feat(api): 统一 LLM JSON 调用层 llm_json_call，按域 Schema 迁移 chat/memoir agents

api/app/agents/memoir/batch_phase1_prep.py

@@ -4,30 +4,22 @@ Phase1 批处理：一次 LLM 调用完成多段的抽取 + 章节分类（与
 
 from __future__ import annotations
 
-import json
 from dataclasses import dataclass
 from typing import Any, Dict, List
 
 from app.agents.memoir.prompts import get_batch_memoir_phase1_prep_prompt
+from app.agents.memoir.schemas import BatchPhase1LLMOutput
 from app.agents.state_schema import MemoirStateSchema
+from app.agents.stage_constants import STAGE_SLOT_KEYS
 from app.core.config import settings
-from app.core.json_utils import extract_json_payload
-from app.core.langchain_llm import invoke_json_object
+from app.core.llm_call import LLMCallError, llm_json_call
 from app.core.logging import get_logger
 from app.features.conversation.models import Segment
 
 logger = get_logger(__name__)
 
 STAGE_ALLOWED_SLOTS: Dict[str, frozenset[str]] = {
-    "childhood": frozenset(
-        {"place", "people", "daily_life", "emotion", "turning_event"}
-    ),
-    "education": frozenset({"school", "city", "motivation", "challenge", "change"}),
-    "career": frozenset({"job", "environment", "decision", "pressure", "growth"}),
-    "family": frozenset(
-        {"relationship", "conflict", "support", "responsibility", "change"}
-    ),
-    "belief": frozenset({"value", "regret", "pride", "lesson"}),
+    k: frozenset(v) for k, v in STAGE_SLOT_KEYS.items()
 }
 
 
@@ -73,32 +65,35 @@ def run_batch_phase1_prep(
         slots_snapshot=_slots_snapshot(state),
         segment_items=items,
     )
-    raw = invoke_json_object(
-        llm,
-        prompt,
-        max_tokens=int(settings.memoir_phase1_batch_llm_max_tokens),
-        agent="BatchPhase1Prep.run",
-    )
-    parsed = json.loads(extract_json_payload(raw))
-    rows = parsed.get("segments") or []
-    if not isinstance(rows, list):
-        raise ValueError("batch phase1: segments must be a list")
+    try:
+        parsed = llm_json_call(
+            llm,
+            prompt,
+            BatchPhase1LLMOutput,
+            max_tokens=int(settings.memoir_phase1_batch_llm_max_tokens),
+            agent="BatchPhase1Prep.run",
+        )
+    except LLMCallError as e:
+        logger.warning("batch phase1 LLM 解析失败: {}", e)
+        raise ValueError("batch phase1: llm parse failed") from e
+
+    rows = parsed.segments
+    if not rows:
+        raise ValueError("batch phase1: segments must be a non-empty list")
 
     by_id: Dict[str, BatchPhase1SegmentRow] = {}
     for row in rows:
-        if not isinstance(row, dict):
-            continue
-        sid = str(row.get("id", "")).strip()
+        sid = str(row.id).strip()
         if not sid:
             continue
-        ds = str(row.get("detected_stage", "") or "").strip().lower()
-        slots_raw = row.get("slots") or {}
-        slots: Dict[str, str] = {}
-        if isinstance(slots_raw, dict):
-            for k, v in slots_raw.items():
-                if k and isinstance(k, str):
-                    slots[k] = v if isinstance(v, str) else str(v)
-        cat_raw = str(row.get("chapter_category", row.get("category", "")) or "")
+        ds = str(row.detected_stage or "").strip().lower()
+        slots_raw = row.slots or {}
+        slots = {
+            k: v if isinstance(v, str) else str(v)
+            for k, v in slots_raw.items()
+            if k and isinstance(k, str)
+        }
+        cat_raw = str(row.chapter_category or "")
         by_id[sid] = BatchPhase1SegmentRow(
             detected_stage=ds or (state.current_stage or "childhood"),
             slots=slots,