ccswarm

The deciding trait is how little it asks of you. During a run the only keys that do anything are y and n. That sounds like a gimmick until you have spent an evening babysitting an agent through a dozen prompts, each phrased differently, each demanding a decision you were not ready to make. Two keys is a stance about attention.

The flip side is that you steer through a YAML file before the run and barely at all during it. Pick it if you like deciding once. Pick an interactive agent if you like arguing.

The dependency is total. This drives an external command line and the row says plainly that code changes come from that program, not this one. Output parsing against a tool whose interface belongs to somebody else is the most fragile contract in software, and nothing here records a supported version, a compatibility matrix, or what happens when the format shifts under it.

What it does right is the doctor command, which probes for the programs it depends on before a run instead of discovering the absence halfway through one.

1. Naming review and fix as distinct stages means the design treats a first attempt as a draft, which is the correct assumption and is rarely made explicit. 2. Inserting a consensus step before implementation borrows a deliberation metaphor; the row does not say how agreement is decided or what happens when it fails, which is where such schemes usually break.

2. No evaluation is offered and no capability claim is made, so nothing needs defending. A stage diagram is an argument about process, and it is presented as one.

Thin layers above popular tools are the most common shape in this category and the least defensible. The layer is valuable exactly until the tool underneath ships the same idea, and workflow orchestration is an obvious thing for a well-funded vendor to absorb. A hundred and fifty stars gives no negotiating position and no revenue to defend.

Moat: none. Likely acquirer: nobody buys the wrapper; the platform ships the feature. Likely path: useful for a year, then redundant. Position: use it, and expect to stop.

A replayable line-delimited record of what happened is the artefact my incident reviews actually need, and rollback being a documented operation rather than a git lecture is worth more than any demo. It runs unattended, so it becomes a measurable stage, and it can open a pull request into the review process we already have.

The licence costs nothing across sixty engineers; the provider subscriptions underneath cost plenty, and that is the number to budget. No identity layer, no central policy. Approved with conditions: pipeline use only, credentials issued centrally.

Permissive licence, Rust source, and installation is a build from the repository rather than a curl piped into a shell, which I appreciate more than the authors probably realise. A fork is viable and the workflow definitions are plain files I can version alongside the code they act on.

Where it loses me is autonomy. My own weights are not a destination, and there is no protocol port for the servers I already run, so the interesting half of the system is a program I did not write and cannot change.