Your AI agents finally get colleagues

What it does

Sno Station Skills installs a bundle of skills into the agents you already run — Claude Code, Codex, OpenClaw, Hermes Agent. The pitch is that it fixes the team, not the model: one agent builds, a second from a different vendor reviews the work cold, in a fresh session with none of the author’s conversation. When the working agent’s quota runs low, it writes a handover brief, wakes an agent from the other vendor, and hands off at an exact commit. Every task can carry written success checks, each answered pass or fail by a real run — “done” has to be proven, not declared.

The interesting bit

The nightly loop (rem-reflect) reads the day’s sessions, finds repeated mistakes, and proposes lessons and skill changes — but nothing changes until you accept it. That’s a self-improvement mechanism with a human gate, which is the right shape for it.

Key highlights

• 31 skills in 22 units, filed under six team abilities: peer review, build, handoff, reporting, delegation, retrospective
• peer-review runs in the other vendor’s CLI when available; with one CLI it still runs fresh and labels itself same-vendor
• rotate-agent + handoff keep long builds alive across quota limits — the README documents a real overnight shift change with published evidence
• less-is-more cuts agent over-building: it lists unneeded additions first, before suggesting fixes
• Every release is self-tested on a clean copy; a skill that fails isn’t published

Caveats

• The evidence stories come from the authors’ own machine (the Sno Station repo), not independent users — take the numbers as anecdotes, not benchmarks.
• The README is heavy on narrative and light on what each of the 31 skills actually does beyond the handful showcased.

Verdict

Worth a look if you run long agent sessions overnight or want a genuinely independent second opinion on agent-written code. If you use one agent for short tasks with your eyes on it, this is machinery you don’t need.