Can I vibecode Devin?

Build a comprehensive, autonomous coding agent assistant inspired by the core features of Devin AI, designed for end-to-end software development lifecycle management. The application should act as an independent software engineer capable of planning, executing, testing, and debugging complex projects from start to finish.

**Tech Stack**:
- Backend: Python 3.12, FastAPI for API routing, SQLite (with SQLAlchemy) for local state and memory management.
- Execution Environment: Docker for sandboxing agent execution and ensuring safe, isolated operations.
- Frontend: A modern React-based web interface (using Vite) to monitor the agent's progress, view terminal output, and interact with the planned steps.

**Core Requirements**:

1. **Autonomous End-to-End Workflow & Dynamic Planning**:
   - The agent must be able to take a high-level user prompt and decompose it into a step-by-step actionable plan.
   - Implement dynamic re-planning: if a step fails (e.g., a compilation error or test failure), the agent should autonomously adjust its plan and attempt to resolve the issue.

2. **Secure Sandboxed Environment**:
   - Integrate Docker to run agent commands, compile code, and execute tests within an isolated container.
   - Provide the agent with a virtualized toolchain: a simulated terminal for shell commands, an integrated file editor to read and write repository files, and basic browser access to read documentation or search for solutions.

3. **Multi-Step Memory & Context Management**:
   - Implement a robust persistence layer to store conversation history, file changes, terminal outputs, and the current state of the plan.
   - The agent must maintain long-term memory across multi-step tasks to retain context and avoid hallucinating past actions.

4. **Self-Healing Code & Iteration**:
   - Whenever code fails to compile or tests fail, the agent should automatically capture the error logs from the sandbox.
   - The agent must feed these logs back into the LLM context to iterate on the codebase and autonomously fix the issues without human intervention for every error.

5. **User Interface & Interaction**:
   - The interface should provide a split-view dashboard: one side showing the agent's step-by-step plan and progress, and the other side showing a live feed of the terminal, diffs, and file changes.
   - Include clear controls to pause, resume, or abort execution, and a chat interface for additional guidance.

6. **Safety & Security Constraints**:
   - The application must be single-user, local-first, and highly secure.
   - All sensitive data and API keys must be strictly managed via `.env` files.
   - Untrusted input must be validated, and all agent-executed commands must be restricted to the sandboxed workspace.

Create a robust `README.md` containing setup instructions. Provide clear scripts for installation, testing, and local development. Start from an empty folder, initialize the full project structure, write focused unit tests for the planning and sandbox execution modules, and ensure the core loop functions correctly.