Can I vibecode Factory?

text
Build the closest honest personal consolation tool inspired by Factory, but do not claim to replace its massive structural moat, proprietary models, or enterprise infrastructure. You must strictly use this technology stack: Python 3.12, Typer for building the Command Line Interface (CLI), and SQLite for local metadata and state tracking. Do not offer alternative technology stacks.

The primary job is to build a repository-local software development assistant (an AI agent) that operates entirely via the terminal. The agent must index exactly one specified local codebase, interact with a single configurable LLM API (e.g., Claude or OpenAI), propose actionable code diffs based on a user prompt, automatically run the local test suite, and record every applied change and its context in the local SQLite database.

Start from an empty folder and create the complete, working CLI application. The default operating mode must be single-user and private to the local machine. Store all metadata (such as prompt history, file indices, and diff records) locally using SQLite. Do not integrate analytics, telemetry, advertisements, or third-party authentication services. All necessary secrets, including the LLM API keys, must be stored securely in a `.env` file; provide a comprehensive `.env.example` and ensure API keys are never hardcoded or logged.

Provide a simple script to generate realistic sample data or index a dummy repository so the core loop can be tested easily. Implement a highly polished, minimal terminal interface (using libraries like Rich if helpful, alongside Typer) that successfully completes the core agent loop (Prompt -> Index -> Diff -> Test -> Commit) end-to-end. The CLI must have clear output states for empty repositories, loading/processing animations, validation warnings, success confirmations, and formatted error tracebacks.

Support an export command (e.g., exporting the agent's action history to CSV or JSON) so the user is not locked into the tool. Ensure the terminal output is accessible, using sensible contrast and clear textual labels. Validate all untrusted input (e.g., user prompts or file paths) and guarantee that API keys or private file contents are never accidentally printed to standard output or log files.

Deliberately exclude paid-tier features of the original product, such as frontier custom coding models, massive cross-repository context windows, cloud-hosted scalable infrastructure, and IDE-wide GUI polish. Do not fake integrations with platforms like GitHub, Jira, or proprietary network effects. Ensure the tool degrade gracefully if the LLM API is unavailable, allowing the user to at least search the local index or view history.

Write focused unit tests for the core file indexing logic and the database data model. Add one end-to-end smoke test that proves the CLI successfully runs a prompt and proposes a diff on a dummy file. Create a detailed README covering setup, required OS permissions, architecture, data storage locations, and limitations. Include scripts (or a Makefile) for install, dev, test, and execution. Finally, run all tests and fix any errors before concluding.