Security audit
Security checks for vulnerabilities and agentic risk
This skill is a context-health reminder that only gives session-management suggestions and does not include code, credential handling, persistence, or destructive behavior.
Before installing, be aware that auto-mode may occasionally interrupt with context-management suggestions because its trigger boundaries are not tightly specified. This is a productivity and conversation-control concern, not evidence of malicious behavior.
The file states that auto-mode is enabled by default, but does not define when the skill should trigger, what scope it can inspect, or what exclusions and safeguards apply. For a context-inspection skill, ambiguous auto-activation can cause unintended invocation across unrelated conversations, increasing the chance of privacy leakage, prompt interference, or disruptive behavior.
The README says auto mode uses "response-layer triggers when signals detected" and later "Automatically triggers when signals detected" without clearly defining activation boundaries or exclusion conditions. This ambiguity could cause unintended invocation because readers are not told exactly when the skill should trigger versus remain inactive.
El README indica que el modo automático se activa "cuando se detectan señales" y luego enumera condiciones heurísticas amplias como "sin progreso", "deriva de tema" o "confusión de memoria" sin delimitar claramente cuándo sí o no debe invocarse. Esa falta de criterios negativos o alcance preciso puede causar activaciones no deseadas en conversaciones normales.
The README states that the automatic mode triggers in the response layer when a 'signal' is detected, but it does not define precise decision boundaries, precedence, or safeguards against false positives. In an agent skill, ambiguous auto-invocation criteria can cause the skill to activate unexpectedly, interrupt normal workflows, and influence conversation control without clear user intent.
The listed auto-trigger signals include subjective conditions such as 'no progress,' 'topic drift,' and 'memory confusion' without concrete measurement criteria or negative examples. That ambiguity can lead to over-triggering, inconsistent behavior across implementations, and accidental context switching recommendations that degrade reliability or can be abused to steer the session away from the user's intended task.
该 README 是 markdown 文件，适用 SQP-1。文档说明“响应层检测到信号时触发”，并列出如“无明显进展”“主题漂移”“记忆模糊”等自然语言条件，但没有给出明确判定标准、优先级或不触发的反例，这些条件带有较强主观性，容易与普通长对话或探索性讨论重叠。
对 markdown 文件应检查触发说明是否足够具体。这里仅说明自动模式默认启用，但未明确启用后是在所有会话、所有项目还是特定环境中生效，也未说明如何禁用或限制范围，容易造成用户对技能何时介入的预期不清。
The manual trigger keywords include generic phrases like '上下文检查', 'context check', and 'context hygiene', which may appear in ordinary discussion rather than as an invocation. This ambiguity can cause accidental activation of the skill, creating unexpected behavioral changes or response hijacking when the user is merely talking about context management conceptually.
The skill defines automatic activation in a response layer using broad, repeated criteria without a precise implementation contract or guardrails. In practice, this can cause unintended triggering during normal conversation, leading to prompt injection of unsolicited guidance, response disruption, or policy interference across unrelated tasks.
The auto-mode example shows a Chinese-language message as the default behavior example, but the surrounding usage text does not state that output language follows user preference or can be selected. This may conflict with language/locale policy expectations because the skill appears to emit a fixed language without explicit opt-in.
The trigger condition tied to '确认式句式（anti-sycophancy 触发）' is underspecified and broad enough to match common user phrasing. That can make the skill intervene in many normal exchanges, increasing the chance of unnecessary output, user confusion, or interaction with other safety/behavior layers in unpredictable ways.
No suspicious patterns detected.
