++

backend/app/api/agents.py

@@ -1,8 +1,32 @@
 from __future__ import annotations
+import re
 from fastapi import APIRouter, Request, HTTPException
 from pydantic import BaseModel
 
 from app.services.llm import LLMClient
+from app.services.i18n import (
+    locale_from_request,
+    translate,
+    language_directive as _language_directive,
+    DEFAULT_LOCALE,
+)
+
+_THINK_BLOCK = re.compile(r"<think>.*?</think>", re.DOTALL)
+
+
+def _strip_thinking(text: str, locale: str) -> str:
+    """Strip LLM thinking blocks. The actual response content is what we
+    return to the client; the model's internal reasoning is never shown.
+    Falls back to a localised canned line when nothing is left after
+    stripping (rare, but happens when the model burns all tokens on thinking).
+    """
+    cleaned = _THINK_BLOCK.sub("", text or "").strip()
+    if not cleaned:
+        return {
+            "en": "(no response)",
+            "zh": "(没有回应)",
+        }.get(locale, "(no response)")
+    return cleaned
 
 router = APIRouter(prefix="/api")
 
@@ -21,7 +45,7 @@ async def list_agents(sim_id: str, request: Request):
 async def get_agent(sim_id: str, agent_id: int, request: Request):
     agent = await request.app.state.store.get_agent(sim_id, agent_id)
     if not agent:
-        raise HTTPException(404, "Agent not found")
+        raise HTTPException(404, translate("agent_not_found", locale_from_request(request)))
     return agent.model_dump()
 
 
@@ -29,15 +53,18 @@ async def get_agent(sim_id: str, agent_id: int, request: Request):
 async def interview(sim_id: str, agent_id: int, req: InterviewRequest, request: Request):
     store = request.app.state.store
     llm: LLMClient = request.app.state.llm
+    locale = locale_from_request(request)
 
     agent = await store.get_agent(sim_id, agent_id)
     if not agent:
-        raise HTTPException(404, "Agent not found")
+        raise HTTPException(404, translate("agent_not_found", locale))
 
     world_state = await store.get_world_state(sim_id)
     world_name = world_state.blueprint.name if world_state else "the settlement"
 
+    head, tail = _language_directive(locale)
     system = (
+        f"{head}\n\n"
         f"You are {agent.name}, a {agent.age}-year-old {agent.role} in {world_name}.\n"
         f"Your personality: honesty={agent.personality.honesty:.1f}, "
         f"empathy={agent.personality.empathy:.1f}, "
@@ -47,7 +74,8 @@ async def interview(sim_id: str, agent_id: int, req: InterviewRequest, request:
         f"Your emotional state: {agent.emotional_state}\n\n"
         "Someone approaches and asks you a question. Respond in character. "
         "Be authentic to your personality. Reference specific events from your memory. "
-        "Keep it to 2-3 sentences."
+        "Keep it to 2-3 sentences.\n\n"
+        f"{tail}"
     )
 
     if agent.life_state:
@@ -81,5 +109,9 @@ async def interview(sim_id: str, agent_id: int, req: InterviewRequest, request:
             system += "\n\nYOUR LIFE SITUATION:\n" + "\n".join(life_lines)
             system += "\n\nWhen answering, let your life situation color your responses naturally. Reference family, pressures, or your past when relevant."
 
-    response = await llm.generate(system=system, user=req.question, max_tokens=200)
-    return {"response": response, "emotional_state": agent.emotional_state}
+    response = await llm.generate(system=system, user=req.question, max_tokens=400)
+    cleaned = _strip_thinking(response, locale)
+    return {"response": cleaned, "emotional_state": agent.emotional_state}
+
+
+# moved up to module top so it can be imported by tests / shared utilities