService();
+  off.onKeyboardEvent('KEY', 'Ctrl+T', Date.now());
+  assert.equal(off.snapshotKeyContext().recentShortcut, 'Ctrl+T');
+});
+
+// ---- source-level guards ----------------------------------------------------
+
+const fs = require('node:fs');
+const path = require('node:path');
+const ROOT = path.resolve(__dirname, '..', '..');
+
+test('the Windows keyboard hook only emits characters when opted in', () => {
+  const src = fs.readFileSync(path.join(ROOT, 'app/capture.js'), 'utf8');
+  // The CHAR emission must be guarded by the opt-in flag threaded into C#.
+  assert.match(src, /CaptureTypedText = \$\{captureTypedText\}/);
+  assert.match(src, /else if \(CaptureTypedText\) \{/);
+});

tests/unit/capture-fixes.test.js

@@ -0,0 +1,120 @@
+'use strict';
+
+const test = require('node:test');
+const assert = require('node:assert/strict');
+
+const CaptureService = require('../../app/capture');
+
+function makeService({ settings: settingsOverrides, powerPolicy } = {}) {
+  const settingsData = {
+    'capture.mode': 'fullscreen',
+    'capture.delayMs': 0,
+    ...settingsOverrides,
+  };
+  return new CaptureService({
+    store: {},
+    settings: { get: (k) => (k in settingsData ? settingsData[k] : null) },
+    getWindow: () => null,
+    notify: () => {},
+    powerPolicy,
+    screenApi: {
+      getCursorScreenPoint: () => ({ x: 0, y: 0 }),
+      getAllDisplays: () => [],
+    },
+  });
+}
+
+// ---- region rect clamping (bug: out-of-bounds / negative drags) -------------
+
+test('overlayRectToImageRect scales, clamps, and normalizes selections', () => {
+  const svc = makeService();
+  const display = { bounds: { x: 0, y: 0, width: 100, height: 100 } };
+  const imgSize = { width: 200, height: 200 }; // 2x DPI
+
+  // Simple selection: display px -> image px (2x).
+  assert.deepEqual(
+    svc.overlayRectToImageRect({ x: 10, y: 20, w: 30, h: 40 }, display, imgSize),
+    { x: 20, y: 40, width: 60, height: 80 }
+  );
+
+  // Negative-size drag (drawn up/left) normalizes to a positive rect.
+  assert.deepEqual(
+    svc.overlayRectToImageRect({ x: 40, y: 40, w: -20, h: -20 }, display, imgSize),
+    { x: 40, y: 40, width: 40, height: 40 }
+  );
+
+  // Selection larger than the screen is clamped to the image bounds.
+  const clamped = svc.overlayRectToImageRect({ x: -10, y: -10, w: 200, h: 200 }, display, imgSize);
+  assert.deepEqual(clamped, { x: 0, y: 0, width: 200, height: 200 });
+
+  // Degenerate selections return null instead of an out-of-bounds crop.
+  assert.equal(svc.overlayRectToImageRect({ x: 0, y: 0, w: 0, h: 0 }, display, imgSize), null);
+  assert.equal(svc.overlayRectToImageRect({ x: 999, y: 999, w: 10, h: 10 }, display, imgSize), null);
+  assert.equal(svc.overlayRectToImageRect(null, display, imgSize), null);
+});
+
+// ---- power ownership follows recording state --------------------------------
+
+test('the power blocker is held only while actively recording', () => {
+  const calls = [];
+  const powerPolicy = { setRecording: (on) => calls.push(on) };
+  const svc = makeService({ powerPolicy });
+
+  // startSession begins PAUSED — must not hold power.
+  svc.startSession('g1', { intervalSec: 0 });
+  assert.deepEqual(calls, [], 'a paused new session must not start the power blocker');
+
+  // Resume records -> power on. (togglePause(false) arms recording.)
+  svc.togglePause(false);
+  assert.deepEqual(calls, [true]);
+
+  // Pause -> power off. This is the tray/second-instance path that used to leak.
+  svc.togglePause(true);
+  assert.deepEqual(calls, [true, false]);
+
+  // Resume again -> on, finish -> off.
+  svc.togglePause(false);
+  svc.finishSession();
+  assert.deepEqual(calls, [true, false, true, false]);
+});
+
+test('finishSession releases power even if it was recording', () => {
+  const calls = [];
+  const svc = makeService({ powerPolicy: { setRecording: (on) => calls.push(on) } });
+  svc.startSession('g1', { intervalSec: 0 });
+  svc.togglePause(false);
+  calls.length = 0;
+  svc.finishSession();
+  assert.deepEqual(calls, [false]);
+});
+
+// ---- explicit click-source reporting ----------------------------------------
+
+test('click source is unavailable outside a session and after stop', () => {
+  const svc = makeService();
+  assert.equal(svc.state().clickSource, 'unavailable');
+  assert.equal(svc.state().clickCapture, undefined); // no session -> no field
+  svc.clickSource = 'evdev-x11';
+  svc.stopClickWatcher();
+  assert.equal(svc.clickSource, 'unavailable');
+});
+
+test('state reports clickCapture true for a non-process (evdev) source', () => {
+  const svc = makeService();
+  svc.session = { guideId: 'g', paused: false, count: 0, intervalSec: 0 };
+  // evdev has no child process; the old Boolean(clickWatcher) reported false.
+  svc.clickSource = 'evdev-wayland';
+  const st = svc.state();
+  assert.equal(st.clickCapture, true);
+  assert.equal(st.clickSource, 'evdev-wayland');
+});
+
+// ---- drain never hangs quit -------------------------------------------------
+
+test('drainPendingClicks resolves within the deadline even if the queue hangs', async () => {
+  const svc = makeService();
+  svc.clickQueue = new Promise(() => {}); // never settles
+  const start = Date.now();
+  await svc.drainPendingClicks(60);
+  assert.ok(Date.now() - start < 1000, 'drain must not block on a hung queue');
+});

tests/unit/capture.test.js

@@ -68,7 +68,9 @@ function makeFrame(name, ageMs = 0, overrides = {}) {
 // ---- fresh-shot fallback path ----------------------------------------------
 
 test('click-triggered session capture uses the low-latency hide pause', async () => {
-  const service = makeService();
+  // The fresh-shot fallback only runs in non-strict (balanced) mode; strict
+  // mode skips rather than storing a post-click shot.
+  const service = makeService({ settings: { 'capture.strictClickFrames': false } });
   service.session = { guideId: 'guide-1', paused: false, count: 0, intervalSec: 0 };
 
   let seenOptions = null;
@@ -1006,7 +1008,9 @@ test('a buffered frame from a different display is ignored for click capture', a
 });
 
 test('a stale buffered frame is not reused — the click falls back to a fresh shot', async () => {
-  const service = makeService();
+  // Balanced (non-strict) mode: a stale frame is rejected and the click takes
+  // the fresh-shot fallback. (Strict mode skips instead — see below.)
+  const service = makeService({ settings: { 'capture.strictClickFrames': false } });
   service.session = { guideId: 'guide-stale', paused: false, count: 0, intervalSec: 0 };
   service.latestFrame = makeFrame('stale-png', 10_000);
 
@@ -1022,12 +1026,12 @@ test('a stale buffered frame is not reused — the click falls back to a fresh s
   assert.equal(shootCalled, true, 'a stale buffered frame must not be reused');
 });
 
-test('strict mode: a frame whose grab started after the click is rejected', async () => {
-  // This replaces the old "idle click waits for the imminent loop frame"
-  // behavior: a grab that begins after the click can already show the
-  // click's effects, so strict mode takes the explicit fresh-shot fallback
-  // instead of passing it off as the click-time screen.
-  const service = makeService();
+test('strict mode: no pre-click frame is skipped, never stored as a post-click shot', async () => {
+  // A grab that begins after the click can already show the click's effects.
+  // Strict mode's promise is that a stored step never shows the post-click
+  // screen, so when no pre-click frame qualifies it SKIPS with a diagnostic
+  // rather than taking a fresh (post-click) shot.
+  const service = makeService(); // strict is the default
   service.session = { guideId: 'guide-strict', paused: false, count: 0, intervalSec: 0 };
   service.frameLoopRunning = true;
   service.frameLoopInFlight = false; // nothing in flight at click time
@@ -1041,11 +1045,18 @@ test('strict mode: a frame whose grab started after the click is rejected', asyn
     shootCalled = true;
     return { ok: true, step: { stepId: 'fresh-step' } };
   };
+  const diagnostics = [];
+  service.notify = (channel, payload) => {
+    if (channel === 'capture:diagnostic') diagnostics.push(payload);
+  };
 
   const result = await service.sessionCapture('click', { x: 1, y: 1 }, { at: clickAt });
 
-  assert.equal(result.ok, true);
-  assert.equal(shootCalled, true);
+  assert.equal(result.ok, false);
+  assert.match(result.reason, /strict/i);
+  assert.equal(shootCalled, false, 'strict mode must not store a post-click shot');
+  assert.equal(diagnostics.length, 1);
+  assert.equal(diagnostics[0].kind, 'strict-click-skipped');
 });
 
 test('balanced mode keeps the legacy slack: an imminent post-click frame is accepted', async () => {
@@ -1183,7 +1194,9 @@ test('click frames come from the stream backend when it is active', async () =>
 });
 
 test('a stream backend with no qualifying frame falls through to the fresh-shot path', async () => {
-  const service = makeService();
+  // Balanced (non-strict) mode: the fresh-shot fallback runs. Strict mode
+  // would skip rather than store a post-click shot.
+  const service = makeService({ settings: { 'capture.strictClickFrames': false } });
   service.session = { guideId: 'guide-stream-miss', paused: false, count: 0, intervalSec: 0 };
   service.streamBackend = {
     isActive: () => true,
@@ -1279,6 +1292,8 @@ test('click capture marks the click-time cursor position', async () => {
     if (key === 'capture.clickMarker') return true;
     if (key === 'capture.clickMarkerColor') return '#E5484D';
     if (key === 'editor.focusedViewDefaultForNewSteps') return false;
+    // Exercise the fresh-shot fallback: strict mode would skip instead.
+    if (key === 'capture.strictClickFrames') return false;
     return null;
   };
   service.session = { guideId: 'guide-4', paused: false, count: 0, intervalSec: 0 };