💎 Gem Team

Multi-agent orchestration framework for spec-driven development and automated verification.

"Turning Model Quality into System Quality."

🚀 Quick Start

Install Now for Code

See all installation options below.

🤔 Why Gem Team?

• ⚡ 4x Faster — Parallel execution with wave-based execution
• 🏆 Higher Quality — Specialized agents + TDD + verification gates + contract-first
• 🔒 Built-in Security — OWASP scanning, secrets/PII detection on critical tasks
• 👁️ Full Visibility — Real-time status, clear approval gates
• 🛡️ Resilient — Pre-mortem analysis, failure handling, auto-replanning
• ♻️ Pattern Reuse — Codebase pattern discovery prevents reinventing wheels
• 📏 Established Patterns — Uses library/framework conventions over custom implementations
• 🪞 Self-Correcting — All agents self-critique at 0.85 confidence threshold
• 🧠 Context Scaffolding — Maps large-scale dependencies before the model reads code, preventing context-loss in legacy repos
• ⚖️ Intent vs. Compliance — Shifts the burden from writing "perfect prompts" to enforcing strict, YAML-based approval gates
• 📋 Source Verified — Every factual claim cites its source; no guesswork
• ♿ Accessibility-First — WCAG compliance validated at spec and runtime layers
• 🔬 Smart Debugging — Root-cause analysis with stack trace parsing + confidence-scored fixes
• 🚀 Safe DevOps — Idempotent operations, health checks, mandatory approval gates
• 🔗 Traceable — Self-documenting IDs link requirements → tasks → tests → evidence
• 📚 Knowledge-Driven — Prioritized sources (PRD → codebase → AGENTS.md → Context7 → docs)
• 🛠️ Skills & Guidelines — Built-in skill & guidelines (web-design-guidelines)
• 📐 Spec-Driven — Multi-step refinement defines "what" before "how"
• 🌊 Wave-Based — Parallel agents with integration gates per wave
• 🗂️ Verified-Plan — Complex tasks: Plan → Verificationn → Critic
• 🔎 Final Review — Optional user-triggered comprehensive review of all changed files
• 🩺 Diagnose-then-Fix — gem-debugger diagnoses → gem-implementer fixes → re-verifies
• ⚠️ Pre-Mortem — Failure modes identified BEFORE execution
• 💬 Constructive Critique — gem-critic challenges assumptions, finds edge cases
• 📝 Contract-First — Contract tests written before implementation
• 📱 Mobile Agents — Native mobile implementation (React Native, Flutter) + iOS/Android testing

🚀 The "System-IQ" Multiplier

Raw reasoning isn't enough in single-pass chat. Gem-Team wraps your preferred LLM in a rigid, verification-first loop, fundamentally boosting its effective capability on SWE-benchmarks:

• For Small Models (e.g., Qwen 1.7B - 8B): The framework provides the "executive brain." Task decomposition and isolated 50-line chunks can up to double their localized debugging success rates.
• For Reasoning Models (e.g., DeepSeek 3.2): TDD loops and parallel research stabilize their native file I/O fragility, yielding up to a +25% lift in execution reliability.
• For SOTA Models (e.g., GLM 5.1, Kimi K2.5): The gem-reviewer acts as a noise-filter, pruning verbosity and enforcing strict PRD compliance to prevent over-engineering.

🔄 Core Workflow

Phase Flow: User Goal → Orchestrator → Discuss (medium|complex) → PRD → Research → Planning → Plan Review (medium|complex) → Execution → Summary → (Optional) Final Review

Error Handling: Diagnose-then-Fix loop (Debugger → Implementer → Re-verify)

Orchestrator auto-detects phase and routes accordingly. Any feedback or steer message is handled to re-plan.

📦 Installation

🏗️ Architecture

flowchart
    USER["User Goal"]

    subgraph ORCH["Orchestrator"]
        detect["Phase Detection"]
    end

    subgraph PHASES
        DISCUSS["🔹 Discuss"]
        PRD["📋 PRD"]
        RESEARCH["🔍 Research"]
        PLANNING["📝 Planning"]
        EXEC["⚙️ Execution"]
        SUMMARY["📊 Summary"]
        FINAL["🔎 Final Review"]
    end

    DIAG["🔬 Diagnose-then-Fix"]

    USER --> detect

    detect --> |"Simple"| RESEARCH
    detect --> |"Medium|Complex"| DISCUSS

    DISCUSS --> PRD
    PRD --> RESEARCH
    RESEARCH --> PLANNING
    PLANNING --> |"Approved"| EXEC
    PLANNING --> |"Feedback"| PLANNING
    EXEC --> |"Failure"| DIAG
    DIAG --> EXEC
    EXEC --> SUMMARY
    SUMMARY --> |"Review files"| FINAL
    FINAL --> |"Clean"| SUMMARY

    PLANNING -.-> |"critique"| critic
    PLANNING -.-> |"review"| reviewer

    EXEC --> |"parallel ≤4"| agents
    EXEC --> |"post-wave (complex)"| critic

🤖 The Agent Team (Q2 2026 SOTA)

Agent File Skeleton

Each .agent.md file follows this structure:

---                                    # Frontmatter: description, name, triggers
# Role                                 # One-line identity
# Expertise                            # Core competencies
# Knowledge Sources                    # Prioritized reference list
# Workflow                             # Step-by-step execution phases
  ## 1. Initialize                     # Setup and context gathering
  ## 2. Analyze/Execute                # Role-specific work
  ## N. Self-Critique                  # Confidence check (≥0.85)
  ## N+1. Handle Failure               # Retry/escalate logic
  ## N+2. Output                       # JSON deliverable format
# Input Format                         # Expected JSON schema
# Output Format                        # Return JSON schema
# Rules
  ## Execution                         # Tool usage, batching, error handling
  ## Constitutional                    # IF-THEN decision rules
  ## Anti-Patterns                     # Behaviors to avoid
  ## Anti-Rationalization              # Excuse → Rebuttal table
  ## Directives                        # Non-negotiable commands

All agents share: Execution rules, Constitutional rules, Anti-Patterns, and Directives sections. Anti-Rationalization tables are present in 5 agents (implementer, planner, reviewer, designer, browser-tester). Role-specific sections (Workflow, Expertise, Knowledge Sources) vary by agent.

📚 Knowledge Sources

Agents consult only the sources relevant to their role. Trust levels apply:

🤝 Contributing

Contributions are welcome! Please feel free to submit a Pull Request. CONTRIBUTING for detailed guidelines on commit message formatting, branching strategy, and code standards.

📄 License

This project is licensed under the MIT License.

💬 Support

If you encounter any issues or have questions, please open an issue on GitHub.