Spaces:
Sleeping
Sleeping
Commit History
-
fix: centralize Supabase client on HTTP/1.1 to stop RemoteProtocolError 500s
Random 500s ("I seem to have lost my train of thought") came from httpx.RemoteProtocolError: ConnectionTerminated NO_ERROR — an HTTP/2 GOAWAY on the pooled PostgREST connection. Every module also built its own client (~10 pools). - db_client.py: one shared supabase client built with a custom httpx client (http2=False + sane timeouts/keepalive), plus db_execute() retry helper for residual transient disconnects - repoint all runtime modules to import the shared client - wrap chat-history load/save (the calls in the traceback) with db_execute Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>6d936a1
-
fix: character attribution in history, past-dialogue scoping, and declarative question suppression
- app_nn.py: pass character_id to search_past_dialogues so each character only retrieves its own past dialogues (not Socrates' by default) - db_5_process_session.py: label FAISS-archived turns with character name (SCHOPENHAUER/CAMUS/...) instead of generic ASSISTANT so long-term memory preserves character attribution - dialogue_declarative_helper.py: prepend absolute no-question override before persona block for stages A/B/C/E/F; only Stage D allows one genuine question; upgraded soft 'no question marks' rules in B/C/E to hard-rule language Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>393b1f9
-
feat: long-term chat history — time-based window + multi-vector FAISS db5
- db_5_process_session: time-based short-term window (10 day min / 21 day max, 40 msg floor / 200 msg ceiling); sliding-window chunker (8 msgs, 2 overlap); per-chunk LLM summary (Tier C); two FAISS docs per chunk (raw + summary) with rich metadata (chunk_id, timestamps, character, turn_range) - classify_memory_reference: Tier C classifier — detects "do you remember"-style queries; extracts time_hint and character_hint for metadata filtering - util_retrieve: db5 excluded by default; searched only on recalls_past=True; chunk_id deduplication; soft metadata boost for time/character hints; force_db5 flag for internal pivot/nudge calls - app_nn: classify_memory_reference wired before RAG retrieval; pivot calls use force_db5=True to bypass the memory-reference gate Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>1775907
-
refactor: systematic rename of all modules + SQL reorganisation
Python renames (prefix-based naming convention): Agent_* dialogue agents → dialogue_maieutic / dialogue_pragmatic (+ helpers) Agent_* classifiers → classify_chat / classify_perception / classify_life_event supabase_*.py → db_socratic / db_pragmatic / db_life_events / db_user *_utils.py → util_loop_hierarchy / util_philosophy_threads / util_story_chapters / util_topic_log / util_cross_character / util_geo / util_push / util_translate / util_retrieve / util_agent / util_llm proactive_pipeline → pipeline_proactive Prompt_building → prompt_building ask_llm_final_prompt → prompt_final Classification_parameters → classify_parameters db6_topic_logger → db_topic_logger Deleted: Agent_Socratic_Dialogue.py (legacy, superseded by dialogue_maieutic.py) Deleted: dialogues_socrates.json (unused) SQL reorganisation:…a5a8164
-
debug proactive
1b41399
-
correct issue with calling db4 and db6
3bf0778
-
debug socratic dialogue and adjust memories
74663ff
-
remove fet_user_id and get_username
0b170f9
-
first push
433f3f1