openswarmlisted
Install: claude install-skill surlymochan/swrm
# swrm
**Division of labour:** Claude decomposes work, writes acceptance criteria, reviews, and integrates.
Executors write code: **Codex** is the only backend in the default automatic routing order. **OpenCode Go** and **DeepSeek** are opt-in through `--backend` or by adding them to `routing.default_order`; when enabled, OpenCode Go handles cheap/mechanical work, non-GPT second opinions, overflow and large context, while DeepSeek is metered. **Pi** (any provider it is configured for, such as MiniMax) and **Cursor Agent** are opt-in, as one-shot backends or resident lanes.
Claude Pro is the tightest quota here, so keep Claude's own tokens for planning and review and send the typing to executors.
CLI: `~/.claude/skills/openswarm/bin/openswarm` (alias it as `XS` in your commands). Runtime data lives in `~/.claude/openswarm/`.
## 0. Preflight (once per session, and again before any big fan-out)
```bash
XS=~/.claude/skills/openswarm/bin/openswarm
$XS status --refresh
```
This shows plans, quota windows (used %, reset time, burn pace), the live model catalogue with tiers, open breakers, the **recommended concurrency** and the **Claude review mode**.
If the Claude line says `STALE` and a usage tool is available in your client (for example Claude Desktop's `get_usage`), call it. When it returns plan percentages, record them:
`$XS ingest-claude --five-hour <pct> --seven-day <pct>`. The statusline hook refreshes this automatically in terminal sessions.
## 1. Decompose: one task