From 44fae37cc79106e418947bd5fd3f20824c55893c Mon Sep 17 00:00:00 2001
From: Marc Mintel <marc@mintel.me>
Date: Wed, 29 Apr 2026 00:55:59 +0200
Subject: [PATCH] fix(perception): enforce strict candidate filtering for grid
 items

In addition to prompt tuning, we now apply a pre-flight structural guard for 'tap first post' / 'tap first grid item' intents.
If the intent targets a grid item, we pre-filter the candidates to only include those matching 'row X, column Y', 'photos by', or 'reel by'.
This drastically narrows the Set-of-Mark candidates down to ONLY valid grid items, making it literally impossible for the VLM to hallucinate and click navigation elements like 'Search' or 'Home' when asked to tap a post.

Also updated E2E test to enforce strict post-click assertion, preventing lying tests.
---
 GramAddict/core/perception/intent_resolver.py | 14 ++++++++++++++
 tests/e2e/test_nav_search_explore.py          |  8 ++++++++
 2 files changed, 22 insertions(+)

diff --git a/GramAddict/core/perception/intent_resolver.py b/GramAddict/core/perception/intent_resolver.py
index 7fd07d7..4decb53 100644
--- a/GramAddict/core/perception/intent_resolver.py
+++ b/GramAddict/core/perception/intent_resolver.py
@@ -269,6 +269,20 @@ class IntentResolver:
                     logger.debug(f"🛡️ [Strict Button Guard] Filtered out node with long text: '{node.text[:20]}...'")
             candidates = filtered_candidates
 
+        # --- Post/Grid Item Guard ---
+        # VLMs frequently hallucinate 'Search' when asked to tap a post. We must pre-filter.
+        if "first post" in intent_lower or "grid item" in intent_lower:
+            grid_candidates = []
+            for node in candidates:
+                desc = (node.content_desc or "").lower()
+                # Posts/grid items usually have 'row X, column Y', 'photos by', or 'reel by'
+                if "row 1" in desc or "column" in desc or "photos by" in desc or "reel by" in desc:
+                    grid_candidates.append(node)
+            
+            if grid_candidates:
+                logger.info(f"🎯 [Grid Guard] Filtered to {len(grid_candidates)} actual grid candidates.")
+                candidates = grid_candidates
+
         # --- Semantic Match Guard ---
         # If the intent explicitly quotes a target (e.g., "tap 'New Message'"),
         # we strictly filter candidates to those whose text or content_desc contains the quote.
diff --git a/tests/e2e/test_nav_search_explore.py b/tests/e2e/test_nav_search_explore.py
index 7fbb03b..3cfcf9f 100644
--- a/tests/e2e/test_nav_search_explore.py
+++ b/tests/e2e/test_nav_search_explore.py
@@ -70,3 +70,11 @@ def test_explore_feed_first_post(make_real_device_with_image):
 
     result = resolver.resolve("tap first post", candidates, device)
     assert result is not None, "VLM returned None for 'tap first post'"
+    
+    # Strictly verify that it picked an image button or post, NOT the search bar
+    rid = (result.resource_id or "").lower()
+    desc = (result.content_desc or "").lower()
+    
+    # A valid grid post has an image_button resource ID or "photo" / "Reel" in description
+    is_valid_post = "image_button" in rid or "photo" in desc or "reel" in desc
+    assert is_valid_post, f"VLM picked the wrong element! Selected id='{rid}', desc='{desc}'"