A harness plugin measures the tool-call tax and puts one file read at 4.44 ms

ToolRush is an MIT-licensed execution layer for the Hermes Agent harness that publishes measured before-and-after latency for four tool-call paths. The GitHub API returned 95 stars and 5 forks on 6 September 2026 against a repository created on 2 September, so the traction is four days old and small. The reason to read it is the measurement, not the star count.

The number that carries the story

The README states four lane results, described as measured on the real installed harness with no mocks or fixtures. A native file read goes from 255.23 ms to 4.44 ms. A warm persistent terminal goes from 285 ms to 12.1 ms. Search transport, running rg directly, goes from a range of 183 to 455 ms down to 27 to 97 ms. Batched parallel remote calls go from 108 ms sequential to 53 ms batched.

These are the author's own figures on their own machine, so they are a claim rather than an audit. What makes them worth quoting is that the README says what they are not.

The project states its own limit

The project states the framing itself: these are tool-operation wall times, not model-inclusive turn speed. In its own words, tool-heavy turns get faster and chat-heavy turns barely move. It also discloses one regression, that trivial native reads gain nothing from threading.

A performance claim that names the case where it does not apply is doing something most vendor benchmarks avoid. The assumption it argues against is that agent speed is a tokens-per-second problem. If a read costs 255 ms of transport before any token is generated, an agent that reads forty files has spent ten seconds on plumbing.

What it costs to adopt

The constraints are real and narrow the audience sharply. It targets Hermes Agent specifically, the badge on the repository says Windows and MSYS, and the code is Python. This is not a general speed-up for any coding agent.

The safety design is stated as fail-closed: work that the admission classifier refuses still runs, just sequentially through the normal path. The batching lane is limited to read operations on up to four workers, with no writes and no terminal. The project reports 206 regression cases passing with zero failures, and five negative controls that each break when the corresponding fix is reverted, which is the check that stops a test suite from passing vacuously.

Four days and 95 stars is not adoption. Nobody outside the author has reproduced these numbers, and the harness-specific patching approach, which restores twenty-five compatibility patches in memory after an upstream update, is the kind of thing that works until the day it does not.

The transferable part is the method. Any team running a coding agent can time its own tool calls separately from its model calls, and find out which half of a slow turn is actually slow. That measurement costs an afternoon and does not require installing anything from this repository. If the answer is that transport dominates, the fix is local: fewer round trips, a persistent shell, batched reads. If the answer is that the model dominates, then the harness was never the problem and a plugin like this would have changed nothing.

Source

OnlyTerp/toolrush, GitHub. All latency figures are quoted from the project README. Star count, fork count, licence and creation date read from the GitHub API on 6 September 2026.