refactor: update max_tokens handling and enhance structured JSON generation

- Modified LLMControlService to dynamically set max_tokens based on input or default value.
- Introduced StructuredJSONGenerateResult class to encapsulate results and error handling in structured JSON generation.
- Updated TensionScoringService to utilize the new structured JSON generation method, improving error reporting and handling.
- Increased max_tokens limit across various configurations and services to accommodate larger outputs.
- Added unit tests for new functionalities and ensured existing tests validate the updated behavior.

Author: shenminglinyi

application/ai/llm_control_service.py

@@ -150,7 +150,7 @@ def _row_to_profile(self, row: dict) -> LLMProfile:
             api_key=row['api_key'] or '',
             model=row['model'] or '',
             temperature=row['temperature'],
-            max_tokens=DEFAULT_MAX_OUTPUT_TOKENS,
+            max_tokens=int(row.get('max_tokens') or DEFAULT_MAX_OUTPUT_TOKENS),
             timeout_seconds=row['timeout_seconds'],
             extra_headers=json.loads(row.get('extra_headers') or '{}'),
             extra_query=json.loads(row.get('extra_query') or '{}'),

application/ai/structured_json_pipeline.py

@@ -7,7 +7,8 @@
 import asyncio
 import json
 import logging
-from typing import List, Optional, Tuple, Type, TypeVar
+from dataclasses import dataclass
+from typing import Generic, List, Optional, Tuple, Type, TypeVar
 
 from json_repair import repair_json
 from pydantic import BaseModel, ValidationError
@@ -27,6 +28,17 @@
 
 T = TypeVar("T", bound=BaseModel)
 
+
+@dataclass(frozen=True)
+class StructuredJSONGenerateResult(Generic[T]):
+    payload: Optional[T]
+    failure_stage: str = ""
+    errors: Tuple[str, ...] = ()
+
+    @property
+    def ok(self) -> bool:
+        return self.payload is not None
+
 # 校验或解析失败后的额外重试轮数；总次数 = 1 + 该值，且不超过全局上限。
 DEFAULT_MAX_RETRIES = LLM_MAX_TOTAL_ATTEMPTS - 1
 
@@ -118,6 +130,25 @@ async def structured_json_generate(
     max_retries: int = DEFAULT_MAX_RETRIES,
 ) -> Optional[T]:
     """调用 LLM 获取结构化 JSON 输出，并经过清洗、修复、校验管线。"""
+    result = await structured_json_generate_with_status(
+        llm=llm,
+        prompt=prompt,
+        config=config,
+        schema_model=schema_model,
+        max_retries=max_retries,
+    )
+    return result.payload
+
+
+async def structured_json_generate_with_status(
+    llm: LLMService,
+    prompt: Prompt,
+    config: GenerationConfig,
+    schema_model: Type[T],
+    *,
+    max_retries: int = DEFAULT_MAX_RETRIES,
+) -> StructuredJSONGenerateResult[T]:
+    """Like structured_json_generate, but preserves failure category."""
     current_prompt = prompt
     last_errors: List[str] = []
     total_attempts = min(1 + max(0, max_retries), LLM_MAX_TOTAL_ATTEMPTS)
@@ -151,7 +182,11 @@ async def structured_json_generate(
                 )
                 await asyncio.sleep(delay)
                 continue
-            return None
+            return StructuredJSONGenerateResult(
+                payload=None,
+                failure_stage="llm_generate",
+                errors=tuple(last_errors),
+            )
 
         cleaned = sanitize_llm_output(raw)
         data, parse_errors = parse_and_repair_json(cleaned)
@@ -187,7 +222,7 @@ async def structured_json_generate(
                 )
                 if attempt > 0:
                     logger.info("结构化 JSON 管线重试成功 (attempt=%d)", attempt)
-                return instance
+                return StructuredJSONGenerateResult(payload=instance)
             last_errors = parse_errors + schema_errors
             recorder.record_span(
                 phase="schema_validated",
@@ -249,4 +284,8 @@ async def structured_json_generate(
             "total_attempts": total_attempts,
         },
     )
-    return None
+    return StructuredJSONGenerateResult(
+        payload=None,
+        failure_stage="parse_or_schema",
+        errors=tuple(last_errors),
+    )

application/analyst/services/tension_scoring_service.py

@@ -14,7 +14,7 @@
     tension_scoring_payload_to_domain,
     tension_scoring_response_format,
 )
-from application.ai.structured_json_pipeline import structured_json_generate
+from application.ai.structured_json_pipeline import structured_json_generate_with_status
 from infrastructure.ai.generation_profiles import generation_config_from_profile
 from infrastructure.ai.prompt_contract import PromptContract
 from infrastructure.ai.prompt_gateway import PromptGatewayError, get_prompt_gateway
@@ -82,17 +82,22 @@ async def score_chapter(
         )
 
         try:
-            payload = await structured_json_generate(
+            parsed = await structured_json_generate_with_status(
                 llm=self._llm,
                 prompt=prompt,
                 config=config,
                 schema_model=TensionScoringLlmPayload,
             )
+            payload = parsed.payload
         except Exception as e:
             logger.warning("张力评分管线异常: %s", e)
             payload = None
+            parsed = None
 
         if payload is None:
+            if parsed is not None and parsed.failure_stage == "parse_or_schema":
+                logger.warning("张力评分解析失败，使用中性 50: %s", list(parsed.errors[:4]))
+                return TensionDimensions.neutral()
             return TensionDimensions.unevaluated()
 
         dims = tension_scoring_payload_to_domain(payload)

application/engine/dag/models.py

@@ -199,7 +199,7 @@ class NodeConfig(BaseModel):
     max_retries: int = Field(default=1, ge=0, le=5)
     timeout_seconds: int = Field(default=60, ge=10, le=600)
     temperature: float = Field(default=0.7, ge=0.0, le=2.0)
-    max_tokens: Optional[int] = Field(default=None, ge=100, le=16000)
+    max_tokens: Optional[int] = Field(default=None, ge=100, le=120000)
 
 
 # ─── DAG 定义模型 ───

config/generation_profiles.yaml

@@ -1,81 +1,81 @@
 base_generation_profile:
   provider_role: writing
-  max_tokens: 16000
+  max_tokens: 120000
   temperature: 0.3
   timeout_seconds: 60
   retries: 1
 
 memory_extraction:
   extends: base_generation_profile
-  max_tokens: 16000
+  max_tokens: 120000
   temperature: 0.3
 
 voice_style_analysis:
   extends: base_generation_profile
-  max_tokens: 16000
+  max_tokens: 120000
   temperature: 0.1
 
 voice_baseline_analysis:
   extends: voice_style_analysis
-  max_tokens: 16000
+  max_tokens: 120000
   temperature: 0.1
 
 chapter_summarizer:
   extends: base_generation_profile
-  max_tokens: 16000
+  max_tokens: 120000
   temperature: 0.7
 
 tension_scoring:
   extends: base_generation_profile
-  max_tokens: 16000
+  max_tokens: 120000
   temperature: 0.3
 
 tension_diagnosis:
   extends: base_generation_profile
-  max_tokens: 16000
+  max_tokens: 120000
   temperature: 0.2
 
 review_json:
   extends: base_generation_profile
-  max_tokens: 16000
+  max_tokens: 120000
   temperature: 0.3
 
 summary_act:
   extends: base_generation_profile
-  max_tokens: 16000
+  max_tokens: 120000
   temperature: 0.4
 
 summary_volume:
   extends: base_generation_profile
-  max_tokens: 16000
+  max_tokens: 120000
   temperature: 0.5
 
 summary_part:
   extends: base_generation_profile
-  max_tokens: 16000
+  max_tokens: 120000
   temperature: 0.45
 
 summary_checkpoint:
   extends: base_generation_profile
-  max_tokens: 16000
+  max_tokens: 120000
   temperature: 0.5
 
 planning_macro:
   extends: base_generation_profile
-  max_tokens: 16000
+  max_tokens: 120000
   temperature: 0.7
 
 planning_repair:
   extends: base_generation_profile
-  max_tokens: 16000
+  max_tokens: 120000
   temperature: 0.5
 
 planning_act:
   extends: base_generation_profile
-  max_tokens: 16000
+  max_tokens: 120000
   temperature: 0.7
 
 planning_chapter_preplan:
   extends: base_generation_profile
-  max_tokens: 16000
+  max_tokens: 120000
   temperature: 0.65

domain/ai/services/llm_service.py

@@ -1,11 +1,10 @@
-# domain/ai/services/llm_service.py
 from abc import ABC, abstractmethod
 from typing import AsyncIterator, Dict, Optional
 from domain.ai.value_objects.prompt import Prompt
 from domain.ai.value_objects.token_usage import TokenUsage
 
 
-DEFAULT_MAX_OUTPUT_TOKENS = 16000
+DEFAULT_MAX_OUTPUT_TOKENS = 120000
 
 
 class GenerationConfig:
@@ -22,14 +21,14 @@ def __init__(
         self.temperature = temperature
         self.response_format = response_format
         self.__post_init__()
-        self.max_tokens = DEFAULT_MAX_OUTPUT_TOKENS
 
     def __post_init__(self):
         """验证配置参数"""
         if not (0.0 <= self.temperature <= 2.0):
             raise ValueError("Temperature must be between 0.0 and 2.0")
         if self.max_tokens <= 0:
             raise ValueError("max_tokens must be greater than 0")
+        self.max_tokens = max(int(self.max_tokens), DEFAULT_MAX_OUTPUT_TOKENS)
 
 
 class GenerationResult:

engine/pipeline/base.py

@@ -1155,8 +1155,9 @@ async def _score_tension_via_llm(self, ctx: PipelineContext) -> Optional[int]:
 
         保留给子类/实验路径显式调用；默认管线不再把它作为兜底。
         """
+        import re
+
         try:
-            import re
             from application.ai_invocation.autopilot.factory import get_or_create_autopilot_helper_invoker
             from application.ai_invocation.autopilot.helper_invoker import AutopilotHelperRequest
             from infrastructure.ai.prompt_keys import TENSION_SCORING
@@ -1177,13 +1178,14 @@ async def _score_tension_via_llm(self, ctx: PipelineContext) -> Optional[int]:
                     config={"max_tokens": 10, "temperature": 0.3},
                 )
             )
-            match = re.search(r'(\d+)', content)
-            if match:
-                score = int(match.group(1))
-                return min(score, 10) * 10  # 1-10 → 0-100
         except Exception as e:
             logger.warning("[%s] 单维张力评分失败: %s", ctx.novel_id, e)
-        return None
+            return None
+        match = re.search(r"(\d+)", str(content))
+        if not match:
+            return 50
+        score = int(match.group(1))
+        return min(score, 10) * 10  # 1-10 → 0-100
 
     def _make_result(
         self,

engine/runtime/daemon_host.py

@@ -1585,10 +1585,13 @@ async def _score_tension(self, content: str) -> int:
                     config={"max_tokens": 5, "temperature": 0.1},
                 )
             )
-            score = int(''.join(filter(str.isdigit, raw[:3])))
-            return max(1, min(10, score))
         except Exception:
             return 0
+        digits = "".join(filter(str.isdigit, str(raw)[:3]))
+        if not digits:
+            return 5
+        score = int(digits)
+        return max(1, min(10, score))
 
     async def _stream_llm_with_stop_watch(
         self,

engine/runtime/generation_token_policy.py

@@ -2,4 +2,7 @@
 from __future__ import annotations
 
 
-CHAPTER_PROSE_MAX_TOKENS = 16000
+from domain.ai.services.llm_service import DEFAULT_MAX_OUTPUT_TOKENS
+
+
+CHAPTER_PROSE_MAX_TOKENS = DEFAULT_MAX_OUTPUT_TOKENS

infrastructure/ai/config/settings.py

@@ -42,7 +42,7 @@ def __post_init__(self):
 
         if self.default_max_tokens <= 0:
             raise ValueError("Max tokens must be positive")
-        self.default_max_tokens = DEFAULT_MAX_OUTPUT_TOKENS
+        self.default_max_tokens = max(int(self.default_max_tokens), DEFAULT_MAX_OUTPUT_TOKENS)
 
         if self.timeout_seconds <= 0:
             raise ValueError("timeout_seconds must be positive")