Browser Use Agent

Browser-Use：把 LLM 变成网页操作员的异步 Python 库（Python 3.11+）。Agent 步循环采集 DOM + 截图 → LLM 一次调用产出 thinking / evaluation / next_goal / action[] → 经 CDP 执行。 Browser-Use: a...

Security Scan

The skill claims to be a browser-automation knowledge pack, but its included seed.yaml and runtime instructions contain unrelated finance blueprint preconditions and local system checks (python import checks, ZVT data dir, install triggers), creating mismatches and unexpected filesystem/command actions — review before use.

Purpose & Capability

SKILL.md and human_summary advertise a browser-agent (web automation) skill, but the referenced seed.yaml is compiled as 'finance-bp-133' with ZVT/backtesting preconditions and finance-specific blueprints. That financial blueprint content (zvt checks, MACD defaults, finance rules) does not align with a pure browser automation purpose and suggests the package wires in unrelated capabilities.

Instruction Scope

Runtime instructions require the host AI to 'read references/seed.yaml' and follow an execution_protocol that mandates reloading seed.yaml, running preconditions (python import and filesystem checks), and executing install/verification steps. These instructions direct the agent to run local shell/python commands and reference host filesystem paths (e.g., ZVT_HOME, host_workspace), which is broader than a simple knowledge-only browser doc and risks unexpected local actions.

Install Mechanism

The skill is instruction-only (no install spec), which normally is low-risk. However seed.yaml contains an install_trigger and expects 'resources.host_adapter.install_recipes[]' and import verification; this is inconsistent with 'no installation required' in SKILL.md and could prompt the host to perform installs via its own adapter. No explicit external download URLs are present in the skill bundle itself.

Credentials

Declared requirements list no env vars, yet the execution protocol and preconditions reference environment/state (ZVT_HOME, local writable directories) and run Python import checks. The skill also covers use cases that may handle sensitive data (auto-fill forms, checkout, password manager paths & negative terms mention '1password'). Requesting no credentials while instructing the agent to access local paths and run checks is inconsistent and disproportionate.

Persistence & Privilege

always is false and autonomous invocation is allowed (platform default). The seed.yaml's rule 'On any behavioral decision, agents MUST re-read seed.yaml' is aggressive (forces frequent re-loads) but does not equate to always:true or system-wide config modification. This increases runtime footprint but is not itself an explicit privilege escalation.

Scan Findings in Context

[regex-scan: none] unexpected: The regex scanner found no code patterns (the skill is instruction-only). Absence of matches is not reassuring here because the included seed.yaml contains instructions that will cause the host to run checks/commands and reference local state; those behaviors are visible only in SKILL.md/seed.yaml, not in code.

What to consider before installing

This skill mixes a browser-automation description with a large, unrelated finance blueprint that instructs the host to run local python checks and read/write host paths. Before installing or invoking it: 1) Ask the publisher/author where this package came from and why finance seed.yaml is bundled with a browser agent. 2) Inspect references/seed.yaml fully (especially preconditions and install_trigger) and confirm you are comfortable with any commands it would make your host run. 3) Do not provide secrets or credentials; avoid running in production or on machines with sensitive data. 4) Prefer skills with a clear homepage/source repo and explicit, minimal install steps; if you must test, run in an isolated environment (VM/container) with no access to personal files. If you want, I can list the exact precondition commands and file paths that the seed.yaml would require the host to run so you can review them line-by-line.

这个 skill 适合什么用户？能做哪些任务？

概览

Browser-Use 是把 LLM 变成网页操作员的异步 Python 库（github.com/browser-use/browser-use）。Agent 步循环：(1) 通过 14 个 watchdog 围绕 bubus.EventBus 采集 BrowserStateSummary（带数字索引的 DOM、截图、tab 列表、页面状态）；(2) 一次 LLM 调用同时产出 thinking + evaluation_previous_goal + memory + next_goal + action[]；(3) 经 CDP 原语在双层 page-change 守卫下执行动作。

CD...

Doramagic 晶体页: https://doramagic.ai/zh/crystal/browser-use-agent

知识规模

• 40 条约束 (4 fatal + 36 non-fatal)
• 上游源码: browser-use/browser-use @ commit f3878b0e
• 蓝图 ID: finance-bp-133

用法

Host AI（Claude Code / Cursor / OpenClaw）读 references/seed.yaml，按其中的：

• intent_router 匹配用户意图
• architecture 理解项目架构
• constraints 应用 anti-pattern 约束
• business_decisions 参考核心设计决策

FAQ 摘要

这个 skill 适合什么用户？能做哪些任务？

适合做网页自动化的工程师：表单填写、信息抓取、回归测试、跨站点数据采集等。Agent 把视觉理解 + 操作规划合并到一次 LLM 调用，比传统 Playwright 脚本更适合非确定性页面。访问 doramagic.ai/r/browser-use 查看完整用例。

需要准备什么环境？依赖什么？

Python 3.11+，Chromium 系浏览器（local_browser_watchdog 自动启动或通过 cdp_url 接管），至少一个 LLM provider 配置（默认 ChatBrowserUse项目自家微调模型）。要求 async event loop——Agent / BrowserSession 接口是 async-native。

会踩哪些坑？这个 skill 怎么防护？

本 skill 内置 40 条约束（4 条 fatal）。典型踩坑：(1) alert/confirm/beforeunload 自动接受，破坏性确认（如 'Delete this'）也会通过；(2) Agent(sensitive_data=...) 不配 Browser(allowed_domains=[...]) 是 fail-OPEN（只 warning 不 raise），合规场景必须显式

---

完整文档: 见 references/seed.yaml (v6.1 schema). 浏览页: https://doramagic.ai/zh/crystal/browser-use-agent