重构 改为标准化输出

backend/app/settings.py

@@ -29,6 +29,7 @@ class ProviderConfig:
     # Chat Completions uses the legacy flat ``reasoning_effort`` parameter.
     # Keep it provider-scoped because compatibility varies by endpoint/model.
     reasoning_effort: str = ""
+    api_style: str = "chat_completions"
 
     @property
     def configured(self) -> bool:
@@ -43,6 +44,14 @@ class ProviderConfig:
             return {"reasoning_effort": self.reasoning_effort}
         return {}
 
+    @property
+    def request_options(self) -> dict[str, object]:
+        if not self.reasoning_effort:
+            return {}
+        if self.api_style == "responses":
+            return {"reasoning": {"effort": self.reasoning_effort}}
+        return {"reasoning_effort": self.reasoning_effort}
+
 
 @dataclass(frozen=True)
 class Settings:
@@ -59,14 +68,10 @@ class Settings:
     review_provider_id: str = ""
     review_model_id: str = ""
     agent_tool_calls_per_cycle: int = 12
-    agent_candidate_attempts_per_head: int = 3
     agent_consecutive_no_progress_limit: int = 6
     agent_format_error_repeat_limit: int = 3
-    agent_max_features_per_fragment: int = 6
     agent_context_char_limit: int = 14000
     agent_render_cache: bool = True
-    modeling_plan_enabled: bool = True
-    modeling_plan_max_revisions: int = 2
     autonomous_generation: bool = True
     resume_running_tasks_on_startup: bool = True
 
@@ -130,6 +135,13 @@ def _reasoning_effort(value: str) -> str:
     return effort
 
 
+def _api_style(value: str) -> str:
+    style = value.strip().lower() or "chat_completions"
+    if style not in {"chat_completions", "responses"}:
+        raise ValueError("CDSL_*_API_STYLE must be 'chat_completions' or 'responses'")
+    return style
+
+
 def _env_flag(name: str, default: bool) -> bool:
     value = os.getenv(name)
     if value is None or not value.strip():
@@ -167,6 +179,7 @@ def _provider(prefix: str, provider_id: str, label: str, default_base_url: str,
         api_key=api_key,
         models=_models(model_list, vision_models),
         reasoning_effort=_reasoning_effort(os.getenv(f"CDSL_{prefix}_REASONING_EFFORT", "")),
+        api_style=_api_style(os.getenv(f"CDSL_{prefix}_API_STYLE", "responses" if provider_id == "openai" else "chat_completions")),
     )
 
 
@@ -195,14 +208,10 @@ def get_settings() -> Settings:
         review_provider_id=os.getenv("CDSL_REVIEW_PROVIDER", "").strip().lower(),
         review_model_id=os.getenv("CDSL_REVIEW_MODEL", "").strip(),
         agent_tool_calls_per_cycle=max(1, int(os.getenv("CDSL_AGENT_TOOL_CALLS_PER_CYCLE", "12"))),
-        agent_candidate_attempts_per_head=max(1, int(os.getenv("CDSL_AGENT_CANDIDATE_ATTEMPTS_PER_HEAD", "3"))),
         agent_consecutive_no_progress_limit=max(1, int(os.getenv("CDSL_AGENT_CONSECUTIVE_NO_PROGRESS_LIMIT", "6"))),
         agent_format_error_repeat_limit=max(1, int(os.getenv("CDSL_AGENT_FORMAT_ERROR_REPEAT_LIMIT", "3"))),
-        agent_max_features_per_fragment=max(1, min(6, int(os.getenv("CDSL_AGENT_MAX_FEATURES_PER_FRAGMENT", "6")))),
         agent_context_char_limit=max(4000, int(os.getenv("CDSL_AGENT_CONTEXT_CHAR_LIMIT", "14000"))),
         agent_render_cache=_env_flag("CDSL_AGENT_RENDER_CACHE", True),
-        modeling_plan_enabled=_env_flag("CDSL_MODELING_PLAN_ENABLED", True),
-        modeling_plan_max_revisions=max(1, int(os.getenv("CDSL_MODELING_PLAN_MAX_REVISIONS", "2"))),
         autonomous_generation=True,
         # Production instances recover durable runs by default. Test workers
         # can disable this before startup to guarantee they touch only tasks