pi_agent_rust

The deciding trait is that it is one binary that starts immediately. That sounds trivial until you count how many times a day you open an agent, abandon the thought while it loads, and go back to what you were doing. A tool that is present the moment you ask gets used for the small jobs, and the small jobs are most of them.

What you give up is breadth: this is young, and the feature list is short on purpose. Pick it for speed. Pick Aider if you want the deeper toolbox.

The maintenance surface is the compatibility layer. Endpoints that claim the same shape differ in streaming details, field names and error behaviour, and every one of those differences becomes a case handled here. Providers change without warning. The failure arrives as a stream that stops mid-response against one vendor while everything else works, which is the slowest kind of bug to report.

What it does right is naming the problem out loud. Most tools pretend the compatible endpoints are actually compatible and let the user discover otherwise.

1. The stated motivation is empirical in form and unevidenced in substance. Overhead is a quantity: a distribution of process start times, a resident set, a failure rate at some concurrency. None of those appears. 2. That does not make the rewrite wrong, since the mechanism is plausible and well understood, but plausible is a different category from demonstrated.

2. A single reimplementation is also a fresh source of defects in behaviour the original had already settled, and no comparison is offered.

Fast star growth on a solo rewrite tells you the itch was real and tells you nothing about durability. The asset is a codebase that tracks another project's design decisions, so the roadmap is partly written elsewhere and there is no commercial entity here to negotiate, hire or be bought.

Moat: none. Likely path: the maintainer's interest holds for a year, or the original ships the same performance work and the reason evaporates. Position: use it as a utility, never as a dependency with a support expectation attached.

The install path is the blocker before the product is. Piping a script from a content host into a shell is not something I authorise across an engineering organisation without a packaged artefact and a checksum to point our fleet tooling at. That is a solvable problem and somebody has to be paid to solve it.

Beyond that: no directory integration, no policy surface, no unattended execution and therefore nothing to measure. Zero licence cost against sixty seats does not rescue it. Not yet, and I have no objection to individuals experimenting.

A configuration file listing my own providers is exactly the interface I want, because adding an endpoint is an edit rather than a pull request. Three local servers are named, which means someone actually ran this against weights on their own box instead of adding a checkbox to a page.

The rider on the licence is the part I would read carefully. MIT with conditions attached is not MIT, and whether that matters depends on which vendor you are, which is an odd thing to have to work out before compiling a terminal tool.