Benzi Indexes Code with Tree-Sitter for O(1) AST-Gated Agent Edits

Tree-Sitter Graph Resolution Over Context Dumping

Unlike workflows that grep through files or compute cosine similarity across vector chunks, Benzi parses the entire project prior to answering questions. Using language-specific tree-sitter grammars across ten languages (including Python, TypeScript, Go, and Rust) alongside a dedicated HTML/CSS engine, it constructs a complete map of symbols, inheritance chains, and call sites. Identified data flows link directly to call locations, enabling lookups like backflow or trace_path in $O(1)$ tool execution steps.

Blast Radius Analysis and Syntax Gating

Benzi executes modifications through guarded writes:

• Pre-write blast radius evaluation: The agent examines caller dependencies and references before applying diffs.
• Parser-level syntax checks: Every write is validated against the live AST; broken AST parses trigger an immediate automated revert.
• Runtime trace capture: For Python codebases, an internal execution tracer observes runtime argument values and dynamic dispatch to reconcile ambiguous static edges.

SWE-bench Verified Benchmarks and Token Costs

Evaluating the full 500-instance SWE-bench Verified benchmark on DeepSeek v4-flash showed marked efficiency gains over file-dumping approaches. Benzi resolved issues at roughly $0.01 per bug, while harnesses like Claude Code scaled upward to $0.18 per step as problem complexity increased. The index resolves navigation demands early, keeping line-read volumes low regardless of repository scale.