SourceCraft Code Assistant

The trait that would decide it in daily use is that nothing leaves the platform. Navigation, secret scanning, dependency analysis and the build all sit beside the assistant, so the context it needs is already there and you are never gluing two products together. When that works it is genuinely pleasant.

It also means the tool is unavailable to you unless the whole platform is. Pick it if that decision is already made. Pick GitLab Duo for the same shape with a purchasing path most teams can follow.

The failure mode is at the last step. A single request is described as producing a task, code, tests, a security check, a pull request and a deployment, and the only stated boundary is the pull request itself. Nothing says who approves, what blocks, or what happens when the security check disagrees with the change that just shipped.

What it does right is including that check at all. A scan inside the loop is better than a scan after it, even when the loop's stopping conditions are undescribed.

1. Independence is the whole point of a test. When the artefact under test and the test itself come from one generation process, a shared misunderstanding of the requirement produces a green suite, and the failure is invisible precisely where it matters most.

2. Nothing published addresses this, because nothing published describes the method at all. 3. The entire documentary record is a product page: no model named, no architecture, no benchmark. The documentation is thin enough that every claim here has to be taken on trust.

The company behind this is large, profitable and not going anywhere, so survival is not the question. Reach is. Pricing appears only inside the platform and only for its home market, which means this competes for customers already inside one national cloud ecosystem and is effectively unavailable outside it.

Moat: the platform bundle and the regional position, which is durable and non-expandable. Likely path: continued investment as a strategic asset rather than a growth business. Position: relevant if you operate in that market, irrelevant if you do not, and there is no middle case.

I cannot compute a number. Rates are published inside the platform rather than as a comparable price list, so there is nothing to take to finance and nothing to benchmark against the two alternatives already on my shortlist.

The larger blocker is where our source would sit. Everything runs on the supplier's cloud, which makes residency a legal question rather than a configuration one, and our counsel will answer it before I do. There is no self-hosted option to fall back on. Not yet, and realistically not at all for us.

Everything I test for is absent. There is no repository, so nothing to fork. There is no key substitution, so I cannot pay my own provider. Inference does not run on my machine, and there is no server or client interface for anything I already built to talk to it.

The whole product is a web application, so my configuration options are whatever buttons exist this quarter. I cannot script it, I cannot version it, and I cannot inspect what it did. There is nothing here for me to grudgingly respect.