repo/overviewLIVE · HydraDB

Traverse structure. Stop burning tokens.

Tokens needed to find the same callers

Fewer tokens than opening the files –

Benchmark

Ground truth: calls recorded while the repo's own tests ran, scored on the held-out half of the test files. Recall and precision for the baselines are measured at the token budget CodeGraph used. Precision only judges functions the tests ran (a lower bound). Tokens are characters ÷ 4 for every method.

repo/blast radiusLIVE · HydraDB

What breaks if I change this?

call found in the code call learned from tests colour = file · inner ring = 1 hop
ImpactHydraDB · – ms
Direct–
2 hops away–
Learned edges–
Files to read–
Context cost of this answer
CodeGraph context
Open the files
Affected symbols
Code CodeGraph sends

    
repo/token mapLIVE · HydraDB

Token map

Every tile is a file, sized by the tokens it costs to read whole. Switch agents to see what each one pulls into context for the same question.

Tokens pulled into context
Selected file Click a tile
Whole file–
Read this run–
Read whole by the grep agent
CodeGraph slice: fill height = share of the file sent
Never touched
repo/agent race

Claude Code, with and without CodeGraph

Real headless runs from agent_bench.py: same questions, same model, once with the CodeGraph MCP tools and once with grep and file reads only. A caller counts as found if the answer names it or cites a line inside it.

Every run