Prompt-Router

基于文本匹配的快速路由引擎，为简单任务提供零 LLM 决策的快速路径。支持中英文混合输入，自动匹配技能/工具，低置信度时降级到 LLM 语义路由。

This package appears to implement a useful local routing engine, but it also contains scripts and documentation that will: 1) collect and persist raw prompts and routing logs (which can contain sensitive data), 2) modify other skills' SKILL.md files (fix_skills.py, generate_metadata.py), and 3) provide automation to create PRs/publish and schedule recurring jobs. Before installing or enabling automated workflows: - Review the code paths that write files (fix_skills.py, generate_metadata.py, scripts/*) and decide whether you trust automatic edits to your skills directory. - Disable or inspect any cron/scheduling steps; do not add cron jobs until you verify behavior. - Check for any functions that perform network actions (create_pr, send_to_discord, send_email, create_issue) and ensure they require explicit credentials and you control them. - If you want the router but not the meta-editing/auto-update behavior, run only the router module (scripts/router.py / scripts/integration.py) and do not run fix/generate/auto-update scripts. - Run the skill in a sandbox or non-production workspace first, and audit the logs directory for sensitive content. If you want, I can point to the exact lines/functions that modify files or log prompts so you can inspect them further.

Type: OpenClaw Skill Name: prompt-router Version: 1.0.0 The bundle provides a 'Prompt-Router' utility designed to optimize OpenClaw agent performance by using local text-matching to route tasks to specific skills, bypassing unnecessary LLM inference. It includes core logic for tokenization and scoring (scripts/router.py, scripts/scorer.py) and utility scripts (fix_skills.py, generate_metadata.py) that automatically update metadata in other skill files within the user's workspace (C:/Users/User/.openclaw/workspace/skills). While the modification of other files is a high-privilege action, it is functionally necessary for the tool's stated purpose of indexing the workspace, and no evidence of malicious intent, data exfiltration, or unauthorized network activity was found.