minion

The deciding trait is what it does not send. Most agents fill a large part of the window before you have typed anything, which is invisible when the window is enormous and fatal when it is not. On a modest local model the difference is whether the thing can hold your file at all.

In exchange you get an interface with no ceremony and no settings screen, which some people find restful and others find bare. Pick it if your inference runs at home. Pick a full harness if you are already paying for a frontier model and the overhead never mattered.

The cost of the simplicity is customisation. There is no extension point, no settings file and no hook, which means adapting it to your environment means editing the file, and the next upstream improvement arrives as a diff against a file you have already changed.

That is manageable for one person and it does not scale past them. Nothing documented describes a supported way to carry a modification across versions. What it does right is being small enough that the fork is survivable, which is not something you can say about most projects on this board.

1. The number is reproducible because the conditions are stated. Anyone can send the same greeting and count, which is the difference between a measurement and an advertisement, and it is a low bar most efficiency claims fail to clear. 2. The comparison figure is not produced the same way.

2. The range attributed to other harnesses is quoted without a method, a version or a named tool, so the ratio a reader takes away is one careful number divided by an estimate. The measured half is exemplary. The comparative half is ordinary marketing arithmetic.

This is the purest version of a pattern that recurs across this board. A single file, given away, solving a problem a specific community has, with no mechanism by which effort becomes revenue. The value is real and it accrues entirely to the people who use it.

Moat: none, and being readable in an afternoon means anybody can rebuild it. Likely path is that the author keeps it working while the local-model community keeps needing it, which could be years or could be until the next hobby. Position: copy it into your own repository and stop worrying.

My problem is provenance. A script that arrives from a repository and runs against an endpoint set in an environment variable has no version I can record, no signature I can check and no entry in the inventory my security team reviews every quarter.

It also does not run in delivery, so it never becomes a step anybody measures, and there is no identity, no policy and no record of use. Not yet, and not ever in this shape. It is a personal tool, and the correct governance for a personal tool is a conversation rather than a rollout.

That fallback is the sentence that told me who wrote this. Rough inference servers do not all implement tool calling properly, and every polished harness treats that as the server's problem. This one parses the text and keeps going, which is what you build when you actually run your own weights.

The licence is permissive, the endpoint is an environment variable, and there is no configuration format standing between me and a base URL. No protocol support, which is the one thing I would add, and I would add it myself in the file.