more fixes for aggregation

Author: ndbroadbent

Taskfile.yml

@@ -74,12 +74,7 @@ tasks:
   check-ignores:
     desc: Check for biome-ignore comments (zero tolerance)
     cmds:
-      - |
-        if grep -r "biome-ignore" --include="*.ts" src/ 2>/dev/null; then
-          echo "❌ Found biome-ignore comments - these are forbidden!"
-          exit 1
-        fi
-        echo "✅ No biome-ignore comments found"
+      - ./scripts/check-ignores.sh
 
   check-embeddings:
     desc: Verify all query strings have pre-computed embeddings

scripts/check-ignores.sh

@@ -0,0 +1,8 @@
+#!/usr/bin/env bash
+# Check for forbidden biome-ignore comments in source code
+
+if grep -r "biome-ignore" --include="*.ts" src/ 2>/dev/null; then
+  echo "❌ Found biome-ignore comments - these are forbidden!"
+  exit 1
+fi
+echo "✅ No biome-ignore comments found"

src/classifier/index.test.ts

@@ -70,6 +70,7 @@ function createMockClassifierResponse(
     confidence: number
     city?: string
     country?: string
+    action?: string
   }>
 ): string {
   // Convert to new schema format
@@ -81,16 +82,17 @@ function createMockClassifierResponse(
       int: 0.5,
       cat: item.category,
       conf: item.confidence,
-      gen: true,
-      com: true,
-      act: null,
-      act_orig: null,
+      gen: false,
+      com: false,
+      act: item.action ?? 'try',
+      act_orig: item.action ?? 'try',
       obj: null,
       obj_orig: null,
       venue: null,
       city: item.city ?? null,
       region: null,
-      country: item.country ?? null
+      country: item.country ?? null,
+      kw: []
     }))
   )
 }

src/classifier/index.ts

@@ -50,11 +50,19 @@ function normalizeCategory(category: string): ActivityCategory {
 
 /**
  * Convert a parsed classification to a classified activity.
+ * Returns null if the activity has no action (invalid/incomplete classification).
  */
 function toClassifiedActivity(
   response: ParsedClassification,
   candidate: CandidateMessage
-): ClassifiedActivity {
+): ClassifiedActivity | null {
+  // Action is required - if empty, the classification is invalid
+  if (!response.act || response.act.trim() === '') {
+    console.warn(
+      `[classifier] Discarding activity with empty action: msg=${response.msg}, title="${response.title}"`
+    )
+    return null
+  }
   // Capitalize first letter of title
   const title = response.title ?? candidate.content.slice(0, 100)
   const capitalizedTitle = title.charAt(0).toUpperCase() + title.slice(1)
@@ -143,13 +151,16 @@ export async function classifyBatch(
     const expectedIds = candidates.map((c) => c.messageId)
     const parsed = parseClassificationResponse(responseResult.value, expectedIds)
 
-    // Map responses to candidates
+    // Map responses to candidates, filtering out invalid classifications
     const suggestions: ClassifiedActivity[] = []
 
     for (const response of parsed) {
       const candidate = candidates.find((c) => c.messageId === response.msg)
       if (candidate) {
-        suggestions.push(toClassifiedActivity(response, candidate))
+        const activity = toClassifiedActivity(response, candidate)
+        if (activity) {
+          suggestions.push(activity)
+        }
       }
     }
 

src/classifier/prompt.ts

@@ -198,7 +198,7 @@ Return JSON array with ONLY activities worth saving. Skip non-activities entirel
     "conf": <0.0-1.0 your confidence>,
     "gen": <true if generic, no specific venue/URL>,
     "com": <true if compound/complex activity that one JSON object can't fully represent>,
-    "act": "<normalized action: hike, eat, watch, visit>",
+    "act": "<normalized action: go, hike, eat, watch, play, visit, etc. (always required)>",
     "act_orig": "<original action word>",
     "obj": "<normalized object: movie, restaurant>",
     "obj_orig": "<original object word>",