🚀 V4.5.0: Cross-Session Continuity + Protocol-Native Skills + Action-First Reports

DevSquad V4.5.0 (merging V4.4.3 + V4.4.4 + V4.5.0 changes) delivers 10 new features for cross-session continuity, protocol-native skill architecture, and action-first reporting. The 7-role AI team orchestrates complex engineering tasks with full audit trails and consensus mechanisms. See docs/VISION.md for the project vision.

V4.5.0 — 10 New Features

• ScratchpadHistoryStore: SQLite-backed cross-session Scratchpad search
• AgentIdentity: Deterministic agent ID for cross-session tracking
• WorkflowTrace: Transparent workflow trace in dispatch reports
• GitContext: Git branch/commit context injection into dispatch
• SkillProvider Protocol: Protocol-native skill architecture (Builtin + MCP providers)
• OutputStyle: Action-first report format (from i-have-adhd insights)
• SessionResume CLI: devsquad sessions list + dispatch --resume
• FileBundler: Deterministic file bundling for review mode (from open-code-review)
• SKILL.md Modular Split: 1216→282 lines + 3 reference docs (MODULE_REFERENCE / SUB_SKILLS / VERSION_HISTORY)
• VISION Documents: docs/VISION.md + VISION_ORCHESTRATION.md + VISION_AGENT_COLLABORATION.md

V4.4.0 — P0-P3 Enhancement Modules (5 new modules)

• P0-1 RiskRegister: PMP risk management with 7-role weighted assessment (probability × impact) + 4 response strategies (avoid/transfer/mitigate/accept) + GateType.RISK_CHECK gate (exposure ≥ 0.36 blocks)
• P0-2 ViewpointRegistry: TOGAF architecture viewpoints with 7-role bound formal viewpoints + is_orthogonal() orthogonality check + check_consistency() conflict detection
• P1-1 ErrorBudgetTracker: SRE error budget with SLO 99.9% default + GateType.ERROR_BUDGET P10 gate (budget exhaustion blocks deployment) + burn_rate() consumption rate
• P1-2 GapAnalyzer: TOGAF gap analysis with analyze(current, target) + prioritize() + generate_roadmap() + suggest_scheduler_decision() driving LoopScheduler
• P2-1 DoraMetricsCollector: DORA metrics (Deployment Frequency / Lead Time / Change Failure Rate / MTTR) + GateType.DORA_CHECK P11 gate (CFR > 15% triggers architecture review) + Elite/High/Medium/Low rating

V4.4.1 — External Docs Restructure

• Archived orphan i18n docs (docs/i18n/ → docs/_archive/i18n/)
• Retired CHANGELOG-CN.md (CHANGELOG.md is now SSOT for all languages)
• Consolidated admin credentials to INSTALL.md only (single source of truth)
• Renumbered INSTALL.md methods to continuous 1-7
• Synced version numbers across all external docs (README/SKILL/INSTALL/CLAUDE)

V4.4.2 — Multilingual + Dashboard Enhancement

• Multilingual role prompts (EN/CN/JP) for all 7 roles
• Dashboard 6-tab visibility (Overview/Dispatch/Lifecycle/Metrics/Audit/Settings)
• P2 Kanban evaluation (work-in-progress limits + cycle time tracking)
• P3 ITSM evaluation (incident management + change advisory board simulation)
• 13 E2E tests xpass + anti-ghost counters

Anti-Ghost Feature Guarantee

Every new module includes _call_counter mechanism + E2E anti_ghost test + CI check_module_activation.py verification. Modules must be truly integrated into dispatch pipeline (not just instantiated), with Markdown report sections user-visible.

Test Pyramid Achieved

• Contract tests: 5.2% (target ≥5% ✅)
• Integration tests: 15.1% (target ≥15% ✅)
• Total tests: 8260+ (CI authoritative)
• E2E coverage: 107 e2e + 1244 integration + 13 V4.4.0 anti-ghost

Historical Features (V4.0.0-V4.3.3)

• V4.3.3: P0-P3 enhancement E2E skeletons (xfail TDD for V4.4.0)
• V4.3.2: LLM vs Mock quality gap measurement (calibration gate + thin-slice probe)
• V4.3.0 Phase 3: Quality hardening + user simulation E2E (NPS 9/10)
• V4.3.0 Phase 2: OutputValidator full integration (LLM output safety detection)
• V4.3.0 Phase 1: DependencyHallucinationChecker (anti-slopsquatting supply chain attack)
• V4.3.0 Phase 0: DeploymentComplianceChecker (anti-violation deployment backstop)
• V4.0.0 P1-1 Loop Engineering: Discovery → Handoff → Verification → Persistence → Scheduling
• V4.0.0 P1-2 UI/UX Patrol: 4-dimension audit + PIL pixel diff visual regression
• V4.0.0 P2-1 Adversarial Verification: red team attack + blue team defense + judge arbitration
• V4.0.0 P2-2 DAG Visualization: Mermaid / JSON / DOT three formats
• V4.0.0 P3-1 Autonomous: plan → dev → verify → fix 4-stage autonomous iteration
• V4.0.0 P3-2 Plugin Hot-Loading: 3 loading paths + path traversal 3-layer protection + reload rollback

8260+ tests passing (CI authoritative).

⚡ Quick Start (7 Ways to Invoke DevSquad)

Method 1: TRAE Skill (Recommended — you're already here)

DevSquad is registered as a TRAE Skill. Simply describe your task in the TRAE IDE chat, and the 7-role team will collaborate automatically. No CLI or API setup needed.

Method 2: CLI (Recommended for Terminal Users)

# Interactive setup wizard (1-2 minutes)
python scripts/cli.py init

# Then start collaborating!
devsquad dispatch -t "your task description"

Method 3: MCP Server (For IDE / Tool Integration)

# Start MCP server with stdio transport (for IDE integration)
python3 scripts/mcp_server.py

# Or SSE transport (for remote access)
python3 scripts/mcp_server.py --port 8080

Method 4: Web Dashboard (Recommended for Teams)

# Start Streamlit dashboard with authentication
streamlit run scripts/dashboard.py

# Open http://localhost:8501
# Login with default dev credentials (see INSTALL.md "Default credentials" section).
# Change all defaults in production.

Method 5: REST API (Recommended for Integration)

# Install dependencies
pip install fastapi uvicorn

# Start API server
uvicorn scripts.api_server:app --host 0.0.0.0 --port 8000 --reload

# Access Swagger UI: http://localhost:8000/docs
# Access ReDoc:      http://localhost:8000/redoc

Method 6: Python API (Recommended for Developers)

from scripts.collaboration.dispatcher import MultiAgentDispatcher

dispatcher = MultiAgentDispatcher()
result = dispatcher.dispatch(
    task="Optimize database query performance",
    roles=["architect", "security", "tester"],
)
print(result.report)
print(result.consensus)

Method 7: One-Click Startup Script (V3.9.2+)

# One-click startup — 4 phases: env check → DB init → frontend build → service start
./scripts/start.sh

# Launch Streamlit dashboard instead of API server
./scripts/start.sh --dashboard

# Override API port
DEVSQUAD_API_PORT=9000 ./scripts/start.sh

# Show help
./scripts/start.sh --help

start.sh is the unified entry point introduced in V3.9.2 (P0-2). It validates the environment, initializes the database, builds the frontend, and starts the service in one command. Use requirements.lock alongside it for reproducible builds (pip install -r requirements.lock). V4.1.0 adds Loop Engineering, UI/UX 巡检, Adversarial 验证, DAG 可视化, Autonomous, and 插件热加载.

👥 7 Core Roles

Auto-match: If no roles specified, the dispatcher automatically matches based on task keywords.

🏗️ Five Capability Domains (Architecture Overview)

DevSquad's 235 modules are organized into 5 capability domains, each solving a specific problem:

🎯 Domain 1: Task Orchestration Engine (Core)

让 7 个角色高效协作的「指挥中心」

Core Workflow:

User Task → [InputValidator] → [RoleMatcher] → [Coordinator Orchestration]
           → [ThreadPoolExecutor Parallel Workers] → [Scratchpad Real-time Sharing]
           → [ConsensusEngine] → [ReportFormatter] → [Structured Report]

🛡️ Domain 2: Quality Assurance System

防止 AI 「偷懒」或「幻觉」

⚡ Domain 3: Performance & Reliability

让系统更快、更稳定、更省钱

📊 Domain 4: Observability & Governance

知道系统在做什么、做得怎么样

🔌 Domain 5: Integration & Extension

融入你的现有工具链

🔬 Cybernetics Enhancement Modules (V3.6.1)

非侵入式包装设计 — 可选开关，零修改现有核心逻辑

The 5 cybernetic modules work independently or together without modifying existing core logic:

User Task
    ↓
[SimilarTaskRecommender] ← Optional: suggest roles from history
    ↓
[AdaptiveRoleSelector]   ← Optional: optimize role selection
    ↓
[MultiAgentDispatcher]
    ↓
[FeedbackControlLoop]     ← Wrap dispatcher for auto-iteration
    ↓ [each worker step]
[ExecutionGuard]          ← Guard each worker execution
    ↓
[PerformanceFingerprint]  ← Record after dispatch completes

1️⃣ FeedbackControlLoop (反馈闭环控制器)

• Closed-loop feedback control with automatic iteration until quality threshold met
• Configurable quality gate (quality_gate) and maximum iterations
• Lightweight quality assessment (no LLM calls), supports dry-run mode

2️⃣ ExecutionGuard (执行守护者)

• Real-time execution monitoring with 4 abort conditions: timeout, output size, token count, critical keywords
• Lightweight checks (<1ms), zero external dependencies
• Dynamically configurable thresholds

3️⃣ PerformanceFingerprint (性能指纹系统)

• Unified execution fingerprint recording (fuses 4 data sources)
• Pure Python TF-IDF implementation (no sklearn/numpy), supports English/Chinese mixed content
• JSON persistence to .devsquad_data/fingerprints/, graceful cold-start degradation

4️⃣ SimilarTaskRecommender (相似任务推荐器)

• TF-IDF-based task similarity search with historical success configuration recommendations
• Intelligent role combination recommendation, intent prediction, execution time estimation
• Confidence scoring (high/medium/low), graceful cold-start degradation

5️⃣ AdaptiveRoleSelector (自适应角色选择器)

• Three-tier selection strategy based on historical success rates
• Configurable minimum success rate and maximum role count
• Supports manual statistics updates and comprehensive role effectiveness reporting

Recommended usage (progressive adoption):

from scripts.collaboration import (
    MultiAgentDispatcher, FeedbackControlLoop,
    ExecutionGuard, PerformanceFingerprint
)

dispatcher = MultiAgentDispatcher()
guard = ExecutionGuard()
fingerprint = PerformanceFingerprint()

# Option 1: Full cybernetics stack
loop = FeedbackControlLoop(dispatcher, quality_gate=0.7)
result = loop.run("Your task here")

# Option 2: Guard only (minimal adoption)
result = dispatcher.dispatch("Your task")
for w in result.worker_results:
    abort, reason = guard.check_abort(w.output, w.duration)
    if abort:
        print(f"Aborted: {reason}")

# Option 3: Learning only
fingerprint.record_execution("task", result, result.timing, result.matched_roles)
similar = fingerprint.find_similar("new task", top_k=3)

All modules are optional switches — DevSquad works perfectly without them.

🏗️ Architecture Overview (Layered Design)

┌─────────────────────────────────────────────────────────────┐
│                    User Access Layer                         │
│  ┌──────────────┐ ┌──────────────┐ ┌──────────────┐        │
│  │ Streamlit    │ │ FastAPI REST │ │ CLI/Notebook │        │
│  │ Dashboard    │ │ API Server   │ │ (Existing)   │        │
│  │ (Auth+HTTPS) │ │ (Swagger)    │ │              │        │
│  └──────┬───────┘ └──────┬───────┘ └──────────────┘        │
└─────────┼───────────────┼───────────────────────────────────┘
          │               │
          ▼               ▼
┌─────────────────────────────────────────────────────────────┐
│                   Business Logic Layer                      │
│  ┌─────────────┐ ┌─────────────┐           │
│  │AuthManager  │ │HistoryMgr   │           │
│  │(RBAC Auth)  │ │(SQLite TSDB)│           │
│  └─────────────┘ └─────────────┘           │
│  ┌─────────────────────────────────────────────┐            │
│  │     LifecycleProtocol (11-Phase Engine)       │            │
│  │     UnifiedGateEngine + CheckpointManager     │            │
│  └─────────────────────────────────────────────┘            │
└─────────────────────────┬───────────────────────────────────┘
                          │
                          ▼
┌─────────────────────────────────────────────────────────────┐
│                    Data Persistence Layer                    │
│  ┌────────────┐ ┌────────────┐ ┌────────────────────────┐  │
│  │ SQLite DB  │ │ YAML Config│ │ Checkpoint Files       │  │
│  │ (History)  │ │ (Deploy)   │ │ (Lifecycle State)      │  │
│  └────────────┘ └────────────┘ └────────────────────────┘  │
└─────────────────────────────────────────────────────────────┘

🧩 Layered Sub-Skill Architecture (V3.6.0)

DevSquad provides 8 atomic sub-skills that can be used independently or together. Each sub-skill is a thin wrapper (~50 lines) importing existing core modules — no duplicated logic.

skills/
├── dispatch/       → DispatchSkill — MultiAgentDispatcher (7-role orchestration)
├── intent/         → IntentSkill   — IntentWorkflowMapper (6 intents × 3 languages)
├── review/         → ReviewSkill   — FiveAxisConsensusEngine (5-axis code review)
├── security/       → SecuritySkill — InputValidator + OperationClassifier + PermissionGuard
├── test/           → TestSkill     — TestQualityGuard + test strategy generation
├── retrospective/  → RetroSkill    — RetrospectiveEngine + pattern extraction
├── prototype/      → PrototypeSkill — Rapid prototype scaffolding (V4.5.0)
└── teach/          → TeachSkill     — Knowledge transfer & onboarding (V4.5.0)

Sub-Skill Quick Reference

Usage Examples

# Direct import (recommended for single skill)
from skills.dispatch.handler import DispatchSkill
result = DispatchSkill().run("Fix login bug", roles=["coder", "tester"])

# Via registry (dynamic discovery)
from skills import get_skill, list_skills
print(list_skills())  # ['dispatch', 'intent', 'review', 'security', 'test', 'retrospective']
skill = get_skill("security")
result = skill.scan_input("DROP TABLE users; --")

All sub-skills work without any API key in Mock mode.

📋 Plan C Architecture (Core Engine)

Unified Lifecycle Architecture - Resolves CLI 6 commands vs 11-phase lifecycle:

CLI View Layer (6 commands)          Core Engine (11 phases)
┌─────────────────────┐            ┌──────────────────────────┐
│ spec → P1, P2       │───View ──→│ P1: Requirements         │
│ plan → P7           │   Mapping │ P2: Architecture         │
│ build → P8          │            │ P3: Technical Design     │
│ test → P9           │            │ ...                      │
│ review → P8,P6      │            │ P10: Deployment          │
│ ship → P10          │            │ P11: Operations          │
└─────────────────────┘            └──────────────────────────┘
        ↓                                    ↓
  UnifiedGateEngine                   CheckpointManager
  (Phase + Worker gates)              (Lifecycle state persistence)

Core Components:

• ✅ LifecycleProtocol - Abstract interface for unified lifecycle management
• ✅ UnifiedGateEngine - Integrates VerificationGate + Phase transition gates
• ✅ FullLifecycleAdapter - Complete 11-phase lifecycle with dependency resolution
• ✅ Enhanced CheckpointManager - Auto save/restore lifecycle state across sessions

📦 Installation

Prerequisites

• Python 3.10+ (3.10, 3.11 supported, tested in CI)
• pip or pipenv for package management

Option A: PyPI Install (Recommended)

# Install from PyPI — zero setup, ready to use
pip install devsquad

# With optional dependencies
pip install "devsquad[api]"    # FastAPI + Streamlit dashboard
pip install "devsquad[all]"    # All optional features

Option B: Git Clone + Local Install

git clone https://github.com/lulin70/DevSquad.git
cd DevSquad

# Install core package (minimal dependencies)
pip install -e .

# Ready to use!
devsquad dispatch -t "Design user authentication system"

Verify Installation

# Check version
devsquad --version
# Expected: devsquad 4.3.0

# Run tests
pytest tests/ -v --tb=short
# Expected: 7681 passed

⚙️ Configuration

Create .devsquad.yaml in your project root:

quality_control:
  enabled: true
  strict_mode: true
  min_quality_score: 85

llm:
  backend: auto
  base_url: ""  # Set via DEVSQUAD_OPENAI_BASE_URL env var
  model: ""     # Set via DEVSQUAD_OPENAI_MODEL env var
  timeout: 120

Or use environment variables (higher priority):

# Default: auto tries real backends first, then falls back to mock
export DEVSQUAD_LLM_BACKEND=auto
export DEVSQUAD_OPENAI_BASE_URL=https://api.openai.com/v1
export DEVSQUAD_OPENAI_MODEL=gpt-4
export DEVSQUAD_OPENAI_API_KEY=sk-...

Environment Variables Reference:

🧪 Testing

Quick Smoke Test (< 30 seconds)

python3 scripts/cli.py --version       # Expected: DevSquad 4.1.0
python3 scripts/cli.py status          # Expected: System ready
python3 scripts/cli.py roles           # Expected: 7 core roles listed

Full Test Suite

# Run all tests (7681 tests passing)
python3 -m pytest tests/ -q --tb=line

# With coverage report
python3 -m pytest tests/ --cov=scripts --cov-report=term-missing

Test Layering Strategy

Total: 7681 CI tests / 266 e2e (7681 collected)

Run by priority:

# P0 only (critical path, < 10s)
python3 -m pytest tests/ -k "anti_ratif or verification or intent_workflow or auth" -q

# P0 + P1 (quality + enhancement, < 30s)
python3 -m pytest tests/ -k "anti_ratif or verification or intent or auth or five_axis or operation" -q

# Full suite
python3 -m pytest tests/ -q --tb=line

📚 Documentation

🗺️ Roadmap

V4.3.0（进行中 — 7-Role 共识达成，文档先行）

版本策略: V4.3.0 预发布（全部代码+文档+E2E 验证）→ 用户确认 → V4.3.0 正式版

整合三方面输入:

1. 技术债持续治理（todo_drift_monitor + CI 阻塞）
2. pickle→JSON 迁移（dead code 删除 + fallback 安全收紧 + 移除）
3. 上游 TraeMultiAgentSkill v2.6-v2.8 精细化启发（Ponytail 双模式 / LoopKernel 回退 / UIUX 审计 / Dashboard 可视化）

V4.3.0 范围（9 项）:

7-Role 共识: 7/7 APPROVE_WITH_CONCERNS，按 10 项调整修订后达成共识。详见 V43_ROADMAP_PROPOSAL.md v1.2。

项目生命周期: 按 11-Phase 模型推进（P1 需求 → P2 架构 → P3 技术设计 → P7 测试计划 → P8 实施 → P9 测试执行 → P10 部署发布）

测试金字塔保障: unit ≥60% / integration 15-25% / e2e ≤10% / contract 5-10% / smoke ≤5%

🤝 Contributing

1. Fork the repository
2. Create feature branch (git checkout -b feature/amazing-feature)
3. Commit changes (git commit -m 'Add amazing feature')
4. Push to branch (git push origin feature/amazing-feature)
5. Open a Pull Request

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

⭐ 如果 DevSquad 对你有帮助，请给个 Star！⭐
让更多开发者享受到「AI 团队协作」的力量

🙏 Acknowledgments
Inspired by TraeMultiAgentSkill upstream project
Built with ❤️ by the DevSquad team

Last updated: 2026-08-03 | Version: V4.5.0 (cross-session continuity + protocol-native skills + action-first reports — 10 new features) | V4.4.0 (5 enhancement modules: RiskRegister / ViewpointRegistry / ErrorBudgetTracker / GapAnalyzer / DoraMetricsCollector — see CHANGELOG.md)