Merge pull request #218 from PredicateSystems/tweak_count

Distinguish noun 'count' from verb 'count' to fix extraction task rou…

Author: rcholic

src/agents/planner-executor/extraction-keywords.ts

@@ -230,7 +230,13 @@ export function isTextExtractionTask(task: string): boolean {
   // Form-fill negative signal: if the task clearly involves filling a form,
   // it's not extraction even if it contains extraction-like keywords
   // (e.g., "Display name", "email" are field labels, not extraction targets)
-  if (FORM_FILL_SIGNALS.some(signal => taskLower.includes(signal))) {
+  const hasFormFillSignal = FORM_FILL_SIGNALS.some(signal => {
+    if (signal.endsWith(' ')) {
+      return taskLower.includes(signal);
+    }
+    return new RegExp(`\\b${escapeRegExp(signal)}\\b`).test(taskLower);
+  });
+  if (hasFormFillSignal) {
     return false;
   }
 
@@ -321,7 +327,10 @@ export function isExtractionTask(task: string): boolean {
     taskLower.includes('content of') ||
     taskLower.includes('headline of') ||
     taskLower.includes('rating of') ||
-    taskLower.includes('review of')
+    taskLower.includes('review of') ||
+    taskLower.includes('summarize') ||
+    taskLower.includes('note the') ||
+    taskLower.includes('list the')
   );
 }
 
@@ -332,17 +341,34 @@ export function isExtractionTask(task: string): boolean {
  * is detected as an extraction task, instructing the planner to
  * use EXTRACT instead of CLICK for data that is already visible.
  */
-export function getExtractionDomainGuidance(): string {
+export function getExtractionDomainGuidance(includeCounting = false): string {
+  const countingSection = includeCounting
+    ? `
+
+STEP 3 - COUNTING ACROSS FULL PAGE:
+If the task asks to COUNT items (e.g., "how many listings", "number of results", "count the products", "total number of entries"):
+- Use SCROLL_AND_COUNT instead of EXTRACT
+- Set "countTarget" to describe what to count (e.g., "listings", "products", "articles")
+- The system will scroll through the entire page and sum up counts
+- Do NOT use EXTRACT for counting tasks — EXTRACT only sees the current viewport
+
+Example - count all listings:
+Goal: "note how many listings are available"
+Current URL: alibaba.com/search?SearchText=smartphones (correct page)
+{"action":"SCROLL_AND_COUNT","countTarget":"product listings","goal":"Count total product listings","verify":[]}
+`
+    : '';
+
   return `
 
 IMPORTANT: Extraction Task Planning Rules
-========================================
+=======================================
 
 STEP 1 - CHECK CURRENT URL:
 Before choosing an action, compare the Current URL to the goal.
 - Does the current page contain the data requested?
 - If the goal mentions a specific section/page (e.g., "show hn", "top stories", "/show"), check if the URL matches.
-- If you are NOT on the right page, NAVIGATE to the correct URL first.
+- If you are NOT on the right page, NAVIGATE or CLICK to navigate to the correct page first.
 
 STEP 2 - EXTRACT VISIBLE DATA:
 If the data is VISIBLE in the page context:
@@ -352,7 +378,7 @@ If the data is VISIBLE in the page context:
 CRITICAL: Do NOT click on links to external sites when extracting.
 - Post/article titles often link to EXTERNAL sites
 - To extract a title that is visible, use EXTRACT directly on the current page
-- Only click if you need to navigate to a detail page (e.g., for comments)
+- Only click if you need to navigate to a detail page to access the data (e.g., nutritional info on a recipe page)
 
 Example - wrong page, need to navigate first:
 Goal: "extract the title of the first showhn post on hackernews show"
@@ -369,18 +395,12 @@ Goal: "find the price of the first laptop"
 Current URL: store.com/laptops (correct page, prices visible)
 {"action":"EXTRACT","target":"price of first laptop","goal":"Extract the price of the first laptop listing","verify":[],"reasoning":"prices are visible in listing elements"}
 
-STEP 3 - COUNTING ACROSS FULL PAGE:
-If the task asks to COUNT items ("how many", "number of", "count", "total"):
-- Use SCROLL_AND_COUNT instead of EXTRACT
-- Set "countTarget" to describe what to count (e.g., "listings", "products", "articles")
-- The system will scroll through the entire page and sum up counts
-- Do NOT use EXTRACT for counting tasks — EXTRACT only sees the current viewport
-
-Example - count all listings:
-Goal: "note how many listings are available"
-Current URL: alibaba.com/search?SearchText=smartphones (correct page)
-{"action":"SCROLL_AND_COUNT","countTarget":"product listings","goal":"Count total product listings","verify":[]}
- `;
+Example - data on a detail page, need to click first:
+Goal: "summarize the calorie count from a recipe's nutritional information"
+Current URL: allrecipes.com/search?q=cookies (search results, not a recipe page)
+{"action":"CLICK","intent":"first recipe link","input":"Best Chocolate Chip Cookies","verify":[],"reasoning":"need to navigate to recipe detail page for nutritional info"}
+${countingSection}
+`;
 }
 
 // ---------------------------------------------------------------------------
@@ -392,22 +412,83 @@ const COUNTING_PHRASES: readonly string[] = [
   'how much',
   'number of',
   'count the',
-  'count of',
+  'count all',
+  'count each',
+  'count every',
   'total number',
-  'total count',
+  'total count of',
   'how numerous',
 ];
 
-const COUNTING_VERBS: readonly string[] = ['count', 'tally', 'enumerate'];
+// NOTE: "count of" is intentionally excluded from COUNTING_PHRASES because
+// it is ambiguous: "count of items" (verb phrase) vs "word count of the
+// article" (noun compound). The "number of" phrase covers the counting
+// semantics of "count of". Bare "count" is handled by Tier 3 below.
+
+const COUNTING_VERBS: readonly string[] = ['tally', 'enumerate'];
+
+// Words that can syntactically precede "count" when it is used as a VERB
+// (imperative, infinitive, or after an auxiliary/modal). This set is finite
+// and well-defined in English grammar. Any word NOT in this set, appearing
+// immediately before "count", indicates a noun compound like "calorie count",
+// "word count", "error count" — regardless of what the modifier noun is.
+const COUNT_VERB_PRECEDERS = new Set([
+  // Infinitive marker
+  'to',
+  // Modals and auxiliaries
+  'can',
+  'could',
+  'will',
+  'would',
+  'shall',
+  'should',
+  'must',
+  'may',
+  'might',
+  'do',
+  'did',
+  'does',
+  'have',
+  'has',
+  'had',
+  // Polite / adverbial markers
+  'please',
+  'just',
+  'also',
+  'even',
+  'still',
+  'not',
+  'never',
+  'only',
+  // Conjunctions that continue an action sequence
+  'and',
+  'or',
+  'but',
+  'then',
+  // Verbs that take infinitive complements
+  'going',
+  'try',
+  'want',
+  'need',
+  'help',
+  'let',
+  'plan',
+  'attempt',
+  'aim',
+  'start',
+  'begin',
+]);
 
 export function isCountingTask(task: string): boolean {
   if (!task) return false;
   const taskLower = task.toLowerCase();
 
+  // Tier 1: Unambiguous counting phrases (substring match)
   if (COUNTING_PHRASES.some(phrase => taskLower.includes(phrase))) {
     return true;
   }
 
+  // Tier 2: Unambiguous counting verbs (word boundary match)
   if (
     COUNTING_VERBS.some(verb =>
       new RegExp(`\\b${escapeRegExp(verb)}(s|ed|ing)?\\b`).test(taskLower)
@@ -416,6 +497,29 @@ export function isCountingTask(task: string): boolean {
     return true;
   }
 
+  // Tier 3: Context-aware "count" — distinguish verb from noun compound.
+  //
+  // In English, "[noun] count" is a compound noun meaning "the count of
+  // [noun]" (e.g., "calorie count", "word count", "page count", "error
+  // count"). As a VERB, "count" appears in imperative position (start of
+  // clause) or after auxiliaries/modals/infinitive markers.
+  //
+  // Strategy: find every "count" preceded by a word. If ALL preceding words
+  // are NOT in COUNT_VERB_PRECEDERS, every occurrence is a noun compound and
+  // this is NOT a counting task. If any occurrence has no preceding word
+  // (imperative) or is preceded by a verb preceder, it IS a counting task.
+  if (/\bcount(s|ed|ing)?\b/i.test(taskLower)) {
+    const precedingMatches = [...taskLower.matchAll(/\b([a-z]+)\s+count(s|ed|ing)?\b/g)];
+    if (precedingMatches.length > 0) {
+      const allAreNounCompounds = precedingMatches.every(m => !COUNT_VERB_PRECEDERS.has(m[1]));
+      if (allAreNounCompounds) {
+        return false;
+      }
+    }
+    // "count" with no preceding word (imperative) or preceded by a verb marker
+    return true;
+  }
+
   return false;
 }
 

src/agents/planner-executor/planner-executor-agent.ts

@@ -1320,7 +1320,9 @@ export class PlannerExecutorAgent {
             isExtractAction &&
             finalOutcome.status === StepStatus.SUCCESS &&
             finalOutcome.extractedData &&
-            (isTextExtractionTask(task) || isCountingTask(task)) &&
+            (plannerAction.action === 'SCROLL_AND_COUNT'
+              ? isCountingTask(task)
+              : isTextExtractionTask(task) || isCountingTask(task)) &&
             (!taskHasInteractionLocal || hasNonExtractActionLocal)
           ) {
             if (this.config.verbose) {
@@ -1847,7 +1849,9 @@ export class PlannerExecutorAgent {
             isRetryExtractOrCount &&
             finalOutcome.status === StepStatus.SUCCESS &&
             finalOutcome.extractedData &&
-            (isTextExtractionTask(task) || isCountingTask(task)) &&
+            (plannerAction.action === 'SCROLL_AND_COUNT'
+              ? isCountingTask(task)
+              : isTextExtractionTask(task) || isCountingTask(task)) &&
             (!taskHasInteraction || hasNonExtractAction)
           ) {
             success = true;

src/agents/planner-executor/prompts.ts

@@ -6,7 +6,11 @@
  */
 
 import type { ActionRecord } from './plan-models';
-import { isExtractionTask, getExtractionDomainGuidance } from './extraction-keywords';
+import {
+  isExtractionTask,
+  isCountingTask,
+  getExtractionDomainGuidance,
+} from './extraction-keywords';
 
 // ---------------------------------------------------------------------------
 // Stepwise Planner Prompt (ReAct-style)
@@ -63,7 +67,7 @@ Actions:
 - TYPE: Type text into a SINGLE form field. Prefer FILL_FORM for forms with multiple fields.
 - TYPE_AND_SUBMIT: Type text into a search box and submit. Set "input" to the SEARCH QUERY from the goal (NOT the element label).
 - SCROLL: Scroll page. Set "direction" to "up" or "down".
-- SCROLL_AND_COUNT: Scroll through the ENTIRE page and count items. Use when the task asks "how many", "number of", "count", or "total". Set "countTarget" to describe what to count (e.g., "listings", "products", "articles"). The system scrolls viewport-by-viewport, counts matching items at each position, and sums the total.
+- SCROLL_AND_COUNT: Scroll through the ENTIRE page and count items. Use ONLY when the task asks to enumerate items (e.g., "how many listings", "number of results", "count the products"). Do NOT use when "count" is a data value to read (e.g., "calorie count", "word count" = use EXTRACT instead). Set "countTarget" to describe what to count.
 - WAIT: Wait for content to appear when a follow-up verification is needed.
 - EXTRACT: Extract the requested information from the current page when the task is data collection.
 - STUCK: Use only when the page state is blocked and you cannot make safe forward progress.
@@ -124,7 +128,9 @@ RULES:
 14. Treat history results "success", "skipped", and "vision_fallback" as already satisfied. Do not repeat those steps; choose the next incomplete part of the goal.`;
 
   // Inject extraction-specific guidance when the goal is an extraction task
-  const extractionGuidance = isExtractionTask(goal) ? getExtractionDomainGuidance() : '';
+  const extractionGuidance = isExtractionTask(goal)
+    ? getExtractionDomainGuidance(isCountingTask(goal))
+    : '';
 
   // NOTE: /no_think MUST be at the START of user message for Qwen3 models
   const user = `/no_think