Schaltwerk

The deciding trait is that the work starts from a written spec you keep. When an agent wanders, you do not coax it back: you throw the attempt away, sharpen the sentence that misled it, and run again from a clean tree. That converts a frustrating hour of correction into two minutes of editing, and the spec is still there for the next attempt.

It suits people who can write a brief and dislike negotiating with a model. Pick something conversational if the plan only becomes clear while you talk.

The feedback loop is manual at its most important point. A reviewer marks a problem, then copies that text into a session and hopes the agent reads it in the context it was written about. Nothing carries the file, the line or the surrounding diff automatically, so precision depends on whoever is doing the copying, and precision is exactly what a correction needs.

What it does right is the escape hatch. A session that went wrong is deleted as a worktree, with no cleanup and nothing left behind in the branch you care about.

1. Persisting the instruction separately from the session is the design decision that makes iteration meaningful. If the prompt is discarded with the attempt, every retry is a new experiment with an uncontrolled variable; if it survives and is edited deliberately, the change between runs is known. 2. Reuse of the same specification across attempts also permits comparison between agents under identical instructions.

2. No evaluation is published, and none is claimed, which keeps the argument architectural.

The star count is the honest number here: small, early, and not yet evidence of anything except taste. There is no entity, no funding and nothing to sell, which means the eighteen-month question reduces entirely to whether one person keeps finding this interesting after the novelty passes.

Moat: none, and the concept is copyable in a weekend by anyone who wants it. Likely path: quiet maintenance, or the pattern showing up inside a funded product with better distribution. Position: use it freely, and keep your specs in the repository so leaving costs nothing.

Support matrices decide adoption, and this one splits my engineers into two tiers. The people on Linux are told their platform is beta and the ones who work through WSL are told nothing works, which means either an exception process or an unhappy third of the department. Neither is worth it for a free tool.

There is also no unattended execution, so it never becomes a measurable step, and no directory integration to point at in an audit. Not yet: revisit if Linux leaves beta.

Shipping an MCP server rather than only consuming one is the inversion I keep waiting for. It means the application is addressable: an agent I already trust can open sessions here, coordinate parallel work and report back, and the orchestration logic lives in my prompt rather than in somebody's product roadmap.

Installation is one brew command, agents can point at a private or self-hosted endpoint instead of a public API, and the licence keeps a fork legal. That is close to everything I ask for from a desktop tool.