Metis

The deciding trait is that it takes a piped diff as input. That single choice makes it a citizen of the shell rather than a destination you visit: the output of one command becomes the subject of the next, and the agent slots into scripts and habits you already have instead of asking you to move into its window.

Around that the experience is ordinary and the project is young. Pick it if your workflow is already a chain of pipes. Pick something more finished if it is not.

Recursion plus roles is a cost multiplier before it is a capability. Each role may invoke the set again, and the row records no maximum depth, no budget guard and no rule for detecting two roles handing work back and forth. The bill for that arrives later and the failure is silent while it happens.

There is no container beneath any of it either, so a recursive branch that decides to run something does so on the host. What it does right is naming the roles, which at least makes a transcript readable.

1. Self-run comparisons are not worthless, but they are not results either. The party publishing the figure chose the harness, the model, the opponent and the moment, and any of those four decisions can produce the observed gap without the tool being better. 2. A single backing model means the number describes one pairing, not the agent.

2. The honest form would be a published harness and a reproduction by someone else. Until then this is a claim with arithmetic attached, and the project's own note concedes the verification gap.

Weekly installs are the only number on this board I treat as close to revenue, because somebody has to actually run a command to produce one. Four figures a week from a project with a modest star count means the users are quieter than the audience, which is the healthier of the two imbalances.

Moat: none yet; the memory and role structure are reproducible. Likely acquirer: nobody buys at this size, but distribution like this is how a solo project attracts a first hire. Position: small stake, watch the install curve for two quarters.

The commercial terms are simple because there are none: no licence to buy, no contract, no supplier to escalate to when it misbehaves during a release week. That last part is what my risk register cares about.

Operationally it is a desktop tool. No identity integration, no directory sync, no central record of which repository an agent touched, and no unattended mode, so it never becomes a stage I can gate a release on. Onboarding is quick and the model spend lands on keys we issue. Not yet.

The licence does the heavy lifting: permissive terms mean I can read the source, patch the parts that annoy me, and keep the fork if the author stops. For a project this young that is the only guarantee worth having, and it is the right one to have.

Everything else points outward. Inference always leaves the machine, and the tools I have already built cannot be attached over a protocol, so extending it means editing this codebase. Which the licence permits, and which I would rather not have to do.