Skill Router BZai

Cost-effective skill selector for maximizing ROI on AI operations (增收降本版 v1.0.0). Use when the user needs to accomplish a task and wants the optimal skill ch...

Security Scan

The skill's code, commands, and runtime instructions are coherent with its stated purpose (discover, score, recommend, and optionally install/run other skills), but there are some implementation choices and shortcuts that reduce trustworthiness and should be reviewed before use.

Purpose & Capability

The skill claims to discover local skills and search clawhub, evaluate them on quality/token/security/speed, present top-3 options, and execute a chosen skill. The included scripts call the expected CLIs (openclaw, clawhub), estimate costs, and implement evaluation logic. No unrelated binaries, env vars, or external credentials are requested.

Instruction Scope

SKILL.md instructs the agent to list local skills, search clawhub, run a security review, present recommendations, and—after user confirmation—install and execute a selected skill. That behavior is consistent with the router's purpose, but it involves installing and running third-party skills (potentially untrusted). The scripts perform downloads (clawhub download --dry-run) and run external CLIs; the skill defers final execution to user confirmation, which reduces risk. Also note the code contains placeholder/partial implementations for security checks and parsing, so the actual runtime checks may be weaker than the doc implies.

Install Mechanism

No install spec is provided (instruction-only skill plus bundled scripts). There are no network-download install steps embedded in the skill package itself. The only external interactions are via existing CLIs (openclaw, clawhub), which is appropriate for a router helper.

Credentials

The skill declares no required environment variables or credentials. It does read/write a history file under the user's home (~/.openclaw/workspace/skill-router-history.json) for cost/history tracking — this is reasonable for a router but is persisted to the user's filesystem. The scripts rely on external CLIs which may themselves use credentials; that is expected for a tool that searches/installs skills.

Persistence & Privilege

always:false (default) — no forced global presence. The skill writes a history file in ~/.openclaw/workspace, which is limited per-user but persistent. It may install and execute other skills (with user confirmation). This is expected functionality for a skill manager but expands its effective privileges by invoking third-party code; consider sandboxing or stricter verification before letting it run autonomously.

Assessment

This skill appears to do what it says: discover, score, and recommend skills from local installs and clawhub. Before installing or enabling it, consider the following: 1) It will call your local openclaw and clawhub CLIs and may install/run third‑party skills after you approve — only approve skills you trust or have inspected. 2) It persists history at ~/.openclaw/workspace/skill-router-history.json; if you prefer no local traces, remove/redirect that path. 3) The bundled security checks are partially placeholder-ish (e.g., evaluate_clawhub/check_security has minimal implementation and evaluate_local_skill assumes local skills are vetted). Don’t rely solely on this router’s automatic security score—manually review any skill flagged as high risk. 4) Run the router in a sandbox or test environment first if you expect strict security controls. If you want stronger guarantees, request that its security-check implementations perform concrete static scans, parse downloaded code, and refuse installs that fail those checks.

Skill Router (增收降本版 v1.0.0)

智能技能选择器 - 用于增收降本的 AI 操作优化工具。

自动选择并执行最优技能，在保证质量的前提下最小化 Token 成本。

When to Use

• User wants to accomplish something but doesn't know which skill to use
• Multiple skills could solve the same problem - need to pick the best one
• Want to minimize token cost while maximizing quality
• Need to evaluate skills from clawhub.com before installation

Workflow

1. Parse Task

Analyze user request and decompose into atomic subtasks if needed.

2. Discover Skills

Search for candidate skills:

• List local installed skills: openclaw skills list
• Search clawhub.com: clawhub search <keywords>

3. Evaluate Candidates

For each candidate skill, score on:

Dimension	Weight	Evaluation Criteria
Quality/Utility	35%	User ratings, downloads, functionality match
Token Cost	30%	Estimated input/output tokens based on skill complexity
Security/Reliability	20%	Code audit, permissions, update frequency, author trust
Speed	15%	API response time, execution efficiency

Scoring Algorithm:

final_score = (quality × 0.35) + (token_score × 0.30) + (security × 0.20) + (speed × 0.15)

4. Generate Recommendations

Present Top-3 ranked options:

• Rank, skill name, final score
• Breakdown by dimension
• Estimated tokens and time
• Reasoning for recommendation
• Security assessment summary

5. User Confirmation

Wait for user to:

• Select option (1, 2, or 3)
• Request more details about a skill
• Cancel or modify the task

6. Execute

After confirmation:

• If skill not installed: clawhub install <skill>
• Execute the skill with original user request
• Record actual metrics vs estimates

Security Assessment

Before recommending any skill from clawhub:

1. Code Review: Check for suspicious patterns (network calls, file system access, credential harvesting)
2. Permission Analysis: Verify requested permissions match functionality
3. Author Verification: Prefer verified authors, established projects
4. Update Frequency: Recently updated skills are preferred
5. Community Trust: Ratings, issues, download count

Red flags that disqualify a skill:

• Requests excessive permissions for its stated purpose
• Contains obfuscated code
• Makes unexpected network calls
• Has no recent updates and low community engagement

Token Estimation

Estimate token costs based on:

• Skill description length/complexity
• Historical usage data (if available)
• Number of API calls required
• Output format verbosity

Store actual vs estimated for continuous improvement.

Scripts

• scripts/evaluate_skill.py - Score a skill across all dimensions
• scripts/search_clawhub.py - Search and fetch skill metadata from clawhub
• scripts/calculate_cost.py - Estimate token and time costs

References

• references/evaluation-rubric.md - Detailed scoring criteria
• references/security-checklist.md - Security audit checklist