gsd-build/gsd-2
Gsd 2
This project is an autonomous AI software development framework designed to plan, code, test, and commit software milestones without human intervention. It functions as a state-machine-driven agent loop that orchestrates development through a recurring cycle of research, execution, and verification.
The system distinguishes itself through a git-isolated task runner that executes milestones in separate worktrees and branches, ensuring changes are squash-merged into a linear commit history. It features a multi-model routing gateway that assigns different LLM providers to specific workflow phases to balance output quality against budget limits and operational costs.
The framework covers a broad range of capabilities, including spec-driven project bootstrapping, context engineering via compression and database-backed state recovery, and the orchestration of specialized subagents for research or codebase reconnaissance. It integrates with Model Context Protocol servers and external tools to extend agent capabilities, while providing real-time steering and monitoring dashboards to track progress.
The project is implemented in TypeScript.
Features
- Autonomous Agent Loops - Orchestrates autonomous software development through a recurring state-machine loop of research, planning, execution, and verification.
- Autonomous Agent Frameworks - Provides a complete autonomous agent framework for planning, coding, and verifying software through a state-machine loop.
- Autonomous Software Engineering - Provides a full framework for building complete software projects autonomously through AI-driven planning, coding, and testing.
- Worktree Isolation - Uses isolated git worktrees to maintain separate file states for different autonomous development milestones.
- Autonomous Agent Orchestration - Provides a framework for managing the lifecycle and persistent memory of autonomous agents over long-running development sessions.
- Context Compression - Reduces token consumption and manages memory by compressing context windows through a preference-driven pipeline.
- Context Memory Management - Implements a pipeline for long-term project memory using context compression and database-backed state management.
- Model Routing Gateways - Assigns different LLM providers or models to each development phase with automatic fallback and budget-aware cost control.
- Multi-step Goal Execution - Uses LLMs to independently plan and execute multi-step sequences of research, coding, and verification.
- Provider Response Routing - Implements logic to route requests to specific LLM providers based on the requirements of different development phases.
- Subagent Orchestrations - Orchestrates specialized helper agents with isolated contexts to perform research and codebase reconnaissance.
- Agentic Workflow Isolation - Isolates AI-generated changes within dedicated git worktrees and branches to maintain a clean commit history.
- Construction - Builds complete software projects autonomously by cycling through research, planning, and coding.
- Git Worktree Management - Manages multiple concurrent git worktrees to provide isolated environments for developing different software milestones.
- Git Workspace Isolations - Executes development milestones in isolated git worktrees to maintain a clean and linear squash-merged commit history.
- Development Cycle Orchestration - Orchestrates AI-driven development cycles involving research, coding, and verification phases.
- Branch-Based Isolation - Uses separate git branches to isolate AI-generated changes before squash-merging them into the main history.
- Agent State Persistence - Persists session state and milestone progress in a database to enable crash recovery and project inspection.
- Persistent Project Artifacts - Maintains project requirements and learned patterns through persistent markdown artifacts and a database.
- Task Outcome Enforcement - Ensures autonomous workflows conclude by running static checks and behavioral tests to verify requirements.
- Behavior Guidance - Loads coding standards and architectural rules from markdown files to guide the agent's implementation behavior.
- Human-in-the-Loop Steering - Enables humans to interact with and redirect active autonomous agent flows and milestones in real-time.
- MCP Server Integrations - Integrates with Model Context Protocol servers to provide agents with specialized tools and external context.
- Autonomous Agent Orchestrators - Coordinates autonomous agents by integrating Model Context Protocol servers and external tools for complex task execution.
- External Tool Integrations - Connects AI agents to external utilities and servers to expand their operational capabilities beyond simple text generation.
- MCP Protocol Integrations - Provides a standardized way for agents to interact with external tools and context servers via the Model Context Protocol.
- Model Gateways - Provides a centralized gateway to manage communication and routing across multiple AI model providers.
- Per-Component LLM Routings - Implements routing logic to assign different LLM models to specific workflow phases based on cost and quality requirements.
- Step-by-Step Task Plans - Implements multi-step task plans that are submitted for manual approval before agent execution.
- Phase-Specific Model Overrides - The project assigns specific model providers to different project phases and sets automatic fallbacks for failed requests.
- LLM Session Recovery - Provides the ability to restore the state and progress of AI agent sessions after a system failure.
- Spec-Driven Generation - Guides a structured discussion to capture project vision and then generates the initial codebase and roadmap.
- Spec-Driven Project Initialization - Captures project vision and constraints through a guided discussion to create the initial structure and roadmap.
- Agent Interaction Monitors - Provides tools for observing and interacting with live AI agent execution streams from a separate terminal.
- Agent Steering Interfaces - Synchronizes state between background autonomous processes and foreground terminals for live steering and monitoring.
- Research Progress Dashboards - Displays a real-time dashboard showing task progress, elapsed time, and cost projections for autonomous runs.
- Token Cost Optimizations - Optimizes AI operational costs via context compression and the selection of cheaper model tiers.
