- name
SKILL.md
技能檔案保留在發布者的儲存庫中。
在終端機中貼上複製的指令。它會將這個 SKILL.md 安裝到對應的代理中。
npx skills add galleonlabs/hypergrok-trading-desk --skill hyperliquid-websocket -a cursor -y在終端機中貼上複製的指令。它會將這個 SKILL.md 安裝到對應的代理中。
npx skills add galleonlabs/hypergrok-trading-desk --skill hyperliquid-websocket -a claude-code -y在終端機中貼上複製的指令。它會將這個 SKILL.md 安裝到對應的代理中。
npx skills add galleonlabs/hypergrok-trading-desk --skill hyperliquid-websocket -a codex -y在終端機中貼上複製的指令。它會將這個 SKILL.md 安裝到對應的代理中。
npx skills add galleonlabs/hypergrok-trading-desk --skill hyperliquid-websocket -a grok -y在 Grok Bot 聊天中傳送複製的訊息。Bot 會從 GitHub 取得這個 SKILL.md,並加入你所有 Bot 共用的技能中。
Install this skill: https://github.com/galleonlabs/hypergrok-trading-desk/tree/HEAD/skills/hyperliquid-websocket來自發布者
Read-only, no key. Endpoints: wss://api.hyperliquid.xyz/ws (mainnet), wss://api.hyperliquid-testnet.xyz/ws (testnet). Prefer WebSocket over polling for anything continuous: fills, order updates, book and price watches.
Subscribe: {"method": "subscribe", "subscription": {...}}. Unsubscribe with "method": "unsubscribe". The server acknowledges with {"channel": "subscriptionResponse", ...} then streams {"channel": "<type>", "data": ...}. Send {"method": "ping"} periodically (the SDKs do it for you; the server expects activity within about a minute) and expect {"channel": "pong"}.
Subscription types the desk uses:
Limits per IP: up to 10 connections, 30 new connections per minute, 1000 subscriptions, 10 distinct users across user subscriptions, 2000 messages per minute. The server closes a connection silent for 60 seconds. One connection per watch process is plenty.
You can also send /info requests over the socket: {"method":"post","id":1,"request":{"type":"info","payload":{"type":"allMids"}}} returns {"channel":"post","data":{"id":1,"response":{...}}}. Useful inside a watch to avoid mixing REST and WS.
import os, json, sys, time, signal
from hyperliquid.info import Info
from hyperliquid.utils import constants
NETWORK = os.environ.get("HYPERLIQUID_NETWORK", "testnet")
BASE = constants.MAINNET_API_URL if NETWORK == "mainnet" else constants.TESTNET_API_URL
ADDR = os.environ.get("HYPERLIQUID_ACCOUNT_ADDRESS")
LOG = open("/workspace/trading-desk/watch/ws.log", "a")
def on_msg(msg):
line = json.dumps({"ts": time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime()), "channel": msg.get("channel"), "data": msg.get("data")})
LOG.write(line + "\n"); LOG.flush()
if msg.get("channel") == "userFills" and not msg["data"].get("isSnapshot"):
for f in msg["data"]["fills"]:
print(f"FILL {f['coin']} {f['side']} {f['sz']} @ {f['px']} fee {f['fee']} oid {f['oid']} cloid {f.get('cloid')}", flush=True)
info = Info(BASE) # skip_ws=False starts the socket thread
info.subscribe({"type": "allMids"}, on_msg)
info.subscribe({"type": "l2Book", "coin": "ETH"}, on_msg)
if ADDR:
info.subscribe({"type": "userFills", "user": ADDR}, on_msg)
info.subscribe({"type": "orderUpdates", "user": ADDR}, on_msg) # one orderUpdates/userEvents subscription per Info
signal.signal(signal.SIGTERM, lambda *_: (info.disconnect_websocket(), sys.exit(0)))
while True:
time.sleep(60)
The SDK's manager pings for you but does not reconnect on drop; run it under a supervisor (see below) and treat a silent log as a dead watch. It also only routes these subscription types to your callback: allMids, l2Book, trades, candle, bbo, userEvents, userFills, orderUpdates, userFundings, userNonFundingLedgerUpdates, activeAssetCtx, activeAssetData. Others in the table (clearinghouseState, openOrders, twapStates, userTwapSliceFills, webData3, notification) are acknowledged by the server but silently dropped by the Python SDK; use the raw socket or the TypeScript client for those. The SDK routes one more type, webData2, but do not subscribe to it: the live server rejects that frame on mainnet and testnet with {"channel":"error"} and the watch never starts. webData2 survives only as an /info request; the frontend snapshot stream is webData3, which SDK 0.24.0 cannot route at all.
Run in the background from the desk computer:
mkdir -p /workspace/trading-desk/watch
nohup python3 /workspace/trading-desk/watch/ws_watch.py >> /workspace/trading-desk/watch/ws_watch.out 2>&1 &
echo $! > /workspace/trading-desk/watch/ws_watch.pid
Heartbeat check for a routine: tail -1 /workspace/trading-desk/watch/ws.log should be recent; if the pid is gone or the log exceeds the watch's staleness bound, mark the watch unavailable, restart it and complete the recovery below. A new PID or subscription acknowledgement does not prove that missed account activity has been recovered.
@nktkas/hyperliquid)import { SubscriptionClient, WebSocketTransport } from "@nktkas/hyperliquid";
const isTestnet = (process.env.HYPERLIQUID_NETWORK ?? "testnet") !== "mainnet";
const transport = new WebSocketTransport({ isTestnet }); // auto-reconnect and re-subscribe by default
const subs = new SubscriptionClient({ transport });
const user = process.env.HYPERLIQUID_ACCOUNT_ADDRESS as `0x${string}`;
await subs.allMids((d) => console.log("mids", d.mids.ETH));
await subs.l2Book({ coin: "ETH" }, (d) => console.log("book", d.levels[0][0], d.levels[1][0]));
await subs.userFills({ user }, (d) => { if (!d.isSnapshot) console.log("fills", d.fills); });
const s = await subs.orderUpdates({ user }, (u) => console.log("orders", u), { onError: (e) => console.error(e) });
// await s.unsubscribe(); transport.close();
The callback examples above are live notifications, not a complete account journal. They suppress snapshot alerts to avoid announcing old fills as new, but snapshots still need reconciliation and de-duplication before they can support risk decisions.
Treat account events as prompts to reconcile, not proof of a gap-free history. On a disconnect, stale heartbeat, process restart or uncertain event order, record the last good UTC time, alert unavailable, and pause new-risk sends. Retain pending risk reservations and the daily-stop latch; restarting does not clear either.
/info account state, frontendOpenOrders, and order statuses for every outstanding ticket after reconnect. Recover userFillsByTime from an overlap before the last good time through the new reads, with pagination. At desk startup use the last durable reconciliation time; if it is absent or outside the retained history, say that coverage is unavailable.(time, coin, tid) as the fill identity. Include snapshot fills not already processed. Preserve equal-timestamp boundaries when paging; a saturated page with no forward progress is incomplete evidence, not the end of history.desk-execution-protocol; silence never resolves them.For a public price/book watch, refresh that market's snapshot and reset its condition baseline before reporting an unchanged result. A recovered price feed alone cannot restore account readiness.
websocat wss://api.hyperliquid-testnet.xyz/ws <<'EOF'
{"method":"subscribe","subscription":{"type":"allMids"}}
EOF
A watch is a condition plus an alert (desk-monitoring). Structure every watch as: subscribe, log everything to a file, evaluate the condition on each message, post the alert once (with value, threshold, source, UTC time), then either exit or keep watching, and never call /exchange.
userFills message as new fills; it is a snapshot (isSnapshot: true).T reached) unless you want intrabar updates.@<index> on the wire (except a few like PURR/USDC); resolve via spotMeta.Subscribe to live Hyperliquid data over WebSocket from the desk computer - mids, order book, trades, candles, best bid/offer, and per-account fills, order updates and events - with raw JSON, Python SD
---name: hyperliquid-websocketdescription: Subscribe to live Hyperliquid data over WebSocket from the desk computer - mids, order book, trades, candles…---# skills/hyperliquid-websocket/SKILL.md總覽
Subscribe to live Hyperliquid data over WebSocket from the desk computer - mids, order book, trades, candles, best bid/offer, and per-account fills, order updates and events - with raw JSON, Python SDK and TypeScript examples, plus how to run a supervised watch that logs to a file and alerts. Read-only. Use for monitoring, fill notifications and any watch that polling would make expensive.
常見問題
從發布者下載 skills/hyperliquid-websocket/SKILL.md,或複製路徑後開啟儲存庫。PlugMCP 只收錄這項技能,不代管檔案。
發布者將它存放在 https://github.com/galleonlabs/hypergrok-trading-desk。
Claude、Cursor、Codex 和 Grok Build 可以讀取你從發布者儲存庫加入的 SKILL.md。在聊天中傳送技能的 GitHub 資料夾連結,Grok Bot 就會安裝它。