  BrunoGaoSZandClaude Sonnet 4.6
|
2360e788d3
|
feat: implement P10 — SOP profiles, integration tests, smoke gate
TASK 1 — Profile Schema (core/profile/schema.py)
已存在：EvidenceThresholds / ModelRouting / WorkflowStep / ApprovalRule /
Permissions / CachePolicy / Profile，全部 frozen=True，含 tool-reference 验证器。
TASK 2 — 三个行业 SOP Profile YAML
- profiles/presales/profile.yaml 企业预售方案生成
- profiles/security_analysis/profile.yaml 安全事件分析（host_isolation 需审批）
- profiles/knowledge_qa/profile.yaml 知识问答（严格模式，阈值全 0.90）
TASK 3 — loader / registry / 测试
- core/profile/loader.py + registry.py（线程安全单例 + 热更新）
- tests/unit/test_profiles.py（19 个用例）
- tests/unit/test_scorer.py（12 个用例，覆盖 EvidenceScorer 全路径）
TASK 4 — 端到端集成测试 + 成本追踪
- tests/integration/test_full_pipeline.py 7 个 mock-based 测试
- tests/integration/test_cost_tracking.py 3 个成本测试
- scripts/smoke_test.sh 更新为与最终门禁一致的 ruff 规则
Bug fixes（使全部测试通过）：
- LangGraph: ask_node 显式初始化 error/pending_approval，确保 ainvoke()
结果 dict 中始终包含这两个字段
- GraphInterrupt 不再被 act_node 的 except Exception 吞掉
- Citation.model_dump(mode="json") 修复 Decimal JSON 序列化问题
- act_node 直接调用 audit_logger.write_tool_call（集成测试 mock executor 时仍可审计）
- 修复 mypy 错误：config.py 重复字段、encryptor __init__ 返回值、
registry 属性声明、sanitizer.py 类型注解
- 修复 47 处 ruff lint（F401/F841/E501/I001）
- 添加 pytest-timeout>=2.3.0 依赖
最终门禁全部通过：
ruff E,F,I 0 errors | mypy 0 errors
143 unit tests, 80.38% coverage | 11 integration tests
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
|
2026-02-27 09:21:43 +01:00 |
|
 brunoandClaude Sonnet 4.6
|
87443535ff
|
feat: implement core platform layers (P5-P8)
Implement complete agent platform with four major layers:
P5 - Engine Layer (LangGraph):
- AgentState with JSON serialization validation
- 5 async nodes (ask, plan, act, verify, report) with error boundaries
- Template resolution system for {{user_input}} and {{steps.xxx}}
- StateGraph with conditional edges and interrupt_after for approval flow
- 17/17 tests passing, 90% coverage
P6 - LLM Layer:
- ClaudeClient with Decimal-based cost calculation (ROUND_HALF_UP)
- Custom exceptions (ClaudeAPIError, ClaudeTimeoutError, ClaudeRateLimitError)
- PromptBuilder with cache_control support
- ModelRouter with sensitivity-based routing (local vs Claude)
- 18/18 tests passing, 90% coverage
P7 - Audit Layer (Three-tier):
- Sanitizer: mask phone/ID card/API keys, contains_sensitive, mask_dict
- Encryptor: AES-256-GCM with keyring support, 12-byte IV
- AuditLogger: L1 (DB), L2 (OTel), L3 (Encrypted payloads)
- 24/24 tests passing, 86% coverage
P8 - API Layer (FastAPI):
- Lifespan management with 8-step startup sequence
- Middleware: TracingMiddleware, RateLimitMiddleware (Redis), CORS
- Auth: JWT decode with python-jose, RBAC with permission matrix
- Routes: health, chat (SSE), approve, trace
- 13/13 tests passing, 91% coverage
All gate checks passing:
- pytest with 85%+ coverage
- mypy type checking
- ruff style checking
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
|
2026-02-26 22:00:23 +08:00 |
|