halo-agent 1.3.1 → 1.3.2

package/orchestrator.js

@@ -381,13 +381,51 @@ async function runJob(queueItem, chromeConn, config, reportStatus) {
     const confirmScreenshot = await page.screenshot({ type: 'jpeg', quality: 70 });
     const confirmKey = await uploadScreenshot(config, confirmScreenshot, `confirm_${queueId}.jpg`);
 
+    // Verify-then-DONE: trusting waitForURL alone was wrong (the Chalk bug —
+    // Ashby rendered "Missing entry for required field: Name, Email, ..."
+    // inline without a URL change, and we marked DONE on a failed submit).
+    // The backend re-fetches the page through Firecrawl + extract and tells
+    // us if the submit actually went through. On Firecrawl failure or no
+    // key, the endpoint returns submitted:true so we don't deadlock — we're
+    // strictly more correct than before, not less.
+    const verdictUrl = page.url();
+    console.log(`[orchestrator] Verifying submission at ${verdictUrl}...`);
+    let verdict = { submitted: true, error_message: null, confirmation_text: null, source: 'unavailable' };
+    try {
+      const vRes = await fetch(`${config.apiUrl}/agent/verify-submit`, {
+        method: 'POST',
+        headers: { 'Content-Type': 'application/json', Authorization: `Bearer ${config.token}` },
+        body: JSON.stringify({ queue_id: queueId, page_url: verdictUrl }),
+      });
+      if (vRes.ok) verdict = await vRes.json();
+    } catch (e) {
+      console.warn(`[orchestrator] verify-submit unavailable: ${e.message}`);
+    }
+
+    if (!verdict.submitted) {
+      const reason = verdict.error_message || 'Submission did not confirm — form may still have errors';
+      console.warn(`[orchestrator] Submission NOT verified. Reason: ${reason}`);
+      await reportStatus('NEEDS_ATTENTION', {
+        review_screenshot_r2_key: confirmKey || null,
+        needs_attention_reason: `Submit clicked but not confirmed: ${reason}`,
+        intervention_type: 'submit_failed',
+        step: 'VERIFY',
+        step_detail: reason.slice(0, 200),
+        fields_filled: cumulativeFilled,
+      });
+      // Do NOT clearCheckpoint — user may dismiss + re-queue, and a stale
+      // checkpoint would resume into the same failed state. The dismiss flow
+      // clears it (DELETE /apply-queue/:id sets form_checkpoint_json = NULL).
+      throw new Error(`Submission failed verification: ${reason}`);
+    }
+
     await reportStatus('DONE', {
       confirmation_screenshot_r2_key: confirmKey || null,
       fields_filled: cumulativeFilled,
     });
     await clearCheckpoint(config, queueId);
 
-    console.log(`[orchestrator] Done: ${queueItem.company} - ${queueItem.title}`);
+    console.log(`[orchestrator] Done (verified): ${queueItem.company} - ${queueItem.title}${verdict.source === 'firecrawl' ? ' · firecrawl-verified' : ' · unverified'}`);
 
     // Post fill session data to backend for learning loop
     await postFillSession(config, {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "halo-agent",
-  "version": "1.3.1",
+  "version": "1.3.2",
   "description": "HALO local apply agent — auto-fills job applications using your real Chrome session",
   "main": "index.js",
   "bin": {

package/scanPage.js

@@ -207,8 +207,27 @@ const CONSENT_PATTERNS = [
   /privacy\s*policy/i,
 ];
 
+// Normalize a label so trailing decorations don't break strict regex matches.
+// Ashby/Greenhouse often render labels like "Name *", "Email (required)",
+// "LinkedIn Profile — Required", "Phone *Required". Without this, the
+// PROFILE_PATTERNS (which use ^anchored regexes for short fields like name)
+// silently miss and the field falls through to 'custom', which means the
+// agent skips filling it with profile data.
+function normalizeLabel(raw) {
+  if (!raw) return '';
+  return raw
+    .toLowerCase()
+    .replace(/[*†‡]/g, ' ')                              // markers
+    .replace(/\((required|optional|mandatory)\)/g, ' ')   // "(required)"
+    .replace(/\b(required|optional|mandatory)\b/g, ' ')   // "required"
+    .replace(/[—–-]+\s*(required|optional)\s*$/g, ' ')   // "— required"
+    .replace(/[:?]+$/g, '')                               // trailing : ?
+    .replace(/\s+/g, ' ')
+    .trim();
+}
+
 function classifyField(field) {
-  const label = field.label.toLowerCase().trim();
+  const label = normalizeLabel(field.label);
 
   // Consent checkboxes
   if (field.inputType === 'checkbox' && CONSENT_PATTERNS.some(r => r.test(label))) {
@@ -222,7 +241,7 @@ function classifyField(field) {
     if (regex.test(label)) return 'profile:' + name;
   }
 
-  // Ashby system fields by name attribute
+  // Ashby system fields by name attribute (legacy form schema)
   if (field.name && field.name.startsWith('_systemfield_')) {
     const sfName = field.name.replace('_systemfield_', '');
     if (['name', 'email', 'phone', 'resume', 'linkedin', 'website'].includes(sfName)) {
@@ -323,11 +342,12 @@ async function scanAshby(page) {
       if (sfField === 'resume' || f.inputType === 'file') return { ...f, category: 'file:resume' };
       if (profileMap[sfField] !== undefined) return { ...f, category: 'profile:' + profileMap[sfField] };
     }
-    // UUID-named fields are custom questions
-    if (f.name && /^[0-9a-f-]{36}$/.test(f.name)) {
-      return { ...f, category: 'custom' };
-    }
     if (f.inputType === 'file') return { ...f, category: 'file:resume' };
+    // Modern Ashby uses UUID names for ALL fields (Name, Email, Phone, custom).
+    // Classify by LABEL first — that's the only signal that distinguishes
+    // a profile field from a custom question when the name attr is opaque.
+    // Don't shortcut UUID-named fields to 'custom' — that's exactly the
+    // Chalk bug where Name/Email/Phone/LinkedIn all fell through unfilled.
     return { ...f, category: classifyField(f) };
   }).filter(f => f.category !== 'ignore');
 }