Docs index

Where everything lives

Where everything lives — migrated from the README on 2026-08-31 and verified against the code at tag v0.1.4; if code and wiki ever disagree, the code wins and this page gets a PR.

Docs

• Security model — the run tool's jail, timeout, and (optional) container seam, stated as a threat model: what is isolated, what is NOT, and which surface refuses what by default.
• Benchmarks — urllib3 — the third dataset, fully regenerable from scripts/benchmark.sh.
• Benchmarks — psf/requests — the same methodology re-run on a well-known public repo this project didn't write, not a self-referential proof case; a real win and a real recall gap, both reported.
• Benchmarks — urllib3 — the third dataset, fully regenerable from scripts/benchmark.sh.
• Tool-selection accuracy — does having 81+ tools instead of ~14 cost an agent selection accuracy? Measured, not asserted: 14/14 correct in both conditions, one run — the hypothesis that breadth costs accuracy didn't hold up here.
• Benchmarks — real tool-call/response-size measurements against a real codebase, published honestly (including where it didn't win). (The competitor head-to-head — CodeGraphContext/Serena actually installed and run — is maintained locally by the author and is not part of this public tree; its published claims are footnoted in the remaining benchmark docs.)
• Adding a language — a complete, verified LanguageAdapter for a language cie has never seen, no tree-sitter grammar or LSP involved.