Build a Codebase Knowledge Graph

• ✓Parse a Python repo into a navigable call/import graph using only the stdlib ast module
• ✓Resolve calls honestly — same-class and unique-name heuristics, everything else reported as ambiguous or external, never guessed
• ✓Build bounded-depth graph traversal (callers, callees, related-within-N-hops) — the retrieval unit a coding agent actually wants
• ✓Combine graph proximity with embedding similarity into one retriever, and prove with a benchmark that combining actually helps — or doesn't
• ✓Swap a deterministic hashing-trick embedder for a real local embedding model and see what changes
• ✓Package the habit as a reusable Claude Code skill

Give a coding agent the repo context it needs before it edits code it didn't write — who calls this, what does it depend on, is this actually reachable — the layer underneath any disciplined plan.