Keen Code

The deciding trait is restraint, and it is unusual enough to be a reason. This sits in the same shape as the well-known terminal agents and deliberately declines most of what they added after the first release, so the thing you interact with is a loop, a tool list and nothing negotiating for your attention.

Whether that is a virtue depends entirely on which of those additions you actually used. Pick it if your terminal agent is already mostly ignored menus. Pick the established one if you would miss the parts you did not notice you relied on.

An unpublished exclusion list is the problem. Minimalism as a principle is defensible; minimalism as an undocumented boundary means the only way to learn what is missing is to need it mid-task and find nothing there. Every other agent's absent feature is at least a filed issue. Here it is a design position, which is harder to argue with and slower to change.

What it does right is keep the orchestration honest. Sub-agents exist and are described plainly as delegation, not dressed up as a team of specialists with job titles.

1. Committing the interactions that produced the code turns a provenance claim into an artefact. A reader can see which instructions produced which module, which is the closest thing to a reproducibility statement this category has offered. 2. It also serves as evidence for the minimalism claim, since the omissions are visible in the record rather than only asserted.

2. No evaluation of coding performance is published, and none is claimed. The argument being made is about harness size, and harness size is directly observable.

The install-to-star ratio is healthier than most rows on this board, which tells me the people who find it actually run it. That is the good news and it is small news, because six hundred a week finances nothing and there is no entity, no hosted surface and no plan described anywhere.

Moat: none available to a minimal harness, since the thing that makes it good is also the thing that makes it a weekend to reproduce. Likely path: quiet volunteer maintenance until the author's interest moves. Position: use it, pin the version, own the fork.

Finance has no objection and security has every objection, which is the standard shape here. Sixty installations, sixty provider keys held by developers, and no central place that records what any of them executed against our source. The smallness that engineers like removes the surfaces I would otherwise use to govern it.

Nothing runs without a person present, so it produces no measurable throughput and never enters delivery reporting. Onboarding is short for anyone who has used a terminal agent before. Not yet, and the first condition would be centrally issued keys.

This is the size of thing I can actually own. Permissive licence, a source tree I can hold in my head, MCP servers attaching so the tools I already run are available, and skills as editable units rather than vendor magic. When it misbehaves I find the function, not a support form.

The gap is local weights. Multiple providers with my key is fine, but nothing points at the machine under my desk, so inference always leaves the building whatever else I control. Everything else here is the right shape, and grudgingly that is enough to keep it installed.