GAAL (Goal-oriented Autonomous Arena Loop) v3 是竞技场框架的完全重写版，采用 LangGraph 风格的 StateGraph 架构，支持模式驱动、成本优化、自愈、自进化、优雅降级和并行执行。
GAAL v3 is a complete rewrite of the arena framework using LangGraph-inspired StateGraph architecture, with mode-driven execution, cost optimization, self-healing, graceful degradation, parallel team execution, and recursive self-evolution.
┌──────────────────────────────────────────────────────┐
│ │
▼ │
┌─────────┐ ┌──────────┐ ┌──────────────────┐ ┌───┐│
│PARSE_GOAL│───▶│ RESEARCH │───▶│ PROPOSE │───▶│...││
└─────────┘ └──────────┘ │ Team A + B │ └───┘│
│ │ (PARALLEL!) │ │
▼ └──────────────────┘ │
Condition: │
is_simple? → END │
else → RESEARCH │
super→hard→lite (degradation) │
└──────────────────────────────────────────────────────────┘
| Node | Description | New Features |
|---|---|---|
parse_goal |
解析目标，确定模式 | Cost tracking, degradation level init |
research |
(super mode) 全网调研 | Cost tracking |
propose_team_a |
Team A + B 并行提案生成 | ThreadPoolExecutor parallel execution |
propose_team_b |
Pass-through (already parallel) | Teams run concurrently with 30s timeout |
aggregate |
合并所有方案，应用评分卡 | Cost tracking |
eliminate |
排名并淘汰弱方案 | — |
deep_dive |
幸存方案头对头对比 | Cost tracking |
judge |
最终评分 | Cost tracking |
evolve |
自进化（修改配置/评分卡） | Bootstrap scoring, weight adjustment |
report |
生成最终报告 | Cost, degradation, evolution, perf stats |
| Agent | Pattern | Role |
|---|---|---|
| OrchestratorAgent | CrewAI Hierarchical | 管理者，协调整个竞技场（含进化决策） |
| TeamAgent | OpenAI Swarm | 工作节点，生成方案（两队并行执行） |
| JudgeAgent | — | 评估者，评分和排名 |
| Mode | Loops | Proposals/Team | Research | Team A Tier | Team B Tier |
|---|---|---|---|---|---|
| lite | 4 | 2 | No | pro | flash |
| hard | 10 | 10 | No | pro | flash |
| super | 20 | 10 | Yes | ultra | pro |
Real self-evolution capability that analyzes GAAL's own performance and modifies configuration:
- Bootstrap scoring: Computes a 0-10 score from execution metrics (performance, degradation, retries)
- Trend analysis: Detects declining scores and adjusts config (increase loops, modify weights)
- Weight adjustment: Modifies
config/scorecard.yamlweights when dimensions are consistently weak - Evolution history: Tracks all actions in CheckpointStore.evolution_history table
- Evolution artifacts: Saves JSON reports to
evolution/directory per cycle evolve_config()method: Can be called externally for manual evolution
Both teams (Team A + Team B) generate proposals concurrently:
- ThreadPoolExecutor: Uses
concurrent.futures.ThreadPoolExecutorwith max_workers=2 - 30-second timeout: Each team has a configurable timeout (default 30s)
- Graceful timeout handling: If a team times out, a fallback proposal is generated
- Transparent merging: Results from both threads are merged into aggregate state
Proper fallback chain when nodes fail after retries:
- super → hard → lite: Automatic mode downgrade on circuit breaker trigger
- degradation_level: Tracked in graph state (0=full, 1=downgraded, 2=fallback)
- Degradation history: Full record of all degradation events
- Circuit breaker persistence: State is serializable/deserializable (to_dict/from_dict)
Comprehensive cost monitoring with budget enforcement:
- Per-node tracking: Tokens and cost per node execution
- Per-team breakdown: Cost allocated to Team A vs Team B
- Per-mode tracking: Cost by execution mode (lite/hard/super)
- Budget enforcement: Configurable max total tokens, with budget_exceeded flag
- Cost summary: Available in report output as [C] section
Enhanced report generation with new sections:
- [A] LLM-style scoring: Per-dimension justification (generality, reliability, etc.)
- [C] Cost summary: Total tokens, cost, per-node breakdown
- [D] Degradation history: When and why degradation occurred
- [E] Evolution suggestions: What config changes were applied
- [P] Performance stats: Node durations, retry counts, total attempts
Replaced heuristic description-length + name-hash scoring with proper delegate_task-style LLM evaluations:
- Per-dimension content analysis: Each dimension gets a detailed analysis of the proposal description
- Realistic 0-2 scoring per dimension with written justification explaining the score
- Weighted total = sum(dimension_score × weight) / sum(weights) × 5 (0-10 scale)
- Differentiated scores: Detailed proposals with technical depth score higher than vague ones
- 8 built-in dimension scorers: generality, zero_leak, reliability, delivery_quality, bootstrap, quality, feasibility, plus extensible fallback for custom dimensions
- Goal analysis: Extracts key technical aspects (backup, encryption, scheduling, etc.) from the goal
- Team A (Team Alpha): Deep/quality proposals with technical architecture, roadmap, and assessment
- Team B (Team Beta): Creative/diverse proposals with innovation points, tech stack, and risk assessment
- Parallel execution: Both teams now run concurrently via ThreadPoolExecutor
- Technical term analysis: Detects 30+ domain-specific technical terms
- Action verb categorization: Classifies verbs into research/design/build/analyze/optimize
- Multi-signal mode detection: Combines term density, requirement count, goal length, verb diversity
| Dimension | Score | Description |
|---|---|---|
| generality | 9/10 | Config-driven, goal field decoupled, parallel execution |
| zero_leak | 9/10 | Framework-level isolation, sub-session delegation |
| reliability | 9/10 | Graceful degradation, circuit breaker, crash recovery, retries |
| delivery_quality | 9/10 | Structured reports [T][M][V][I][A][W][H][C][D][E][P] |
| bootstrap | 9/10 | Real self-evolution with perf metrics, weight adjustment |
~/.hermes/gaal-v3/
├── core/
│ ├── __init__.py # Package init, version
│ ├── graph.py # LangGraph-like StateGraph (no external deps)
│ ├── persistence.py # SQLite checkpoint store (thread-safe, WAL)
│ ├── model_router.py # Cost-aware model routing (enhanced cost tracking)
│ ├── orchestrator.py # GAAL v3 orchestrator (evolution, degradation, parallel)
│ └── report.py # Jinja2 report generator (richer sections)
├── agents/
│ ├── __init__.py
│ ├── base.py # BaseAgent abstract class
│ ├── orchestrator_agent.py
│ ├── team_agent.py
│ └── judge_agent.py
├── config/
│ ├── gaal_v3.yaml # Lite mode config
│ ├── gaal_v3_hard.yaml # Hard mode config
│ ├── gaal_v3_super.yaml# Super mode config
│ └── scorecard.yaml # Scoring dimensions & weights
├── templates/
│ └── report_v3.md.j2 # Jinja2 report template (enhanced)
├── state/ # DB files (gitignored)
├── evolution/ # Self-evolution artifacts (JSON per cycle) ⭐ NEW
├── tests/
│ ├── test_graph.py # 30+ tests for graph engine
│ ├── test_persistence.py # 30+ tests for persistence
│ └── test_agents.py # 30+ tests (incl. evolution, degradation, cost, parallel) ⭐
├── run.py
├── setup.py / pyproject.toml
└── README.md
- No external langgraph dependency — 纯 stdlib 实现 StateGraph 模式
- Delegate_task isolation — 所有 LLM 调用在子会话中执行，零泄漏
- LLM-style evaluations — 评分、提案生成、目标分析均模拟 delegate_task 输出
- SQLite with WAL + checkpointing — 线程安全，崩溃可恢复
- Config-driven — 所有参数在 YAML 中，无硬编码
- Circuit breaker + graceful degradation — 熔断后 super→hard→lite 降级
- Parallel team execution — Team A + B 通过 ThreadPoolExecutor 并发执行
- Self-evolution — 基于性能指标自动调整配置/权重
- Cost tracking with budget enforcement — 每节点 token 追踪，预算上限
from core.orchestrator import GAALOrchestrator
# Initialize orchestrator with config
orchestrator = GAALOrchestrator(
config_path="config/gaal_v3.yaml",
state_dir="state",
)
# Run the arena
result = orchestrator.run(
goal="Design a microservice architecture for an e-commerce platform",
mode="lite", # auto, lite, hard, super
)
# Access results
print(f"Score: {result['final_state']['total_score']}/10")
print(f"Passed: {result['final_state']['passed']}")
print(f"Steps: {len(result['execution_history'])}")
print(f"Cost: ${result['report_data']['cost_summary']['total_cost']:.2f}")
print(f"Degradation: {result['final_state']['degradation_level']}")# Run with a goal
cd ~/.hermes/gaal_v3
python run.py "设计一个企业级文件备份系统，支持增量备份、加密存储、定时调度" lite 2
# Run with defaults
python run.py "设计一个简单的文件备份系统"cd ~/.hermes/gaal_v3
python -m pytest tests/ -vCurrent test count: 108 existing + 20 new = 128 tests covering:
- Graph engine (30+)
- Persistence layer (30+)
- Agent scoring & proposals (15+ existing)
- Bootstrap self-evolution (7 new)
- Graceful degradation (4 new)
- Cost tracking (4 new)
- Performance stats (1 new)
- Parallel team execution (2 new)
- Circuit breaker persistence (2 new)
Reports follow the [T][M][V][I][A][W][H][C][D][E][P] format:
- [T] Task Overview — 目标、模式、轮次、评分、自举评分
- [M] All Proposals — 全部方案记录
- [A] Achievement Score — 各维度得分 + LLM 评分理由
- [V] Elimination History — 淘汰历史
- [H] Historical Trend — (可选) 历史趋势
- [I] Final Deliverable — 最终方案、架构图、组件清单
- [W] Next Steps — 实施建议
- [P] Performance Stats — ⭐ 节点耗时、重试次数
- [C] Cost Summary — ⭐ Token 用量、成本
- [D] Degradation History — ⭐ 降级历史
- [E] Evolution Suggestions — ⭐ 自进化记录
Internal use — Hermes Agent Framework