The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The script scripts/unpack.ts uses the new Function constructor to evaluate and deobfuscate JavaScript code provided as input, specifically when handling Packer and AAEncode patterns. Because this tool is intended to process untrusted or potentially malicious code samples, this capability allows for arbitrary code execution within the agent's process context. A --no-eval flag is provided as a mitigation, but the core functionality remains high-risk.
[REMOTE_CODE_EXECUTION]: The script scripts/wakaru-normalize.ts performs dynamic execution of external code by invoking npx @wakaru/cli@1.5.0. This process downloads and runs a package from the npm registry that is not included in the skill's official dependency list.
[DATA_EXFILTRATION]: The import resolution mechanism in scripts/build-import-graph.ts uses path.resolve on user-provided import strings without restricting access to a safe directory. A malicious input file could use relative path segments like ../ to perform a path traversal attack, causing the agent to read and expose sensitive files from the host system.
[PROMPT_INJECTION]: The skill has an indirect prompt injection surface because it ingests untrusted JavaScript source code and uses it as context for automated logic and agent-based review.
Ingestion points: scripts/build-import-graph.ts and scripts/extract.ts (processing user-provided source files).
Boundary markers: Present in scripts/acceptance-checklist.md (=== FILE: ... ===).
Capability inventory: Subprocess execution via spawnSync in multiple scripts, new Function in scripts/unpack.ts, and various file system write operations.
Sanitization: Absent; the skill relies on structural parsing which does not prevent natural language instructions embedded within source code components from influencing downstream agents.
[COMMAND_EXECUTION]: The script scripts/format.ts executes the well-known prettier utility via npx or bunx to format code deliverables.

deobfuscate-javascript