Late

The deciding trait is that there is nothing to install around it. One compiled binary, no Python environment to keep alive, no Node version to argue with, and it runs from whatever directory you are already standing in. If you have ever lost an evening to a broken virtualenv before writing a line of code, that alone is the pitch.

What you give up is the ecosystem: no plugin catalogue, no extension marketplace, no community recipes to copy. Pick it if you want a small tool that starts instantly. Pick Aider if you would rather have five years of accumulated documentation behind you.

It runs shell commands and edits files, and there is no sandbox and no git integration anywhere in the row. So the safety net is whatever you set up yourself before you start. That matters more here than usual, because the thing making the edit is not the thing that planned it, and it exits as soon as it is done.

There is no unattended mode, so every recovery is manual and interactive by construction. The design does earn one thing: because each edit is scoped to its own worker, a bad run tends to damage one file rather than wander through the tree.

1. The stated failure hypothesis is specific and testable: analysis output, compiler errors and file contents accumulate in a single window until planning quality falls. Naming the mechanism is better than the usual appeal to context length. 2. The remedy is structural rather than prompted. Subagents are enforced, one per atomic edit, so tool output terminates in a process that is discarded instead of returning to the planner.

2. Nothing is measured. The design argument is coherent, and coherence is not evidence; a context-degradation claim is unusually easy to test and untested here.

One author, no company, no entity of any kind: this is a personal project with 415 stars and no mechanism for turning attention into time. That is not a criticism of the code, it is a description of the risk, and the risk is a single calendar.

Moat: none, and none available at this scale. Nothing here is acquirable except the author, which is how most tools in this shape end, quietly, as a hiring event somewhere else. Pricing power: zero by design, since there is nothing to charge for. Position: use it, vendor the binary, and do not assume next year's release exists.

Distribution is the easy part: a Homebrew formula and a curl script put it on sixty laptops in an afternoon, and the licence line costs nothing. Then it stops being easy. There is no directory integration, no audit trail, no documented retention policy, and no console that tells me which of sixty engineers ran what against which provider.

It does not execute in a pipeline, so it cannot be a gate on anything I measure. Onboarding is genuinely cheap, one binary and a key. Not yet: I would need a central key issuance story and something written down about where prompts go before this passes review.

The licence is the one thing I cannot check off. GitHub reports a non-standard file rather than an SPDX identifier, which means I do not know what a fork is allowed to be until someone reads it properly. Everything else here is exactly my shape: my key, my endpoint, my choice of provider.

Local models are the point rather than an afterthought, and the whole thing is Go, so reading the source is a single tree with no framework underneath it. There is no MCP client, so servers I run stay unreachable from here. I would take that trade for a binary I can grep.