Knowledge Graph

Baton indexes your repo into a knowledge graph (via the external graphify CLI) and a deterministic CODEBASE.md map, so agents navigate the code with targeted queries instead of broad, expensive file scans.

Initialize

baton kb init

This:

• detects sub-projects (each gets its own graph) plus a merged cross-project graph,
• installs a git hook that rebuilds incrementally on commit,
• generates a CODEBASE.md map per project, and
• writes MCP config so agents can query the graph.

baton kb status      # projects, node/edge counts, last build
baton kb rebuild     # incremental; --full to re-extract

CODEBASE.md — the token-cheap map

baton kb init / rebuild generates a CODEBASE.md (under ~2k tokens) per project: stack, file tree, the highest-centrality symbols, and pointers for querying the graph. It carries a staleness footer tied to the graph's commit.

Agents are told to read CODEBASE.md first — orienting from a 2k-token map is dramatically cheaper than re-reading the whole repo every session. CODEBASE.md is meant to be committed.

Agents query it

The generated graphify-* MCP servers expose:

• query_graph(query) — search the graph (symbols, files, relationships)
• get_node(id) — details on one node, with source locations

Share it

Avoid re-indexing on every clone:

baton kb export                 # → a .tar.gz pack
baton kb import <pack|kb/>       # re-anchored, staleness-checked
baton kb share on                # commit a kb/ dir so teammates clone-and-go