Nezha

The deciding trait is the badge. When a session stops and wants a human, you are told, and that single behaviour is what turns three parallel agents from a source of anxiety into something you can actually supervise. Everyone who has left an agent waiting for twenty minutes on a yes-or-no question knows the value of it.

What you should not expect is a place to write code. The bundled editor is deliberately small and you will keep your real one open beside it. Pick it to watch agents. Pick Nimbalyst if you want the editors too.

The exposure is upstream. Sessions are discovered rather than owned, which means the format, the on-disk layout and the lifecycle all belong to somebody else and can change in a release nobody here was consulted about. A supervisor that stops seeing its subjects is worse than no supervisor, because you trust it until the moment it goes quiet.

What it does right is the review surface. Finished work is rendered for reading and can be resumed, so a session that went wrong is examined rather than guessed at.

1. The symlink choice is the most considered thing in the design. Duplicated instruction files diverge silently, and divergence in the material an agent reads produces behaviour differences that look like model variance and are not. A single source with links is the cheapest correct answer to that problem. 2. Worktrees give each session its own checkout, so concurrent work is isolated by the version control system rather than by convention.

2. No evaluation accompanies the tool, and its claims are about management, which does not need one.

The structural fact is that this is a complement, not a product. Its value moves with the tools it wraps, and if either of those ships its own multi-session manager the reason to run this narrows overnight. Nearly two thousand stars proves an audience and finances nothing.

Moat: none available to a wrapper. Likely path: a hobby project that stays excellent while the author stays interested, or gets outrun by a first-party feature. Position: fine to use, wrong to standardise a workflow on, and the second one is the mistake people make.

Licensing costs nothing and the operational cost is sixty machines. There is no administrative console, no single sign-on, no directory and no audit record, so configuration and version drift fall to whatever desktop management we already run. Every workstation also needs a working agent installation before this is worth opening.

More decisively, it does not execute unattended, so it cannot become a pipeline step or a metric. That makes it a personal productivity choice rather than infrastructure. Not yet: I have no objection to individuals installing it, and no case for standardising on it.

A copyleft licence on a supervisor is the right call, because the thing watching my agents should not become a closed box in someone else's release. Credentials never enter this app; they stay where I put them, in the underlying tool, which means the attack surface it adds is a UI rather than a secret store.

The quarantine flag on the macOS build is a two-second fix and a signal: this is somebody's project, not a signed corporate artefact. I prefer the honest version.