sno-ai/sno-station-skills

Sno Station Skills – What they are

Sno Station Skills is a collection of ready‑to‑use skills (small, self‑contained programs) that you install into existing AI‑coding agents such as Claude Code and Codex CLI. The skills give those agents the habits of a real software‑engineering team – peer review, hand‑offs, quota monitoring, nightly retrospectives, etc. The repository does not contain a new model; it augments the behaviour of the agents you already run.

---

How it works (high‑level)

1. Install the sno command (a tiny helper script). It discovers the agents you have (Claude, Codex, OpenClaw, Hermes) and copies the skill binaries into each agent’s environment.
2. Each skill declares the resources it needs (shell, background process, a hook that adds context before every turn). During sno setup the helper either installs the full skill, a reduced version, or skips it if the agent cannot satisfy the requirements.
3. At runtime the agents call the skill via the sno <name> CLI. Skills can read/write a shared workspace, send messages through Sno Reach (a lightweight inbox system), and invoke the other vendor’s CLI to get an independent review.
4. Updates are automatic – the sno binary checks GitHub for newer releases and upgrades itself.

---

What you get

Category	Example skill	What it adds to an agent
J – Peer Review & Audit	peer-review	After a change, the opposite‑vendor agent runs a fresh session, checks the code (or spec) against real inputs, and returns a structured report.
	less-is-more	Finds the smallest implementation that satisfies the tests, removing unnecessary scaffolding.
	owner‑intent‑audit	Verifies that a plan or diff actually matches the original intent, catching hidden drift.
M – Build & Rollout	medic	One‑line health check of tools, installed skills, quota visibility, and temporary‑space availability.
	sno-cli	Provides the agent with an up‑to‑date command reference so it can install, update, or configure itself without hard‑coding flags.
S – Communication & Handoff	handoff	Captures the full state of a running task (uncommitted changes, decisions, limits) and hands it to a fresh agent, guaranteeing only one owner edits at a time.
	rotate-agent	Monitors subscription quota; when it falls below a threshold it automatically spawns the other vendor’s agent and transfers the work.
	heartbeat	Sends periodic progress messages into the agent’s conversation window, so you see live updates instead of digging through logs.
	reach / join-talk	Lightweight messaging layer that lets agents address each other by a simple address, without a persistent server.
H – Reporting & Visibility	catch-report	Counts how many defects each agent caught, giving you a quantitative view of cross‑vendor review effectiveness.
T – Delegation & Work Orders	write‑charter (example)	Generates a one‑page brief that defines the scope, success criteria, and cost of a task; any agent can later deliver it and prove each check with a real run.
R – Retrospective & Improvement	rem‑reflect	Nightly loop that reads the day’s sessions, extracts repeat mistakes, proposes rule changes, and waits for your explicit acceptance before applying them.

---

Typical workflow

1. You finish a change → say “peer‑review this” → the opposite‑vendor skill runs, returns a report, you iterate until clean.
2. You start a long build before sleep → “arm rotate‑agent” → the skill watches quota; if it drops, it writes a hand‑off brief and wakes the other agent.
3. You return → “away brief” → a concise 60‑line summary of what was done, what is blocked, and quota usage.
4. You assign work → “write charter” → a structured charter file is created; any agent can later deliver it, and each success check is proven by an actual run.
5. Nightly → rem‑reflect reads all sessions, surfaces recurring bugs, and suggests skill updates; you accept with sno rem‑reflect accept <id>.

---

Who might use this

• Solo developers who rely on a single LLM and want a second‑opinion safety net.
• Teams that already use Claude Code, Codex, or other vendor CLIs and need automated hand‑offs when a subscription quota runs out.
• Ops / MLOps engineers looking for reproducible, auditable pipelines that can survive agent crashes.
• Researchers interested in “dual‑brain” setups – two independent models reviewing each other.

---

Limitations & assumptions (as described in the README)

• Requires the vendor’s CLI to be installed and logged in (Claude Code, Codex, OpenClaw, Hermes).
• Handoff and rotation need a shared filesystem so the progress record can be read by the new agent.
• Skills are self‑tested before each release, but they rely on the underlying agents behaving as documented.
• The system does not create new models; it only orchestrates existing ones.

---

License

Apache‑2.0 – you can use, modify, and redistribute the skills under the same terms.

---

Bottom line: Sno Station Skills is a toolbox that turns LLM‑driven coding agents into a coordinated, self‑checking “team” by adding peer review, quota‑aware hand‑offs, health checks, and nightly retrospectives. It is a genuine software project aimed at improving reliability and safety of AI‑assisted software development.