The Agent Skills Directory

[COMMAND_EXECUTION]: The skill's primary function involves generating temporary Python scripts (.py) at runtime and executing them using a Bash shell. These scripts utilize the provided coze_yaml_builder.py to assemble workflow YAMLs and build_coze_zip.py to package them. Subprocess execution of dynamically generated code is a significant capability that requires careful oversight.
[PROMPT_INJECTION]: The skill ingests natural language descriptions from the user to define workflow nodes and parameters. These strings are interpolated into the generated Python scripts. There is a risk of indirect prompt injection where maliciously crafted user input could attempt to escape string boundaries in the generated Python code (e.g., using triple quotes or escape characters) to achieve arbitrary code execution within the agent's runtime environment.
Ingestion points: Natural language workflow descriptions processed in SKILL.md.
Boundary markers: The _esc function in scripts/coze_yaml_builder.py provides escaping for backslashes and double quotes, but the agent remains responsible for the integrity of the script it writes.
Capability inventory: The skill has the capability to write files to /tmp and the current directory, and to execute shell commands.
Sanitization: Basic escaping is implemented in the builder script, but comprehensive validation of user-controlled strings during the script generation phase is critical.

coze-workflow