The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides instructions for using the vectorcraft-cli and a local JSON control channel for automation.
Evidence: SKILL.md describes various CLI subcommands (mcp, run, convert) and a --control flag that enables a local listener on 127.0.0.1 (L256-267).
[INDIRECT_PROMPT_INJECTION]: The skill has a broad attack surface for indirect prompt injection because it is designed to ingest and process a wide variety of untrusted external file formats.
Ingestion points: SKILL.md (L186-194) lists 31 supported extensions for opening or placing files, including SVG, PDF, and DOCX.
Boundary markers: The skill explicitly recommends using --automation-read-root and --automation-write-root to confine the agent's file system access (L330-335).
Capability inventory: The tool provides an extensive engine command set (780 commands), file system operations, and 32 MCP tools as documented in SKILL.md and evidence.md.
Sanitization: Plugin execution is isolated via a WebAssembly sandbox (ABI v1) with resource caps (L274-279).
[DYNAMIC_EXECUTION]: The application supports the runtime loading and execution of external code through a plugin system.
Evidence: SKILL.md describes the use of sandboxed WebAssembly modules as filters, effects, or exporters, noting that they have no host imports and are subject to fuel and memory caps (L274-279).

vectorcraft