The Agent Skills Directory

[COMMAND_EXECUTION]: The skill implements the dspy.ProgramOfThought component, which dynamically generates and executes Python code to solve math or data tasks. While the documentation indicates this execution is sandboxed, the pattern inherently involves running code generated from language model outputs that can be influenced by untrusted inputs.
[PROMPT_INJECTION]: Several modules described in the skill, such as ComplexQuestionSolver and PlanningAssistant, ingest untrusted data (customer questions, data descriptions, or project scenarios) and interpolate them directly into reasoning prompts. This creates an indirect prompt injection surface.
Ingestion points: Variable inputs such as question, context, data_description, and scenario found in SKILL.md and examples.md.
Boundary markers: The provided code snippets do not utilize explicit boundary markers (like XML tags or specific delimiters) to isolate untrusted data from the system's reasoning instructions.
Capability inventory: The skill possesses the capability to execute generated Python code through the dspy.ProgramOfThought module.
Sanitization: No evidence of sanitization, filtering, or validation of the input data is present before it is used to generate reasoning plans or executable code.

ai-reasoning