A source-aware software agent framework. It parses an existing codebase into a structured program model — a call graph plus a semantic index — so an LLM agent can understand and safely modify unfamiliar code instead of guessing.
Most coding agents operate on raw text. This framework front-loads a mapping pass so the agent reasons over structure:
core/mapper/
ast_parsers/ → language front-ends (Python, Fortran)
call_graph.py → who-calls-what across the codebase
semantic_embedder.py → embedding index for retrieval
program_model.py → the unified model the agent queries
doc_scraper.py / sdk_mapper.py → pull in external API/SDK knowledge
core/executor/
com_bridge.py → drive host applications (e.g. Office COM) as tools
core/tools/ → agent-facing tools over the program model
plugins/ → domain packs (e.g. aermod)
app.py— entrypoint- Plugin-based: a domain plugin injects vocabulary and tools for a specific problem area
- Multi-language by design — Python and Fortran parsers ship in-box
Python · AST analysis · embeddings/semantic retrieval · LLM agent · COM bridge
Research framework, active development. .env.example documents required
configuration; no secrets are committed.