A smaller coordinator input is only one part of the cost
Ruflo released version 3.70.0 on November 1 with an optional mechanism for handing large analysis reads to a worker model. The main agent receives findings with citations, while checks compare the cited lines with a source snapshot identified by a SHA-256 hash.
The maintainer reports a 97.6% reduction in coordinator payload on a fixed test, alongside a 3.8% increase in total bytes crossing the tested boundary. All 300 test answers matched their quoted source. These are maintainer-reported protocol results, not independently verified production measurements.
The accompanying explanation explicitly says no live provider call had succeeded in that test effort. Lower actual model bills, response speed and answer quality therefore remain unproven.
Existing installations keep their default behavior
Delegation is disabled by default. An observation mode records a routing decision while continuing to read directly; an enforcement mode must be deliberately enabled to delegate. Editing, debugging, architecture work and small or targeted reads remain direct, according to the release notes.
CGH perspective: evaluate this kind of optimization with both coordinator and worker usage included. A source-matching citation can establish that text came from the supplied file; it cannot establish that the worker found every relevant fact or drew the right conclusion. Measure missed information as well as traffic and price.
Release and commercial boundaries
The tagged software is available under the repository's MIT license. Model-provider charges and eligibility are separate from that software license; the reviewed release does not establish a universal operating price or regional service availability. GitHub displays a November 1 release time without a verified timezone in the retrieved text. CGH has not installed the package or reproduced the benchmark.