UmaDev

The deciding trait is proportion. A small edit stays a small edit, and the machinery only appears when the work is big enough to deserve it, which is the opposite of every framework that makes you fill in a plan before changing a constant. You get the ceremony when you need it and silence when you do not.

You are the wrong buyer if your day is uniform, because then you are paying for a decision that always comes out the same way. Pick it for variety of task size. Pick a plain agent for consistency.

The cost model is the exposure. Every role runs through the base tool you are already paying for, so a task that triggers the full expansion multiplies your usage against a quota somebody else sets, and the threshold for triggering it is described as judgement rather than as a rule. You find out afterwards, on a page that shows a limit reached.

What it does right is refuse to touch git. It leaves your history alone, which for a tool this eager is the correct restraint.

1. Communication between roles is bounded and structured rather than a shared transcript, which keeps a delegation from inheriting everything the parent ever thought and keeps the exchange inspectable. 2. The completion signal is the important design decision: failed and partial work is reported as such, which is the rarest honesty in this category.

2. A system that can say it did not finish is measurable. One that reports success for everything cannot be evaluated at all. No benchmark is published, and after that admission, the absence is forgivable.

257 stars, and the project states openly that it owns no model endpoint. That is honest architecture and a precarious position: five upstream vendors each hold the ability to change an interface, a quota or a licence and remove the product's ability to function.

Moat: none, and the orchestration idea is legible to any of those five. Pricing power: none, at zero. Likely path: one of the base tools ships role orchestration and this becomes a preference. Position: use it, and keep the base tool usable without it.

An audit trail exists, produced by the component that schedules the work, which means a run leaves a record without me building one. That is unusual at this size and I will take it.

What it does not do is connect that record to a person in my directory, because there is no SSO and no identity model at all. The licence is free and the real cost is sixty engineers consuming their base subscriptions faster than the plan assumed. Approved with conditions: usage reviewed after one month against the previous month's baseline.

One binary, no runtime beside it, and it uses the login I already have on the base tool, so nothing new holds a credential. The only model it can serve itself is a small embedding model for retrieval, which is a modest and honest use of local inference rather than a checkbox.

There is no MCP in either direction, so the tools available are whichever ones the base CLI carries, and my servers stay outside. MIT means I can read how the roles are prompted, which is the part I actually wanted to see.