new RPG system

services/rpg_facts.py

@@ -20,22 +20,37 @@ Return ONLY valid JSON (no markdown), as an array of objects:
 Rules:
 - Return at most 5 NEW facts per turn. If nothing new, return [].
 - Do NOT repeat or rephrase facts already listed under "Already known".
-- Facts must be durable (names, relations, inventory, locations, lasting world state).
+- Facts must be DURABLE world/character state only:
+  traits, relationships, inventory, locations, secrets revealed, lasting abilities/rules.
+- NEVER store plot events or scene narration (no "they went", "they decided", "they hugged",
+  "they started a new arc", "they stepped through the portal").
 - Skip momentary emotions unless they permanently change a relationship.
 - text <= 120 chars each.
 - rp_day: in-world time label (день 1, второй день, та же ночь, через год). Use RP time hint when unclear."""
 
 FACTS_COMPRESS_SYSTEM = """You consolidate RPG session memory for a long-running chat.
-Return ONLY valid JSON (no markdown): an array of {"text": "...", "rp_day": "..."}.
+Return ONLY valid JSON (no markdown): an array of {{"text": "...", "rp_day": "..."}}.
 
 Goals:
+- Use ONLY information from the input facts. NEVER invent or infer new facts.
 - Aggressively MERGE near-duplicates (same topic in RU/EN, Rin/Рин, Grigo/Григорий).
 - Keep ONE best fact per topic; combine rp_day if needed (e.g. "день 1–2").
-- DROP redundant, trivial, or superseded facts.
-- Keep: names, relationships, key locations, lasting magic/rules, inventory, unresolved threads.
+- DROP redundant, trivial, superseded, and ALL one-off narrative/event facts.
+- DROP facts that describe a single scene action (went, decided, hugged, called for help, stepped into portal).
+- KEEP durable state only: names, nicknames, relationships, inventory, home items, locations,
+  lasting abilities, secrets/identity, unresolved mysteries.
 - Target at most {target} facts (fewer is better). Each text <= 120 chars.
 - rp_day = in-world labels only."""
 
+_NARRATIVE_EVENT_RE = re.compile(
+    r"(?:"
+    r"отправил(?:ись|а|и)?|решили|начали|обнялись|шагнули|вызвали|стали ближе|"
+    r"передали|раскрыла|начали новую арку|вместе шагнули|тактическ(?:ое|и) отступлени|"
+    r"went to|decided to|hugged|stepped through|called for help|started a new arc"
+    r")",
+    re.IGNORECASE,
+)
+
 _NAME_ALIASES = (
     ("grigoriy", "григорий"),
     ("grigo", "григо"),
@@ -135,6 +150,38 @@ def facts_are_similar(a: str, b: str) -> bool:
     return overlap >= 0.32
 
 
+def is_likely_narrative_event(text: str) -> bool:
+    """One-off scene actions — not durable memory."""
+    t = (text or "").strip()
+    if not t:
+        return True
+    if _NARRATIVE_EVENT_RE.search(t):
+        return True
+    if "новую арку" in t.lower() or "new arc" in t.lower():
+        return True
+    return False
+
+
+def filter_durable_facts(facts: list[dict]) -> list[dict]:
+    return [f for f in facts if not is_likely_narrative_event(f.get("text", ""))]
+
+
+def validate_compressed_against_source(
+    original: list[dict], compressed: list[dict]
+) -> list[dict]:
+    """Reject LLM-hallucinated facts not grounded in the input list."""
+    if not compressed:
+        return []
+    out: list[dict] = []
+    for c in compressed:
+        text = (c.get("text") or "").strip()
+        if not text or is_likely_narrative_event(text):
+            continue
+        if any(facts_are_similar(text, o.get("text", "")) for o in original):
+            out.append(c)
+    return out
+
+
 def dedupe_facts_fuzzy(facts: list[dict]) -> list[dict]:
     out: list[dict] = []
     for f in facts:
@@ -243,8 +290,10 @@ async def compress_facts(
                 if entry:
                     out.append(entry)
             if out:
-                logger.info("compress_facts: %d -> %d", len(facts), len(out))
-                return dedupe_facts_fuzzy(out)[:FACTS_STORE_LIMIT]
+                out = validate_compressed_against_source(facts, out)
+                if out:
+                    logger.info("compress_facts: %d -> %d", len(facts), len(out))
+                    return dedupe_facts_fuzzy(out)[:FACTS_STORE_LIMIT]
     except json.JSONDecodeError:
         logger.warning("compress_facts JSON parse failed. Raw=%.400s", raw)
     return dedupe_facts_fuzzy(facts)[-target:]
@@ -263,6 +312,7 @@ async def merge_facts_persist(
         existing_json, new_facts, rp_day_default=rp_day_default
     )
     facts = dedupe_facts_fuzzy(parse_facts_list(merged_json))
+    facts = filter_durable_facts(facts)
     if len(facts) > FACTS_DEDUP_THRESHOLD:
         facts = await compress_facts(
             facts,
@@ -340,6 +390,8 @@ async def extract_facts(
                     continue
                 if any(facts_are_similar(entry["text"], k["text"]) for k in known):
                     continue
+                if is_likely_narrative_event(entry["text"]):
+                    continue
                 if not entry["rp_day"] and hint:
                     entry["rp_day"] = hint[:80]
                 out.append(entry)
@@ -350,7 +402,7 @@ async def extract_facts(
 
 
 def facts_to_prompt(facts_json: str, max_items: int = FACTS_PROMPT_MAX) -> str:
-    facts = dedupe_facts_fuzzy(parse_facts_list(facts_json))
+    facts = filter_durable_facts(dedupe_facts_fuzzy(parse_facts_list(facts_json)))
     if not facts:
         return ""
     recent = facts[-max_items:]