Rewrite Question

补全上下文并重写用户提问

Security Scan

Skill behavior mostly matches its rewrite purpose but it silently inherits and exposes the agent's environment (.env and all env vars) to a spawned Python process and performs network calls — this mismatch in declared vs. actual environment access is a concern.

Purpose & Capability

The skill's name/description (rewrite an incomplete user query using conversation history) aligns with the included code: index.js invokes rewrite_question.py which implements LLM-driven rewrite logic and QA matching. It reasonably accepts optional Gemini API parameters. Minor inconsistency: SKILL.md frames the skill as instruction-only but the package actually contains runnable code that will be executed.

Instruction Scope

The SKILL.md and code instruct the agent to run a local Python script that performs LLM calls and QA SQL time-updates. index.js also reads a parent .env file (../.env) if present and will pass the full process environment into the child process, which means the Python code can access any environment variables available to the agent. That behavior (silent loading/exposure of environment secrets) expands scope beyond the simple rewrite task and is not declared in the skill metadata.

Install Mechanism

There is no install spec or remote download; the skill is delivered as local code files and executed via spawning the system Python. No archive downloads or package installs were found in the manifest.

Credentials

The skill declares no required env vars, yet index.js will use GEMINI_API_URL / GEMINI_API_KEY / GEMINI_TOKEN if present and will also load ../.env and forward the entire process.env to the Python process. This implicit, wide environment access (including unrelated secrets) is disproportionate to a straightforward rewrite task and not surfaced in metadata.

Persistence & Privilege

The skill is not marked 'always:true' and does not request persistent system-level privileges. It communicates via stdio when invoked by the agent; no evidence it modifies other skills' configs or requests permanent presence.

What to consider before installing

This skill appears to implement the advertised rewrite functionality, but it will: (1) read a parent ../.env file if present, and (2) pass the whole process environment into the spawned Python process, giving the Python code access to any environment variables or secrets available to the agent. The Python code also imports httpx and will make network calls (expected for LLM/API usage). Before installing, consider: - Inspect rewrite_question.py fully to see what external endpoints it calls and whether it transmits user data. - Do not run it in an environment that contains unrelated secrets (AWS keys, DB credentials, tokens). Prefer creating a minimal environment exposing only the explicit GEMINI_* variables required. - Remove or sanitize any ../.env file you don't want read, or patch index.js to avoid auto-loading parent .env. - If you must keep secrets in the agent environment, avoid installing this skill or run it in an isolated container. - If you want higher assurance, ask the author to explicitly declare required env vars and to stop inheriting the entire process.env/.env files.

Skill: rewrite_question

• Description: 补全上下文并重写用户提问。
• Inputs: [query, history]
• Outputs: [rewritten_query]
• ID: rewrite_question
• Role: 语境重构专家
  • 功能描述：分析 history 上下文，消除 query 中的指代不明（它、上周、那边）并补全省略项，输出自包含的标准查询。
  • 输入参数：
    • query (string): 原始输入。
    • history (list): 往返对话记忆。
  • 输出结果：rewritten_query (string): 补全业务背景后的独立请求。
  • 执行策略：代词消解 → 时间对齐 → 实体补全。

独立运行说明

# 基本用法（输出写入 skills/.workflow/rewrite_output.json）
python rewrite_question.py --query "今天汉河店的成交额"

# 全量重跑（清空本步及所有后续中间文件）
python rewrite_question.py --query "今天汉河店的成交额" --clean

输出文件（skills/.workflow/rewrite_output.json）

{
  "original_query": "今天汉河店的成交额",
  "final_query": "2026-03-11 汉河店的成交额",
  "is_rewritten": true,
  "confidence": 0.95,
  "thought": "...",
  "is_qa_matched": false,
  "matched_sql": null
}

下一步

python ../recognize-intent/recognize_intent.py