data/llm/__init__.py

"""LLM 集成模块。

支持 openai/dashscope/anthropic 三种供应商，通过统一的 OpenAI-compatible API 调用。
"""

from .client import LLMClient as LLMClient
from .client import LLMError as LLMError
from .client import LLMResponse as LLMResponse
from .prompts import (
    build_audit_prompt as build_audit_prompt,
    build_cwb_prompt as build_cwb_prompt,
    build_full_plan_prompt as build_full_plan_prompt,
)

__all__ = [
    "LLMClient",
    "LLMResponse",
    "LLMError",
    "build_audit_prompt",
    "build_cwb_prompt",
    "build_full_plan_prompt",
]
-												feat(llm): 内置 LLM 供应商配置与审核主链接入骨架

配置层:
- Settings 新增 llm_provider/api_key/base_url/model/timeout/max_tokens
- 生产 fail-closed: GAOKAO_LLM_PROVIDER=none 禁止, provider!=none 且 API key 为空禁止
- .env.docker.example / .env.payment.example 补 LLM 变量
- payment_readiness_doctor 将 LLM 配置纳入 readiness 检查

基础设施:
- 新增 data/llm/client.py: OpenAI-compatible LLMClient
- 新增 data/llm/prompts.py: audit/cwb/full_plan prompt 模板
- 新增 data/llm/tests/test_llm.py: 12 个单元测试

主链接入:
- ReviewResultContract 新增 llm_generated / llm_summary / llm_cwb_suggestions
- _start_review_result 优先尝试 LLM 生成审核结果, 失败时回退到原规则默认逻辑
- cwb 页面优先展示 LLM 生成的三档建议

测试适配:
- conftest / health / app / p2_4 tests 注入默认 LLM 测试配置，避免被新 fail-closed 提前拦截

验证:
- data/llm/tests 12 passed
- 核心 prod settings tests 62 passed

											
										
										
											2026-06-28 12:11:43 +08:00
+								"""LLM 集成模块。
 								支持 openai/dashscope/anthropic 三种供应商，通过统一的 OpenAI-compatible API 调用。
 								"""
-												chore: fix ruff F401 re-export + mypy type errors + smoke e2e path handling

- data/llm/__init__.py, tests/__init__.py: 改用显式 `X as X` re-export 满足 ruff F401
- data/llm/prompts.py: 移除未使用的 json import
- admin/routes/web_public.py: 修复 LLM 集成新增代码的 5 处 mypy 类型错误
  - escape() 参数显式 str() 转换
  - int() 参数显式 str() 转换避免 AnyStr 报错
  - risk_level 字面量类型注解
- scripts/score_range_fullchain_100_e2e.py: 修复 batch_json/csv 绝对路径导致 relative_to() 崩溃

验证: GAOKAO_SKIP_INSTALL=1 bash scripts/dev-verify.sh
=> 1330 passed, 3 skipped, coverage 90.44%, core 100%, all checks passed

											
										
										
											2026-06-29 07:37:49 +08:00
+								from .client import LLMClient as LLMClient
 								from .client import LLMError as LLMError
 								from .client import LLMResponse as LLMResponse
-												feat(llm): 内置 LLM 供应商配置与审核主链接入骨架

配置层:
- Settings 新增 llm_provider/api_key/base_url/model/timeout/max_tokens
- 生产 fail-closed: GAOKAO_LLM_PROVIDER=none 禁止, provider!=none 且 API key 为空禁止
- .env.docker.example / .env.payment.example 补 LLM 变量
- payment_readiness_doctor 将 LLM 配置纳入 readiness 检查

基础设施:
- 新增 data/llm/client.py: OpenAI-compatible LLMClient
- 新增 data/llm/prompts.py: audit/cwb/full_plan prompt 模板
- 新增 data/llm/tests/test_llm.py: 12 个单元测试

主链接入:
- ReviewResultContract 新增 llm_generated / llm_summary / llm_cwb_suggestions
- _start_review_result 优先尝试 LLM 生成审核结果, 失败时回退到原规则默认逻辑
- cwb 页面优先展示 LLM 生成的三档建议

测试适配:
- conftest / health / app / p2_4 tests 注入默认 LLM 测试配置，避免被新 fail-closed 提前拦截

验证:
- data/llm/tests 12 passed
- 核心 prod settings tests 62 passed

											
										
										
											2026-06-28 12:11:43 +08:00
+								from .prompts import (
-												chore: fix ruff F401 re-export + mypy type errors + smoke e2e path handling

- data/llm/__init__.py, tests/__init__.py: 改用显式 `X as X` re-export 满足 ruff F401
- data/llm/prompts.py: 移除未使用的 json import
- admin/routes/web_public.py: 修复 LLM 集成新增代码的 5 处 mypy 类型错误
  - escape() 参数显式 str() 转换
  - int() 参数显式 str() 转换避免 AnyStr 报错
  - risk_level 字面量类型注解
- scripts/score_range_fullchain_100_e2e.py: 修复 batch_json/csv 绝对路径导致 relative_to() 崩溃

验证: GAOKAO_SKIP_INSTALL=1 bash scripts/dev-verify.sh
=> 1330 passed, 3 skipped, coverage 90.44%, core 100%, all checks passed

											
										
										
											2026-06-29 07:37:49 +08:00
+								    build_audit_prompt as build_audit_prompt,
 								    build_cwb_prompt as build_cwb_prompt,
 								    build_full_plan_prompt as build_full_plan_prompt,
-												feat(llm): 内置 LLM 供应商配置与审核主链接入骨架

配置层:
- Settings 新增 llm_provider/api_key/base_url/model/timeout/max_tokens
- 生产 fail-closed: GAOKAO_LLM_PROVIDER=none 禁止, provider!=none 且 API key 为空禁止
- .env.docker.example / .env.payment.example 补 LLM 变量
- payment_readiness_doctor 将 LLM 配置纳入 readiness 检查

基础设施:
- 新增 data/llm/client.py: OpenAI-compatible LLMClient
- 新增 data/llm/prompts.py: audit/cwb/full_plan prompt 模板
- 新增 data/llm/tests/test_llm.py: 12 个单元测试

主链接入:
- ReviewResultContract 新增 llm_generated / llm_summary / llm_cwb_suggestions
- _start_review_result 优先尝试 LLM 生成审核结果, 失败时回退到原规则默认逻辑
- cwb 页面优先展示 LLM 生成的三档建议

测试适配:
- conftest / health / app / p2_4 tests 注入默认 LLM 测试配置，避免被新 fail-closed 提前拦截

验证:
- data/llm/tests 12 passed
- 核心 prod settings tests 62 passed

											
										
										
											2026-06-28 12:11:43 +08:00
+								)
 								__all__ = [
 								    "LLMClient",
 								    "LLMResponse",
 								    "LLMError",
 								    "build_audit_prompt",
 								    "build_cwb_prompt",
 								    "build_full_plan_prompt",
 								]