fix

backend/app/engine/context.py

@@ -58,7 +58,10 @@ async def build_orchestrator_messages(
 
     recent = visible_msgs[-recent_n:] if visible_msgs else []
 
-    # Build orchestrator system prompt with current state
+    # Build orchestrator system prompt with current state.
+    # NOTE: prompts are always English (system content convention). The LLM
+    # produces player-facing text in world.language when relevant (the
+    # step-writer prompt interpolates world_language explicitly).
     defn = world.definition or {}
     system_prompt_template = get_prompt("orchestrator", world.language)
     player_state = world.state.get("player", {}) if world.state else {}
@@ -69,14 +72,14 @@ async def build_orchestrator_messages(
         current_time=world.current_time or "",
         player_state=json.dumps(player_state, ensure_ascii=False)[:600],
         plot_rails=json.dumps(defn.get("plot_rails", {}), ensure_ascii=False)[:400],
-        summary=summary_text or "(нет сводки)",
+        summary=summary_text or "(no summary yet)",
     )
 
     messages: List[Dict[str, Any]] = [{"role": "system", "content": system_prompt}]
 
     # Add summary as a system note if present
     if summary_text:
-        messages.append({"role": "system", "content": f"Сводка прошлого:\n{summary_text}"})
+        messages.append({"role": "system", "content": f"Past summary:\n{summary_text}"})
 
     # Add recent visible messages
     for m in recent:
@@ -86,7 +89,7 @@ async def build_orchestrator_messages(
             messages.append({"role": "assistant", "content": m.content})
 
     # The current action
-    messages.append({"role": "user", "content": f'Действие игрока: "{action_text}"'})
+    messages.append({"role": "user", "content": f'Player action: "{action_text}"'})
 
     return messages, settings_map
 
@@ -100,7 +103,11 @@ async def _maybe_compress(
     world: World,
     settings_map: Dict[str, Any],
 ) -> None:
-    """If history exceeds threshold, summarize older messages into a single summary message."""
+    """If history exceeds threshold, summarize older messages into a single summary message.
+
+    Uses the `submit_summary` tool (tool-calling-first design) to get structured
+    output from the summarizer LLM.
+    """
     visible = [m for m in all_msgs if not m.hidden]
     if len(visible) <= recent_n + summary_n:
         return
@@ -110,45 +117,61 @@ async def _maybe_compress(
     if not to_summarize:
         return
 
-    # Build summarization input
+    # Build summarization input (English labels — system content convention)
     summary_input_lines = []
     for m in to_summarize:
         prefix = {
-            "player_action": "Игрок",
-            "narrative_step": "Сцена",
-            "summary": "Сводка",
+            "player_action": "Player",
+            "narrative_step": "Scene",
+            "summary": "Summary",
             "orchestrator_plan": "GM",
-            "technical_offscreen": "За кадром",
+            "technical_offscreen": "Offscreen",
         }.get(m.kind, m.kind)
         summary_input_lines.append(f"{prefix}: {m.content[:300]}")
     summary_input = "\n\n".join(summary_input_lines)
 
     llm = LlmClient(settings_map)
     system_prompt = get_prompt("summarizer", world.language)
+    from app.engine.tools.tools import SUMMARIZER_TOOL_SCHEMAS
     response = await llm.chat(
         messages=[
             {"role": "system", "content": system_prompt},
             {"role": "user", "content": summary_input[:4000]},
         ],
+        tools=SUMMARIZER_TOOL_SCHEMAS,
         temperature=float(cast_setting("llm.summary_temperature", settings_map.get("llm.summary_temperature", 0.3))),
-        max_tokens=300,
+        max_tokens=400,
         purpose="summary",
         session_id=session_id,
         db=db,
     )
 
-    # Parse summary response
-    summary_text = response.text
+    # Extract from submit_summary tool call; fall back to text parse.
+    summary_text = response.text or ""
     facts: List[Dict[str, Any]] = []
-    import re as _re
-    json_match = _re.search(r"\{[\s\S]*\}", response.text)
-    if json_match:
-        try:
-            data = json.loads(json_match.group(0))
-            summary_text = data.get("summary", response.text)
-            facts = data.get("facts", [])
-        except json.JSONDecodeError:
-            pass
+    extracted = False
+    for tc in (response.tool_calls or []):
+        if tc.get("function", {}).get("name") == "submit_summary":
+            args_str = tc.get("function", {}).get("arguments", "{}")
+            try:
+                data = json.loads(args_str) if args_str else {}
+                summary_text = data.get("summary", response.text or "")
+                facts = data.get("facts", []) or []
+                extracted = True
+            except json.JSONDecodeError:
+                pass
+            break
+    if not extracted:
+        # Fallback: extract JSON from text response (older models).
+        import re as _re
+        json_match = _re.search(r"\{[\s\S]*\}", response.text or "")
+        if json_match:
+            try:
+                data = json.loads(json_match.group(0))
+                summary_text = data.get("summary", response.text or "")
+                facts = data.get("facts", []) or []
+            except json.JSONDecodeError:
+                pass
 
     # Create summary message
     next_seq = (max((m.seq for m in all_msgs), default=0)) + 1
@@ -207,7 +230,11 @@ async def build_step_writer_messages(
     outcome: str,
     narrative_prompt: str,
 ) -> List[Dict[str, Any]]:
-    """Build messages for the step writer LLM call."""
+    """Build messages for the step writer LLM call.
+
+    The step-writer prompt is in English (system content convention) but
+    instructs the LLM to produce the narrative in world.language.
+    """
     defn = world.definition or {}
     player_state = world.state.get("player", {}) if world.state else {}
     system_prompt = get_prompt("step_writer", world.language).format(
@@ -216,6 +243,7 @@ async def build_step_writer_messages(
         player_state=json.dumps(player_state, ensure_ascii=False)[:400],
         outcome=outcome,
         narrative_prompt=narrative_prompt[:600],
+        world_language=world.language or "en",
     )
     return [{"role": "system", "content": system_prompt}]