The Agent Skills Directory

[DATA_EXFILTRATION]: The skill is designed with a strict offline-first policy. It stores all data (boss profile, project history, and ideas) in a local directory (~/.mishu or MISHU_HOME). Both the system instructions and the bundled scripts explicitly state that no network requests are made, and it includes clear rules against storing passwords, API keys, or other sensitive personal information.
[REMOTE_CODE_EXECUTION]: All executable components, including Shell installation scripts and Python metadata parsers, are local to the skill package. These scripts use only standard system libraries and perform specific, documented tasks such as initializing directories or reading agent session metadata to reconstruct project status. There is no evidence of fetching or executing remote payloads.
[COMMAND_EXECUTION]: The skill uses local shell commands for setup and as integration hooks for supported agent platforms (e.g., Claude Code). These commands are transparently defined in the source code, operate on local paths, and do not incorporate unsanitized user input into shell execution.
[INDIRECT_PROMPT_INJECTION]: While the skill reads project-specific markdown files (NOW.md, HANDOFF.md) and agent session metadata, which represents a potential ingestion surface for untrusted content, it mitigates this risk by requiring the agent to present a candidate summary for user review and correction before committing any changes to the project index (SHELF.md).

mishu