Security audit
Security checks for vulnerabilities and agentic risk
This skill is a coherent task-memory helper, but it can automatically create durable workspace records of user task context with weak consent and data-minimization boundaries.
Install only if you want workspace-level task memory. Avoid storing secrets, credentials, private account details, or sensitive personal information in tracked tasks, and review or delete the `tasks/` files regularly.
The front-matter description defines very broad activation conditions such as any new task, progress update, status question, or channel switch, which can cause the agent to persist conversation details during ordinary interactions without a clear user request. In a persistence skill, overbroad triggering increases the chance of silent data retention and accidental storage of sensitive context.
The skill explains that task files persist across sessions and channels, but it does not provide a clear user-facing warning that conversation details will be written to workspace files and retained over time. Users may therefore disclose information assuming normal ephemeral chat behavior when the skill actually creates durable records.
The 'When This Skill Fires' table instructs automatic file creation and updates based on conversational cues, including at session start and when the user signals leaving, without clearly distinguishing read-only preparation from write actions requiring consent. This ambiguity can lead to automatic persistence of user content beyond expectations, creating privacy and integrity risks in the workspace.
The file includes Chinese-only example trigger phrases such as '先这样' and '回家再说', and much of the task template and index structure uses Chinese labels, but it does not state that this is a region-specific skill or offer language choice. That creates a locale policy concern because the skill appears to assume a specific language without user opt-in.
The skill explicitly says persistent task files should keep 'core facts' including accounts, which encourages storing sensitive credentials or account-identifying data in long-lived workspace files unrelated to strict task execution needs. Because this skill is designed for cross-session persistence, any over-collection becomes durable and easier to expose through later reads, sync, backup, or compromise of the workspace.
The workflow explicitly requires adding tasks to a persistent index, confirming that the skill maintains cross-session memory in workspace files. In this context, session persistence is risky because it creates durable records of user activity and task context, and the skill does not pair this persistence with strong consent, minimization, or sensitivity restrictions.
### Creating a New Task
1. Create `tasks/{task-name}.md` using the template above
2. Add entry to `_INDEX.md` under 进行中
3. Confirm creation with user
### Updating a Task
The archive section allows files to be moved automatically after 30 days of inactivity without a user-facing warning or confirmation. While this does not directly exfiltrate data, it can surprise users, affect availability of expected context, and alter retention behavior for persisted records without consent.
The documented handoff phrases, prompt text, and status labels are all written in Chinese, implying the skill expects or steers interaction in a specific language. There is no indication that users may choose another language or that the locale restriction is intentional and justified.
The activation condition 'When user discusses something new that sounds like a project/task' is subjective and lacks clear boundaries or exclusion examples. This could cause the skill to activate during ordinary conversation that merely resembles planning or brainstorming.
No suspicious patterns detected.