The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script subskills/solidworks-fillet-chamfer-cnc/scripts/hold_line_bridge.py dynamically compiles C# source code (HoldLineBridge.cs) and C++ source code (NativeHoldLineAddin.cpp) at runtime using csc.exe and cl.exe to create specialized DLLs and executables for SolidWorks interaction.
[INDIRECT_PROMPT_INJECTION]: The skill implements a RAG system that creates a potential injection surface:
Ingestion points: apps/desktop/cad_workbench/knowledge_retrieval.py reads local .md, .txt, and .yaml files from the skill's reference and subskill directories to build a knowledge context for the AI.
Boundary markers: The prompt compiler in apps/desktop/cad_workbench/agent_contracts.py appends retrieved knowledge chunks to the prompt without utilizing strong delimiters or instructions to the AI to ignore potentially malicious embedded commands.
Capability inventory: The skill possesses high-privilege capabilities, including file system write access, optional network operations for cloud RAG, and the ability to execute arbitrary commands via subprocesses to control CAD software (scripts/sw_connect.py, scripts/fea_analysis.py).
Sanitization: No specific filtering or sanitization of retrieved knowledge snippets is performed before they are interpolated into the agent's execution context.
[COMMAND_EXECUTION]: The skill extensively uses subprocess.run and subprocess.check_call to interact with the host Windows environment, manage toolchains, and execute CAD commands or FEA solvers (e.g., in scripts/sw_preflight.py, scripts/fea_analysis.py, and subskills/autocad-automation/scripts/acad_dotnet_regression.py). The Tauri-based worker process in apps/desktop/cad_workbench/queue_worker.py manages a job queue that spawns subprocesses to run Agent CLI tools such as Codex and Claude Code.

solidworks-automation