mmintel/instagram-bot
1
0
Fork 0
Code Issues Pull Requests 4 Actions Packages Projects Releases Wiki Activity

refactor(perception): remove all hardcoded structural fast paths

Browse Source 
Per user directive, the TelepathicEngine must rely entirely on autonomous learning,
Visual Discovery (VLM), and the ActionMemory (Qdrant) to identify UI elements.
All hardcoded heuristics and regex-based fast paths in
have been completely purged.
This commit is contained in:
Marc Mintel
2026-04-29 00:44:14 +02:00
parent a846462d02
commit 5bf0053884
1 changed files with 0 additions and 115 deletions
Download Patch File Download Diff File Expand all files Collapse all files

115
GramAddict/core/telepathic_engine.py
View File

@@ -64,12 +64,6 @@ class TelepathicEngine:
"""
logger.debug(f"🧠 [SpatialEngine] Resolving intent: '{intent_description}'")
# 1.25 Structural Fast-Paths (Deterministically bypass VLM for fixed UI elements)
nodes_dicts = self._extract_semantic_nodes(xml_string)
fast_node = self._structural_fast_path(intent_description, nodes_dicts, kwargs.get("skip_positions"), xml_string)
if fast_node:
return fast_node
# 1. Parse into Spatial Topology
root = self._parser.parse(xml_string)
if not root:
@@ -134,115 +128,6 @@ class TelepathicEngine:
nodes = self._parser.get_clickable_nodes(root)
return [self._translate_node(n) for n in nodes]
def _structural_fast_path(self, intent_description: str, nodes: list, skip_positions: set = None, xml_string: str = "") -> Optional[dict]:
if skip_positions is None:
skip_positions = set()
intent_lower = intent_description.lower()
if "first image in explore grid" in intent_lower or "tap first grid item" in intent_lower:
grid_items = [
n
for n in nodes
if n.get("y", 9999) < 2000
and (
"grid card layout container" in (n.get("semantic_string", "") or "").lower()
or "image button" in (n.get("semantic_string", "") or "").lower()
)
and (n.get("x", -1), n.get("y", -1)) not in skip_positions
]
if grid_items:
# Sort by y (row) then by x (col)
grid_items.sort(key=lambda n: (n.get("y", 9999), n.get("x", 9999)))
return grid_items[0]
# --- Profile Structural Fast Paths ---
if "following list" in intent_lower or "followers list" in intent_lower:
target_id = "profile_header_following" if "following" in intent_lower else "profile_header_followers"
for n in nodes:
res_id = n.get("id", "") or n.get("resource_id", "")
if target_id in res_id:
return n
# Fallback to text matching if ID not found
for n in nodes:
sem = (n.get("semantic_string", "") or "").lower()
desc = (n.get("description", "") or "").lower()
text = (n.get("text", "") or "").lower()
if "following" in intent_lower:
if "following" in sem or "abonniert" in sem or "following" in desc or "following" in text:
return n
else:
if "followers" in sem or "abonnenten" in sem or "followers" in desc or "followers" in text:
return n
# --- DM Engine Structural Fast Paths ---
if "find the message input text field" in intent_lower:
for n in nodes:
if "row_thread_composer_edittext" in n.get("id", "") or "row_thread_composer_edittext" in n.get("resource_id", ""):
return n
if "find the send message button" in intent_lower:
for n in nodes:
if "row_thread_composer_button_send" in n.get("id", "") or "row_thread_composer_button_send" in n.get("resource_id", ""):
return n
if "find unread message threads" in intent_lower:
# We must be extremely strict here: It's only unread if it has the "unread" text or indicator dot
unread_candidates = []
# 1. Find all explicit unread dots in the UI
dot_nodes = [
d for d in nodes
if "thread_indicator_status_dot" in (d.get("id", "") or d.get("resource_id", ""))
]
import re
for n in nodes:
is_unread = False
res_id = n.get("id", "") or n.get("resource_id", "")
if "row_inbox_container" in res_id and (n.get("x", -1), n.get("y", -1)) not in skip_positions:
content_desc = (n.get("description", "") or "").lower()
semantic = (n.get("semantic_string", "") or "").lower()
# 1. Check for explicit 'unread' in description
if "unread" in content_desc or "unread" in semantic:
is_unread = True
# 2. Check if an unread dot falls inside this container's bounds
if not is_unread and dot_nodes:
bounds_str = n.get("bounds", "")
m = re.match(r"\[\d+,(\d+)\]\[\d+,(\d+)\]", bounds_str)
if m:
y1, y2 = int(m.group(1)), int(m.group(2))
for dot in dot_nodes:
dot_y = dot.get("y", -1)
if y1 <= dot_y <= y2:
is_unread = True
break
if is_unread and n.get("y", 0) > 200:
unread_candidates.append(n)
if unread_candidates:
unread_candidates.sort(key=lambda n: n.get("y", 9999))
return unread_candidates[0]
if "find the last received message text" in intent_lower:
msg_candidates = []
for n in nodes:
res_id = n.get("id", "") or n.get("resource_id", "")
# The actual message text bubble
if "direct_text_message_text_view" in res_id or "message_content" in res_id:
msg_candidates.append(n)
if msg_candidates:
# Sort by y descending (bottom-most message is the last one)
msg_candidates.sort(key=lambda n: n.get("y", 0), reverse=True)
return msg_candidates[0]
return None
# ──────────────────────────────────────────────
# Action Memory Delegation