Skyportal SRE – 开源的 AI 基础设施工程师

一个开源 AI 基础设施工程师，在生产中断前告诉你发生了什么变化。

Skyportal 通过持续观察部署、Kubernetes 事件、GPU 指标、配置变更、日志和基础设施更新，逐步构建你的 AI 基础设施时间线。它会在整个技术栈中关联这些事件，并解释最可能的根本原因。

不用再翻查几十个仪表盘，直接提问：

• 为什么 GPU 利用率突然下降？

• 延迟翻倍之前发生了什么变更？

• 哪个部署导致了这次回归？

• 为什么这个模型突然变慢了？

• 我们以前见过这个事故吗？

  观察基础设施 → 构建变更时间线 → 关联回归 → 解释最可能的原因

Skyportal 将症状与它之前的变更联系起来。一次诊断可以对比本次部署与上一版本，衡量影响，识别最可能的变更，并给出置信度。

需要 Python 3.11 或更高版本。

pip install skyportalai skyportal

在终端内，运行一次 /login，用 /servers 选择服务器，然后询问发生了什么变化：

skyportal [connected] > diagnose the latest deployment

常用命令：

/login 连接你的 Skyportal 账户 /servers 列出可用的基础设施 /server 选择服务器 /status 显示当前上下文 /new 开始新的调查 /resume 继续上一次调查 /help 显示所有命令

当你需要用 Python 启动或自动化调查时，可以使用 SDK：

python from skyportalai import Skyportal

with Skyportal(api_key="sk-...") as client: chat = client.chat.create_chat( "What changed before GPU utilization dropped?", server_id=12, ) result = chat.wait(on_approval=lambda approval: True) print(result.status)

设置环境变量 SKYPORTAL_API_KEY 代替直接传入密钥。客户端也支持通过 SKYPORTAL_BASE_URL 指向自托管部署。

skyportalai 命令为脚本和 CI 提供稳定的 JSON 输出：

skyportalai chat send --server 12 --wait "Diagnose the latest regression" skyportalai --json chat messages 123

运行 skyportalai --help 获取完整的命令参考。

安装采集器依赖并在实验卷上运行 Agent 之前，请阅读部署指南：

pip install "skyportalai[agent]"

参见 Agent 部署与数据处理。

poetry install --all-extras poetry run pytest poetry run ruff check . poetry check --strict

参阅 CONTRIBUTING.md 以参与贡献。请按照 SECURITY.md 私下报告安全问题。

MIT

An open-source AI infrastructure engineer that explains what changed before production breaks.

Skyportal continuously builds a timeline of your AI infrastructure by observing deployments, Kubernetes events, GPU metrics, configuration changes, logs, and infrastructure updates. It correlates those events across your stack and explains likely root causes.

Instead of searching through dozens of dashboards, ask:

• Why did GPU utilization suddenly drop?

• What changed before latency doubled?

• Which deployment caused this regression?

• Why is this model suddenly slower?

• Have we seen this incident before?

  Observe infrastructure → Build a change timeline → Correlate regressions → Explain the likely cause

Skyportal connects a symptom to the changes that preceded it. A diagnosis can compare a deployment with its previous release, measure the impact, identify the most likely change, and report its confidence.

Requires Python 3.11 or newer.

pip install skyportalai
skyportal

Inside the terminal, run /login once, select a server with /servers, and ask what changed:

skyportal [connected] > diagnose the latest deployment

Useful commands:

/login          Connect your Skyportal account
/servers        List available infrastructure
/server <id>    Select a server
/status         Show the active context
/new            Start a new investigation
/resume         Continue the previous investigation
/help           Show every command

Use the SDK when you want to start or automate an investigation from Python:

from skyportalai import Skyportal

with Skyportal(api_key="sk-...") as client:
    chat = client.chat.create_chat(
        "What changed before GPU utilization dropped?",
        server_id=12,
    )
    result = chat.wait(on_approval=lambda approval: True)
    print(result.status)

Set SKYPORTAL_API_KEY instead of passing a key directly. The client also supports SKYPORTAL_BASE_URL for self-hosted deployments.

The skyportalai command provides stable JSON output for scripts and CI:

skyportalai chat send --server 12 --wait "Diagnose the latest regression"
skyportalai --json chat messages 123

Run skyportalai --help for the complete command reference.

Install the collector dependencies and review the deployment guide before running the agent on experiment volumes:

pip install "skyportalai[agent]"

See agent deployment and data handling.

poetry install --all-extras
poetry run pytest
poetry run ruff check .
poetry check --strict

See CONTRIBUTING.md to contribute. Report security issues privately using SECURITY.md.

MIT