A multi-agent dev pipeline that uses Gemini CLI to analyze a codebase, plan, and implement changes.

Given a project directory and a task, it runs specialized agents in sequence -- Product Owner, Business Analyst, Architect, Developer, QA, Testing -- each one owning a stage of the development lifecycle. Tree-sitter builds a "knowledge vault" of the codebase first, so agents reason over actual code structure, not raw text.

pip install -r requirements.txt
cp .env.example .env          # add your Gemini API key

Requires Python 3 and a Gemini API key.

python main.py ./my-project ./vault "add rate limiting to the API"
python main.py ./my-project ./vault "..." --rebuild   # force vault rebuild

The vault (knowledge vault + code-index.json) is reused across runs unless --rebuild is passed.

Phase 0 -- Context: Tree-sitter builds a knowledge vault -- ASTs, signatures, class hierarchies, docstrings, call graph. Output: Markdown brief + code-index.json.

Phase 1 -- Planning: Stack detector classifies the project, then Product Owner and Business Analyst agents frame the task, then Architect proposes an implementation grounded in the vault.

Phase 2 -- Development: Devflow agent breaks the architecture into work, Developer agent implements, QA and Testing agents review.

Every agent subclasses base_agent.py. Output is normalized to JSON via strip_json.py, and query_index.py lets you query the code-index offline.

MIT