CoStrict

The deciding trait is that it follows you. The same assistant ships as an editor extension, a plugin for the other editor and a command line tool, so the half of your team that will not leave its IDE and the half that lives in a shell end up behaving the same way. Most tools make you pick a home.

What you do not get is a second opinion on your architecture; it will follow a bad instruction across both editors at once. Pick it if the toolchain split is real. Pick Cline if everyone is in one editor and would rather tinker.

The failure mode is the caller you did not touch. An index built on resemblance returns passages that look like the diff, and the function that will break at runtime is frequently the least similar thing in the repository: another module, different names, quietly invalidated by a signature change. Nothing about it looks like the diff.

No fallback for that case is described, and no false-positive rate is published for the pass at all. What it does right is scope. The review runs over the whole repository rather than the diff alone, which at least makes the attempt that most review bots skip entirely.

1. Naming the stages is the substantive contribution here. Each one produces an output a human can read and reject before the next begins, which makes the intermediate artefacts reviewable instead of internal. 2. The ordering is conventional software process rather than an invention, and that counts in its favour, since the failure modes are already catalogued.

2. No evaluation accompanies any of it. Nothing published measures whether a staged pipeline produces better changes than a single pass, and the claim is plausible enough that somebody ought to have tried to falsify it. Stated clearly, which is more than most manage, and demonstrated nowhere.

The company is the signal. This comes from an established vendor with an existing enterprise motion, which quietly removes every question at the top of my list: no runway problem, no seed-stage abandonment, no founder who takes an offer next quarter and takes the roadmap with him. That is worth more than any feature on the page.

Pricing power sits with the relationship rather than the assistant, and that is the durable half of the arrangement. Moat: a customer base that already buys infrastructure from the same supplier. Likely path is bundling, not acquisition. Position: safe to depend on, and read the commercial terms before the free version becomes load-bearing.

The deployment model is what gets this through procurement. Source stays on infrastructure we control, which removes the single objection that kills most assistants in review, and sixty desks cost nothing. That is a shorter meeting than I am used to having about this category.

What is missing is everything after the install. No single sign-on, no user provisioning and no audit trail are documented, so I would be running a service with no idea who used it or what it read. It does not execute in a pipeline either, so it never becomes a step I can measure. Approved with conditions: one team first, behind our own identity proxy.

The licence is the permissive kind, which means a fork is a real option rather than a threat I make in an issue thread. The model configuration matters more to me: it takes any endpoint speaking the common API format, so a server on my own hardware is a first-class choice instead of a footnote.

The free models it ships are convenient and the first thing I would turn off, because I would rather know which weights are reading my repository. MCP tools attach directly, so the servers I already run become available without writing glue. Grudging respect for a vendor tool that lets me remove the vendor.