dddcbd24e3
feat: 私董会讨论模式 + 回测集成 + WS持久化修复
...
私董会讨论模式 (Board Meeting Mode):
- BoardRouter: @board 前缀路由, 专家名验证, 模板回退
- BoardTeam: 讨论容器, 状态机 (FORMING->DISCUSSING->CONCLUDING->COMPLETED)
- BoardOrchestrator: 多轮自主循环讨论引擎, 主持人小结, 停止命令检测
- 9个预设名人专家 YAML (马斯克/贝佐斯/张小龙/芒格等)
- 前端 BoardStatusView 群聊式 UI + WebSocket 事件处理
- 后端 chat.py 集成 @board 路由到主聊天流程
回测集成:
- benchmark.py: 新增 board_meeting 维度 (18 tasks, 6 categories)
- benchmark_dataset.py: 新增 BOARD_BENCHMARKS (11 E2E cases)
- test_board_backtest.py: 66 个回测测试 (9 test classes)
Bug 修复:
- resolve_expert_configs: deep-copy 防止 is_lead 修改污染共享模板
- 所有专家名无效时回退到默认模板
- board_router: 非匹配路径 topic 未 strip
- benchmark_dataset: board-name-invalid-001 输入修正
WebSocket 持久化修复:
- chat.py: 三层防御机制确保任务结果不丢失
- chat store: 断线恢复逻辑
部署配置:
- Gitea Actions CI/CD workflow
- docker-compose.deploy.yaml 部署编排
- scripts/deploy.sh 自动化部署脚本
测试结果: 120 单元测试通过, 71 benchmark 测试 100% 通过, ruff 全部通过
2026-06-17 23:52:53 +08:00
1fbfd9d132
refactor: standardize benchmark with industry methodology (P/R/F1, multi-run, baseline)
2026-06-17 12:01:34 +08:00
d00995504d
feat: comprehensive capability benchmark and agentkit benchmark CLI
2026-06-17 11:28:09 +08:00
5374bc8501
refactor: eliminate routing layer, align with industry best practices
...
Phase 1 of architecture optimization (U1/U2/U4/U8):
- U1: Rename SimpleRouter to RequestPreprocessor, route() to preprocess()
Eliminates misleading routing concept; LLM decides autonomously
in REACT agent loop (matches Codex/Claude Code/Trae pattern)
- U2: Delete CostAwareRouter, HeuristicClassifier, SemanticRouter
(~700 lines removed). skill_routing.py: 1688 to 220 lines
- U4: PlanExecEngine defaults to ReActStepExecutor, delete _LLMStepExecutor
(pure LLM calls without tools = no execution capability)
- U8: ReActEngine defaults to ContextCompressor(keep_recent=10)
Supersedes plans 2026-06-15-002/003/004.
New plan: 2026-06-16-006-refactor-architecture-optimization-evolution-plan.md
2026-06-17 10:44:40 +08:00
c4257591d4
refactor(router): replace CostAwareRouter with SimpleRouter and prompt-based tool calling
2026-06-16 03:31:05 +08:00
11e2009cb8
feat(router): improve colloquial/mixed-lang routing, fix low-complexity IntentRouter bypass
...
Key improvements:
- Low-complexity queries (<0.3) now try IntentRouter keyword match
before falling back to DIRECT_CHAT, fixing 0% F1 on keyword_match
- SemanticRouter similarity_low lowered from 0.6 to 0.4
- Short text (<20 chars) uses effective_low = max(0.25, low - 0.15)
- Short text with no semantic match forces LLM classify fallback
- Added colloquial keywords to 7 skill YAMLs
- Fixed code_reviewer.yaml output_schema placement
- Fixed SemanticRouter build in e2e tests
- Fixed base_url detection for bailian-coding API keys
Results: keyword_match F1 0->60.87%, colloquial F1 0->100%, mixed_lang F1 0->100%
2026-06-15 23:54:57 +08:00
fa2a6dece2
feat(router): enable SemanticRouter + upgrade benchmark to L3/L5
...
- Enable SemanticRouter in agentkit.yaml (router.semantic.enabled: true)
- Integrate SemanticRouter into e2e backtest (_build_real_components)
- Add 8 new semantic test cases: 5 colloquial + 3 mixed-lang expressions
- Add L3 output quality evaluation framework (LLM-as-Judge, 1-5 score)
- Add L5 adaptive capability metrics (consistency rate from overfitting data)
- Add OutputQualityObservation model and evaluate_output_quality() method
- Report now includes L3 and L5 sections
Results: 52 tests pass, description_match F1=66.67%, L5 adaptive rate=100%
2026-06-15 23:02:47 +08:00
e984b4c462
feat(router): optimize routing intelligence — ExecutionMode expansion, multi-candidate scoring, quality gate skill match
...
- Expand ExecutionMode enum with REWOO/REFLEXION/PLAN_EXEC
- Add _resolve_execution_mode() to respect skill.config.execution_mode
- Rewrite IntentRouter._match_keywords() for multi-candidate scoring
- Add QualityGate 5th dimension: skill_match validation with warning escalation
- Calibrate HeuristicClassifier: low-complexity signals only when no high signals
- Fix negation regex for Chinese text (avoid matching past punctuation)
- Fix backtest mode_map normalization and .env loading
- Add 61 unit tests (21 HeuristicClassifier + 14 IntentRouter + 13 QualityGate + 13 existing)
Results: execution_mode_accuracy 9.09%→36.36%, skill_routing_F1 66.67%→77.78%
2026-06-15 22:43:13 +08:00
chiguyong