Skip to content

Find code by asking what it does. A CLI for coding agents that uses Jev to discover relevant files and source context.

2.0 KB
Raw Blame

Go/Rust parsing measurements

Three fresh-process trials of the final candidate on darwin/arm64 with Node 24.14.0 are compared with the unchanged ae10a9e baseline trials from development. Baseline trials were not rerun. Cold time includes module import and first inspection; warm time is the median of ten inspections, summarized across the three processes. No provider calls were made.

Language Functions Cold before → after (ms) Warm before → after (ms) Target unit bytes before → after
go 100 143.38 → 281.41 0.06 → 2.65 5,195 → 52
go 1,500 151.98 → 454.15 0.63 → 30.63 23,976 → 54
rust 100 148.12 → 275.91 0.06 → 2.42 5,180 → 52
rust 1,500 152.00 → 257.82 0.62 → 22.19 23,976 → 54

This is a deliberate precision/cost tradeoff. Parsing adds work and worker memory; text fallback had no parser worker. The measured unit contains a known target marker, not a model-selected final excerpt. Smaller units do not establish lower end-to-end cost, better recall, or faster coding-agent tasks. Worker RSS snapshots and source identities are in the raw measurements.

The installed search fixture supplies deterministic relevance answers. It verifies that late methods reach large-file declaration indexes, named selection returns their source and Rust attributes, and distant noise is not returned. This is a capability check, not a live-provider quality benchmark.

Reproduce each workload in prepared baseline/candidate checkouts using the same measurement script, with go or rust and 100 or 1500 functions:

node --experimental-strip-types specs/go-rust-parsing/bench.mjs /absolute/checkout/packages/core/src/source.ts go 1500

Run each arm in alternating fresh processes. The candidate must first prepare its pinned grammar assets through the CLI build. Further extraction fixes may change timings; the recorded source hashes identify this measurement cohort.