功能描述

Use Gemini CLI (@google/gemini-cli) to do web search / fact-finding and return a sourced summary. Use when the user asks “why did X happen today”, “what’s the latest news”, “search the web”, “find sources/links”, or any task requiring up-to-date info. Prefer this over other search tools when Gemini is available but slow; run it with a TTY, wait longer, and verify source quality.

使用说明 (SKILL.md)

Gemini Web Search

Name: Gemini Web Search
Author: fengjiajie

Use Gemini CLI to search the web and produce a concise, sourced answer.

Quick workflow

Formulate a tight query

Include: entity + ticker/name + date/time window + what you need (%, $, cause, quotes, links)
Example: PayPal (PYPL) fell Feb 4 2026: % change, $ change, main catalyst(s), 3 sources

Run Gemini CLI with a TTY and long timeout Gemini CLI can hang or be slow without a pseudo-TTY.

Preferred (OpenClaw tool call):

Use functions.exec with pty: true
Use timeout 300–600s (longer for heavy searches)
Use yieldMs ~10000 then process.poll until completion

Command template:

~/.npm-global/bin/gemini -p "\x3Cprompt>"

If pty:true still behaves poorly, use a pseudo-tty wrapper:

script -q -c "~/.npm-global/bin/gemini -p \"\x3Cprompt>\"" /dev/null

Extract the answer in a structured way Return:

The key numeric facts (e.g., % move, $ move, close/intraday)
2–4 bullets of the main catalyst(s)
Links (always)

Quality control (mandatory)

Prefer: company IR/SEC filing, Reuters, Bloomberg, WSJ/FT, CNBC, reputable outlets.
Avoid relying on low-quality finance blogs/SEO sites.
If sources conflict or look unreliable: say so and ask user for a screenshot/link, or re-run with a stricter prompt.

Prompts that work well

Fast triage: Search the web: \x3Ctopic>. Give 3 bullets + 2 reputable links.
Market move: Search the web: Why did \x3CTICKER> move today (\x3Cdate>)? Provide exact % and $ move (close + intraday if available) and the main catalyst(s). Cite sources with links.
Force better sources: Search the web and prioritize Reuters/company IR/SEC filing. If you cannot find them, say so. Topic: \x3C...>. Provide links.

Failure modes & fixes

Gemini prints “I will search…” then stalls
- Wait longer (it can be slow).
- Ensure TTY: run with pty:true or script -q -c ... /dev/null.
Output has suspicious claims (e.g., odd CEO news)
- Re-run with: “use Reuters/company IR/SEC filing only; otherwise say unknown”.
- Cross-check with at least 2 independent reputable sources.
Need numbers but sources don’t show them
- Ask user for the quote/screenshot from their market data app and reconcile.

Local setup notes

Gemini CLI binary: ~/.npm-global/bin/gemini
Auth: already completed by Jiajie (should work without re-login)

安全使用建议

This skill is internally consistent: it simply runs the local Gemini CLI in a pseudo-TTY to perform web searches and return links. Before installing, verify you (or your environment) have the Gemini CLI installed and authenticated at the path indicated (or set GEMINI_BIN to the correct binary). Clarify the SKILL.md note that a person named 'Jiajie' already completed auth — if you don't have a configured Gemini account locally, searches will fail. Be aware that running the CLI performs network queries (expected) and that the recommended long timeouts / pty usage can spawn long-running processes on the host. If you need stronger guarantees, inspect or run the helper script locally to confirm behavior and consider restricting where the skill may run.

功能分析

Type: OpenClaw Skill Name: gemini-web-search Version: 1.0.0 The skill bundle is benign. The `SKILL.md` provides clear instructions for an AI agent to use the local `gemini` CLI for web search, emphasizing quality control and proper handling of the tool. The `scripts/gemini_prompt.sh` script correctly executes the `gemini` binary, safely escapes user-provided prompts to prevent shell injection, and uses `script -q -c` for legitimate pseudo-TTY emulation, as explained in the documentation. There is no evidence of data exfiltration, malicious execution, persistence, or prompt injection attempts aimed at subverting the agent for harmful purposes.

能力评估

✓ Purpose & Capability

Name/description match the runtime instructions and included script: both instruct running the local Gemini CLI to perform web searches and return sourced summaries.

ℹ Instruction Scope

Instructions stay within the stated purpose (formulate queries, run gemini with a TTY, extract facts and links). Minor oddity: SKILL.md claims 'Auth: already completed by Jiajie', which is a non-general note and may be leftover documentation; otherwise no instructions to read unrelated files or exfiltrate data.

✓ Install Mechanism

No install spec; only an opinionated command path (~/.npm-global/bin/gemini) and a tiny shell wrapper (script uses standard 'script' tool). No remote downloads or archive extraction.

ℹ Credentials

The skill declares no required env vars or credentials. It does rely on an already-installed and authenticated Gemini CLI (local user credentials/token stored by that CLI) — expected for this purpose, but the SKILL.md's reference to a specific person completing auth is ambiguous and should be clarified before deployment.

✓ Persistence & Privilege

always is false, the skill is user-invocable and does not request persistent system changes or modify other skills' configs.

版本历史

v1.0.0

Initial release of Gemini Web Search skill: - Uses Gemini CLI to perform web search and fact-finding, producing concise, sourced summaries. - Optimized for up-to-date info requests such as latest news, daily events, data with sources/links. - Provides detailed workflow: query formulation, CLI execution with TTY/timeout, structured extraction, and strict source quality control. - Includes prompts for various search scenarios and troubleshooting for common failure modes. - Setup and CLI usage instructions included for local deployment.

元数据

Slug gemini-web-search

版本 1.0.0

许可证 —

累计安装 17

当前安装数 17

历史版本数 1

常见问题