The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill generates C++ and Python code for AscendC operators and immediately executes it. Phase 4 involves generating source files (op_host, op_kernel), running a build script (./build.sh), and installing the resulting package via pip. Phase 6 and 7 generate and execute Python scripts for precision and performance testing. While this is the primary purpose of the skill, the runtime compilation and execution of generated code is a significant capability.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user input in the form of 'functional descriptions' and 'operator names' in Phase 0. These inputs are used in Phase 2 to design the operator logic and Phase 4 to generate the actual implementation and registration code. There are no explicit instructions for the agent to sanitize these inputs or use boundary markers to prevent the functional description from injecting malicious logic into the generated code or documentation.
[COMMAND_EXECUTION]: The skill performs several shell operations including sourcing environment scripts (set_env.sh), running project build scripts (build.sh), and executing performance profiling tools (msprof). It specifically instructs the agent to source a script from a path that may be provided by the user if the environment variable is not set.
[EXTERNAL_DOWNLOADS]: Phase 1 references a process for cloning an operator project template if it does not exist. While the template source is not explicitly defined in this orchestration file, it implies a remote fetch operation managed by a sub-skill.

ascendc-operator-dev