The Agent Skills Directory

[SAFE]: File System Interaction: The skill is restricted to reading and writing local markdown files (plan.md and actions/*.md) within a specific project directory structure to maintain action plans and task states.
[SAFE]: No Remote Execution or Downloads: Analysis confirms the absence of remote code execution patterns, external script downloads, or installation of unverifiable packages. All operations are local to the execution environment.
[SAFE]: Absence of Data Exfiltration: The skill contains no network-capable commands (e.g., curl, wget) or exfiltration patterns targeting sensitive system files or credentials.
[SAFE]: User-in-the-Loop Mitigations: The planning flow enforces mandatory manual confirmation checkpoints ('Confirm bảng trước ghi') before the agent updates the filesystem, providing a significant safeguard against unintended or malicious file modifications.
[SAFE]: Indirect Prompt Injection Surface: While the skill ingests user-controlled content from files like 'objective.md', the risk is minimized by the structured output requirements and the manual review process. Ingestion points: objective.md, plan.md, actions/*.md, tools.md; Boundary markers: Manual confirmation steps; Capability inventory: Local file-read and file-write access; Sanitization: Relies on user verification of the proposed plan table.

pk-plan