Byted Las Pdf Parse Doubao

Parses and reads PDF documents into structured Markdown text using Volcengine LAS Doubao AI models. PDF parsing, PDF OCR, and document recognition — extracts...

Security Scan

The skill largely does what it claims (PDF → Markdown via Volcengine LAS), but several inconsistencies and risky behaviors—undeclared required credentials, reliance on a remote wheel URL for installation, and implicit dependence on the lasutil CLI—make the package suspicious and worth caution before use.

Purpose & Capability

The SKILL.md declares this is a wrapper around the lasutil CLI and requires LAS_API_KEY/LAS_REGION and sometimes VOLCENGINE_ACCESS_KEY/VOLCENGINE_SECRET, but the registry metadata lists no required env vars or binaries. The skill implicitly requires lasutil, python3, jq and pip, yet 'required binaries' is empty. This mismatch between declared metadata and actual needs is incoherent and unexpected.

Instruction Scope

Runtime instructions and scripts read .env files, source a local env_init.sh (which may install/upgrade an SDK), instruct the agent to upload local files to TOS and to prompt the user for API keys if missing. The flow requires sending PDF data to Volcengine endpoints (expected for the stated purpose), but also instructs or automates storing/reading keys from local files (env.sh/.env) which increases risk of accidentally persisting secrets. The SKILL.md also mandates specific UI wording for billing statements—an operational constraint but not a security issue.

Install Mechanism

Although the registry lists 'no install spec', scripts/env_init.sh will fetch a remote manifest and pip-install a wheel from https://las-ai-cn-beijing-online.tos-cn-beijing.volces.com/... . That causes arbitrary remote code to be written/executed in the agent environment at runtime. The wheel is hosted on a non-standard third-party URL rather than a widely auditable release host (e.g., PyPI or GitHub releases). This is higher risk than an instruction-only skill that does not fetch code.

Credentials

SKILL.md requires LAS_API_KEY (and LAS_REGION) and sometimes asks for VOLCENGINE_ACCESS_KEY / VOLCENGINE_SECRET_KEY when downloading outputs from TOS. None of these are declared in the registry metadata. Requiring multiple cloud credentials (and instructing the user to write them into a local env file) is disproportionate to a simple parser wrapper unless explicitly documented in metadata and privacy/security policy.

Persistence & Privilege

The skill does not request 'always: true' or system-wide privileges. It does include a background polling script (for long-running jobs), but that is a normal component for an async API wrapper. The skill does create a temporary workdir and may create a virtualenv in project root (.las_venv) if missing—this is standard but worth noting.

Scan Findings in Context

[remote-pip-install-from-unfamiliar-host] unexpected: env_init.sh downloads and pip-installs a wheel from a non-standard domain (las-ai-cn-beijing-online.tos-cn-beijing.volces.com). Installing remote code at runtime is not strictly necessary to 'explain how to call an API' and increases risk. If you trust the vendor, this may be acceptable; otherwise it is unexpected.

[uses-external-cli-labeled-as-required-in-docs] expected: The skill wraps the 'lasutil' CLI (submit/poll/file-upload/pdf-pages), which is expected for a CLI wrapper. However, the registry metadata did not list lasutil or other binaries as required, creating an inconsistency.

[reads-local-env-files-and-requests-secrets] unexpected: The skill instructs sourcing .env or creating env.sh with LAS_API_KEY and will read them automatically. Asking users to put API keys in a file in the current directory is a convenience but increases the chance of credential leakage; this behavior was not declared in registry metadata.

What to consider before installing

This skill appears to implement a legitimate PDF→Markdown wrapper around Volcengine LAS, but there are important mismatches and risky behaviors you should consider before installing or running it: - Metadata mismatch: The registry says no required env vars/binaries, but SKILL.md and scripts require LAS_API_KEY, LAS_REGION (and potentially VOLCENGINE_ACCESS_KEY/SECRET) and rely on the lasutil CLI, python3, jq, and pip. Verify you are comfortable providing those credentials and that lasutil is available. - Remote code install: The env_init.sh script may pip-install a wheel from a vendor-specific URL at runtime. If you don't trust that host, do not run env_init.sh; instead ask the skill author for a vetted, signed release or prefer an SDK from an auditable source (PyPI/GitHub releases). Run the install steps in an isolated environment (container or VM) if you need to test. - Secrets handling: The skill suggests storing API keys in .env or env.sh files. Avoid putting long-lived secrets into chat or non-encrypted files. If possible, create a scoped/temporary API key for testing and rotate it afterward. - Operational safety: The skill uploads local files to Volcengine/TOS for processing; ensure you are allowed to transmit that data to the provider and that no sensitive data leaks. Confirm retention/expiration of pre-signed URLs. - Ask the author: Request clarification about the manifest/wheel URL, why the registry metadata omits required env vars/binaries, and whether the wheel can be obtained from a standard package host. If the author cannot justify these inconsistencies, treat the skill as risky. If you decide to proceed, run it in a sandboxed environment, supply only limited-scope credentials, and inspect the contents of any downloaded wheel before executing it.

LAS AI PDF 解析（las_pdf_parse_doubao）

将 PDF 解析为 Markdown 格式文本，保留标题/段落/表格结构。支持 normal（快速）和 detail（深度分析）两种模式。

设计模式

本 skill 主要采用：

• Tool Wrapper：封装 lasutil CLI 调用
• Pipeline：包含 Step 0 → Step N 的顺序工作流

核心 API 与配置

• 算子 ID: las_pdf_parse_doubao
• API: 异步（submit → poll）
• 环境变量: LAS_API_KEY (必填)

详细参数与接口定义见 references/api.md。

Gotchas

• 模式选择：normal 速度快，推荐默认；detail 适合复杂排版文档，耗时更长且费用翻倍。
• 密钥安全：若聊天框屏蔽密钥，让用户在当前目录创建 env.sh 并写入 export LAS_API_KEY="..."，SDK 会自动读取。
• 免责声明：最终回复结果时必须包含："本方式的计费均为预估计费，与实际费用有差距，实际费用以运行后火山产生的账单为准。计费说明请参考 Volcengine LAS 定价。"，且禁止使用"实际费用"字眼描述预估价。

工作流（严格按步骤执行）

复制此清单并跟踪进度：

执行进度：
- [ ] Step 0: 前置检查
- [ ] Step 1: 初始化与准备
- [ ] Step 2: 预估价格
- [ ] Step 3: 提交任务
- [ ] Step 4: 异步查询
- [ ] Step 5: 结果呈现

Step 0: 前置检查（⚠️ 必须在第一轮对话中完成）

在接受用户的任务后，不要立即开始执行，必须首先进行以下环境检查：

1. 检查 LAS_API_KEY 与 LAS_REGION：确认环境变量或 .env 中是否已配置。
   • 若无，必须立即向用户索要（提示：LAS_REGION 常见为 cn-beijing）。
   • 注意：LAS_REGION 必须与您的 API Key 及 TOS Bucket 所在的地域完全一致。如果用户中途切换了 Region，必须提醒用户其 TOS Bucket 也需对应更换，否则会导致权限异常或上传失败。
2. 检查输入路径：
   • 如果用户要求处理的是本地文件，则需要先通过 File API 上传至 TOS（只需 LAS_API_KEY，无需额外 TOS 凭证）。
   • 如果算子的输出结果存放在 TOS 上，且用户需要下载回本地，则需要 VOLCENGINE_ACCESS_KEY 和 VOLCENGINE_SECRET_KEY。对于仅需要上传输入文件的场景，TOS 凭证不再必须。
3. 确认无误后：才能进入下一步。

Step 1: 初始化与准备

环境初始化（Agent 必做）：

# 执行统一的环境初始化与更新脚本（会自动创建/激活虚拟环境，并检查更新）
source "$(dirname "$0")/scripts/env_init.sh" las_pdf_parse_doubao
workdir=$LAS_WORKDIR

如果网络问题导致更新失败，脚本会跳过检查，使用本地已安装的 SDK 继续执行。

• 处理本地文件时：使用 File API 上传（只需 LAS_API_KEY，无需 TOS 凭证和 Bucket）：

  lasutil file-upload <local_path>
  

  上传成功后返回 JSON，取其中的 presigned_url（HTTPS 预签名下载链接，24 小时有效）传给算子作为输入 URL。

Step 2: 预估价格（⚠️ 必须获得用户确认）

1. 读取 references/prices.md 获取最新计费标准。
2. 获取 PDF 页数：

   lasutil pdf-pages <pdf_url>
   

3. 根据 parse_mode 单价计算总价，将计费单价与预估总价一并告知用户并强制暂停执行，明确等待用户回复确认。在用户明确回复"继续"、"确认"等同意指令前，绝对禁止进入下一步（执行/提交任务）。提示：预估仅供参考，实际以火山账单为准。计费说明请参考 Volcengine LAS 定价。

Step 3: 提交任务 (Submit)

构造基础 data.json：

{
  "url": "<presigned_url>",
  "parse_mode": "normal"
}

data=$(cat "$workdir/data.json")
lasutil submit las_pdf_parse_doubao "$data" > "$workdir/submit.json"

⚠️ 强制反馈：任务提交成功后，必须立即向用户返回生成的 task_id，以便用户跟踪进度或在必要时手动查询。

Step 4: 异步查询 (Poll)

⚠️ 异步任务与后台轮询约束：

• 如果你当前的环境支持后台任务/异步长效运行：你可以利用环境提供的后台能力（例如发起后台轮询任务），并在任务完成后主动将结果返回给用户。
• 如果你当前的环境不支持长效后台任务（如普通的单轮对话沙箱），且直接 sleep 循环会导致超时崩溃：绝对禁止在代码中执行死循环等待！ 此时必须立即向用户输出 Task ID 并结束当前轮次，告知用户："任务已提交，请稍后向我询问进度"。

mkdir -p "./output/{task_id}"
lasutil poll las_pdf_parse_doubao {task_id} > "./output/{task_id}/result.json"
# 若解析完成，提取 Markdown
cat "./output/{task_id}/result.json" | jq -r '.data.markdown // empty' > "./output/{task_id}/parsed.md"

• COMPLETED → 结果已保存至 ./output/{task_id}/parsed.md。
• RUNNING/PENDING → 稍后重试。

Step 5: 结果呈现

处理结果：

# 保存结果到本地
mkdir -p "./output/{task_id}"
cat "./output/{task_id}/result.json" | jq -r '.data.markdown' > "./output/{task_id}/parsed.md"
cat "./output/{task_id}/result.json" | jq '.data.detail' > "./output/{task_id}/pages_detail.json"

上传结果文件（可选）：

# 上传 Markdown 文件供用户下载
lasutil file-upload "./output/{task_id}/parsed.md"
lasutil file-upload "./output/{task_id}/pages_detail.json"

向用户展示：

1. Markdown 内容摘要（前 500 字）
2. 总页数
3. 本地文件路径：./output/{task_id}/
4. 签名下载链接（如上传成功）
5. ⚠️ 图片链接有过期时间，如需保留请立即下载
6. 计费声明

审查标准

执行完成后，Agent 应自检：

1. 环境变量是否正确配置
2. 输入文件是否成功上传
3. 输出结果是否正确呈现给用户
4. 计费声明是否包含