Security audit
Security checks for vulnerabilities and agentic risk
The skill is a mostly coherent token-saving guide, but it requests command execution and changes/caches conversation context without enough scope or retention detail.
Review before installing. This skill does not appear to contain malware or hidden scripts, but only install it if you are comfortable with an agent using broad exec permission and with prior conversation content being summarized or cached under rules that are not fully specified.
Earlier sections repeatedly describe the skill as performing intelligent context compression and importance-based summarization of old messages. This line says the free version does not include LLM-based intelligent summarization, which directly conflicts with the documented core mechanism for compression.
The skill promotes context compression and semantic caching without clearly warning that prior conversation content may be retained, transformed, or reused. In practice, caching and summarization can preserve sensitive data longer than users expect or propagate details into later responses, creating confidentiality and data-minimization risks—especially because the skill is designed to process long-running conversations.
Broad natural-language triggers like asking for 'Token状态' or similar phrases can overlap with ordinary conversation and activate optimization behavior unintentionally. In a context-compressing skill, accidental activation can alter conversation state, summarize prior content, or clear/pivot operational mode without the user intending to invoke the skill, which can affect integrity and retention of important context.
The command/phrase activation range is underspecified, so the agent may interpret ambiguous user text as a control instruction. Because this skill can change compression mode, clear cache, or disable optimization, ambiguous triggering can lead to unintended state transitions, loss of cached material, or degradation of conversation fidelity.
The manifest description scopes the skill to token-cost optimization via context compression and semantic caching. These lines additionally state that the free version uses a specific model route (GPT-4o-mini), which is a distinct operational behavior not described in the manifest's stated purpose.
The skill declares exec access even though its stated purpose is token optimization and conversational context handling. Unnecessary command-execution capability expands the attack surface: a prompt-injected or misaligned agent could invoke shell commands unrelated to optimization, enabling local inspection, network probing, or harmful side effects under the guise of the skill.
文档在版权声明中明确强调“完全中文化表达”，整体说明和触发方式也默认中文，但没有说明这是面向特定中文用户群体的限定技能，也未提供语言选择或英文替代。根据语言/地区策略，强制单一语言而无用户选择可能构成自然语言政策问题。
