terryso/backpack-grid

最近记录的来源活动: 2026年10月4日 04:04

SkillsMP 收录数据更新: 2026年11月12日 12:52

已收集 skills: 1

GitHub 星标: 4
GitHub Forks: 4

用 AI 助手安装

复制这段 Prompt，发给你正在使用的 AI 助手。

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/terryso/backpack-grid.

仓库 README

仓库 README 来源：README.md

查看 README 源文件

展开完整 README

# Backpack Grid Autopilot **[English](README.md) | [中文](README.zh-CN.md)** An autonomous crypto grid-trading system that designs its own strategy, runs on a **live Backpack perpetual-futures account with real money**, and re-balances itself every 15 minutes — no human in the loop. > **📌 Account prerequisite**: Backpack perp grids are **only available on Bot-type subaccounts** — regular subaccounts have no grid feature and cannot be managed by this system. Every grid create/stop call here runs against a Bot subaccount; to replicate, log that bot account into the ego lite browser so the system can reuse its session. <p align="center"> <a href="https://grid.terryso.dev"> <img src="docs/dashboard.png" alt="Live dashboard" width="860"> </a> </p> 🔗 **[Live dashboard](https://grid.terryso.dev)** — real equity, positions, per-grid health and volume progress, updated every 15 minutes. Public, read-only. The screenshot above is re-captured daily at 09:00 by a scheduled job and committed automatically when it changed. > ⚠️ This is a personal live-trading experiment, not financial advice. Perps trading is extremely risky. 🎁 Running this on Backpack yourself? [Sign up with my referral link](https://backpack.exchange/join/suchuanyi) — it costs nothing extra and supports the experiment. ## What it does Backpack provides native grid bots with exchange-side protection. This project adds selection, rotation, account risk rules, and recovery around those bots: - **Runs its own strategy**: scans every USDC perp, scores markets by mean-reversion suitability (multi-window chop × liquidity − drift − funding), and picks up to 4 neutral grids - **Trades with real money**: ~$590 account, 4 grids, automatic rotation when a grid hits +5% take-profit, −6% stop-loss, exits its range, or drifts near liquidation — plus position-level auto-exits at ±150% margin-relative PnL and dwell-tracked range-breach stops - **Protects itself in layers**: exchange-side native backstop on every grid (survives a dead machine) → portfolio-level risk budget → equity-drawdown circuit breaker (90% warn / 100% kill-switch, deliberately pinned to ride through drawdowns) → write-ahead intent ledger so a crashed process never leaves half-done trades - **Reports publicly**: a Cloudflare Worker + KV dashboard updated after every round It has been running unattended since 2026-09-30 — hundreds of autonomous rounds, including riding out a violent market crash (2026-10-08) and its fully automated recovery. Live equity and a certified accounting window are on the dashboard; no performance numbers are frozen in this README. ## Architecture No API keys, no trading library — the system drives Backpack through the **same authenticated session the web app uses** (browser-context `fetch` with cookies), executed by a local launchd job every 15 minutes: ``` observe session-authenticated APIs: grid ledgers, positions, equity, fills decide risk gates (breaker / pending-cleanup / backstop reconciliation / orphans) → per-grid rules (TP / SL / range-exit / liquidation distance) → replacement planning (score gate, ecosystem cap, forward risk budget) act validate → create (with native TP/SL written at creation) → verify orders placed → on any failure: write-ahead ledger keeps the state recoverable verify re-observe; every number in state files is shape-checked, fail-loud ``` Engineering details that survived 14 rounds of adversarial AI code review (~45 real defects found and fixed): - **Two-phase execution**: risk exits execute immediately; replacement creation waits for a re-observation — a stop is never delayed by market analysis - **Write-ahead intent ledger**: disable/delete/create intents persist before any exchange write; a lost response can always be recovered - **Shrink-to-fit risk budget**: if the portfolio stop-loss budget can't fit a full-size grid, it creates a smaller one instead of skipping - **Kernel flock, whole lifetime**: all entries (scheduled, manual, peak probe) share one `fcntl` lock whose fd is passed to the business child (deeper descendants depend on their runtime) — a killed wrapper can't release the lock while the round is still running; corrupt state files quarantine new risk instead of silently resetting - **Budget semantics note**: the forward position budget (≤ equity×80%) and the peak-drawdown breaker are two DIFFERENT mechanisms — the former is a nominal constraint that loosens as profits grow, the latter is the true equity-floor guarantee ## Selection logic Markets are scored by **oscillation suitability** — grid trading earns from round-trips, not trends: ``` score = min(chop, 20) × 0.9 + range24 × 40) × liquidity − drift24 × 25 − funding penalty chop = mean(24h path ÷ 24h net drift, 72h path ÷ 72h net drift) # capped to tame one-day spikes ``` Filters: volume floor, score floor (an empty slot beats a mediocre grid), max 2 grids per ecosystem (no four-correlated-coins blowups), per-grid health badges computed from live distance-to-trigger. ## Honest caveats - **Short track record.** Live since 2026-09-30, through at least one violent crash and its automated recovery. Performance lives on the dashboard (including a certified accounting window) instead of being frozen in this file - **No backtest yet.** Fixed-grid vs. rotation comparison, per-strategy attribution and a proper equity benchmark are the next milestone after the current campaign ends - **Browser-session auth** means the automation dies if the login session dies — by design (no keys to leak); it fails loud and waits ## Testing ```bash npm test # full offline suite: 11 suites, 430+ scenarios, no network ``` Layer A: pure-logic replicas + source canaries. Layer B: **executes the production code** — decide.cjs runs against sandbox fixtures via `BG_ROOT`, and the grid-stop state machine runs against mocked exchange I/O. ## Operations Single-command deployment of the dashboard (`scripts/deploy_dashboard.sh`, Cloudflare OAuth), launchd timers for the 15-minute round and the 1-minute peak sampler plus a resident keep-awake agent (idle sleep would silently stop the whole trading loop), full runbook in-repo — including post-reboot recovery (the ego browser task space is lost; recreate it programmatically, then one manual login re-attaches the session — see the Chinese runbook's "ego 空间丢失恢复"). See [README.zh-CN.md](README.zh-CN.md) for the complete Chinese operations manual. ## Validation and accounting Use Node 22 and `npm ci && npm test`. The offline suites forbid network I/O and execute production decisions, complete mocked act/runner lifecycles, history pagination, and state contracts. Kernel locks are tested with real disposable processes. Historical fills and research refresh independently of risk exits. Equity change is not labeled strategy PnL without reconciled cashflow coverage and an exact baseline timestamp — the homepage publishes a certified accounting window, and every capital/platform event inside the window must be manually attributed in `state/attribution_ledger.json` or the metric honestly falls back to "pending review". Public-API fetches from Node need an explicit proxy (`BG_PROXY` in `.env`) because Node ignores system proxies. Run statistics come from confirmed execution events. Static UI requests bypass Workers; a dated, visibly degraded deployment snapshot is shown if dynamic APIs are unavailable. Upload success requires snapshot readback. See [acceptance and outstanding external gates](docs/project-review-2026-10-02.md). ## Support If this project was useful and you'd like to try Backpack, signing up via my referral link — <https://backpack.exchange/join/suchuanyi> — supports future experiments at no extra cost to you.

正在加载仓库 README…

这个仓库中的 skills

职业分类待补全

skill

职业分类

描述

更新

backpack-grid-ops

未分类

Backpack 网格巡检系统（~/CascadeProjects/backpack_grid）的运维操作——启动/停止/重启定时巡检、部署或更新线上仪表盘、手动跑巡检与测试、处理 OBSERVE_FAILED / ACT_FAILED / 熔断等故障。当用户要求管理 backpack_grid 项目、看巡检状态、启动定时服务、部署网页时使用。

2026-10-04

已展示 1 / 1 个已收集 Skill。