revmux

The deciding trait is what it will look at. A design document, an implementation plan, a written proposal, all go in the same way a branch does, which quietly makes this the only tool on the board that will argue with you before you have written anything. Getting three opinions on a plan is worth more than getting them on the consequences of a bad one.

It changes nothing and fixes nothing, by design. Pick it for second opinions. Pick something else if you wanted the work done.

The failure mode is upstream and silent. Given a context file that omits the relevant module, the panel will read what it was given and produce a fluent report about an incomplete picture, with nothing in the output signalling that something was missing. There is no validation of the brief and no statement of what the reviewers could not see.

What it does right is refusing to touch source or version control at all, which means a wrong review costs you attention and never costs you a working tree.

1. Aggregation is where multi-model panels usually destroy their own value. Collapsing four positions into one answer hides the variance that made the panel worth running, and preserving the arguments hands the reader the evidence instead of a summary of it. 2. That choice also makes the output auditable: a disagreement is visible rather than resolved by a rule nobody documented.

2. No measurement of panel accuracy is offered, and the design is careful enough not to imply one. It reports positions, and positions are not scores.

Small sharp tools have excellent survival characteristics and terrible commercial ones. There is nothing here that could be sold, which also means there is nothing that could be taken away, withdrawn or repriced, and single-purpose utilities from experienced maintainers tend to still compile a decade later.

Moat: none, and none wanted. Likely acquirer: none; this is not an asset, it is a habit. Likely path: sporadic maintenance, indefinite usefulness. Position: adopt freely, because the downside is a Go binary you stop running.

Structured output is what makes a tool a stage rather than a toy. A report my pipeline can parse goes into a gate, a dashboard or a ticket without anyone reading it first, and that is the only way an opinion becomes a control.

Against that: no published install path, so packaging is ours, and two separate provider subscriptions sit underneath every run, which is the real cost across sixty engineers. No identity integration and no audit record of what was reviewed. Approved with conditions: pipeline use, with the reports retained by us.

A tool that does one job, reads what you hand it and prints a result is a tool I can put in a pipeline, a script, a git hook or a makefile without asking anyone's permission. Permissive terms and a compiled binary mean it is mine to keep, and its refusal to grow features is a design decision I wish more projects made.

The catch is what it supervises: two vendor command lines, neither of which I control, and no route to weights on my own hardware. The wrapper is perfect and the engine is rented.