WEIPING_WHALE

The deciding trait is that undo is not aspirational. The workspace is checkpointed around every single turn, so backing out a bad step is one command rather than an archaeology session in your reflog, and that changes how willing you are to let it attempt something you have not fully thought through.

Against that, this is a young reimplementation with a small audience and no published package. Pick it if reversibility is what you have been missing. Pick the established tool if you would rather not be the one finding the bugs.

Loopback is not a security boundary. Any process on the machine, including a browser tab running somebody else's script, can reach a listening port, and this one accepts instructions for a tool that edits files and runs commands. The feature is optional, which helps, and the documentation describes what it does without describing what protects it.

What it does right is gate the shell. Commands sit behind approval prompts rather than executing on the model's word, so the interactive path has a checkpoint the network path does not.

1. Presenting a change as a patch rather than as a rewritten file puts the reviewer in the format they already read, which is a small decision with a measurable effect on how carefully a change is inspected. 2. Honouring ignore files when building the repository map keeps generated output and vendored trees out of context, which is a correctness question and a cost one.

2. Neither is novel and both are frequently omitted, so their presence in a project this small says something about who wrote it.

The missing registry entry is the signal worth reading. Publishing a package is the cheapest possible act of distribution, and a project that has not taken it is a project not yet trying to be adopted, which is fine as a statement of intent and fatal as a forecast. A hundred stars is curiosity, not a base.

Moat: none, and the category is crowded with better-resourced alternatives. Likely path: it stays personal, or it stops. Position: read it for ideas, and copy anything you want to keep.

Distribution stops this before anything else does. Building from a source checkout on every machine is not a rollout, it is sixty small projects, and my endpoint team cannot attest to an artefact nobody signed. That alone means a packaging effort my organisation would own permanently.

Beyond it there is nothing to govern with: no central policy, no usage record, no retention statement, and no way to run it where I could measure it. Not yet, and the first thing that would change my mind is a published, signed release.

Using the reference implementation for the connector layer is the correct decision and one that most small projects skip, which means the servers I already run attach without surprises instead of hitting a subtly wrong message format. Permissive licence, TypeScript I can read, and per-turn routing between tiers of the same family is a genuinely useful knob.

What it will not do is take my hardware. One provider family, no local serving path, so the cheapest configuration still involves somebody's API and a key. That is the part I would patch first.