Commit History
-
fix: debate stance detection + expanded emotional context valid_emotions
Isaac Quarenta- debate_manager.py: improved stance detection with counting-based disambiguation (não alone is not contra, needs context) - emotional_control.py: expanded valid_emotions set to include hostil, hostile, anger, disgust, desaprovaçãofd05160 -
fix: emotional control imports + debate false contradictions + config self-import
Isaac Quarenta- emotional_control.py: fixed 'No module named config' by using relative imports with fallback - debate_manager.py: improved _claims_contradict() to check subjects before flagging contradictions (e.g. 'parecida não é clonar' vs 'clonar é crime' are consistent, not contradictory) - debate_manager.py: cleaned up duplicate entries in DEBATE_SIGNALS - debate_manager.py: increased _topics_overlap threshold to 2+ shared words to reduce false positives - config.py: removed unnecessary self-import of PALAVRAS_RUDES and AGGRESSIVE_EMOTION_LABELS935a4c4
-
fix: add personal insult patterns to DEBATE_SIGNALS for better debate activation
Isaac Quarentac51e370 -
fix: debate position tracking - Akira no longer loses own stance in long debates
Isaac Quarenta- DebateManager: added full_text + speaker_name to ArgumentPosition - DebateManager: expanded claim patterns for philosophical arguments - DebateManager: lowered debate threshold from 2 to 1 - DebateManager: expanded debate signals for philosophical debates - DebateManager: get_debate_context_for_prompt shows full conversation flow with named speakers - ThinkingEngine: strengthened debate consistency rules (MUST read DEBATE_CONTEXT before responding) - ThinkingEngine: auto-consistency rule now prioritizes ORIGINAL position over recent - api.py: passes speaker_name to update_debate_state, enhanced enforcement with Akira history - config.py: added real name + position consistency rules to personabfaaca6
-
Integrate DebateManager into message pipeline
Antigravity Assistant- update_debate_state() tracks user positions per message - get_debate_context_for_prompt() injects context into CoT (ThinkingEngine) and final prompt - track_bot_response() tracks Akira's own replies for self-contradiction detection - Programmatic enforcement when coherence < 0.5 or contradictions detected - Speaker ID consistency: user=numero_usuario, bot='Akira' - Added debate_context param to ThinkingEngine.think() and _generate_dynamic_thought() - Added <DEBATE_CONTEXT> block in system prompt with mandatory rules60d0936 -
fix: resolve merge conflicts + integrate debate_manager into api.py
Antigravity Assistant +1: Claude Opus 4.6- Remove CONFLICT markers from modules/api.py (file was 11k lines with duplicate copy from unresolved upstream/stash merge) - Add debate_state dict at top of processing pipeline - Call update_debate_state() before ThinkingEngine in both unified_context and non-unified code paths - Inject get_debate_context_for_prompt() output into prompt before thinking_analysis = _te.think() - Normalize reply tag prefix: [VOCÊ DISSO a] everywhere - Remove scratch files add_success_check.py and fix_all.py - debate_manager.py: add _time guard import - short_term_memory.py: add decay_minutes param - thinking_engine.py: add anti-hallucination + auto-consistency rules - unified_context.py: add decay_minutes param - All 61 modules/*.py pass py_compile (verified) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>1e0bd95 -
fix(debate_manager): add missing import time
Antigravity Assistant9f57109 -
fix(debate): add consistency tracking and fallacy detection to prevent Akira self-contradiction
Antigravity Assistant- Added AKIRA_STANCE, OPPONENT_STANCE, CONSISTENCY_STATUS, FALACIA_DETECTION fields to THINK_OUTPUT - Added Rule 8 (consistency check) and Rule 9 (anti-strawman) to CoT brain rules - Added Rule 10 (self-consistency before attacking) - Added CONTRADICOES block with self-consistency validation for Akira - Enhanced FALACIA_DETECTION to include strawman and warn against false positives - Added debate_manager.py for debate session management - Added skill error handling to prevent LLM confusion on failure6dd0df1