Waveloom

The deciding trait is that the model chooses its own gear. Routine edits go to the fast tier and the hard thinking goes to the deep one, without you remembering to switch, so the expensive setting is reserved for the turns that need it rather than applied to every file rename you asked for at four in the afternoon.

What you give up is the breadth and the crowd of an older project. Pick it when the meter is what hurts. Pick the established option when you would rather have answers than savings.

A tool engineered around a single vendor's billing mechanics inherits that vendor's decisions. Three other endpoints are supported and none of them is what the loop was tuned for, so a user who switches keeps the interface and loses the reason to have chosen this over anything else, with nothing in the documentation quantifying what is lost.

What it does right is publish the mechanism rather than the outcome. A claim about how the prompt is structured can be checked against the source; a claim about savings cannot.

1. Most agents rebuild the prompt each turn and pay for it, because the ordering that makes a prompt readable is not the ordering that makes it cacheable. Treating prefix stability as a constraint on how context is assembled is a genuine architectural position and an uncommon one.

2. The saving is arithmetic rather than a benchmark, which is the honest way to make this kind of claim. 3. Nothing is published about how often the prefix survives a real editing session, and that number is the one a reader would want.

Building a product on a supplier's billing structure is building on ground you do not own. The provider can change how caching is priced, or make it automatic, and either move erases the reason this exists without anyone consulting the author. A hundred stars is not enough adoption to make that renegotiable.

Moat: none, and rented. Likely path: the advantage is commoditised upstream and the project becomes one more terminal agent among dozens. Position: use it while the arbitrage lasts, and do not write tooling that assumes it.

The packaging is better than most free projects manage. Every desk I have is covered, on both processor families, which means my endpoint team is not building anything and the rollout is a distribution task rather than a project. That is a real and underrated saving.

Everything after that is missing: no central configuration, no usage reporting, no retention statement, no unattended mode that would make it a measurable pipeline step, and no vendor to call. Approved with conditions: managed distribution, provider keys issued centrally, and no customer code until we have logging of our own.

Direct provider connections with no relay in the middle is the property I check first, and here it is stated rather than implied. My code goes from my machine to an endpoint I chose and nowhere else, which removes the question every hosted wrapper leaves open. Permissive licence, compiled artefact, no account.

The gap is hardware. There is no documented path to weights running on my own box, so the cheapest possible configuration is still somebody's API. For a project built entirely around cost that omission is strange, and it is the first thing I would add.