feat: add value-free teacher mode

README.md

@@ -145,6 +145,7 @@ Type `/help` in the panel for full signatures and flags. The most useful ones:
 | `/schedule [prompt]`                                                                   | Create a scheduled task                                                     |
 | `/watch [--keep] [--secs <30-120>] [--long \| --short] <condition and action> [/beep]` | Poll the current page and act when a condition is met                       |
 | `/workflow` · `/workflow --save <name>`                                                | Manage saved workflows or compile the last successful run                   |
+| `/teach --start <name>` · `/teach --end`                                               | Learn a reusable workflow from your demonstrated actions                    |
 | `/memory --add <text>`                                                                 | Save a user preference                                                      |
 | `/screenshot [--full-page]`                                                            | Capture the tab, or the full scrollable page                                |
 | `/record [--transcribe]`                                                               | Record the current tab, optionally saving a transcript                      |

docs/architecture.md

@@ -531,6 +531,14 @@ action CSS selectors, coordinates, query strings, fragments, and typed values. E
 value becomes a declared runtime parameter; unsupported or failed actions are
 skipped and reported to the user as save warnings.
 
+Teacher mode is a second, value-free compiler input for the same schema.
+`content/teacher-capture.js` observes only trusted top-document clicks, field
+completion, checkbox/radio changes, Enter submissions, and navigation while a
+tab-scoped session is active. Field values are never read or sent; a field
+action becomes a runtime parameter marker as soon as it crosses the capture
+boundary. The sanitized session survives navigation in `storage.session` and
+is removed when `/teach --end` compiles and stores the workflow.
+
 Each compiled step contains semantic target metadata (role, accessible name,
 label, field identity, link, or placeholder), an expected postcondition, and
 the origin/path family observed before that action. `/workflow` renders a

docs/fr/slash-commands.md

@@ -30,6 +30,9 @@ d'une espace ouvre l'autocomplétion de ses options disponibles.
 | `/workflow --delete <id>` | Supprimer un workflow enregistré |
 | `/workflow --export <id>` | Télécharger un fichier JSON portable `webbrain-workflow/1` assaini |
 | `/workflow --import --file` | Importer un fichier de workflow portable comme nouveau workflow local |
+| `/teach` | Afficher l’état du mode d’apprentissage et le nombre d’actions capturées dans l’onglet actuel |
+| `/teach --start <nom>` | Commencer à apprendre un workflow à partir de vos clics et modifications de champs |
+| `/teach --end` | Arrêter l’apprentissage et compiler les actions en workflow enregistré sans valeurs saisies |
 | `/allow-api` | **Dérogation de mutation API par conversation.** Voir [plus bas](#allow-api). |
 | `/foreground [invite]` | Exécuter une tâche locale au premier plan pour assurer la compatibilité visuelle |
 | `/dangerously-skip-permissions` | **Contournement global des demandes d'autorisation.** Désactive `Ask before consequential actions` sans ouvrir les Paramètres. WebBrain agira sans demandes par site jusqu'à ce que vous réactiviez le réglage. |

docs/privacy-and-data-flow.md

@@ -135,6 +135,15 @@ postconditions, and parameter descriptors. It does not contain typed field
 values, raw historical `ref_id` values, action CSS selectors, coordinates, URL query strings, or URL
 fragments.
 
+`/teach --start <name>` records a user demonstration in temporary,
+tab-scoped session storage. The page capture code never reads field values:
+it sends only the field's semantic identity, and the compiler immediately
+represents that action as a runtime parameter. Click/field target labels and
+sanitized URL origin/path families are retained because replay needs them to
+find the same controls. `/teach --end` removes the temporary session whether
+compilation succeeds or fails; a successful compilation writes the same
+`webbrain-workflow/1` format described above.
+
 `/workflow --run <id>` collects declared values in a temporary side-panel form
 and sends them directly to the background replay executor. The values are not
 written to the workflow, chat text, retry payload, user memory, replay trace,

docs/slash-commands.md

@@ -30,6 +30,9 @@ for its available flags.
 | `/workflow --delete <id>` | Delete a saved workflow |
 | `/workflow --export <id>` | Download a sanitized portable `webbrain-workflow/1` JSON file |
 | `/workflow --import --file` | Import a portable workflow file as a new local saved workflow |
+| `/teach` | Show the current tab's Teacher mode status and captured action count |
+| `/teach --start <name>` | Start learning a saved workflow from your demonstrated clicks and field edits |
+| `/teach --end` | Stop teaching and compile the captured actions into a value-free saved workflow |
 | `/allow-api` | **Per-conversation API mutation override.** See [below](#allow-api). |
 | `/foreground [prompt]` | Run one local task in the foreground for visual compatibility |
 | `/dangerously-skip-permissions` | **Global permission-prompt bypass.** Turns off `Ask before consequential actions` without opening Settings. WebBrain will act without per-site prompts until you re-enable the setting. |

docs/zh-CN/slash-commands.md

@@ -28,6 +28,9 @@ WebBrain 接受作为输入框某行开头的斜杠命令。在面板内输入 `
 | `/workflow --delete <id>` | 删除已保存的工作流 |
 | `/workflow --export <id>` | 下载经过清洗的可移植 `webbrain-workflow/1` JSON 文件 |
 | `/workflow --import --file` | 将可移植工作流文件导入为新的本地工作流 |
+| `/teach` | 显示当前标签页的示教状态和已捕获操作数 |
+| `/teach --start <名称>` | 开始从您演示的点击和字段编辑中学习工作流 |
+| `/teach --end` | 结束示教，并将捕获的操作编译为不含实际输入值的已保存工作流 |
 | `/allow-api` | **按对话的 API 变更覆盖。** 参见[下文](#allow-api)。 |
 | `/foreground [提示词]` | 为视觉兼容性在前台运行一次本地任务 |
 | `/dangerously-skip-permissions` | **全局权限提示绕过。** 无需打开设置即可关闭「操作前询问」。在你重新启用该设置之前，WebBrain 将不再按站点弹出提示。 |

src/chrome/manifest.json

@@ -48,6 +48,7 @@
       "js": [
         "src/content/rich-text-toolbar-heuristic.js",
         "src/content/accessibility-tree.js",
+        "src/content/teacher-capture.js",
         "src/content/content.js",
         "src/content/agent-visual-indicator.js",
         "src/content/ollama-launch-handoff.js"

src/chrome/src/agent/agent.js

@@ -558,6 +558,7 @@ export class Agent extends LoopDetector {
     this._recentSubmitClicks = new Map(); // tabId -> recent submit click timestamps
     this._formValidationBlocks = new Map(); // tabId -> validation state that must change before another submit
     this._runningTabs = new Set(); // tabIds with an active processMessage/Stream in flight
+    this._runStartGuard = null;
     // Ordinary runs capture and act in their original tab without activating
     // it. Managed cloud runs and the explicit /foreground compatibility
     // override retain the old Page.bringToFront behavior for their lifetime.
@@ -896,6 +897,38 @@ export class Agent extends LoopDetector {
     this._conversationScopeChangeListener = typeof listener === 'function' ? listener : null;
   }
 
+  setRunStartGuard(guard) {
+    this._runStartGuard = typeof guard === 'function' ? guard : null;
+  }
+
+  _runEntryKind(defaultKind, runOptions = {}) {
+    if (runOptions?.cloudRun === true) return 'cloud';
+    if (runOptions?.scheduledRun === true) return 'scheduled';
+    return defaultKind;
+  }
+
+  async assertRunStartAllowed(tabId, defaultKind = 'interactive', runOptions = {}) {
+    await this._runStartGuard?.(tabId, {
+      kind: this._runEntryKind(defaultKind, runOptions),
+      runOptions,
+    });
+  }
+
+  async _claimRunEntry(tabId, defaultKind, runOptions = {}) {
+    if (this._runningTabs.has(tabId)) {
+      throw new Error('An agent run is already in progress for this tab.');
+    }
+    // Claim synchronously before awaiting the external guard so teacher-mode
+    // startup cannot race a run whose persisted teacher-state check is pending.
+    this._runningTabs.add(tabId);
+    try {
+      await this.assertRunStartAllowed(tabId, defaultKind, runOptions);
+    } catch (error) {
+      this._runningTabs.delete(tabId);
+      throw error;
+    }
+  }
+
   isRunning(tabId) {
     return this._runningTabs.has(tabId);
   }
@@ -15577,11 +15610,16 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
   }
 
   async replaySavedWorkflow(tabId, workflow, parameters = {}, onUpdate = () => {}, runOptions = {}) {
-    if (this._runningTabs.has(tabId)) throw new Error('An agent run is already in progress for this tab.');
     if (!workflow?.id || !Array.isArray(workflow.steps) || !workflow.steps.length) {
       throw new Error('Saved workflow is missing or invalid.');
     }
-    await this._hydrate(tabId);
+    await this._claimRunEntry(tabId, 'workflow', runOptions);
+    try {
+      await this._hydrate(tabId);
+    } catch (error) {
+      this._runningTabs.delete(tabId);
+      throw error;
+    }
     // Deterministic workflow replay is an independent task even when it never
     // falls back to processMessage. Clear stale selected-text grounding before
     // any replay action so a successful replay cannot poison the next turn.
@@ -15598,7 +15636,6 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
     this.permissions.beginTurn(tabId);
     this.conversationModes.set(tabId, 'act');
     const completionRunToken = this._beginCompletionInvariant(tabId);
-    this._runningTabs.add(tabId);
     const previousForegroundCapture = this._configureCapturePolicyForRun(tabId, runOptions);
     const startUrl = await this._currentUrl(tabId);
     const conversationId = await this.ensureConversationId(tabId, 'act');
@@ -21948,10 +21985,7 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
   }
 
   async processMessage(tabId, userMessage, onUpdate = () => {}, mode = 'ask', attachments = [], runOptions = {}) {
-    if (this._runningTabs.has(tabId)) {
-      throw new Error('An agent run is already in progress for this tab.');
-    }
-    this._runningTabs.add(tabId);
+    await this._claimRunEntry(tabId, 'interactive', runOptions);
     try {
       // Hydration has to run before the toolbar-ledger reset below, so a
       // persisted obligation cannot outlive the run that cleared it. It is
@@ -21972,7 +22006,6 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
     if (runOptions?.trustedContinuation !== true) this._continuationExecutionEvidence.delete(tabId);
     this._clickAxCdpFallbacks.delete(tabId);
     const completionRunToken = this._beginCompletionInvariant(tabId);
-    this._runningTabs.add(tabId);
     if (runOptions?.trustedContinuation !== true) this.persistenceDegradedTabs.delete(tabId);
     this._runUpdateCallbacks.set(tabId, onUpdate);
     const previousForegroundCapture = this._configureCapturePolicyForRun(tabId, runOptions);
@@ -22874,10 +22907,7 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
    * Process a message with streaming output.
    */
   async processMessageStream(tabId, userMessage, onUpdate = () => {}, mode = 'ask', runOptions = {}) {
-    if (this._runningTabs.has(tabId)) {
-      throw new Error('An agent run is already in progress for this tab.');
-    }
-    this._runningTabs.add(tabId);
+    await this._claimRunEntry(tabId, 'interactive', runOptions);
     try {
       // Hydration has to run before the toolbar-ledger reset below, so a
       // persisted obligation cannot outlive the run that cleared it. It is
@@ -22898,7 +22928,6 @@ Rules: no prose intro, no conclusion, no "this screenshot shows...", no layout d
     if (runOptions?.trustedContinuation !== true) this._continuationExecutionEvidence.delete(tabId);
     this._clickAxCdpFallbacks.delete(tabId);
     const completionRunToken = this._beginCompletionInvariant(tabId);
-    this._runningTabs.add(tabId);
     if (runOptions?.trustedContinuation !== true) this.persistenceDegradedTabs.delete(tabId);
     this._runUpdateCallbacks.set(tabId, onUpdate);
     const previousForegroundCapture = this._configureCapturePolicyForRun(tabId, runOptions);

src/chrome/src/agent/scheduler.js

@@ -585,6 +585,10 @@ export class ScheduledJobManager {
     this.restoreAlarms().catch((e) => console.warn('[WebBrain] restore scheduled alarms failed:', e));
   }
 
+  isRunning(tabId) {
+    return this._runningTabs.has(tabId);
+  }
+
   async _getJobs() {
     const stored = await this.api.storage.local.get(SCHEDULED_JOBS_KEY);
     const jobs = stored?.[SCHEDULED_JOBS_KEY];
@@ -1523,20 +1527,49 @@ export class ScheduledJobManager {
     }
 
     let tabId;
+    let reservedTabId = null;
+    const releaseReservation = () => {
+      if (reservedTabId == null) return;
+      this._runningTabs.delete(reservedTabId);
+      reservedTabId = null;
+    };
+    const reserveTab = async (candidateTabId) => {
+      if (candidateTabId == null) return;
+      const numericTabId = Number(candidateTabId);
+      const resolvedTabId = Number.isFinite(numericTabId) ? numericTabId : candidateTabId;
+      if (this._runningTabs.has(resolvedTabId) || this.agent.isRunning(resolvedTabId)) {
+        throw new Error('An agent run is already in progress for this tab.');
+      }
+      this._runningTabs.add(resolvedTabId);
+      reservedTabId = resolvedTabId;
+      try {
+        await this.agent.assertRunStartAllowed?.(resolvedTabId, 'scheduled', { scheduledRun: true });
+      } catch (error) {
+        releaseReservation();
+        throw error;
+      }
+    };
     try {
+      const candidateTabId = job.kind === 'resume' || job.target?.type === 'current_tab'
+        ? (job.tabId || job.target?.tabId)
+        : job.target?.tabId;
+      await reserveTab(candidateTabId);
       tabId = await this._resolveTab(job);
+      if (reservedTabId !== tabId) {
+        releaseReservation();
+        await reserveTab(tabId);
+      }
       await this._validateConversation(job, tabId);
       await this._validateTaskTarget(job, tabId);
     } catch (e) {
-      await this._markFailed(job, e.message);
-      return;
-    }
-
-    if (this._runningTabs.has(tabId) || this.agent.isRunning(tabId)) {
-      await this._requeue(job, 'The target tab already has an active WebBrain run.');
+      releaseReservation();
+      if (isActiveRunError(e)) {
+        await this._requeue(job, 'The target tab already has an active WebBrain run.');
+      } else {
+        await this._markFailed(job, e.message);
+      }
       return;
     }
-    this._runningTabs.add(tabId);
 
     const running = await this._updateJobIf(job.id, (prev) => (
       ['pending', 'queued'].includes(prev.status)