The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of execFileSync, spawn, and execFile to run local utilities like git and node. Crucially, in scripts/live-copy-edit-agent.mjs, it invokes the codex CLI with the --dangerously-bypass-approvals-and-sandbox flag to automate source code edits. While intended for its 'Manual Edit Apply' feature, this represents a significant privilege level.
[EXTERNAL_DOWNLOADS]: The skill makes network requests to https://impeccable.style/api/roll and https://impeccable.style/api/chosen to fetch design directions and record telemetry. It also contacts https://api.openai.com for image generation. While these are known services for the tool's ecosystem, they involve external data exchange.
[REMOTE_CODE_EXECUTION]: The skill dynamically loads modules from the user's project node_modules directory. In scripts/live/svelte-ast.mjs, it uses createRequire on the project's package.json path to load the Svelte compiler (svelte/compiler). This dynamic loading from computed paths is a medium-severity pattern used here to analyze the project's own frontend code.
[DATA_EXFILTRATION]: Telemetry and 'concept rolls' are fetched from impeccable.style. Although the script documentation claims no sensitive project data or conversation context is transmitted, the network activity to a non-whitelisted domain is a notable finding.
[PROMPT_INJECTION]: Static analysis identified a pattern in reference/craft.md regarding action concealment. However, this appears to be a false positive related to UI terminology guidance (instructing the agent to use natural language instead of technical aliases), rather than an attempt to hide malicious behavior.

impeccable