fix: purge 4 production bugs — resonance null-guard, POST_DETAIL misclassification, VLM JSON response handling

🔴 RED → 🟢 GREEN for 4 critical bugs found in production run 2026-04-29:

1. ResonanceEvaluator: Add null-guard for evaluate_post_vibe() return.
   When VLM returns truncated JSON, the function returns None. The caller
   now handles this gracefully instead of crashing with AttributeError.

2. ScreenIdentity POST_DETAIL: Replace broken 'and not selected_tab'
   condition with structural differentiator using main_feed_action_bar.
   Posts opened from feed retain feed_tab selected, which was causing
   misclassification as HOME_FEED → LLM fallback → OWN_PROFILE hallucination
   → permanent Qdrant cache poisoning.

3. ActionMemory VLM verification: When VLM returns JSON instead of YES/NO,
   treat as inconclusive (fall through to structural delta) rather than
   hard failure. Only return False when response explicitly contains 'no'.

4 new E2E regression tests, 75/75 pass, zero regressions.
This commit is contained in:
2026-04-29 18:05:32 +02:00
parent effb1f5ae1
commit 068a6a616a
4 changed files with 295 additions and 7 deletions

View File

@@ -51,10 +51,15 @@ class ResonanceEvaluatorPlugin(BehaviorPlugin):
logger.info("✨ [Resonance] Performing visual vibe check...")
persona_interests = getattr(ctx.configs.args, "persona_interests", [])
vibe = tele.evaluate_post_vibe(ctx.device, persona_interests)
vibe_score = vibe.get("quality_score", 5) / 10.0
if vibe.get("matches_niche"):
vibe_score = min(1.0, vibe_score + 0.2)
res_score = (res_score * 0.3) + (vibe_score * 0.7)
if vibe is None:
logger.warning(
"✨ [Resonance] VLM vibe check returned None (truncated JSON?). Keeping neutral score."
)
else:
vibe_score = vibe.get("quality_score", 5) / 10.0
if vibe.get("matches_niche"):
vibe_score = min(1.0, vibe_score + 0.2)
res_score = (res_score * 0.3) + (vibe_score * 0.7)
ctx.shared_state["res_score"] = res_score
logger.info(f"📊 [Resonance] Post Score: {res_score:.2f}")

View File

@@ -192,11 +192,18 @@ class ActionMemory:
if response and "yes" in response.lower() and "no" not in response.lower():
logger.debug(f"🧠 [ActionMemory] VLM visually confirmed success for '{intent}'.")
return True
else:
elif response and "no" in response.lower() and "yes" not in response.lower():
logger.warning(
f"⚠️ [ActionMemory] VLM visual verification FAILED for '{intent}'. VLM replied: '{response}'"
)
return False
else:
# VLM returned ambiguous response (JSON, mixed signals, etc.)
# Don't treat as hard failure — fall through to structural delta verification
logger.info(
f"🧠 [ActionMemory] VLM response for '{intent}' was not YES/NO "
f"(got: '{response[:80]}...'). Falling through to structural verification."
)
except Exception as e:
logger.error(f"Failed to query VLM for visual verification: {e}")
# Fallthrough to structural delta if VLM crashes

View File

@@ -193,8 +193,14 @@ class ScreenIdentity:
except KeyError:
pass
if "row_feed_button_like" in ids and "row_feed_photo_profile_name" in ids and not selected_tab:
return ScreenType.POST_DETAIL
# POST_DETAIL vs HOME_FEED: Both have row_feed_* markers. The differentiator
# is that HOME_FEED has the main_feed_action_bar (top bar with 'Instagram' title).
# POST_DETAIL lacks this because it shows a single expanded post.
# Note: We MUST NOT use `not selected_tab` here — posts opened from feed
# retain the feed_tab as selected, which previously caused misclassification.
if "row_feed_button_like" in ids and "row_feed_photo_profile_name" in ids:
if "main_feed_action_bar" not in ids:
return ScreenType.POST_DETAIL
# Story view structural markers — present in full-screen story viewer.
# Stories hide the navigation tab bar, so selected_tab is always None.