Measured, not modeled β a real Claude Code agent solving real GitHub issues, run with and without Lettuce.
0 of 16 supported languages at the 30-run target.
Real repos, never tuned against. 32 open-source repositories β 16 languages, 2 repos per language β none of which Lettuce was built or tuned for. 160 tasks in total, each a real GitHub issue (bug reports and code questions, picked automatically, not hand-curated).
Same agent, same model, two arms. For every issue the identical Claude Code agent (Claude Opus 4.8) located the relevant code twice: once with only grep + file reads, once with Lettuce. Same task, same model β the only thing that changes is whether the agent can call Lettuce.
How Lettuce is used. Lettuce indexes the repo into a graph once, then exposes two MCP tools β understand and read_snippet. The agent calls understand once and reads only the lines it actually needs, instead of grepping and reading whole files. This is the exact same two-tool setup and prompt a real user gets.
Measured, not modeled. Each arm ran twice; the median is reported. Token counts come straight from the agent's own usage meter β not an estimate. We also confirm the agent genuinely calls understand rather than silently falling back to grep.
Last run 6/14/2026.